Nature Biotechnology 08 2010 (magazine journal; August 2010)

volume 28 number 8 august 2010

e d i tor i a l s

© 2010 Nature America, Inc. All rights reserved.

761 761

A computer-generated representation of HIV on the surface of a T lymphocyte. Holt et al. block the entry of HIV into blood cells by using zinc finger nucleases to knock out CCR5 in hematopoietic stem cells (p 839). Credit: ANIMATE4.com/ SciencePhotoLibrary

Wrong numbers? MAQC-II: analyze that!

n ews 763 Industry makes strides in melanoma 765 Firms combine experimental cancer drugs to speed development 767 FDA transparency rules could hit small companies hardest 767 Supremes rule on Bilski 768 Lawsuits rock Jackson 769 Food firms test fry Pioneer’s trans fat–free soybean oil 769 Anti-CD20 patent battle ends 769 EU states free to ban GM crops 770 GM alfalfa—who wins? 770 Biofuel ‘Made in China’ 771 data page: 2Q10—spreading the wealth 772 News feature: Drugmakers dance with autism

B i oe n trepre n eur B u i l d i n g a bus i n ess 775

At ground level Julian Bertschinger

op i n i o n a n d comme n t

Jackson Lab’s legal woes, p 768

778 779 780

C O R R E S P O ND E N C E Waking up and smelling the coffee Genetic stability in two commercialized transgenic lines (MON810) Distances needed to limit cross-fertilization between GM and conventional maize in Europe

Nature Biotechnology (ISSN 1087-0156) is published monthly by Nature Publishing Group, a trading name of Nature America Inc. located at 75 Varick Street, Fl 9, New York, NY 10013-1917. Periodicals postage paid at New York, NY and additional mailing post offices. Editorial Office: 75 Varick Street, Fl 9, New York, NY 10013-1917. Tel: (212) 726 9335, Fax: (212) 696 9753. Annual subscription rates: USA/Canada: US$250 (personal), US$3,520 (institution), US$4,050 (corporate institution). Canada add 5% GST #104911595RT001; Euro-zone: €202 (personal), €2,795 (institution), €3,488 (corporate institution); Rest of world (excluding China, Japan, Korea): £130 (personal), £1,806 (institution), £2,250 (corporate institution); Japan: Contact NPG Nature Asia-Pacific, Chiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843. Tel: 81 (03) 3267 8751, Fax: 81 (03) 3267 8746. POSTMASTER: Send address changes to Nature Biotechnology, Subscriptions Department, 342 Broadway, PMB 301, New York, NY 10013-3910. Authorization to photocopy material for internal or personal use, or internal or personal use of specific clients, is granted by Nature Publishing Group to libraries and others registered with the Copyright Clearance Center (CCC) Transactional Reporting Service, provided the relevant copyright fee is paid direct to CCC, 222 Rosewood Drive, Danvers, MA 01923, USA. Identification code for Nature Biotechnology: 1087-0156/04. Back issues: US$45, Canada add 7% for GST. CPC PUB AGREEMENT #40032744. Printed by Publishers Press, Inc., Lebanon Junction, KY, USA. Copyright © 2010 Nature America, Inc. All rights reserved. Printed in USA.

i

volume 28 number 8 august 2010 C O M M E N TA R Y 783

case study: India’s billion dollar biotech Justin Chakma, Hassan Masum, Kumar Perampaladas, Jennifer Heys & Peter A Singer

784 DNA patents and diagnostics: not a pretty picture Julia Carbone, E Richard Gold, Bhaven Sampat, Subhashini Chandrasekharan, Lori Knowles, Misha Angrist & Robert Cook-Deegan

feature Rapid bacterial engineering, p 812

793

Public biotech 2009—the numbers Brady Huggett, John Hodgson & Riku Lähteenmäki

801

Bilski v. Kappos: the US Supreme Court broadens patent subject-matter eligibility William J Simmons

806

Recent patent applications in proteomics

State

N E W S A ND V I E W S H3K14ac H3K23ac H4K12ac H2AK9ac H4K16ac H2AK5ac H4K91ac H3K4ac H2BK20ac H3K18ac H2BK120ac H3K27ac H2BK5ac H2BK12ac H3K36ac H4K5ac H4K8ac H3K9ac PolII CTCF H2AZ H3K4me3 H3K4me2 H3K4me1 H3K9me1 H3K79me3 H3K79me2 H3K79me1 H3K27me1 H2BK5me1 H4K20me1 H3K36me3 H3K36me1 H3R2me1 H3R2me2 H3K27me2 H3K27me3 H4R3me2 H3K9me2 H3K9me3 H4K20me3

© 2010 Nature America, Inc. All rights reserved.

pate n ts

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26

807

Can HIV be cured with stem cell therapy? see also p 839 Steven G Deeks & Joseph M McCune

810

Microarrays in the clinic see also p 827 Guy W Tillinghast

812

Shaking up genome engineering see also p 856 Kim A Tipton & John Dueber

813

The expanding family of dendritic cell subsets Hideki Ueno, A Karolina Palucka & Jacques Banchereau

816

Research highlights

Epigenetic marks define chromatin states, p 817

computat i o n a l b i o l ogy a n a lys i s 817 Discovery and characterization of chromatin states for systematic annotation of the human genome Jason Ernst & Manolis Kellis

0.982

0.910

0.845

0.748

0.575

0.557

0.311

0.323

0.244

0.973

0.918

0.829

0.792

0.493

0.437

0.322

0.306

0.307

0.202

0.965

0.801

0.816

0.652

0.514

0.349

0.383

0.360

0.217

0.243

0.193

0.991

0.752

0.750

0.778

0.509

0.483

0.345

0.305

0.295

0.193

0.973

0.869

0.825

0.755

0.403

0.413

0.321

0.275

0.193

0.266

0.982

0.762

0.823

0.702

0.533

0.557

0.284

0.203

0.143

0.257

0.982

0.871

0.445

0.728

0.472

0.249

0.429

0.353

0.295

0.293

0.930

0.838

0.805

0.773

0.542

0.386

0.345

0.289

0.225

0.181

0.982

0.847

0.835

0.737

0.488

0.344

0.118

0.324

0.110

0.176

0.057

0.243

0.973

0.860

0.829

0.690

0.371

0.376

0.344

0.229

0.956

0.815

0.847

0.773

0.491

0.202

0.185

0.385 −0.014 0.187

0.982

0.847

0.780

0.755

0.377

0.423

0.313 −0.042 0.198

0.725

0.782

0.824

0.770

0.531

0.344

0.168

0.349 −0.096 0.165

0.982

0.707

0.782

0.466

0.499

0.184

0.271

0.000 −0.062 0.203

0.636

0.761

0.454

0.748

0.247

0.377

0.062

0.324

0.856

0.054

0.709

0.751

0.455 −0.213 −0.078 0.114

0.982

0.830

0.595

0.544

0.036 −0.090 −0.027 0.336 −0.143 −0.030

0.973

0.830

0.816

0.748

0.491

0.376

0.311

0.306

0.193

0.193

0.982

0.891

0.829

0.732

0.403

0.479

0.429

0.301

0.217

0.162

0.043

0.241

0.085

0.479 −0.096

research ARTICLES 827

The MicroArray Quality Control (MAQC)-II study of common practices for the development and validation of microarray-based predictive models MAQC Consortium see also p 810

839

Human hematopoietic stem/progenitor cells modified by zinc-finger nucleases targeted to CCR5 control HIV-1 in vivo N Holt, J Wang, K Kim, G Friedman, X Wang, V Taupin, G M Crooks, D B Kohn, see also p 807 P D Gregory, M C Holmes & P M Cannon

Evaluating microarray classifiers, p 827

nature biotechnology

iii

volume 28 number 8 august 2010 848

Cell type of origin influences the molecular and functional properties of mouse induced pluripotent stem cells J M Polo, S Liu, M E Figueroa, W Kulalert, S Eminli, K Yong Tan, E Apostolou, M Stadtfeld, Y Li, T Shioda, S Natesan, A J Wagers, A Melnick, T Evans & K Hochedlinger

856

Rapid profiling of a microbial genome using mixtures of barcoded oligonucleotides J R Warner, P J Reeder, A Karimpour-Fard, L B A Woodruff & R T Gill see also p 812 l etters

Epigenetics of iPS cells, p 848

863 Implications of the presence of N-glycolylneuraminic acid in recombinant therapeutic glycoproteins D Ghaderi, R E Taylor, V Padler-Karavani, S Diaz & A Varki

© 2010 Nature America, Inc. All rights reserved.

868

Global analysis of lysine ubiquitination by ubiquitin remnant immunoaffinity profiling G Xu, J S Paige & S R Jaffrey

careers a n d recru i tme n t

nature biotechnology

875

Second quarter biotech job picture Michael Francisco

876

people

v

in this issue

© 2010 Nature America, Inc. All rights reserved.

MAQC-II: evaluating microarray classifiers Building on its original work assessing the technical performance of DNA microarray technology (http:// www.nature.com/nbt/focus/maqc/ index.html), the Microarray Quality Control (MAQC) consortium, a partnership of research groups from the US Food and Drug Administration (FDA), academia, industry and other government agencies, has set out to investigate the capabilities and limitations of microarray data analysis with respect to disease diagnosis or choice of therapies. Although numerous methods for analyzing microarray data have been developed, there remains a lack of consensus regarding best practices in terms of their use in identifying gene signatures that are representative of a pathological condition. Such practices are becoming increasingly important, especially as the FDA receives many proposals to use microarrays to support medical product development and testing. In the present paper, 36 data analysis teams applied a variety of analytic methods to build classifiers to predict the toxicity of chemicals in rodent models and to predict clinical outcomes in human patients with breast cancer, multiple myeloma or neuroblastoma. The experience gained during this large project may be useful for developing classifiers for data from other high-throughput assays. This is important in light of the study’s finding that microarrays perform poorly at making certain clinical predictions, suggesting that technologies that assay additional aspects of human physiology may be needed to formulate better clinical treatment plans. [Articles, p. 827; News and Views p. 810] CM

Engineered stem cells control HIV Cannon and colleagues present an anti-HIV strategy in which human hematopoietic stem/progenitor cells are modified with zinc-finger nucleases to knock out C-C chemokine receptor 5 (CCR5), the principal co-receptor for HIV. CCR5 has been a target of exceptional interest ever since the 1996 discovery that a homozygous 32-bp deletion in the gene confers resistance to HIV infection without any apparent ill effects on health. Most previous work has used small molecules, ribozymes or siRNA to inhibit CCR5 protein or mRNA. In contrast, Cannon and colleagues  nucleofect plasmids expressing  two zincfinger nucleases into human CD34+ stem/progenitor cells to permanently knock out the CCR5 gene. The modified cells are transplanted into irradiated, immunodeficient mice and allowed to engraft for 8–12 weeks before the mice are challenged with CCR5-tropic HIV. Although human T cell counts initially decline, by week 8 they have recovered to their original levels. By weeks 10 and 12, HIV RNA in Written by Kathy Aschheim, Markus Elsner, Michael Francisco, Peter Hare, Craig Mak, & Lisa Melton

nature biotechnology volume 28 number 8 august 2010

the intestine is undetectable. Because hematopoietic stem cells can reconstitute the entire hematopoietic system, the authors propose that modified CD34+ cells could provide long-term HIV resistance in all the lymphoid and myeloid cell types that the virus infects. In support of this hypothesis, a transplant of allogeneic CCR5Δ32 hematopoietic stem cells in an HIV+ individual with acute myeloid leukemia may have cured the HIV infection (N. Engl. J. Med. 360, 724–725, 2009). [Articles, p. 839; News and Views, p. 807] KA

Epigenetic marks stand together With over 100 known histone modifications that can occur in thousands of possible combinations, it is challenging to identify specific combinations that have distinct biological functions. Ernst and Kellis describe an algorithm that deduces chromatin states (reoccurring, spatially coherent combinations of epigenetic marks) from experimental data on the distribution of different modifications. Using a multivariate Hidden Markov Model to analyze data on the position of 41 different marks in human T cells, they define 51 distinct chromatin states. The authors correlate these states with prior genome annotation and find that individual states are associated with specific functional regions such as gene promoters, transcriptionally active genes, large-scale repressed regions or intergenic active regions. The identification of chromatin states will facilitate genome annotation, the discovery of functional elements, and mechanistic studies of gene regulation by epigenetic marks. [Analysis, p. 817] ME chr1:

242959000

242959500

State 2 State 3 State 5 State 37 State 38 Coding Exon Spliced ESTs

242960000

242960500

242961000

242961500

low-expression promoter state

new exon prediction

Mammalian Conservation

Faster trait-to-gene mapping Gill and colleagues describe an approach for creating rationally modified collections of Escherichia coli in which every strain contains the same defined mutation but in a different gene. Such collections are valuable tools for mapping the genetic basis of traits, but until now have been labor intensive to construct. The method creates thousands of modified strains in parallel by transforming bacteria with pools of oligonucleotides that each recombine with a single gene to introduce a mutation. Barcode sequence tags uniquely identify each oligo and thus each strain. The collection of strains is grown in a condition of interest that selects for genetic modifications that confer fitness advantages. Fitter strains are recovered and identified by sequencing or by microarray detection of their barcodes. To demonstrate the method, Gill and colleagues created collections of E. coli with strains in which single genes were either up- or downregulated. Growing these strains in cellulosic hydrolysates—a toxic intermediate of biofuel processing—or in the presence of valine, d-fucose or methyglyoxal revealed unexpected genes that influenced growth in these industrially relevant conditions. The identified genes could form the basis for subsequent combinatorial genetic engineering. [Articles, p. 856; News and Views, p. 812] CM vii

in t h is issue

© 2010 Nature America, Inc. All rights reserved.

Ubiquitination sites in the crosshairs Immunoaffinity-based approaches have been key to enabling proteome-wide analysis of GG post-translational modifications such as phosK phorylation. However, attempts to selectively purify ubiquitinated peptides on a large scale have been frustrated by the difficulty of isolating and identifying peptides tagged with the 76-amino-acid ubiquitin protein. Jaffrey and colleagues simplify such analyses by generating a monoclonal antibody that selectively recognizes sites of protein ubiquitination. When protein lysates are digested with trypsin, ubiquitin adducts are trimmed to a diglycine stub. The ability of the antibody to recognize these ubiquitin remnants conjugated to the side chains of ubiquitinated lysines in a range of sequence contexts enables the authors to enrich for peptides carrying sites of ubiquitination and then identify them using tandem mass spectrometry. Working with cells expressing hexahistidine-tagged ubiquitin, the authors use this strategy to extend the catalog of mammalian ubiquitinated proteins and further illustrate the strength of the approach by demonstrating differential regulation of ubiquitination at distinct sites within the same protein. [Letters, p. 868] PH

Neu5Gc content and biologics Much effort has been devoted to reducing the immunogenicity of protein biologics caused by peptide epitopes. However, far less attention has been

Patent Roundup The US Food and Drug Administration is proposing new transparency rules to increase the information it discloses about product applications. The rules could compromise trade secret protection and put small companies at a competitive disadvantage. [News Analysis, p. 767] LM The US Supreme Court’s long-awaited decision on Bilski v. Kappos rules against patenting only inventions transformed by a machine. But the ruling leaves several questions unanswered, especially with regard to the eligibility of patents for diagnostic methods. [News in brief, p. 767] LM The not-for-profit Jackson Laboratory has been caught up in patent disputes, for the first time in its 80-year history. If the expense of such litigation escalates, the lab may have to cover its costs by charging researchers higher prices for access to mouse strains in its repository. [News in brief, p. 768] LM A four-year dispute over a European patent for an anti-CD20 monoclonal antibody to treat rheumatoid arthritis has ended in favor of Trubion, based in Seattle, and against Genentech and Biogen Idec. The decision frees up the patent space for anyone contemplating a CD20 program, according to Trubion. [News in brief, p. 769] LM Both sides are claiming victory following the US Supreme Court’s verdict in Monsanto v. Geerston Seed Farms over future sales of Roundup Ready alfalfa seeds. Monsanto (St. Louis, MO) cheered the court’s decision to reverse a previous injunction banning the transgenic alfalfa, but the seeds’ commercialization is still subject to an environmental impact statement by the US Department of Agriculture. [News in brief, p. 770] LM The US Supreme Court recently broadened the definition of patent-eligible subject matter. In this issue, Simmons parses Bilski v. Kappos and what the far-reaching decision means for biotech and pharmaceutical patent seekers. [Patent Article, p. 801] MF Recent patent applications in proteomics. [New Patents, p. 806]MF

viii

paid to the possibility of untoward effects caused by immune reactions to glycans on glycoprotein therapeutics. Varki and colleagues present evidence suggesting that it may be necessary to revisit whether the presence of the sialic acid N-glycolylneuraminic acid (Neu5Gc) on certain glycoprotein drugs may influence their immunogenicity and half-lives in vivo. Unlike other mammals studied to date, humans lack the ability to make Neu5Gc. Nonetheless, recent studies have revealed that most of us have variable—and sometimes relatively high—levels of circulating antibodies against Neu5Gc. The authors demonstrate the presence of Neu5Gc on only one of two clinically approved monoclonal antibodies directed against the same target. In vitro, antibodies or antisera against Neu5Gc from healthy humans generate immune complexes only in the presence of the Neu5Gc-containing drug. Moreover, antibodies to Neu5Gc in mice with a human-like defect in Neu5Gc synthesis promote the clearance of only the Neu5Gc-containing drug. Injection of this drug also promotes the production of preexisting antibodies against Neu5Gc. If further studies support the possibility that antibodies against Neu5Gc might influence the immunogenicity and efficacy of therapeutic glycoproteins in humans, production using cultured human cells may not resolve the issue, as Neu5Gc could still be incorporated from animal-derived products in culture media. Varki and colleagues show that a better solution would be to displace Neu5Gc from being incorporated into recombinant proteins by inclusion of an excess of the human sialic acid N-acetylneuraminic acid in culture media. [Letters, p. 863] PH

Epigenetic memory in iPS cells All induced pluripotent stem (iPS) cells from different tissues are not created equal. That is the conclusion of a study comparing mouse iPS cells derived from four tissues— tail-tip fibroblasts, splenic B cells, bone marrow–derived granulocytes and skeletal muscle precursors. Hochedlinger and colleagues use a ‘secondary’ system for reprogramming (Nat. Biotechnol. 26, 916–924, 2008) so that all iPS cells have identical integrations of the four transgenes, eliminating this confounding variable. They find that early-passage iPS cells retain an epigenetic memory of their cell type of origin and that this memory alters the cells’ gene expression and differentiation potential. Notably, these epigenetic, transcriptional and functional differences can be attenuated by extended passaging. Several lines of evidence suggest that this erasure of epigenetic memory occurs not though the selection of rare, fully reprogrammed cells but through gradual epigenetic changes in the majority of cells. Epigenetic memory in iPS cells can be considered desirable or not depending on one’s experimental goals. In studies aimed at producing a specific cell type, it could be beneficial—suggesting, for example, that a project to generate blood cells should begin by reprogramming blood cells rather than an unrelated cell type. [Articles, p. 848] KA

Next month in • Castor bean genome • Benchmarking dynamic mass redistribution • Measuring protein-DNA interactions at equilibrium • Metabolic modeling made easier

volume 28 number 8 august 2010 nature biotechnology



www.nature.com/naturebiotechnology

EDITORIAL OFFICE [email protected] 75 Varick Street, Fl 9, New York, NY 10013-1917 Tel: (212) 726 9200, Fax: (212) 696 9635 Chief Editor: Andrew Marshall Senior Editors: Laura DeFrancesco (News & Features), Kathy Aschheim (Research), Peter Hare (Research), Michael Francisco (Resources and Special Projects) Business Editor: Brady Huggett Associate Business Editor: Victor Bethencourt News Editor: Lisa Melton Associate Editors: Markus Elsner (Research), Craig Mak (Research) Editor-at-Large: John Hodgson Contributing Editors: Mark Ratner, Chris Scott Contributing Writer: Jeffrey L. Fox Senior Copy Editor: Teresa Moogan Managing Production Editor: Ingrid McNamara Senior Production Editor: Brandy Cafarella Production Editor: Amanda Crawford Senior Illustrator: Katie Vicari Illustrator/Cover Design: Kimberly Caesar Senior Editorial Assistant: Ania Levinson

© 2010 Nature America, Inc. All rights reserved.

MANAGEMENT OFFICES NPG New York 75 Varick Street, Fl 9, New York, NY 10013-1917 Tel: (212) 726 9200, Fax: (212) 696 9006 Publisher: Melanie Brazil Executive Editor: Linda Miller Chief Technology Officer: Howard Ratner Head of Nature Research & Reviews Marketing: Sara Girard Circulation Manager: Stacey Nelson Production Coordinator: Diane Temprano Head of Web Services: Anthony Barrera Senior Web Production Editor: Laura Goggin NPG London The Macmillan Building, 4 Crinan Street, London N1 9XW Tel: 44 207 833 4000, Fax: 44 207 843 4996 Managing Director: Steven Inchcoombe Publishing Director: Peter Collins Editor-in-Chief, Nature Publications: Philip Campbell Marketing Director: Della Sar Director of Web Publishing: Timo Hannay NPG Nature Asia-Pacific Chiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843 Tel: 81 3 3267 8751, Fax: 81 3 3267 8746 Publishing Director — Asia-Pacific: David Swinbanks Associate Director: Antoine E. Bocquet Manager: Koichi Nakamura Operations Director: Hiroshi Minemura Marketing Manager: Masahiro Yamashita Asia-Pacific Sales Director: Kate Yoneyama Asia-Pacific Sales Manager: Ken Mikami DISPLAY ADVERTISING [email protected] (US/Canada) [email protected] (Europe) [email protected] (Asia) Global Head of Advertising and Sponsorship: Dean Sanderson, Tel: (212) 726 9350, Fax: (212) 696 9482 Global Head of Display Advertising and Sponsorship: Andrew Douglas, Tel: 44 207 843 4975, Fax: 44 207 843 4996 Asia-Pacific Sales Director: Kate Yoneyama, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746 Display Account Managers: New England: Sheila Reardon, Tel: (617) 399 4098, Fax: (617) 426 3717 New York/Mid-Atlantic/Southeast: Jim Breault, Tel: (212) 726 9334, Fax: (212) 696 9481 Midwest: Mike Rossi, Tel: (212) 726 9255, Fax: (212) 696 9481 West Coast: George Lui, Tel: (415) 781 3804, Fax: (415) 781 3805 Germany/Switzerland/Austria: Sabine Hugi-Fürst, Tel: 41 52761 3386, Fax: 41 52761 3419 UK/Ireland/Scandinavia/Spain/Portugal: Evelina Rubio-Hakansson, Tel: 44 207 014 4079, Fax: 44 207 843 4749 UK/Germany/Switzerland/Austria: Nancy Luksch, Tel: 44 207 843 4968, Fax: 44 207 843 4749 France/Belgium/The Netherlands/Luxembourg/Italy/Israel/Other Europe: Nicola Wright, Tel: 44 207 843 4959, Fax: 44 207 843 4749 Asia-Pacific Sales Manager: Ken Mikami, Tel: 81 3 3267 8765, Fax: 81 3 3267 8746 Greater China/Singapore: Gloria To, Tel: 852 2811 7191, Fax: 852 2811 0743 NATUREJOBS [email protected] (US/Canada) [email protected] (Europe) [email protected] (Asia) US Sales Manager: Ken Finnegan, Tel: (212) 726 9248, Fax: (212) 696 9482 European Sales Manager: Dan Churchward, Tel: 44 207 843 4966, Fax: 44 207 843 4596 Asia-Pacific Sales & Business Development Manager: Yuki Fujiwara, Tel: 81 3 3267 8765, Fax: 81 3 3267 8752 SPONSORSHIP [email protected] Global Head of Sponsorship: Gerard Preston, Tel: 44 207 843 4965, Fax: 44 207 843 4749 Business Development Executive: David Bagshaw, Tel: (212) 726 9215, Fax: (212) 696 9591 Business Development Executive: Graham Combe, Tel: 44 207 843 4914, Fax: 44 207 843 4749 Business Development Executive: Reya Silao, Tel: 44 207 843 4977, Fax: 44 207 843 4996 SITE LICENSE BUSINESS UNIT Americas: Tel: (888) 331 6288 Asia/Pacific: Tel: 81 3 3267 8751 Australia/New Zealand: Tel: 61 3 9825 1160 India: Tel: 91 124 2881054/55 ROW: Tel: 44 207 843 4759

[email protected] [email protected] [email protected] [email protected] [email protected]

CUSTOMER SERVICE www.nature.com/help Senior Global Customer Service Manager: Gerald Coppin For all print and online assistance, please visit www.nature.com/help Purchase subscriptions: Americas: Nature Biotechnology, Subscription Dept., 342 Broadway, PMB 301, New York, NY 100133910, USA. Tel: (866) 363 7860, Fax: (212) 334 0879 Europe/ROW: Nature Biotechnology, Subscription Dept., Macmillan Magazines Ltd., Brunel Road, Houndmills, Basingstoke RG21 6XS, United Kingdom. Tel: 44 1256 329 242, Fax: 44 1256 812 358 Asia-Pacific: Nature Biotechnology, NPG Nature Asia-Pacific, Chiyoda Building, 2-37 Ichigayatamachi, Shinjuku-ku, Tokyo 162-0843. Tel: 81 3 3267 8751, Fax: 81 3 3267 8746 India: Nature Biotechnology, NPG India, 3A, 4th Floor, DLF Corporate Park, Gurgaon 122002, India. Tel: 91 124 2881054/55, Tel/Fax: 91 124 2881052 REPRINTS [email protected] Nature Biotechnology, Reprint Department, Nature Publishing Group, 75 Varick Street, Fl 9, New York, NY 10013-1917, USA. For commercial reprint orders of 600 or more, please contact: UK Reprints: Tel: 44 1256 302 923, Fax: 44 1256 321 531 US Reprints: Tel: (617) 494 4900, Fax: (617) 494 4960

E d i t o r ia l

Wrong numbers? With biotech infiltrating multiple industries and fewer life science ventures listing on stock exchanges, what do we really learn from surveying the set of public biotech companies?

© 2010 Nature America, Inc. All rights reserved.

E

ach year, Nature Biotechnology trawls through the accounts of publicly quoted biotech companies and pulls out some numbers that characterize this part of the commercial life science landscape. Perhaps the most surprising statistic this year was that most of the companies that appeared in last year’s survey are still there. The current straitened circumstances took their toll, of course, but total revenues were up 10%, R&D was only down 4% and the group collectively was profitable for another year. But what, if anything, does the survey tell us about the general health of the innovative life science sector? Back in the 1990s, the answer seemed clear. Thanks to much freer flows of capital then, the annual audit measured the progress of a specialized, self-reliant and relatively independent industrial endeavor. It assessed the rapid churn of companies listing newly on exchanges. Companies could float much earlier; some were even able to go public without products in human trials. Buoyant stock markets took valuations to ecstatic heights and poured money into the sector. Product for product and dollar for dollar, biotech companies were valued much more highly than ‘traditional’ pharma companies. That differential was unsustainable. As Amgen and Genentech and Biogen Idec and others climbed up the pharmaceutical league standings, reality dawned. Innovators metamorphosed into drugmakers. And as the pharma sponge absorbed more biotech, the boundaries between the two spheres faded. The consequence of this merging is that much, if not most, of the biological products and biological techniques now resides outside the group of independent public companies that we survey. Pharma spends $65 billion a year on R&D, 25–40% of it either devoted to biological products or using the techniques of biotech. Thus, pharma outspends ‘biotech,’ even on biotech R&D. Furthermore, biotech processes extend far beyond the pharmaceutical segment: political imperatives and technological capability have expanded industrial biotech for biofuels production, waste management and green chemistry. Geographically, biotech is no longer a Western province: China, India, South Korea and elsewhere are prominent actors in follow-on biologic drugs, diagnostics and clinical testing. Our public company survey reflects none of these changes: pharma companies, biogenerics firms, diagnostic and device providers all fall outside the definitions of our survey. In Asia, successful biotech companies (see p. 783) have only restricted access to mature public capital markets. Overall, the survey is now less a gauge for innovative life science and more a pointer to the shape of the Western healthcare market. To measure life sciences’ impact more broadly, other indicators are needed. To quantify innovation, we need to look, too, at activities within small private companies and, increasingly, at the early translational work in the public sector. These data are exponentially more difficult to gather than nature biotechnology volume 28 number 8 august 2010

data from publicly quoted firms. Accordingly, policymakers, governments and industry associations need to devote much more effort and resources to collecting them.

MAQC-II: analyze that! The MAQC consortium’s latest study suggests that human error in handling DNA microarray data analysis software could delay the technology’s wider adoption in the clinic.

F

ollowing up on its publications in Nature Biotechnology four years ago (http://www.nature.com/nbt/focus/maqc/index.html), the Microarray Quality Control (MAQC) consortium publishes the results of its second phase of assessment (MAQC-II) on p. 827, in conjunction with ten accompanying papers in The Pharmacogenomics Journal (http://www.nature. com/tpj/journal/v10/n4/index.html). The new work assesses the capabilities and limitations of microarray data analysis methods—so-called genomic classifiers—in identifying gene signatures representative of a specific pathological condition. All in all, >30,000 genomic classifier models were built by combining one of 17 different data preprocessing and normalization methods, with one of 9 methods for filtering out problematic data, with one of >33 techniques for picking ‘signature’ genes, with one of >24 algorithms for discerning patterns from those genes, and with one of 6 methods for testing the robustness of the results. Thirty-six research teams sought gene signatures within 6 massive microarray datasets derived from toxicological studies of chemicals on rodents and expression profiles of human cancer patients that predict 13 ‘endpoints’ potentially relevant to preclinical or clinical applications. As discussed on p. 810, one key finding of MAQC-II is that the classifier models are remarkably similar in predicting outcome, irrespective of the approach used. On the other hand, the overall success of the classifiers in predicting endpoints depends on the endpoints themselves. For example, predictions were in general much worse for breast cancer and multiple myeloma, which have highly heterogenous genetic backgrounds, than for liver toxicology or neuroblastoma. Perhaps most striking of all, some data analysis teams were consistently better at predictions than others. This may relate to simple errors associated with manipulating such large datasets. But insufficient tuning of the parameters used in a classifier model is also a likely contributor. In this sense, MAQC-II was as much an exercise in sociology as in technology. The human element in classifier implementation is key. Thus a key take-home message is that classifier protocols need to be more tightly described and more tightly executed. In this respect, regulatory agencies and scientific journals can promote good practice. A clear need exists for greater meticulousness both in documenting the parameters of a particular classifier model used and in detailing the procedures for normalization, batch effect correction, quality control and reduction of quality control flaws. Greater attention to detail will not only enhance reproducibility of research—it will also facilitate the progression of this technology toward the clinic. 761

news in this section Investigational cancer agents tested in pairs

Transparency rules challenge small firms p767

GM soybeans for trans fat–free oil p769

p765

© 2010 Nature America, Inc. All rights reserved.

Industry makes strides in melanoma After decades of continuous failures, the treatment of metastatic melanoma is finally advancing. This year’s American Society for Clinical Oncology (ASCO) annual meeting heralded a breakthrough antibody therapy for the disease. Top-line, phase 3 results for Bristol-Myers Squibb’s humanized monoclonal antibody (mAb) ipilimumab showed a survival benefit in patients with advanced cancer—the first ever phase 3 trial to do so. These results contrast with a litany of letdowns from cancer vaccines, cytokine therapies, adoptive T-cell therapies as well as several targeted therapies that all have failed to improve on standard chemotherapy, which itself achieves a meager 15% response rate with negligible survival benefit. “Those of us in the melanoma business have felt like we’ve been in a long, dark tunnel,” said oncologist Vernon Sondak, of the H. Lee Moffitt Cancer Center in Tampa, Florida, at the ASCO meeting. The ipilimumab data, released by New York–based Bristol-Myers Squibb in June, have changed all that. The 676 individuals included in the study had unresectable, metastatic melanoma and had previously undergone chemotherapy for the disease. Those receiving ipilimumab, with or without the synthetic peptide vaccine glycoprotein 100 (gp100), had a median survival of about 10 months, against 6.4 months for the vaccine alone. Ipilimumab, which targets cytotoxic T-lymphocyte antigen 4 (CTLA4), nearly doubled the rates of survival at 12 months (46% versus 25%) and 24 months (24% versus 14%) after treatment compared with the peptide. “This is really a benchmark for the field,” says John Kirkwood, a melanoma researcher at the University of Pittsburgh. “We finally have a randomized controlled trial that is positive.” Finalized phase 1 results of a BRAF inhibitor, developed by the Berkeley, California–based Plexxikon, are at least as dramatic. The small molecule PLX4032 (also RG7204), which Plexxikon is co-developing with Roche of Basel, specifically inhibits the V600E mutant BRAF, a constitutively active kinase present in more than half of metastatic melanomas. The drug produced an 81% response rate among 32 patients receiving the therapeutic dose. “The early effects are [as] profound, reliable and gratifying as

B7

Antigenpresenting cell

Ipilimumab

MHC

Ag

CTLA4

TCR

B7

T cell

CD28

T cell activated

Figure 1 Ipilimumab stimulates antitumor immunity by blocking CTLA4, a natural brake on T cells, and allowing their unimpeded ‘costimulation’. Ipilimumab is the first agent to extend survival in metastatic melanoma patients in phase 3.

one could ever want out of a cancer therapy,” says trial principal investigator Keith Flaherty of Massachusetts General Hospital in Boston. PLX4032 is now in phase 3. Although both compounds will almost certainly become approved drugs, they have limitations. Ipilimumab extends median survival but, strangely, has only an 11% overall response rate. And almost all patients on PLX4032 relapse, most within a year. Nevertheless, the two drugs have revitalized melanoma research. By using ipilimumab and PLX4032 in combination with a variety of standard and investigational agents —or with each other—researchers hope to push long-term survival of metastatic melanoma patients up from the roughly 10% combined cure rate now achievable with ipilimumab monotherapy and interleukin-2 (IL-2) monotherapy. “We’re going to move the cure rate of melanoma progressively up,” predicts melanoma researcher Mario Sznol, of Yale University in New Haven, “to what could be a very respectable 30, 35, 40% of patients, over the course of the next several years.”

nature biotechnology volume 28 number 8 AUGUST 2010

Anti-CTLA4 therapy has succeeded where other immunotherapies failed because, instead of trying to indirectly stimulate T cells by presenting tumor antigen to overcome immune tolerance, it activates T cells directly, by disabling a brake on T-cell activity. Normally, when a T cell is activated after CD28 binding of the B7 receptor on antigen-presenting cells, CTLA4 acts as a brake, trafficking from the T-lymphocyte cytosol to the surface to bind B7 molecule with high affinity. Thus CTLA4 turns the T cell off. When the ipilimumab mAb is present it blocks CTLA4, keeping the T lymphocyte activated. The mAb also promotes unfettered binding of the T-cell CD28 receptor to the antigen-presenting cell receptor B7, together with antigen presentation to the T-cell receptor (Fig. 1). Such ‘co-stimulation’ is necessary for T-cell activation, and antitumor immunity. Unfortunately, ipilimumab also triggers autoimmune side effects, some severe. A few patients have died from colitis-related bowel perforations, for example. But Kirkwood points out, “[for] the vast majority of patients, we can 763

NEWS

© 2010 Nature America, Inc. All rights reserved.

Table 1 Selected phase 3 trials in metastatic melanoma Company (location)

Product

Description

Bristol-Myers Squibb

Ipilimumab (MDX-010)

Fully human antibody targeting the CTLA-4 receptor on T cells

Plexxikon/Roche

PLX4032

Small-molecule inhibitor of V600E mutant BRAF kinase

Abraxis Bioscience (Los Angeles)

Abraxane Nanoparticle albumin-bound paclitaxel (Taxol) (nab-paclitaxel, ABI-007)

Eli Lilly (Indianapolis)

Tasisulam (LY573636)

Acyl sulfonamide, generates reactive oxygen species and induces apoptosis

Biovex OncoVEX (Woburn, Massachusetts)

Oncolytic herpes simplex virus type-1 encoding granulocyte macrophage colony stimulating factor; selectively replicates in tumor cells, recruits dendritic cells

Novartis (Basel)

Tasigna (nilotinib, AMNN-107)

Small molecule oral c-kit kinase inhibitor for c-kit mutant melanoma

GlaxoSmithKline

Astuprotimut-r (MAGE-A3 ASCI)

Protein subunit vaccine based on melanoma-associated antigen A3 (MAGE-A3), specific for tumor cells

Vical (San Diego)

Allovectin-7

DNA plasmid/lipid complex containing human leukocyte antigen B7 and beta-2 microglobulin DNA sequences that together form major histocompatibility class I; improves antigen presentation

Source: BioMedTracker & Nature Biotechnology

manage the side effects fairly easily, once you know how to look for them.” The one controversy in the phase 3 trial was the choice of the gp100 peptide vaccine, developed by the Bethesda, Maryland–based National Cancer Institute, as the active control arm for the study. The combination of this HLAA0201–restricted peptide vaccine with highdose IL-2 resulted in higher response rates and improved progression-free survival in an earlier randomized trial. Thus the choice of gp100 for the control arm. Some researchers speculate that the vaccine may have hurt patients, thus giving ipilimumab an artificial statistical boost. (Certain vaccines have reduced survival in melanoma trials). Kirkwood disagrees, because gp100 did not appear to cause harm in its other trials. “The issues regarding the control are, in my book, non-issues,” he says. The question remains, why did ipilimumab succeed whereas tremelimumab, a similar antiCTLA4 antibody from Pfizer, failed? It is possible that tremelimumab didn’t really fail. “[Pfizer] analyzed the trial early,” says Sznol. “You need to wait for the events to develop.” Sznol points out that some patients treated with anti-CTLA4 mAbs experience progression of their cancers initially, followed by regression, and that other patients have most of their lesions disappear while a few continue growing. All are classified as nonresponders, but some may live for a long time. It’s also possible, Sznol says, that the company used the wrong drug dose and schedule. Kirkwood agrees that Pfizer was probably too quick to analyze the data. Pfizer defended the tremelimumab phase 3 trial dose and schedule in an e-mail, noting that phase 2 results (using the same dose and schedule as in phase 3) were very similar to ipilimumab’s despite the different dose regi764

mens. Long-term phase 3 follow up did show a survival advantage for the tremelimumab arm, but not enough to justify US Food and Drug Administration registration. Many patients in the tremelimumab trial control arm went on to receive ipilimumab in a compassionate use program, which could have decreased tremelimumab’s apparent effect. So circumstances, not biology, may have defeated tremelimumab. Any lingering ipilimumab doubts may disappear with a second completed phase 3 trial, comparing ipilimumab plus dacarbazine chemotherapy to dacarbazine alone. Patient accrual ended more than two years ago, and results have not yet been reported. The delay suggests to many a successful trial, but no one knows for sure. No efficacy doubts exist for PLX4032. All agree the drug works, and works quickly, in the vast majority of patients with mutant BRAF tumors. Because PLX4032 targets the mutant form of the protein encoded by the BRAF oncogene, this allows very high doses to be given without adverse effects on normal cells. Data from several groups show, in fact, that PLX4032 paradoxically activates BRAF signaling in normal cells. This pathway activation enhances the therapeutic window, but also probably leads to the appearance of skin lesions known as keratoacanthomas in many patients. They are benign, but raise the theoretical possibility that longterm treatment could promote the growth of other cancers. But the main downside of PLX4032 is relapses. Median duration of response in phase 1 was about nine months. By historical standards, this is excellent, and a few patients have had complete responses lasting two years or more (they remain on the drug). But the relapses indicate a still-unknown form of drug

resistance. Some residual BRAF signaling in tumor cells persists, despite treatment, and there are new data that the mitogen-activated protein (MAP) kinase signaling pathway is reactivated downstream of BRAF. In either case, combining a BRAF inhibitor with an inhibitor of MAP kinase kinase (MEK), which is immediately downstream of BRAF, could overcome resistance and prolong survival. Such a trial is now underway with GSK2118436—a small-molecule inhibitor of the V600E mutant BRAF—and MEK inhibitor GSK1120212, both from GlaxoSmithKline in London and soon to be in phase 2/3 studies. Meanwhile, PLX4032 is moving forward quickly. An already completed phase 2 trial will “we all believe … likely be enough for FDA approval next year,” says Flaherty. Phase 3 will definitively show whether PLX4032 changes the natural history of the disease and extends survival. The list of agents in phase 3 trials is growing (Table 1), although none of them displayed the efficacy of ipilimumab and PLX4032 in phase 2. One comparable compound, however, is Bristol-Myers Squibb’s humanized anti-PD-1 mAb, MDX-1106. PD-1, or programmed cell death-1, is a T-cell molecule that, like CTLA4, downregulates T-cell activity. It appears to be at least as powerful as CTLA4, and may function at the later stages of the immune response to shut down T cells. In phase 1, MDX-1106 treatment led to 15 confirmed responses among 46 metastatic melanoma patients. As of June, none of the responders had relapsed, with more than a year passing in several cases. “This is one of the most promising starts I’ve seen for any drug,” said Sznol, the trial’s principal investigator. “It’s the kind of thing where we can’t sleep because we want to offer this to our next patient.” Autoimmune side effects occur, but fewer than with ipilimumab. A combination trial with ipilimumab has begun (see p. 765). The most anticipated combination is ipilimumab and PLX4032. This would bring together the quick responses of PLX4032 with ipilimumab’s ability to deliver cures. “The two are made for one another,” says Kirkwood. Tumor cells killed by PLX4032 should release antigen, enhancing ipilimumab’s ability to activate antitumor T cells. Flaherty says that the two sponsoring companies have agreed to collaborate on a large randomized combination trial, which should begin next year. Individually, ipilimumab and PLX4032 have ended the futility and nihilism that have long dominated melanoma treatment. It will take time to sort out the best combinations and the best way to apply them. “But at least the cupboard is not bare any more,” said Sondak. Ken Garber Ann Arbor, Michigan

volume 28 number 8 AUGUST 2010 nature biotechnology

news

The next generation of cancer treatments Merck’s ridaforolimus-dalotuzumab program, could be approved in pairs, at least judging which is due to enter phase 2 trials later this by a growing trend among drug makers to year, is a key initiative and is being closely combine drugs early in development and the scrutinized. It exemplifies a science-based US Food and Drug Administration’s (FDA) approach to combining investigational drugs willingness to regulate that may offer limited them. On 2 June, the potential as single agents, FDA opened its public but which may offer synerconsultation into the gistic effects when adminformulation of guidistered together, as well as ance for combinations of reducing the risk of drug investigational therapies. resistance. Trials of sevIn the same week, Merck, eral other combinations of of Whitehouse Station, new types of agents are also New Jersey, reported at underway (Table 1). the annual American Although combination Society of Clinical therapy in cancer—and Oncology meeting in other indications—is not Chicago that a combinaa new theme, it has develtion of ridaforolimus, an oped historically through oral inhibitor of mamtrial and error. “Our knowlmalian target of rapamyedge of biological pathways cin (mTOR) developed Tackling breast cancer. Drug and networks is so superfiwith Ariad of Cambridge, developers are starting to combine cial it really is hard to come Massachusetts, and dalo- novel, unapproved agents in search of up with a strong rationale,” tuzumab, an antibody synergistic activity. says Alan Ashworth, protargeting the insulinfessor of molecular biology like growth factor 1 receptor (IGFR1), led at the Institute of Cancer Research in London. to responses in a cluster of patients with The ridaforolimus-dalotuzumab combination highly proliferative, estrogen-receptor- emerged from an unbiased screen of a colon positive breast cancers in a phase 1b trial. cancer cell line in which individual genes were Collaborations between different sponsors to systematically switched off using short haircombine drugs very early in development are pin RNAs, whereas each of the two drugs was unusual and pose new issues for regulators tested in turn in a cell proliferation assay. This compared with oversight of combinations of kind of synthetically lethal screen can unveil agents already on the market. dependencies between related pathways and The FDA initiative is not limited to can- overcome compensatory mechanisms that cer—it also covers infection, seizure disor- cancer cells switch to when only one target is ders and cardiovascular disease. But cancer hit. “Those types of approaches couldn’t be drug makers, in particular, are grappling with done before,” says Eric Rubin, vice president some thorny questions as they attempt to of clinical oncology at Merck. The upcoming translate their rapidly expanding knowledge phase 2 trial will recruit around 200 breast of tumor biology into therapies that offer sig- cancer patients, who will be assigned to one nificant improvements on what is now avail- of four treatment arms, comprising either able. Foremost among their concerns is how ridaforolimus as monotherapy, dalotuzumab to accelerate clinical development to deliver as monotherapy, the two drugs in combinasolid efficacy data without compromising tion or exemestane, the active comparator. patient safety. “We’ve talked to the FDA about The key question is whether that kind of design specific combinations and have received guid- would need to be replicated in a large-scale regance on an ad hoc basis,” says Pearl Huang, vice istration trial of a new combination comprising president and oncology franchise integrator at two investigational compounds. “What we have Merck. “For us, the burning issue is if we dem- proposed—and others have as well—is to do this onstrate great activity for the combination, are in a more limited setting,” Rubin says. Balancing we obligated to demonstrate lack of activity for regulators’ requirements for statistical power the single agent alone?” with patients’ needs for effective therapy is not Some claim combinations of investigational a straightforward task, particularly if some trial drugs could accelerate clinical development. participants are to receive single agents that are Sebastian Kaulitzki/iStockphoto

© 2010 Nature America, Inc. All rights reserved.

Firms combine experimental cancer drugs to speed development

nature biotechnology volume 28 number 8 AUGUST 2010

765

NEWS

© 2010 Nature America, Inc. All rights reserved.

Table 1 Selected targeted experimental combination cancer therapies in development Company

Combination

Mechanism

Indication

Status

AstraZeneca (AZ)

Cediranib maleate (AZD2171) + olaparib

Vascular endothelial growth factor (VEGF) receptor inhibitor + poly(ADP-ribose) polymerase inhibitor

Recurrent ovarian cancer

Phase 1/2

AZ & Merck (Darmstadt, Germany)

Cediranib maleate + cilengitide

VEGF receptor inhibitor + integrin inhibitor

Recurrent glioblastoma

Phase 1b

GlaxoSmithKline (London)

GSK1120212 + GSK2141795

MEK inhibitor + Akt kinase inhibitor

Solid tumors

Phase 1b

Novartis & GlaxoSmithKline

BKM120 + GSK1120212

Phoshphoinositide-3-OH kinase inhibitor + MEK inhibitor

Solid tumors

Phase 1b

AZ & Roche (Basel)

Cediranib maleate + RO4929097 VEGF receptor inhibitor + γ-secretase inhibitor

Solid tumors

Phase 1

Bristol-Myers Squibb (New York) Ipilimumab + MDX-1106 & Ono Pharma (London)

Cytotoxic T-Lymphocyte antigen 4 (CTLA-4) inhibitor + Programmed Melanoma death-1 receptor (PD-1) inhibitor

Phase 1

Merck & Ariad

Dalotuzumab + ridaforolimus

Insulin-like growth factor receptor 1 (IGFR1) inhibitor + mTOR inhibitor

Neoplasms

Phase 1

Merck & AZ

MK-2206 + selumetinib

Akt inhibitor + MEK1/2 inhibitor

Solid tumors

Phase 1

Pfizer (New York)

Figitumumab + PF-00299804

IGFR1 inhibitor + HER tyrosine kinase inhibitor

Solid tumors

Phase 1

Pfizer

Crizotinib + PF-00299804

Met tyrosine kinase inhibitor + HER tyrosine kinase inhibitor

Non-small cell lung carcinoma

Phase 1

Roche

GDC-0449 + RO4929097

Hedgehog antagonist + γ-secretase inhibitor

Breast cancer Sarcoma

Phase 1 Phase 1/2

Source: http://www.ClinicalTrials.gov

unlikely to confer any benefit, while at the same time, the duration of combination trials is significantly extended. Ashworth says that more innovative trial designs and early use of biomarkers can help—but only if there is already a solid case for moving a particular therapy into the clinic in the first place. “You need a very strong biological basis for your combination treatment,” he says. “If you need 4,000 patients to prove your hypothesis, I’m sorry mate, you’ve got the wrong hypothesis.” There is some precedent for rapid approval of investigational therapies based on a strong phase 2 efficacy signal, particularly when it is backed by a solid understanding of the underlying biological mechanism. For example, Novartis, of Basel, gained FDA approval for Gleevec (imatinib mesylate) in chronic myeloid leukemia on the basis of a phase 1b dose-escalating trial (New. Engl. J. Med. 344, 1031–1037, 2001). “If in a phase 2 trial, you’ve figured out the right dose and the correct schedule for a combination, and you get a dramatic change in efficacy, for example in a directed patient population, a path for that combination could be very straightforward,” says Bill Sellers, global head of oncology research at the Novartis Institutes for Biomedical Research, in Cambridge, Massachusetts. Head-to-head studies against the existing standard of care would also smooth the path toward approval—and combination therapies, he says, should aim for curative levels of efficacy rather than small, incremental improvements. “A major change in the rate of complete response or partial response to a therapeutic says you’ve killed a lot of the cancer.” Many of the combinations being tested target different kinase enzymes. Merck’s Huang 766

says the combination of their investigational anti-cancer agent MK-2206, that inhibits Akt (a component of the phosphatiyliositol-3 kinase pathway), with London-based AstraZeneca’s selumetinib (AZD6244), an inhibitor of the enzyme MEK, was chosen because each target is part of a canonical signal transduction pathway, downstream from a receptor tyrosine kinase. “They’re in parallel, but they also cross-talk,” she says. “They are not the cancer’s mutational drivers, they’re more the downstream effectors.” Even so, insights into tumor biology do not always yield significant clinical benefits. “In oncology, what we think works and what [actually] works are two different things, and that’s why we need to do big studies,” says Justin Stebbing, a physician scientist based at Imperial College London. “The initial promise of biomarkers doesn’t hold up to scrutiny, ultimately.” Matthew Ellis, professor of medicine at Washington University in St. Louis, Missouri, who recently published genomic analyses of cancer and normal tissues taken from an individual with breast cancer (Nature, 464, 999–1005, 2010), has a different take: “My guess is we can solve the companion diagnostic problem by making full-genome sequencing of cancer the primary screen.” “We’re beginning to understand cancer genomes at a much more fundamental level than we ever have before,” he adds. “What we’re seeing, I think, is a great deal of complexity, much more complexity than was ever appreciated before.” This complexity is accompanied by an appreciable degree of heterogeneity—no two cancers appear the same. “We’re [starting] to classify them and put them into different buckets,” says James Zwiebel, chief of the

investigational drug branch at the National Cancer Institute, in Rockville, Maryland. “That’s really only scratching the surface. When you get down to it, every patient is going to have some unique characteristics.” That could make life more difficult for drug developers, he notes. This genome-level view of cancer, rather than the classic assumption of cancer as a disease affecting a particular organ, is turning our understanding of cancer on its head. Breast cancer perfectly illustrates the point. “When you do the genetics, what you see is a constellation of rare diseases,” Ellis says. In contrast, gastrointestinal stromal tumors, for example, seem to have a more uniform genetic profile. “You’ve got rare diseases defined by a common mutation, and we’re making progress,” he says. “We haven’t worked out how to handle the reverse situation, a common disease defined by multiple rare mutations.” Although the cost of individual genome sequencing is falling, Sellers says that full cancer genome sequencing may not be necessary to identify the dominant mutations that drive a particular cancer: partial approaches, based on techniques such as hybrid capture, targeted resequencing and high-throughput genotyping, may be sufficient. But even with the correct genomic information at hand, clinical progress will remain difficult, as combining two investigational agents correctly is not a straightforward task. “This is probably the biggest challenge: finding the effective and tolerated dose and, importantly, the schedule,” Sellers says. “I think this is probably a bigger challenge than the FDA regulatory challenge.” Cormac Sheridan Dublin

volume 28 number 8 AUGUST 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

news

FDA transparency rules could hit small companies hardest

in brief

The US Food and Drug Administration (FDA) is considering changing how much information it discloses about product applications—news that biotechs have greeted with a mixture of trepidation and hope. The agency is proposing to make publicly available ‘complete response’ and ‘refuse-to-file’ letters for drugs and ‘not approvable’ letters for devices. From opinions gathered in advance of the final decision, it seems the smallest biotechs stand to lose the most. The proposed changes are wide-reaching and include some things most experts agree are good. On the upside, they say, this is an opportunity to make more information about what FDA does available to the public and ensure that data sources are more user-friendly. The downside, however, is the proposal to disclose information early in the approval process, including Investigational New Drug (IND) applications, holds and IND withdrawals. Few can see how revealing more information at the product application stage can be reconciled with trade secrets protection. The Biotechnology Industry Organization (BIO) wants more details about how these proposed regulations would be implemented. “They [FDA] define trade secrets [in the document], but oddly there is no definition of what constitutes competitive information,” explains Andrew Emmett, director for science and regulatory affairs at BIO, based in Washington, DC. The organization also wants clarification around who will decide what remains secret. Under current Freedom of Information Act regulations, Emmett says, companies have five days to determine whether documents that are going to be made public contain trade secrets that should be redacted. “We need to know exactly what the role of the sponsor will be in deciding what information is going to be shared,” he says. Otherwise, companies could be put at competitive disadvantage or become victims of wild speculation. The confidentiality issue is particularly critical for small biotechs. “When a small public company has a clinical trial pending, hedge fund managers do everything they can to get a sense of what the outcome might be,” says Alan Mendelson, senior partner at Los Angeles– headquartered law firm Latham and Watkins. If every pause in the clinical trial process gets announced to the public, it could lead to stock trading based on misleading or inadequate information. “It’s bad enough today,” he says, “But at least now people are commenting on

The US Supreme Court has ruled on a long-awaited and controversial patent litigation case, a decision greeted with relief by the biotech industry but vague enough that both sides can Biotech welcomes claim victory. The ruling. Bilski v. Kappos case was closely watched by the biotech community after the US Court of Appeals for the Federal Circuit ruled in 2008 that only methods tied to a machine or transformed into a different state are patentable, a standard which appeared to exclude crucial aspects of medical diagnostics. Commentators feared a restrictive ruling could have severely limited the ability to obtain patents on methods that use genes, proteins and metabolites to diagnose disease. Instead, the Supreme Court struck down patent claims on narrow grounds. “The Court was clearly conscious of the potential negative and unforeseeable consequences of a broad and sweeping decision,” stated Washington, DC–based Biotechnology Industry Organization president and CEO Jim Greenwood. The court ruled on two issues. First, it ruled against patenting only those inventions that are “tied to a particular machine” or those that transform “a particular article into a different state or thing.” Second, the court held that the word “process” as used in the US Patent Act should be read broadly to include modern day inventions. The ruling does not address the eligibility of patents for diagnostic methods, however, which leaves a number of questions unanswered with regard to a string of pending cases, including the closely watched dispute against Myriad Genetics and its breast cancer gene patents. Dan Ravicher of the Public Patent Foundation, a co-plaintiff with the American Civil Liberties Union in the suit against Myriad Genetics, believes “the opinion reinforces the line of case law that Judge Sweet relied upon in his decision striking down gene patents [in the Myriad case]. It rejects the argument that ‘anything’ is patentable.” Justices Stevens, Breyer, Ginsburg and Sotomayor would have struck down not only the specific Bilski business method claims, but all business method patents on historical grounds that this class of patents was never contemplated by the framers of the US Constitution. The same argument would be difficult to support in biotech-specific cases as there is ample evidence that Thomas Jefferson, who reformed the Patent Act of 1793, considered medicine a “useful art” as was originally stated, a language later changed to “process.”  Kenneth Chahine & Javier Mixco

nature biotechnology volume 28 number 8 AUGUST 2010

Lee Pettet/istockphoto

definitive data, not just a signal that might prove to be nothing.” Wayne Kubick, a vice president in safety at Waltham, Massachusetts–based PhaseForward, says companies with “limited products” are also going to be at greatest risk of competitive disadvantage. Competitors will be able to use some types of information better than others. Says Gregory Conko, senior fellow at the Competitive Enterprise Institute in Washington, DC, “It’s less important with complete response or rejection letters, but with a new drug application, a hold, or a withdrawal, that is where tipping off competitors is a much bigger concern.” Smaller companies are already at a disadvantage in the review process. In comments it filed in April, BIO pointed out that a recent study from the law firm Booz Allen Hamilton found that small firms had only a 48% first-cycle approval rate for products in the priority review category, compared with a 78% rate for larger companies. In a survey of 168 of its members (http://www. bio.org/letters/20100412b.pdf), BIO also found that “early, frequent and explicit communication with the FDA” was felt to be the most helpful means for first-time filers to improve their success rates. The transparency initiative could help shore up this communication weakness. “A variety of leaders have been pushing for more open and straightforward dialog with the agency for years,” says J. Donald deBethizy, president and CEO of Winston-Salem, North Carolina–based Targacept. “This initiative could provide a means for that.” Greater transparency could also put pressure on FDA to provide rationales for rejections, which critics charge are sometimes based on “petty” issues, according to Conko. Overall, such changes may not necessarily translate to better decision making, Conko warns. “FDA’s political incentives are still poorly aligned. Even when their rationale is weak, they still don’t have to pay a price for it,” he says. On the other hand, transparency is not necessarily a bad thing. “The world is very different already in 2010” says Kubick. “We have clinicaltrials.gov and a lot of other information already available.” But it means companies will face more instances where study data is used out of context. “You have to protect yourself against people who data mine and then hold up a little data nugget as the truth,” deBethizy says. Many are watching closely as the next phase of the initiative rolls out. “This is by no means a done deal,” says Kubick. “Some [of the proposed] things are going to happen, but not everything will.” Others are very skeptical, like Jack McLane,

Supremes rule on Bilski

767

NEWS

in brief Lawsuits rock Jackson

768

JASON REED/Reuters/Corbis

© 2010 Nature America, Inc. All rights reserved.

Lee Pettet/istockphoto

The Jackson Laboratory has unwittingly found itself ensnared in patent disputes. In June, the nonprofit laboratory mouse developer located in Bar Harbor, Maine, was cleared of a patent Litigation over models infringement may inflate prices. allegation—the first in the laboratory’s 80-year history—and now faces a second allegation by another party. Jackson’s mission of making its repository of more than 5,000 mouse strains available to researchers at affordable prices could be challenged if it is forced to continue defending itself in expensive lawsuits, says David Einhorn, the laboratory’s in-house attorney. In Jackson’s first scuffle, the Central Institute for Experimental Animals (CIEA), a Kawasaki, Japan–based nonprofit, in 2008 sued Jackson for distributing a mouse model particularly useful for grafting human tissue. Both groups in the 1990s separately developed these immunodeficient mice by starting with a strain of nonobese diabetic mouse (NOD), crossing those with mice carrying the scid mutation for immunodeficiency, and crossing them again with mice whose gene for a key immune signaling molecule, interleukin-2 receptor γ, was knocked out. Jackson has distributed the mouse to more than 1,000 research groups worldwide, says Einhorn. But the laboratory didn’t patent its mouse, whereas CIEA did. On June 1, a US District Court judge ruled that the Jackson Laboratory had not infringed CIEA’s patent. What ultimately swayed the judge to side with Jackson was that the CIEA, in its patent application, described the mouse but didn’t claim it. In his decision the judge cited the Guidelines for Nomenclature of Mouse and Rat Strains, which state mice inbred for more than 20 generations can be considered a different strain, and Jackson’s mouse line had been separately inbred many times. Michael Rader, attorney with Wolf, Greenfield & Sacks in Boston, who represented Jackson, says this was likely the first time nomenclature rules have been used to help decide a lawsuit. Now Jackson faces another lawsuit involving transgenic mice with mutations useful in Alzheimer’s disease research. The Alzheimer’s Institute of America in February sued Jackson and six biotech and pharma companies for patent infringement. Despite the high costs of the two lawsuits, Einhorn says Jackson won’t alter its mission of making laboratory mice accessible. But he notes that if the suing trend continues, “the most obvious way to recoup the costs is to charge more for mice.” He adds: “That falls on the backs of scientists who do the research.” Emily Waltz

The FDA’s Transparency Task Force is proposing to increase access to the agency’s decision letters about products or drugs. Such a move would challenge small biotech.

vice president of clinical and regulatory affairs at Hudson, Massachusetts–based Clinquest. McLane points out that releasing more data earlier will also stretch the agency’s resources because there will be pressure to analyze many more signals quickly and thoroughly. “It’s a tremendous overreach,” he says. “A lot of people do not think this will go through.” McLane says he’d rather see the agency bring their transparency rules in line with the Sarbanes-Oxley Act of 2002, which set new standards for US boards, management and accounting firms. “A lot of what the FDA is asking for here is competitive information,” he says.

The agency was accepting comments through July 20. In the autumn, the task force will consider the public comments as well as the “priority, operational feasibility, and resource requirements” of each proposal, according to Afia K. Asamoah, director of the FDA’s transparency initiative. BIO submitted one set of comments in April, and Emmett says the group will submit more before the deadline. Even if the agency decided to go through with all the proposals, though, some of the changes could not be implemented without new legislation. Malorye Allison Acton, Massachusetts

in their words “They have grown so fast and so suddenly that people are still skeptical. But we should get used to it.” Rasmus Nielsen, a geneticist at the University of California at Berkeley, who collaborates with Chinese colleagues, on China’s sudden boom in sequencing output. (The Washington Post, 28 June 2010) “Until the capacity issues can be addressed, this will not be an effective agent.” Chris Logothetis, head of prostate cancer research at the University

of Texas MD Anderson Cancer Center in Houston, on the year-long wait patients currently face for Dendreon’s prostate cancer vaccine Provenge. (Pharmalot, 28 June 2010) “Everyone can claim victory, except of course Mr. Bilski himself.” Dan Ravicher, of the Public Patent Foundation, the organization leading the attack on Myriad, on the Supreme Court’s decision in Bilski v. Kappos. (GoozNews, 28 June 2010) “Now that the full integration has taken place, it’s the Genentech guys who are being promoted and getting the key positions.” Allianz Global Investors’ Joerg de Vries-Hippen on how Genentech is the strongest in the marriage with Roche. (Bloomberg Businessweek, 1 July 2010)

volume 28 number 8 AUGUST 2010 nature biotechnology

news

The US Department of Agriculture (USDA) market has gone from 76% in 2005 to 64% has approved for environmental release one of today, according to the US Census Bureau. the first biotech crops aimed at the food indus- “We hope to recapture that space [for soytry. The new crop, a genetically modified soy- beans],” says Pioneer’s Russ Sanders, director bean with an altered fatty acid profile, yields oil of enhanced oils. that is more stable at high frying temperatures Pioneer’s new soybean oil has an oleic fatty and has a longer shelf life than commodity soy- acid content of >75%, a property that gives it bean oil. It was developed by Pioneer Hi-Bred frying and shelf stability comparable to that in Johnston, Iowa, a Dupont company. The of palm, high oleic acid canola and hydrogecompany received marketing approval for the nated soybean oils. It also contains 20% less biotech soybean in June and aims to commer- saturated fat than commodity soybean oil. cialize it by 2012. St. Louis–based Monsanto Pioneer dubbed the crop “Plenish high-oleic is following close behind, with two soybean soybeans.” Overproduction of oleic acid and products with modified oil profiles in its pipe- decreased levels of linoleic and linolenic acids line. in Plenish arise from The new soybean transgenic expression traits may help the of a fragment of the biotech industry soybean microsomal deliver on a twoomega-6 desatudecade-long promise: rase gene (FAD2-1) to develop crops with under the control improved nutritional of soybean Kunitz value. Until now, trypsin inhibitor most commercialized gene promoter, which biotech crops have silences endogenous been engineered with The success of Pioneer’s recently approved soy omega-6 desatusuch traits as pest bean, which has been engineered to cut down on rase. The transgenic trans fats, will depend on how well it is received resistance and herbi- by the food industry. soybean also carries cide tolerance—traits the S-adenosyl-lthat mostly benefit methionine synfarmers rather than the food industry or con- thetase as a marker to enable initial selection sumers. “Heat stability and longer shelf life: in the laboratory by acetolactate synthase these are the things that can light up the food (ALS)-inhibiting herbicide. industry, not reduced pesticides,” says Tom The success of the Plenish soybean will Hoban, a professor of food science at North depend on how well it is received by the food Carolina State University in Raleigh. industry. Pioneer has already set up testing Pioneer is marketing its new soybean oil as agreements with a dozen undisclosed food an alternative to partially hydrogenated veg- companies, says Sanders. The companies will etable oils. For decades, food producers have run consumer taste tests, frying tests and shelf relied on partially hydrogenated soybean oil life tests—just about anything a food company because it retains its flavor at high cooking would normally do with a new ingredient. temperatures and for extended periods on the Food companies can already choose from an grocery store shelf. But the process of partial array of oils with modified fatty acid contents hydrogenation produces trans fatty acids, or developed with conventional breeding. “The trans fats, which are known to increase ‘bad’ hard reality will be how producers of liquid low-density lipoprotein (LDL) cholesterol and vegetable oils compete,” says Terry Etherton, increase risk of coronary heart disease. professor of animal nutrition at Penn State in In 2006, the US Food and Drug University Park, Pennsylvania. Administration began requiring food manuFood industry representatives say they welfacturers to label food with trans fats, and come the new oil option, but see it as a “trial measures to alert the public of the health risks situation,” says Jeffrey Barach, vice president of trans fats ensued. Food producers turned of science policy at Grocery Manufacturers to alternatives, such as palm oil and certain Association in Washington, DC .“Each comkinds of canola oil, that have more stable fry- pany has to try it out and do some experimening and shelf life characteristics than those tal work,” he says. of unhydrogenated soybean oil. As a result, Although Pioneer received the full go-ahead soybean oil’s share of the edible fats and oils from regulators, the company doesn’t plan to John Lee/iStockphoto

© 2010 Nature America, Inc. All rights reserved.

Food firms test fry Pioneer’s trans fat–free soybean oil

nature biotechnology volume 28 number 8 AUGUST 2010

in brief Anti-CD20 patent battle ends On June 1, a four-year dispute over a European patent for anti-CD20 drugs to treat rheumatoid arthritis came to an end, with Seattle-based Trubion winning the dispute. This result frees up the space for anyone with a CD20 program, says Jeff Pepe, associate general counsel at Trubion. Multiple oppositions had been filed against the patent (European Patent 1176981) held jointly by Genentech of S. San Francisco, California, and Biogen Idec of Cambridge, Massachusetts. Trubion was joined by MedImmune, GenMab, Centocor, the Glaxo Group and Merck Serono, all pursuing anti-CD-20 programs at one time. In 2008, the Opposition Division of the European Patent Office ruled that, as filed, the patent did not meet the necessary requirements, favoring Trubion. Genentech and Biogen appealed in 2009. Finally, at an oral hearing this June, the original ruling was upheld, and no further appeals will be allowed. Ironically, around the time of the hearing, New York–based Pfizer, which acquired Trubion’s CD20 programs when it bought Wyeth in 2009, announced they would drop Trubion’s lead anti-CD20 compound (TRU015) though retaining the biotech’s second generation anti-CD20 monoclonal antibody also in rheumatoid arthritis. For Genentech/Roche “the decision does not impact our expectations with respect to protection against Rituxan [rituximab, anti-CD20 chimeric monoclonal antibody],” says company spokesperson Rubin Snyder.  Laura DeFrancesco

EU states free to ban GM crops In July, the European Commission (EC) officially proposed to give member states the freedom to veto cultivation of genetically modified (GM) crops without having to back their decision with scientific evidence on new risks. The reform’s goal is to hand back responsibility to individual states and speed up pending authorizations. Anti-GM countries can now choose to opt out whereas biotech-friendly countries can cultivate new GM varieties. However, there is no guarantee it will work. “We are not against freedom for member states, the problem is how the principle is articulated,” says Carel du Marchie Sarvaas, director for agricultural biotech at EuropaBio. The proposal stands on two legs: an amendment to directive 2001/18 that must gain the approval of the council of ministers and the European Parliament, and an EC recommendation on coexistence, already effective. The first legalizes national or local bans on growing, the second one achieves the same result by conceding that countries wanting to keep ‘contamination’ levels well below the labeling threshold can enforce wide isolation distances between GM and conventional or organic fields. “It’s a Pandora’s box. We are concerned it will create legal uncertainty and unpredictability for farmers and operators,” says du Marchie Sarvaas. The reform doesn’t target imports of GM material for food or feed, whose approvals are also stalled.  Anna Meldolesi

769

NEWS

in brief

© 2010 Nature America, Inc. All rights reserved.

GM alfalfa—who wins? Both sides are claiming victory following the Supreme Court’s verdict issued June 21 in Monsanto v. Geerston Seed Farms over the future sale of Roundup Ready (RR) alfalfa seeds. The Supreme Court repealed a lower court injunction issued in 2007 banning the biotech seeds nationwide (Nat. Biotechnol. 28, 184, 2010). Monsanto’s business lead for the crop, Steve Welker, says the St. Louis–based company has plenty of RR alfalfa seeds “ready to deliver,” although their release is subject to a pending environmental impact statement (EIS) by the US Department of Agriculture (USDA). “Our goal is to have everything in place for growers to plant in fall 2010,” Welker adds. Not so fast, says lawsuit opponent Andrew Kimbrell of the Center for Food Safety in Washington. He points out that the Supreme Court “just took away the injunction, and USDA still has to comply with NEPA [the National Environmental Policy Act] and complete an EIS” before the crop can be deregulated. Although USDA appears poised to complete its EIS and fully deregulate RR alfalfa, the Center for Food Safety could renew its challenge of USDA’s decision. This lingering uncertainty has agitated many members of Congress. Seven senators and 49 representatives have asked agriculture secretary Tom Vilsack to retain regulated status for RR alfalfa, whereas two other senators have urged Vilsack to “mount vigorous defenses against lawsuits that seek to upend sciencebased regulatory decisions.” Jeffrey L Fox

Biofuel ‘Made in China’ Collaboration between the Danish enzyme producer Novozymes of Bagsvared, Beijingbased China Petroleum and Chemical and Cofco, the state-run agriculture company, will produce three million gallons of ethanol a year for local consumption, using corn stalks and leaves from northeastern China’s corn belt. The demonstration plant will test novel technologies, including Novozymes’ new Cellic CTec2 enzymes, with a view to launch a commercial facility by 2013. Cofco has been running a small pilot plant in Heilongjiang province for four years, but as a precondition for commercialization “we need more capacity to optimize our design and operation,” says Guo Shunjie, general manager of Cofco’s bio-energy and biochemical department. One remaining hurdle is the inability to break down five-carbon sugars abundant in lignocellulose, which make up 20–40% of the plant biomass. The new process could cut costs considerably, as it requires half the dose of enzymes needed by other treatments to break down plant waste. The partners’ goal is to produce cellulosic ethanol at $2.25 a gallon, a price further pushed down by government tax credits to be competitive with corn-based ethanol, currently at $1.50–1.60 a gallon. “Since the trend to lower carbon emissions is here to stay, it won’t be long before we break even,” says Shunjie. Daniel Grushkin

770

Table 1 USDA-approved soybeans modified for improved trans fat content Product

Company

Description

DP-305423

Pioneer Hi-Bred International

High oleic acid soybean produced by inserting extra copies of a portion of the gene encoding omega-6 desaturase, gm-fad2-1, resulting in silencing of the endogenous omega-6 desaturase gene (FAD2-1).

DD-026005-3

DuPont

High oleic acid soybean produced by inserting a second copy of a portion of the gene encoding omega-6 desaturase, gm-fad2-1, resulting in silencing of the endogenous omega-6 desaturase gene (FAD2-1).

OT96-15

Agriculture & Agri-Food Canada

Low linolenic acid soybean produced through traditional crossbreeding to incorporate the trait from a naturally occurring fan1 gene mutant that was selected for low linolenic acid.

Source: AGBIOS

commercialize Plenish soybeans until the first quarter of 2012, after food players have had time to determine what food applications, if any, they want to pursue with Plenish soybeans. “We’re being fairly conservative in our commercialization schedule,” Sanders says. The time to market also depends on Pioneer’s ability to secure regulatory approval in key global markets, such as Europe, Japan, China, Taiwan and South Korea, Sanders says. The soybean is already approved in Canada and Mexico. Global regulatory hurdles hampered Dupont’s earlier development of a different high oleic acid soybean (Table 1). In 1997, the USDA approved, or deregulated, DD-026005-3 —a Dupont soybean with an oleic acid content of 85%. This variety was modified with an extra copy of soybean Δ12-fatty acid dehydrogenase under the control of the soybean β-conglycinin promoter, which triggered silencing of the transgene and its counterpart endogenous gene. But the product fizzled after the company encountered global regulatory complexities associated with the crop’s marker technology, says Sanders. Markers are used by crop developers to test whether genetic material is successfully transferred to the host crop. In this case, DD-026005-3 contained the Escherichia coli uidA gene, encoding β-glucuronidase as a colorimetric marker, and the bla gene, encoding the enzyme β-lactamase as a selective marker that confers resistance to β-lactam antibiotics (such as penicillin and ampicillin). Pioneer’s new high oleic soybean targets the same oleic acid pathway as the 1997 version, but it is hoped that use of a different marker gene, one imparting tolerance to an ALSinhibitor herbicide, will smooth the regulatory path. (The plant will not be tolerant to ALS-inhibitor herbicides at the levels used in the field.) Sanders says he is “optimistic” about the 2012 regulatory goals. On Pioneer’s regulatory heels are two Monsanto soybean products with modified oil profiles, one with omega-3 fatty acids for

nutrition and the other with enhanced texture and functionality, called high stearic acid soybeans. Monsanto has submitted to the USDA petitions for deregulation of both products. Still in the discovery phase, Dow AgroSciences in Indianapolis, Indiana is developing omega-9 canola and sunflower oils. With one nutritionally altered crop approved and a handful in the pipeline, the public may finally get what it has been promised for two decades. But whether high oleic acid soybeans directly benefit consumers enough to boost public opinion of biotech crops is doubtful, say agriculture experts. “Companies already have methods of removing trans fats” from food, says Jane Rissler, a senior scientist with the Union for Concerned Scientists in Washington, DC. Pioneer is “offering an alternative to those existing methods” without much added benefit to consumers, she says. Alan McHughen, a plant biotechnologist at the University of California, Riverside, notes that: “Those who already despise [genetic modification] will continue to do so, those who accept GM will continue to do so, and most others won’t even notice it, as it’s not a high-profile whole food with immediate consumer-recognized benefit.” In the US, food companies aren’t required to label food derived from genetically engineered crops, and generally don’t voluntarily do so. An April 2010 survey of 750 US consumers asked this question: “All other things being equal, how likely would you be to buy a food product made with oils that had been modified by biotechnology to avoid trans fats?” Seventy-four percent said they were either very likely or somewhat likely to buy this kind of biotech food. However, in a separate question, only 32% of those respondents said they had a favorable impression of biotech food. The survey was conducted by the International Food Information Council Federation in Washington, DC. Emily Waltz Nashville, Tennessee

volume 28 number 8 AUGUST 2010 nature biotechnology

data page

2Q10—spreading the wealth Walter Yang Although biotech stocks, along with the general markets, performed poorly last quarter, more companies were able to access capital, more than in each of the previous four quarters. Excluding US partnership monies, 219 companies pulled in $8.1 billion (compared with 157 firms raising $5.3

billion in 2Q09), 39% of which originated from debt deals by Genzyme (Cambridge, MA) and Teva Pharmaceuticals (Petah Tikva, Israel). Venture funding was up 36% from 2Q09; ten companies launched initial public offerings (IPOs), raising $342.9 million.

Stock market performance

Global biotech venture capital investment

The BioCentury 100 and the NASDAQ Biotechnology were down 11% and 15%, respectively, similar to other major indices.

Venture money raised was up 36% to $1.7 billion from $1.2 billion in 2Q09.

9 /2 0 12 09 /2 00 1/ 9 20 1 2/ 0 20 1 3/ 0 20 1 4/ 0 20 1 5/ 0 20 1 6/ 0 20 10

09

00

1,400 1,200

$9 $180 $1,035

Partnership

Venture

Debt and other financing

Follow-on

PIPE

IPO

8.5, 5.0, 1.7, 0.6, 0.4, 0.3

2Q10

14.8, 3.1, 1.6, 2.3, 0.7, 0.3

4Q09 3Q09

9.4, 2.3, 1.2, 2.4, 0.6, 0.7

2Q09

8.0, 2.6, 1.2, 0.8, 0.7, 0.0 0

5

10

15

20

25

Amount raised ($ billions)

600 400

Global biotech initial public offerings 15 7 635

500

50 151 70

400 300 100 0

31 0 364

50 85 208

Asia-Pacific

4Q09

2Q09

3Q09

4Q09

1Q10

2Q10

43

49

60

60

76

Europe

14

14

32

30

28

1

1

1

1

1

Asia-Pacific

Table indicates number of venture capital investments and includes rounds where the amount raised was not disclosed. Source: BCIQ: BioCentury Online Intelligence

Amount Venture capital raised Company (lead investors) ($ millions) AiCuris (Santo Holding) 74.9 Achaogen (Frazier Healthcare) 56.0 Pacific Biosciences (Gen-Probe) 50.0 OptiNose (Avista Capital) 48.5 Agile Therapeutics (Investor Growth Capital,Care Capital) 45.0 Tetraphase (Excel Venture) 45.0 Anaphore3 (5AM Ventures, Versant, Apposite Capital) 38.0

1Q10

Target OSI Pharma Valeant Abraxis Wuxi PharmTech

Company (lead underwriters) Codexis Alimera Lansen Pharma Tengion GenMark Aposense

Round number 2 3 6 NA 2 3 1

Date closed 14-Apr 7-Apr 17-Jun 8-Jun 14-Jun 1-Jun 14-May

Acquirer Astellas Biovail Celgene Charles River

Value ($ million) 4,000 3,200 2,900 1,500

Date announced 17-May 21-Jun 30-Jun 26-Apr

Amount raised ($ millions) 78.0 72.1 50.2 30.0 27.6 24.8

Change in stock price since offer –33% –32% 3% –26% –26% –11%

Date completed 22-Apr 22-Apr 30-Apr 9-Apr 28-May 7-Jun

Value ($ millions) Deal description $1,100 Exclusive, worldwide rights, excluding the Middle East and North Africa, to develop and commercialize small-molecule glucokinase activators Regulus Sanofi-aventis >$750 Discover, develop and commercialize microRNA therapeutics for up to four targets Diamyd Johnson & $625 Exclusive rights to Diamyd diabetes vaccine outside Nordic Johnson countries Neurocrine Abbott $595 Exclusive, worldwide rights to develop and commercialize endometriosis compound elagolix OncoMed Bayer >$500 Discover and develop antibodies, proteins and small molecules targeting the Wnt signaling pathway to treat cancer Source: BCIQ: BioCentury Online Intelligence Researcher TransTech

Americas

3Q09

2Q10

Licensing/collaboration

Europe

0 0 0 2Q09

3Q09 4Q09 1Q10 Financial quarter

Americas

IPOs

Ten companies raised $342.9 million through IPOs last quarter versus none in 2Q09.

600

2Q09

Mergers and acquisitions

Partnership figures are for deals involving a US company. Source: BCIQ: BioCentury Online Intelligence, Burrill & Co.

700

Americas Europe Asia

800

Notable Q2 deals

6.1, 2.1, 1.3, 1.3, 0.5, 0.4

1Q10

$24 $331 $939

1,000

0

Excluding partnership monies, 2Q10 funding was up $8.1 billion, 53% on 2Q09, largely through debt deals, which shot up 97%.

$6 $104 $1,064

$0 $458 $1,210

200

11

/2

09

20

10

09

20

8/

9/

09

20

09

20

6/

7/

09

5/

20

09

20

09

20

4/

20

2/

3/

8 00

20

09

BioCentury 100 Dow Jones S&P 500 NASDAQ NASDAQ Biotech Swiss Market

1/

/2 12

Amount raised ($ millions)

1,600

Global biotech industry financing

200

$9 $479 $1,065

1,800

Month

Amount raised ($ millions)

© 2010 Nature America, Inc. All rights reserved.

Index

1,700 1,600 1,500 1,400 1,300 1,200 1,100 1,000 900 800 700 600 500

2Q10

Financial quarter 2Q09

3Q09

4Q09

1Q10

2Q10

Americas

0

2

2

4

4

Europe

0

1

2

0

5

Asia-Pacific

0

1

2

2

1

Table indicates number of IPOs. Source: BCIQ: BioCentury Online Intelligence

nature biotechnology volume 28 number 8 AUGUST 2010

Investor Forest

Walter Yang is Research Director at BioCentury

771

N E W S f e at u r e

Drugmakers dance with autism

In June, the Autism Research Project published the largest genetic study of autism so far, identifying 226 gene mutations that are found in people with the syndrome1. Children with autism are 20% more likely to carry one of these rare mutations, though they are not inheriting them; they are present in less than 6% of the parents of autistic children. This study adds to the growing list of genes that could serve as starting points for research on autism therapies. Whereas the pharmaceutical industry increasingly has been shying away from psychiatric disorders, such as schizophrenia and depression, interest in autism has intensified. Together with an increasing number of autism cases diagnosed each year, there is a dearth of effective treatments. As a result, “autism seems to be a relatively hot area,” says Manuel Lopez-Figueroa of Bay City Capital, a venture capital firm in San Francisco, and scientific liaison for the Pritzker Neuropsychiatric Disorders Research Consortium. Not only is the pharmaceutical sector ploughing R&D resources into the condition, but several smaller companies are pioneering therapies, one of which is an enzyme replacement therapy already in phase 3 human testing (Table 1 and Box 1). What’s more, progress in drug discovery programs aiming to target proteins associated with Mendelian neurodevelopmental disorders may pave the way for expansion into broader spectrum autism conditions. Repurposed drugs Current estimates indicate that 1 in 110 children in the United States have an autism spectrum disorder defined by three core symptoms: deficits in social interactions, problems with communication and repetitive behaviors. Although twin and family studies have established a strong genetic basis for autism, no clear genetic cause has emerged. In addition to complex genetics, the disorder is phenotypically diverse: individuals with an autism spectrum diagnosis may be intelligent and high functioning (e.g., those with Asperger’s syndrome) or have severe mental deficits. The large variation in phenotypes and 772

Mike Agliolo/Corbis

© 2010 Nature America, Inc. All rights reserved.

With monogenetic neurodevelopmental disorders similar to autism serving as starting points for several drug discovery programs, smaller biotechs are now joining big pharma in pursuing therapies to tackle this perplexing condition. Sarah Webb reports.

Trouble at the synapse. The genetics of autism is pointing toward malfunctioning at the synapse.

high concordance in monozygotic twins suggests many genetic and environmental biasing factors are involved. A diagnosis of autism brings along a slew of unmet medical needs, including anxiety, sleep disturbances, and metabolic and gastrointestinal issues. Initial moves by industry into autism therapeutics have involved applying existing drugs to alleviate some of these symptoms, says Sophia Colamarino, vice president for research at Autism Speaks, a patient advocacy group based in New York. “In the short term, that’s where many of the pharmaceutical companies will be able to have an immediate impact,” she says. Two atypical antipsychotics have been approved by the US Food and Drug Administration (FDA) for treating irritability in autistic children. Johnson & Johnson’s Risperdal (risperidone) was approved in late 2006, followed by Abilify (aripiprazole) from Bristol-Myers Squibb in New York, and Otsuka in Princeton, New Jersey, in 2009. Selective serotonin reuptake inhibitors such as low-dose Prozac (fluoxetine) are approved for use in adults and children for obsessive compulsive disorder and have been tested in children with autism. Anticonvulsives such

as valproate (Stavzor, Depakene, Depacon) may serve the same sort of purpose for some patients, says Eric Hollander, director of the Compulsive, Impulsive and Autism Spectrum Disorders Program at Albert Einstein College of Medicine and Montefiore Medical Center in New York. Treating these related symptoms gives patients and their caregivers an improved quality of life, making it more likely that an individual with autism can live at home rather than in a care facility, Hollander adds. Improving those related symptoms can also make patients more responsive to behavioral therapies, says Robert Ring, who is heading up Pfizer’s autism research unit in Groton, Connecticut. At least one repurposed drug is targeting the imbalance between excitatory and inhibitory signaling suspected to be part of the basis of autism. New York-based Forest Laboratories is testing Namenda (memantine), an Alzheimer’s drug and N-methyl d-aspartate receptor (NMDA) receptor modulator, in a phase 2 trial in autism patients. Abnormal synaptic connectivity Because this spectrum of disorders has a clear genetic basis but no clear genetic cause, researchers are chewing on the question of how so many different mutations could lead to a similar phenotype, says Luca Santarelli, head of Roche’s central nervous system exploratory development in Basel. Genetic studies are important, but they don’t tell a complete story. “Identifying genes and coming up with gene candidates is really just a first step in gaining confidence in a potential genetic target that could be druggable,” says John Spiro, a research director at the Simons Foundation Autism Research Initiative in New York City. “There are not many genes that you can be really, really confident are accounting for any significant portion of autism.” Though researchers remain hopeful that the genes might converge into a single meaningful pathway, he adds, “for the most part in autism, it’s not clear yet that’s going to be the case.” Nonetheless, some patterns are emerging that may help researchers devise new therapeutic strategies. A genome-wide survey of a group of autistic and mentally retarded individuals revealed a set of mutations (point mutations and copy number variants) in a gene, SHANK2, that controls synaptic structure, defects in which could lead to problems in neuronal communication2. Mutations in another family of genes involved with synapse formation, the neuroligins, which code for adhesion molecules that cluster on the receiving side

volume 28 number 8 august 2010 nature biotechnology

news f e at u r e

Box 1 Enzyme replacement for autism?

© 2010 Nature America, Inc. All rights reserved.

Unlike other emerging treatment strategies for autism that target genes or neurochemical pathways, Rye New York’s Curemark is working on an enzyme replacement therapy comprising a mixture of several digestive enzymes (Table 1). In clinical work with children who showed symptoms of autism, Curemark’s founder and CEO, Joan Fallon, noticed that several of these patients restricted their diets by their own choice, preferring carbohydrateladen foods such as crackers and pasta. Searching for an explanation, she found that these patients had low fecal levels of the protease chymotrypsin (fecal chymotrypsin levels have also served as a diagnostic indicator of cystic fibrosis). Children with autism without a known genetic cause, often had these low enzyme levels, Fallon says. Administering high-protease enzymes, the physicians observed behavioral changes in the children. Fallon filed patents in 1999 and formed a biotech company in 2005. The company’s protease-based treatment, CM-AT, is currently being tested in a phase 3 study with 170 children ages 3–8 in 12 locations around the United States.

of the synapse, may account for up to 6% of autism cases, according to Nils Brose, director of the Department of Molecular Neurobiology at the Max Planck Institute of Experimental Medicine, in Göttingen, Germany. Neuroligins 3 and 4 localize to glutamatergic synapses, and loss-of-function mutations in these genes segregate in certain pedigrees with mental retardation, autism and Asperger’s syndrome. These molecules are likely operating as the organizational point for information coming into the postsynaptic space, recruiting signaling receptors. In mouse knockouts of two of these neuroligins, Brose says, “the synapses are intrinsically operational, but they lack normal receptors and as a consequence don’t function properly.” But just noting a connection between these genes and synaptic structures isn’t enough for developing drug candidates, Spiro adds. “You don’t know. Is it too much? Is it too little? Are [the structures] in the wrong place during development? There are just a million questions that need to be ironed out before you can think about a pharmaceutical intervention.” Santarelli’s group at Roche is trying to get at some of these questions, in collaboration with Peter Scheiffele, a professor of cell and developmental neurobiology at the University of Basel and a leader in the neuroligin research area. “We’d like to understand the common downstream effects of different genetic alterations that lead to autisms and whether there are common mechanisms that could lead to treatments,” Santarelli says. Clues from rare single-gene disorders The increasing understanding of some of the molecular mechanisms of autism is providing one avenue forward. The second breakthrough, according to Colamarino, is coming through animal studies of single-gene disorders such as fragile X3 and Rett’s syndromes4, which are

found in a disproportionate number of individuals who meet the criteria for autism spectrum disorders. Since 2007, a handful of studies of animal models with inducible mutations have shown that animals can develop to adulthood with these disorders, and then recover after proper gene function is switched back on. That ability to reverse the symptoms in animals with advanced disease has been a major breakthrough, says Spiro. With clear genetic causes coupled with the opportunity to build animal models of these disorders, “it may be very reasonable to say that the pathway to drug discovery in autism may be paved by a careful focus on these rarer syndromes,” Ring says. Fragile X syndrome provides a case study in this approach that weds treatment strategies for a rare disorder with the possibility of understanding the underpinnings of autism. This genetic disorder, which affects 1 in 4,000 males and 1 in 6,000 females (http:// www.fraxa.org/), leads to learning disabilities and even mental retardation, anxiety and seizures. Up to 20% of individuals with fragile X also meet the criteria for an autism diagnosis. As a result of a single gene mutation, these individuals do not make the fragile X mental retardation protein (FMRP). Mark Bear of the Massachusetts Institute of Technology in Cambridge and his colleagues found that the lack of FMRP leads to dysregulation of signaling through the metabotropic glutamate receptors (mGluR). The mGluR5 receptor is highly expressed in regions of the brain critical for learning and memory. FMRP serves as a brake on this signaling pathway, says Randall Carpenter, CEO and president of Seaside Therapeutics, a Cambridge, Massachusetts, biotech company co-founded by Bear. “When it’s not there then there’s overactivation of the signaling pathway. The brain can’t discriminate between important information and noise and it doesn’t develop normally.” In mice

nature biotechnology volume 28 number 8 august 2010

with the fragile X mutation, Bear and his colleagues found that knocking down expression of mGluR5 to 50% rescued the learning deficits, stopped seizures and increased other measures of plasticity in the brain. Confident that they’re targeting the appropriate pathways, Seaside Therapeutics has licensed a series of small-molecule compounds from Merck to target glutamate signaling in general and mGluR5 signaling specifically, Carpenter says. They recently completed a phase 2 clinical trial of a general γ-aminobutyric acid (GABA) B agonist, STX209, in fragile X patients, and will soon complete a phase 2 trial of the same compound in individuals with autism spectrum disorders. A specific antagonist of the mGluR5 receptor is currently in repeat-dose phase 2 trials, and Seaside expects to start phase 2 trials with fragile X patients by early 2011. Mutations in glutamate receptor genes GRIN2A and GRIK2 and multiple GABA receptor genes have been associated with autism. Two pharma companies also see promise in the mGluR5 receptor strategy for treating fragile X patients. Novartis in Basel recently completed a phase 2 clinical trial of their compound AFQ 056 at sites in Europe and is planning their next study, which is scheduled to open later in 2010, says spokesman Jeffrey Lockwood in an e-mail. Roche’s small-molecule mGluR5 antagonist is being tested in phase 2 clinical trials in five locations in the United States, says Santarelli. Their results are “encouraging so far,” he says. This growing understanding of these specific, related genetic disorders, Santarelli adds, provides a pathway to think about possible extrapolations to the more sporadic types of autism. Peptide hormone targets The peptide oxytocin and its related receptors are emerging as a pathway that could prove useful for treating a variety of neuropsychiatric disorders including autism. Animal studies have pointed to the importance of oxytocin in social behavior; in voles, for example, oxytocin and its counterpoint hormone vasopressin appears to have a role in pair bonding. Karen Parker and her colleagues at Stanford University in California observed seasonal differences in the way females and males who are raising young interacted. In the laboratory, they tracked these differences, caused by purely environmental cues to the locations of oxytocin receptors in the animals’ brains. Changes based on environmental cues have led researchers to consider oxytocin therapies for treating social dysfunctioning in humans. Such tests are already being done in humans. Hollander has given intravenous oxytocin 773

© 2010 Nature America, Inc. All rights reserved.

N E W S f e at u r e to higher functioning patients with autism and Asperger’s syndrome and has observed improved social cognition. Patients were better able to lay down social memories or recognize emotions in spoken language, he says. Such treatments also decreased the severity of repetitive behaviors and self-stimulatory behaviors such as hand clapping, rocking and head banging. Patients treated with intranasal oxytocin showed similar improvements. Earlier this year, researchers at the Center for Cognitive Neuroscience in Bron, France, found that adults diagnosed as high functioning on the autism spectrum who received doses of intranasal oxytocin were better able to recognize cooperative play than adults with a similar diagnosis who had not received oxytocin. Those who had received oxytocin also spent more time looking at the face of their virtual playmates5. But teasing out the importance of oxytocin isn’t easy. The French study shows variation in individual responses to oxytocin. “We don’t have good biomarkers of oxytocin levels,” Parker says. Funded by a grant from the Simons Foundation, she and her colleagues are trying to measure plasma oxytocin levels, various mutations and social phenotypes among individuals with autism and their siblings and compare them with controls matched for age and gender. Oxytocin and the related response pathways represent “one of the most exciting biologies in the autism space today,” says Pfizer’s Ring and could have implications for other psychiatric areas as well. In research Ring carried out at Wyeth, he developed the first nonpeptide oxytocin receptor agonist6. “The oxytocin receptor is a priority target for the field, but a very challenging target to develop traditional smallmolecule chemistry for.” Cellceutix, a biotech company in Beverly, Massachusetts, is also testing a preclinical compound for autism, KM-391, in a rodent model of autism developed by researchers at the Kennedy Krieger Institute in Baltimore. The autism-like symptoms are induced by injecting the chemical 5,7-dihydroxytryptamine (5,7-DHT) into the forebrain of newborn rat pups, leading to neonatal serotonin depletion, reduced brain plasticity and abnormal behaviors. In an initial study, KM-391 given over 90 days restored normal behaviors, and near-normal serotonin levels and increased brain plasticity relative to a nontreatment group and a group given Prozac. Another study measuring serotonin levels in three regions of the rat brain has confirmed the restoration of normal serotonin levels. Another small study added an oxytocin antagonist to the mix. The antagonist alone intensified the autism-related behaviors, such as 774

Table 1 Selected companies with autism targets in clinical development Stage of development

Company

Target

Drug candidate

Curemark

Protease deficiency

CM-AT (a mixture of amylase, protease, chymotrypsin, trypsin, papain and papaya in a 4–10:1 ratio with lipase, derived from animal, plant, microbial or synthetic sources)

Phase 3

Novartis

mGluR5

AFQ 056 (small molecule)

Phase 2

Roche

mGluR5

RO4917523 (small molecule)

Phase 2

STX209 (R-isomer of baclofen) STX107 (2-methyl-1,3-thiazol-4-yl) ethynylpyridine)

Phase 2 Phase 1

Seaside GABA B Therapeutics mGluR5 Forest Laboratories

NMDA receptor Namenda (memantine) modulator

repetitive behaviors and sensitivity to touch, but when given with KM-391, the frequency and intensity of these behaviors were reduced. Measuring outcomes Fueled by academic research and increased funding from the US National Institutes of Health, nonprofit and advocacy organizations, the field is moving forward. But even as some drug candidates are moving into the clinic, a number of challenges remain for the field as a whole. Above all is the problem of the heterogeneity of the disorder, according to Colamarino. “We’re calling it one thing when it’s really probably more than one.” That heterogeneity can pose a challenge in choosing appropriate study subjects. The field is also struggling with finding appropriate outcome measures, particularly those that can be measured within the time frame of a clinical study. Without sensitive measures of changes in the core symptoms, researchers need to identify what the focus should be within a particular trial. In many cases researchers have depended on parental reporting of behavioral changes, Colamarino says, leading to a large placebo effect. Although no biomarkers have been established for autism, some sort of biological measure of change in connection with autism’s core symptoms, would be particularly attractive. Some clinical trials have failed because of methodological issues, she adds. “That’s why we need to address this sooner rather than later.” To bring researchers together to discuss these challenges, Autism Speaks and Pfizer are co-sponsoring a translational research meeting to improve clinical study methodology and design, tentatively scheduled for later this year. “There’s no better investment for us externally than to bring together all the key experts in this area and have a discussion with FDA present and try to iron out a framework to address this challenge together,” Ring says. The development of the Diagnostic and Statistical Manual of Mental

Phase 2

Disorders (DSM-V), the bible for neurological diseases, scheduled for release in May 2013, could complicate the development of trial endpoints, Bay City’s Lopez-Figueroa adds, depending on how autism disorders and symptoms are classified. A second meeting in early 2011 will look at clinical targets—both their identification and validation—in an attempt to reach a consensus on where therapeutics can bring the most initial benefit to patients. This is something the field is still struggling with, Ring says. “If we had one shot today to demonstrate that this would work, what would be the clinical target that we should take on?” Pfizer and Roche are also developing an autism proposal for the Innovative Medicines Initiative, which coordinates European Union–based public-private partnerships in drug discovery and development. The idea is for companies to join forces to work on research that is not generating intellectual property, Santarelli says, such as the development of animal models, understanding disease mechanisms and physiology, finding biomarkers and developing clinical methodology. Unquestionably, developing therapeutics for a developmental neuropsychiatric disorder with such an early onset presents several challenges. But Autism Speaks’ Colamarino is encouraged by the growth in the field. “Three to five years ago, we wouldn’t have been talking about clinical trials, certainly with respect to novel drug discovery,” she says. Pfizer’s Ring expects industry involvement to continue to grow: “It’s just too large an unmet medical need for companies not to see the opportunity to enter into this research space.” Sarah Webb, Brooklyn, NY 1. Pinto, D. et al. Nature 466, 368–372 (2010). 2. Berkel, S. et al. Nat. Genet. 42, 489–491 (2010). 3. Guy, J. et al. Science 315, 1143–1147 (2007). 4. Dölen, G. et al. Neuron 56, 955–962 (2007). 5. Andari, E. et al. Proc. Natl. Acad. Sci. USA 107, 4389–4394 (2010). 6. Ring, R.H. et al. Neuropharmacology 58, 69–77 (2010).

volume 28 number 8 august 2010 nature biotechnology

building a business

At ground level Julian Bertschinger The hardest—and perhaps loneliest—period of being an entrepreneur might be just after your company is founded.

© 2010 Nature America, Inc. All rights reserved.

I

cofounded Covagen when I was 30 years old. Although my PhD and postdoc work had taught me to think in a focused manner and be product oriented, I was as green as they come concerning the nuts and bolts of launching a company. Picking it up as you go might not be the optimal way to learn, but I’m living proof that it can be done with the right team. Here’s how we did it. Two men and a plan The most important motivating factor, for me, was my education. I did my thesis in Dario Neri’s lab at the Institute of Pharmaceutical Sciences at ETH Zurich. The research group there had just isolated an antibody fragment that binds to a tumor-associated marker, and proof-of-concept data showed that the fragment selectively targeted solid tumors in mice. Neri went on to cofound Philogen, based in Siena, Italy, and develop the antibody in collaboration with Bayer Schering in Berlin. Today, several derivatives of this antibody are in phase 2 trials. Seeing this process firsthand showed me (and Dragan Grabulovski, my cofounder at Covagen, which is based in Zurich) that it was possible to move from the lab to the commercial side. This had our group thinking about products right away, which I believe is crucial when contemplating a biotech company. But the truth is that Covagen never would have been founded without the Venture business plan competition, organized every two years by McKinsey, in Zurich, and ETH Zurich. One of the winners of this competition was Glycart Biotechnology, also in Zurich, which took the prize in 1998 and eventually was acquired by Roche, in Basel, Switzerland, for CHF235 million (US$180 million) in 2005. Grabulovski and I decided to take part in Julian Bertschinger is CEO at Covagen, Zurich, Switzerland. e-mail: [email protected]

the Venture 2006 competition for two reasons: we were eager to learn how to write a business plan (we’d never written one) and we thought it would be interesting precisely because it was so different from the reports and scholarly articles we were used to writing. The competition is divided into two phases. During the first, entrants submit a business idea outlined on a few pages, and the best ten ideas are awarded a prize. In the second, all participants receive free coaching from industry experts and venture capitalists, who then give advice to participants writing their first business plan. The ten best business plans are chosen by a jury and all receive the same prize amount of CHF2,500 (US$2,057). We submitted our business idea, but I didn’t actually expect us to be one of the winners; I was busy applying for postdoc positions abroad. Nevertheless, our idea was chosen out of about 100 applications to be awarded with a CHF2,500 prize. This

surprised me—not because we doubted our entry, which was based on the Fynomer technology (Fig 1; Box 1 and D. Grabulovski et al. J. Biol. Chem. 282, 3196–3204 (2007)) but because we felt that it was too early to found a company on the available results: we had no in vivo data. Looking back, the biggest effect of participating in Venture 2006 was that it let us begin to establish a business network—previously, we’d known only people within academia. At workshops during the second phase of the competition, we met Rudolf Gygax, a managing director of Novartis Venture Fund, who would be a key contact for us later on. He and Neri helped us to draft our first business plan. The prize money was certainly useful, but the large amount of positive feedback we received was even more important. That boosted our confidence, and after winning, I thought for the first time that we really could found our own company.

Box 1 The technology behind Covagen Covagen is built on Fynomer technology (Fig. 1), developed at the Institute of Pharmaceutical Sciences at ETH Zurich. Fynomers are a class of binding proteins derived from the Src homology 3 (SH3) domain of the human Fyn kinase (D. Grabulovski et al. J. Biol. Chem. 282, 3196– 3204 (2007)). The Fyn SH3 domain structure is made up of two anti-parallel β-sheets and two loops—n-src and RT—which are known to be involved in interactions with other ligand proteins. Fynomers can be produced in bacteria at Figure 1 Fyn Src homology 3 (SH3) high yields and are approximately 20 times domain structure. The RT-Src loop is smaller than antibodies. Additionally, they shown in red, and the n-Src loop is shown in green. (Protein Data Bank entry 1M27) have the advantage of being easily assembled in a modular manner to yield bispecific and/or multivalent proteins, which might allow new treatment modalities that are challenging or impossible to explore with traditional antibody formats.

nature biotechnology volume 28 number 8 august 2010

775

b u i l di n g a b u si n ess Box 2 Securing our funding

© 2010 Nature America, Inc. All rights reserved.

I was able to found Covagen with an initial investment (in several tranches) from the Novartis Venture Fund. The first tranche came after signing investment documents, and the following tranches were hinged on attaining research milestones. It was crucial that Novartis Venture Fund was prepared to invest in us at a very early stage. Corporate venture funds are beneficial in this way: they are usually more likely to do early-stage investments than most private venture capitalists because corporate funds can afford longer times to exit. If you’ve hit upon an interesting idea in academia, you might look to corporate venture funds first. In 2009, Covagen was able to attract three other investors: the corporate venture fund MP Healthcare Venture Management, of Boston; Ventech, of Paris; and Edmond de Rothschild Investment Partners, also of Paris. We also have received some funds via our research collaboration with Roche, which was secured in June 2009. To move our interleukin-17A inhibitor into preclinical and clinical development, we are planning to raise additional money this year, so we are seeking one or two venture capitalists to join our existing investors.

Founding Covagen We stayed in contact with Gygax, and he invited us to present our project at the Novartis Venture Fund headquarters in Basel. The fund was interested in investing, and we sat down to negotiate our first term sheet. I had absolutely no idea what the difference was between a binding contract and term sheet, and this was my initiation. I learned what Series A shares are, how to calculate pre-money and post-money valuations, what drag-along and tag-along clauses are, why a high liquidation preference for investors is bad for holders of common shares and how anti-dilution protection for investors can hurt founders in a down round. I was moving into a whole new world. It is very important to understand every word in term sheets and agreements. You should always know what you are signing. To do this, first make sure you find a lawyer who intimately knows relationships between venture capitalists and biotech startup companies, and then be persistent enough to ask your lawyer about every single expression or phrase you do not understand. (You can familiarize yourself somewhat with the terminology by using the internet, in particular http://www. investopedia.com/terms/v/venturecapital.asp, but also ask your lawyer directly.) When we finally signed the term sheet, we found it just meant more paperwork. We still needed to establish a licensing agreement with ETH Zurich and negotiate the investment and shareholder’s agreements. I admit that when I first read the investment document drafts, I thought the beginning definitions weren’t very relevant. But after further reading and questioning our lawyer, I quickly realized that those definitions are actually one of the most important things in a contract. 776

Once all the details were ironed out (Box 2), we founded Covagen in December 2006 and signed the investment agreements with Novartis Venture Fund. The real work was about to start. The lonely lab Grabulovski still had to finish his PhD thesis. This made me Covagen’s only employee from December 2006 until May 2007, and Covagen was a startup in every sense of the word. My first task was to open a bank account so Novartis Venture Fund could transfer in its investment. When that was done, I set up Covagen’s homepage (be sure to check for domain name availability before you decide on a company name). A friend of a friend runs a company offering website design and e-mail hosting services, and he helped me create Covagen’s website. Here’s a tip: make sure that you can administer the website yourself so you will not have to pay a web designer for every small change or update. In addition, I opened a Covagen e-mail account, and here, too, I made sure I could independently set up additional e-mail accounts. But there remained a very big need—work space. We had no laboratory. Unfortunately, ETH Zurich does not offer incubator space for spin-outs. Startup companies usually try to find space within the department they originated from, but in our case there was no room available. After asking around within ETH Zurich, Grabulovski learned of an empty laboratory not attached to any department, and we were able to make an arrangement to allow us to rent this space. In addition, our former institute enabled us to access some rather expensive instruments for an affordable fee. The laboratory was empty, except for benches and desks, and somewhat dusty. On

my second day, I brought rags from home and started cleaning. This wasn’t really what I envisioned a biotech CEO doing, but the truth is, I was excited—I was starting a company from the very bottom! There was no network connection for my computer, no printer, no phone, no fax. However, after making a few calls with my mobile phone, the university’s staff set up all the necessary connections within a few days. This is a benefit of staying within academia: when starting your company, all issues related to infrastructure need only minimal time and management resources. After all that work, I thoroughly appreciated making the first company phone call and sending the first message from my Covagen e-mail account! With communications behind me, I was left with the science. It’s only when you start from scratch that you realize how many different instruments and tools, disposable plastic tubes, glassware, kits, antibodies and chemicals are needed for research, and I had none of it. I also realized how comfortable my life in the academic lab had been, where many instruments were available and I didn’t have to think about budgeting. That was not the case at Covagen, where I became very cost sensitive. Comparison shopping takes time, and it was four months before the last instruments and reagents arrived. This neatly coincided with Grabulovski earning his PhD in May 2007, and he joined Covagen as CSO. I finally had company. Building a biotech Established as Covagen, we now had several target proteins in mind to validate the technology, but we did not have a clear plan on which targets we wanted to focus on for the development of our first Fynomerbased clinical candidate. Choosing a good first target was the most important decision we needed to make because once we made the call, we’d invest most of our resources in that direction. We investigated many different targets to find one that was economically promising and in an area in which Covagen had freedom to operate. We decided to go for inhibition of the cytokine interleukin-17A, which is an attractive emerging target for diseases such as rheumatoid arthritis, psoriasis and uveitis. In early summer 2007, we hired another person to help speed up our research. We had spent less money than we expected in the first half of 2007, so we had sufficient financial resources to hire. We felt that our first employee should be someone we already knew and someone we could trust to be dependable

volume 28 number 8 august 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

b u i l di n g a b u si n ess and competent. As several investors had warned us, not getting along with co-workers is a big reason why many small companies fail. Personal frictions tend to increase even more if a company hits hard times. We asked Simon Brack, an antibody engineering specialist we knew from our time in Neri’s group, to join Covagen. Brack was returning to Switzerland from Oxford, where he’d worked as a postdoc. In October 2007, he became Covagen’s third employee and was a great hire. Even in a company as small as Covagen was then, there were a million administrative things to do, and they occupied a large amount of my time—I was finding it hard to do the necessary work on the bench to develop our technology, not to mention that creating documents and presentations for potential investors takes a lot of time. So at the very least, it felt good to know that if I had to leave the lab, I had four hands working while I was gone. Now, we are up to seven employees. Advancing our technology is the most important task we have at Covagen, just as it was when we started. For this reason, all employees at Covagen are PhD scientists. We are a young and enthusiastic team; none of

us is older than 33. This can be a problem at times: when talking to investors, I realize that we sometimes lack credibility. Quite often, investors do not believe our claims, and mainly that’s because they do not believe I have enough experience. In some ways, they are right—I am a scientist still learning the business side of things. But we have been taught a lot about the varying aspects of drug development through working with Neri, and I believe a young group like us can learn fast if given the right advice. Currently, we’re getting that advice from Ray Hill, who was executive director for licensing in Europe at Merck & Co. and now is a visiting professor in neuroscience and mental health at Imperial College London. Hill sits on our board of directors. We’ve also established an excellent scientific advisory board, which will be of great help and value when bringing our first drug candidate to preclinical development and broadening our research activities.

Conclusions Even as our company grows, things continue to change quickly and will for the foreseeable future. The larger we get, the more important (and time consuming) communicating with employees, investors, our board of directors and our scientific advisory board becomes. My tasks are always shifting as we adapt, improve and complement our skills. But this fluid environment is partially what makes startup companies attractive workplaces. Now, our company doesn’t feel so young anymore. This year, we plan to bring our first drug candidate to good manufacturing practice production and preclinical development. That, of course, will require additional money, and we plan to close a financing round this year. Raising a sizable round is another challenge for me, and it means I’m no longer on the bench. My job is raising money now. In that regard, I’ve graduated to the role of a typical biotech CEO.

To discuss the contents of this article, join the Bioentrepreneur forum on Nature Network:

http://network.nature.com/groups/bioentrepreneur/forum/topics

nature biotechnology volume 28 number 8 august 2010

777

correspondence

© 2010 Nature America, Inc. All rights reserved.

Waking up and smelling the coffee To the Editor: As I pointed out recently on the Patent Docs weblog (http://www.patentdocs. org/), the editorial ‘Sitting up and taking notice’ in the May issue1, announcing Judge Sweet’s 29 March decision in favor of the plaintiffs in Association for Molecular Pathology v. US Patent and Trademark Office, contains several misstatements and promotes the wrong-headed idea that gene patenting is a problem. In describing the case, you begin by making factual errors. Judge Sweet’s decision (summary judgment) does not indicate that “the judge felt that Myriad had no case to argue.” Rather, summary judgment is used when there are no disputed issues of material fact, and the case is decided as a matter of law. I would argue that the prudence of Judge Sweet’s judgment is questionable because he chose to make law by deciding that DNA is not patent eligible for being “the physical embodiment of genetic information.” You then state that “[t]he plaintiffs…won on virtually every count.” In fact, the court refused to consider the US Constitutional issues raised in the complaint, which formed the basis for the breast cancer victims to have standing in the lawsuit. This is not trivial because the court used these constitutional issues not only to deny defendants’ motions to dismiss, but also, politically, to provide the political frisson so attractive to the American Civil Liberties Union (New York) and the Public Patent Foundation (New York). The editorial goes on to mischaracterize the effects of BRCA patents on research, stating that “Myriad’s influence has been particularly pernicious. Its lawyers have issued cease-and-desist letters to genetics laboratories in universities, hospitals and clinics that offered diagnostic services based on the BRCA1 and BRCA2 genes.” Why is enforcing your patent rights pernicious? Use of these patented tests by these institutions constitutes infringement. It doesn’t matter whether the infringer is a university, hospital or clinic, they 778

are still liable for infringement owing to their for-profit, commercial activities. There is no evidence that Myriad Genetics (Salt Lake City, UT, USA) or any other gene patent holder has inhibited basic biological research by threatening patent infringement litigation; indeed, there are several thousand basic research papers in scientific journals that have been published since the BRCA gene patents were granted. The piece also attempts to achieve ‘truth by association’ in citing several groups having “concerns” about gene patents that filed amicus briefs, including the International Center for Technology Assessment, Greenpeace, the Indigenous Peoples’ Council on Biocolonialism and the Council for Responsible Genetics. Their contribution would be more worthwhile if it did not include incorrect statements regarding gene patenting’s consequences, including “the privatization of genetic heritage, the creation of private rights of unknown scope and consequences and the violation of patients’ rights.” The editorial was correct in noting that “[t]he alignment of physicians’ and patients’ groups with what are, in effect, antibiotech lobbyists is a worrying development,” albeit ignoring the fact that not only the biotech sector, but also the public should be worried if these groups get their way. The editorial did supply potentially informative data, that Myriad reported “$326 million in revenue from diagnostic testing against $43 million in costs.” Assuming that these numbers are correct, and reflect only BRCA testing, this could be a measure of the profitability of BRCA testing results (perhaps providing motivation for the “universities, hospitals and clinics” to be so keen on getting into

the business, infringing or no). But even here, the figures are completely out of context. No indication is provided whether these profits are out of the ordinary for a diagnostics company, traditional or genetic, or whether the ‘costs’ include ancillary costs like genetic counseling or physician education (both critical in genetic diagnostics due to the consequences for a patient of receiving a genetic diagnosis). If Myriad’s profits are significantly higher than those at other diagnostic companies, that fact would be relevant. The absence of any comparisons suggests that the absolute numbers were used because they better supported the editorial’s views. Finally, the editorial departs from reality when it decries the patent system for rewarding “only the last inventive step—the small breakthrough that enables a concept to be realized.” Such a statement indicates just how little the writers understand the ‘balance of rights’ that the patent bargain actually strikes. The patent system rewards inventors who disclose how to make and use an invention that is new, useful and nonobvious. Whether the improvement is groundbreaking or incremental, satisfaction of the statutory requirements governs patentability. Thus, if technology becomes obsolescent, new technology takes its place—because patents expire, as indeed Myriad’s patents will begin to expire in 2014. The consistent lack of understanding of innovation and the patent process is illustrated by the suggestion that rights to specific genes in multigene tests be assigned based on “the importance of any specific gene sequence to the utility of the test.” This is something the marketplace can be counted on to do without the government’s help. The last sentence of the piece even acknowledges the editorial idea

volume 28 number 8 AUGUST 2010 nature biotechnology

correspondence is “implausible within the current petrified patent system and commercial infrastructure,” and then adds that this “doesn’t have to stop the dream” or “stop the discussion.” I would counter that the dream of better diagnostics and therapies is being, and has been, realized by 30 years of biotech and protection thereof by an invigorated patent system in the United States (and elsewhere). Changing that now, particularly if based on the wooly-headed arguments (really, sentiments) in the editorial, is the fastest and surest way that those hopes and dreams will be dashed.

© 2010 Nature America, Inc. All rights reserved.

COMPETING FINANCIAL INTERESTS The author declares no competing financial interests.

Kevin E Noonan McDonnell Boehnen Hulbert & Berghoff LLP, Chicago, Illinois, USA. e-mail: [email protected] 1. Anonymous. Nat. Biotechnol. 28, 381 (2010).

Nature Biotechnology replies: We were not making the case that gene patenting itself was a problem, although it is clear that some DNA patents with overly broad claims are cause for concern. We disagree with the contention that “there is no evidence that Myriad Genetics…or any other gene patent holder has inhibited basic biological research by threatening patent infringement litigation.” There are cases where exclusive licensing practices (a particular problem for methods patents) or aggressive license enforcement has stymied research, as is detailed elsewhere in this issue1. The problems also reach beyond basic research: a survey of 132 clinical laboratory heads in the United States found that 53% had “decided not to develop or perform a test/service for clinical or research purposes because of a patent”2. Indeed, one of the plaintiffs in the Association for Molecular Pathology v. US Patent and Trademark Office case is a patient who would like to have their BRCA1 test from Myriad independently verified by another laboratory, but cannot because of Myriad’s aggressive stance that prevents other laboratories performing the test. It might be good business for Myriad, but is it reasonable to enforce intellectual property in such a manner that it is so difficult for a patient to confirm a DNA test in an independent laboratory? The claim that new technology takes the place of ‘obsolescent’ technology because “patents expire” is also moot in relation to

DNA patents. A point we were trying to make in the editorial is that the fields of molecular diagnostics and sequencing are moving so quickly that they are becoming obsolete along much shorter timelines than patent terms of 20 years. Although

it was not trivial to sequence a human gene 20 years ago, it is certainly becoming routine today. 1. Carbone, J. et al. Nat. Biotechnol. 28, 784–791 (2010). 2. Cho, M.K. et al. J. Mol. Diagnostics 5, 3–6 (2003).

Genetic stability in two commercialized transgenic lines (MON810) To the Editor: A letter of correspondence by Dany Morisset and his colleagues1 in the August 2009 issue cites two recent publications2,3 in which “two commercial seed varieties of the MON810 maize genetically modified event (ARISTIS BT and CGS4540) present genetic variation thus hampering the detection by several methods for MON810 (Monsanto, St. Louis).” As representatives of Monsanto Europe (Brussels), Syngenta Crop Protection (Basel) and Limagrain Services Holding (Chappes, France), we would like to correct the scientific record concerning the claimed “variation” of the transgenic insertion in these transgenic hybrids. Upon request for further information, Margarita Aguilera and her colleagues at the European Commission, Directorate General Joint Research Center (JRC) in Ispra, Italy, informed us that the seeds tested were among 26 MON810 varieties provided by the Spanish Instituto Nacional de Investigación y Technología Agraria y Alimentaria (INIA; Madrid). The Spanish agency did not provide the JRC with details of the respective batch numbers for each variety. Our investigation has revealed that the two deviating results were not in fact related to variation of the transgenic insertion, as reported by Aguilera et al.2,3. Instead, our conclusions are that the two varieties (reported as entry 2 and entry 5) were not MON810 maize hybrids at all. Variety CGS4540 (entry 5) is a Bt176 maize hybrid and we do not understand why the seed was provided by INIA as MON810. Entry 2, which was designated as Aristis

nature biotechnology volume 28 number 8 AUGUST 2010

Bt, is most likely Aristis, the conventional counterpart of Aristis Bt (MON810). When we requested INIA to send a sample of Aristis Bt to its official Spanish laboratory CSIC (Consejo Superior de Investigaciones Científicas) for testing, the results were positive for MON810, as expected. Aguilera and her colleagues were not able to provide a correct chain of custody for the samples used in their analyses, which would have allowed resolution of the origin of these deviating results. The seed industry has invested significantly to provide quality products to the market place, which includes selling compliant and stable products. Traits are tested for presence and stability for many generations before release to the market place. We are therefore convinced that there is no scientific evidence of instability in MON810 hybrids. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/.

Sofia Ben Tahar1, Isabelle Salva2& Ivo O Brants3 1Limagrain Services Holding, Quality Assurance, Chappes, France. 2Syngenta Crop Protection AG, Regulatory Affairs, Basel, Switzerland. 3Monsanto Europe SA, Scientific Affairs, Brussels, Belgium. e-mail: [email protected]

1. Morisset, D. et al. Nat. Biotechnol. 27, 700–701 (2009). 2. Aguilera, M. et al. Food Anal. Methods 1, 252–258 (2008). 3. Aguilera, M. et al. Food Anal. Methods 2, 73–79 (2009).

779

correspondence

780

EU countries have decided to establish mandatory separation distances between GM and non-GM maize fields as the preferred single measure to limit cross-fertilization6. An overview of mandatory separation distances adopted by EU member states (Supplementary Table 1) shows a remarkable range of variation, 25–600 m, between the different countries. Although climatic and landscape parameters in maize cultivation (that affect cross-fertilization rates) are variable in the EU, often there is little sciencebased evidence that the distances adopted are proportional to achieve the desired purity standards. To test the proportionality of the separation distances established by EU member states, we perform a statistical analysis of data obtained from a number of recent studies on maize cross-fertilization performed in different European countries. Although the various studies recorded different variables, we analyzed only data on cross-fertilization rates (measured as percentage of seeds in the sample) in the receptor field as a function of distance from the edge of the pollen source. The aim of the analysis was to estimate distances necessary to keep cross-fertilization below different arbitrary tolerance thresholds and with different confidence levels. The results should inform debate on whether current distances between GM and non-GM maize fields stipulated by member states to meet legal EU labeling thresholds are supported by scientific data. 40

We first compiled a database of crossfertilization rates and distance by collating different publications and unpublished studies on maize cross-fertilization, to obtain a total of 1,174 observations covering four European countries (Germany, Italy, Spain and Switzerland). Details on the sources of data used are given in Supplementary Table 2. The database covered studies with a variety of experimental designs (mostly receptor and donor fields side by side, but also donor and receptor fields dispersed in actual agricultural landscapes) and that had been performed in different growing seasons (2001–2006). Data originate from experimental designs representing worst-case scenarios (receptor fields situated downwind from donor fields and coincidence of flowering between donor and receptor fields) in Europe. The relationship between distances and cross-fertilization rates for the database shows a negative relationship between these two variables (Fig. 1). This reciprocal relationship between cross-fertilization rates and distance was pointed out previously by several other authors4,5,7–9. For further analyses, cross-fertilization rates were analyzed for 10 m distance intervals (Supplementary Table 3). Because of the lack of sufficient observations from 50 m upwards, the size of intervals was increased to 20 m. Supplementary Table 3 shows that data on maize cross-fertilization are mostly available for short distances, close to the donor (84.1% of the data set, or 985 observations, are taken between 0 m and 20 m). In contrast, only

5

35

Out-crossing (% seeds)

To the Editor: To avoid the economic consequences of admixtures of genetically modified (GM) and non-GM harvests, and to ensure that agricultural production complies with mandatory labeling provisions, the European Union (EU; Brussels) member states have adopted co-existence measures directed to farmers cultivating GM varieties. For GM maize cultivation, regulators have established mandatory isolation distances, which differ between countries and in some cases have been regarded as disproportionate1,2. Taking advantage of numerous field studies conducted by EU researchers in recent years, we report here a statistical analysis of crossfertilization data in maize, showing that separating fields 40 m is sufficient to keep GM adventitious presence below the legal labeling threshold in the EU set at 0.9%. Currently, insect-resistant maize (engineered to express Bacillus thuringiensis toxin; Bt) and Amflora potato (engineered with antisense against granule-bound starch synthase), which was recently approved3, are the only two GM crops authorized for commercial cultivation in the EU. Bt maize was approved in 1998 and currently covers 1.2% of the total maize area in the EU (Supplementary Notes 1 and 2). Given the legal standards for labeling and/ or purity, the cultivation of GM maize in the EU is associated with mandatory technical coexistence measures designed to reduce the adventitious presence of GM maize in neighboring non-GM maize harvests. Such measures, to be applied by GM maize growers, should be stringent enough to keep adventitious presence below 0.9% so that conventional maize can comply with labeling provisions and avoid any potential price premium losses associated with GM admixtures4,5. Cross-fertilization between neighboring maize fields is the most important ‘biological’ source of admixture between GM and conventional maize4,5. Factors influencing cross-fertilization rates in maize cultivation are well studied and include, among others, the distance between fields, flowering synchrony, weather conditions, the relative positions of donor and receptor fields (with respect to dominant winds in the area) and the size and shape of fields4. Because of the difficulty to control some of these parameters, regulatory bodies from most

Out-crossing (% seeds)

© 2010 Nature America, Inc. All rights reserved.

Distances needed to limit cross-fertilization between GM and conventional maize in Europe

30 25 20 15

4

3

2

1

0

10

0

25

50

75

100

125

150

Distance (m)

5 0 0

50

100

150

200

Distance (m)

Figure 1 Cross-fertilization rates for Bt maize. The figure shows a meta-analysis of maize crossfertilization data. Cross-fertilization rates are represented in relation to the distance from the pollen donor. The upper chart is a magnification of the original chart with a limited scale of the respective axis.

volume 28 number 8 AUGUST 2010 nature biotechnology

correspondence Table 1 Probability of keeping cross-fertilization below a certain threshold level (%) using a gamma distribution Cross-fertilization threshold (% of seeds)1 1.5% Mean (low-high bounds)

0.9% Mean (low-high bounds)

0.5% Mean (low-high bounds)

0.3% Mean (low-high bounds)

(0–10]

49.44 (46.10–52.92)

41.16 (37.80–44.62)

33.11 (29.76–36.66)

27.30 (24.06–30.64)

(10–20]

91.19 (88.58–93.70)

70.89 (67.56–74.38)

41.41 (37.78–45.06)

21.80 (18.20–25.68)

(20–30]

99.86 (99.54–100)

95.62 (92.12–98.44)

66.94 (58.30–75.14)

31.19 (21.52–41.00)

(30–40]

99.99 (99.96–100)

99.61 (98.76–100)

94.14 (87.70–99.44)

77.26 (63.70–91.08)

(40–50]

99.88 (99.56–100)

98.56 (96.10–100)

92.07 (84.12–99.80)

79.38 (66.48–95.34)

(50–70]

99.88 (99.28–100)

99.11 (96.26–100)

95.89 (87.30–99.90)

88.05 (74.54–96.86)

(70–90]

99.98 (99.90–100)

99.58 (98.68–100)

96.08 (91.56–99.94)

86.81 (77.48–97.66)

100 (100–100)

99.96 (99.86–100)

98.58 (97.30–99.54)

90.76 (86.22–94.76)

© 2010 Nature America, Inc. All rights reserved.

Distance (m)

>90 1Numbers

in italics indicate a scenario where separation distance is sufficient to reduce admixture in maize cultivation below different threshold levels (1.5%, 0.9%, 0.5% and 0.3%). Square brackets denote that the upper limit is included in the interval.

4.2% of the measurements are available from distances above 50 m from the donor field. The mean cross-fertilization rate and the standard deviation for each distance interval were calculated using all data points in the interval, and the highest and the lowest values for cross-fertilization rate registered (Supplementary Table 3). The mean and variance of each distance interval were used to calculate the parameters that characterize different probability distributions at those intervals. Once the distribution was obtained, probability of avoiding maize crossfertilization at different thresholds levels was calculated for each distance interval. To ensure robustness of the results obtained, different probability distributions were used following parametric and nonparametric approaches. Both approaches produced similar results. In the parametric approach, the probability distribution used to represent the cross-fertilization level for a given distance interval was the gamma distribution. The parameters of the gamma distribution were determined by the mean and the variance of the data in each interval. The probability distribution of crossfertilization being above a certain threshold level was obtained by conducting bootstrap sampling per interval 1,000 times. Bootstrap sampling allows obtaining a range of values for the parameters of the gamma distribution and therefore we were able to calculate the probability of being above a number of stated cross-fertilization thresholds (e.g., 0.9%; see ‘Gamma parameterization’ in Supplementary Note 3). We also estimated

a beta distribution to analyze the data (Supplementary Note 4). The nonparametric approach, where no distributional parameters are assigned, was based on a bootstrap simulation that consisted in drawing the observed data on cross-fertilization 1,000 times with replacement per interval. Therefore, we obtained 1,000 subsamples per interval. From each of these subsamples, the probability distribution of being above any cross-fertilization threshold can be calculated and mean and confidence intervals for the probability of being above a cross-fertilization threshold can be obtained. Table 1 shows the mean probability of keeping cross-fertilization between maize fields below different arbitrary threshold levels (1.5%, 0.9%, 0.5% and 0.3%) for each separation distance interval, using the gamma distribution. A 95% confidence interval of the mean probability of keeping cross-fertilization below a certain threshold is calculated (see low and high bounds for each distance interval). The results provided in Table 1 are relevant for policy decision-making. For example, implementing a 30 m separation distance would result in a probability higher than 95% (95.62%, see mean probability values in bold in Table 1) to keep crossfertilization values below the 0.9% EU labeling threshold. The probability increases to 99% if a 40 m distance is implemented. However, it is known that cross-fertilization is not the only source of GM adventitious presence in maize harvests. Traces of GM seeds in conventional seeds and machinery

nature biotechnology volume 28 number 8 AUGUST 2010

are considered to be additional contributors to final adventitious presence4,10. Greater distances to the pollen source would be required if lower threshold levels for crossfertilization were to be considered that aim to take into account additional sources of adventitious presence. For example, a distance of 40 m is needed to keep crossfertilization below 0.5% with a probability higher than 90% (94.1%). An analysis of the data in Table 1 also allows the effects of a hypothetical increase in the EU mandatory labeling threshold on segregation practices in maize cultivation to be estimated (countries such as Japan allow as much as 5% tolerance). For example, a 20 m separation distance would be sufficient to achieve a desired threshold level of 1.5% (with a probability of 91.19%). When using a nonparametric approach (bootstrapping simulation) results were quite similar to those obtained for the gamma distributions (Supplementary Table 4). The results presented here (Table 1) clearly show that some of the current mandatory separation distances proposed by several EU countries for maize segregation (Supplementary Table 1) are disproportionate. They are set too high to the objective of keeping cross-fertilization below the legal threshold level in real agricultural landscapes. Our results are robust because the experimental data set considered represents several climatic conditions, field sizes and locations in Europe. A previous study by Sanvido et al.5 looking at separation distances in Switzerland came to similar conclusions. Also, the levels of 781

© 2010 Nature America, Inc. All rights reserved.

correspondence cross-fertilization recorded in our database correspond to individual data points in receptor fields at several distances. Because most of the field points sampled were located at short distances from the donor field, crossfertilization rates at these distances were likely to be higher than cross-fertilization rates computed for an entire field harvested. In an agricultural context, harvest always represents a mixture of different harvested areas. The actual GM content in the harvest is thereby often substantially reduced because zones with higher cross-fertilization rates at the field margin are mixed with zones with lower GM content further within the receptor field. Studies performed in real agricultural landscapes with commercial cultivation of GM and non-GM maize point to distances over 20 m as being sufficient to prevent cross-fertilization below a threshold level of 0.9%11,12. In practice, large mandatory distances restrict farmers’ freedom of choice to grow GM maize in certain agricultural landscapes (especially in those with substantial presence of maize cultivation in small and scattered fields). This imposes important opportunity costs on farmers, reducing the potential net gains in farmers’ gross margins derived from Bt maize cultivation13. In conclusion, we have shown that a separation distance of 40 m is sufficient to reduce admixture in maize cultivation below the legal threshold of 0.9%. However, this is not an endorsement of using separation distances as the single tool to regulate coexistence in maize production. Numerous recent studies have pointed to the need for flexibility in co-existence measures4,14,15. Pollen barriers consisting of non-GM maize, for example, have proven to reduce cross-fertilization rates more effectively than an isolation of the same distance with open ground or low-growing crops. With a maize barrier of 10–20 m, the remaining maize harvest in the field rarely exceeds the threshold of 0.9% GM material11. Buffer zones, discard zones and other measures could therefore be combined or substitute for large, fixed-separation distances in search of a system that increases the real options for farmers to cultivate their crop of choice1. Note: Supplementary information is available on the Nature Biotechnology website.

782

Disclaimer The views expressed are purely those of the authors and may not in any circumstances be regarded as stating an official position of the European Commission. ACKNOWLEDGMENTS The authors thank M. Czarnak-Klos for help in the interpretation of the data sets of maize crossfertilization trials that constitute the database of this analysis and J. Delincé for his useful comments on statistical simulation. The authors wish to express thanks to G. Squire, as coordinator of the gene flow and ecological field studies of the SIGMEA project, for providing SIGMEA data sets on maize crossfertilization trials. Within the SIGMEA partners, many thanks are extended to R. Wilhelm for providing data under German agricultural conditions, A. Vogler for Swiss data and J. Messeguer for data from Spain. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests.

Laura Riesgo1, Francisco J Areal1, Olivier Sanvido2 & Emilio Rodríguez-Cerezo1 1European Commission, Joint Research Centre

(JRC), Institute for Prospective Technological Studies (IPTS), Edificio Expo, Avda. Inca Garcilaso, Seville, Spain. 2Agroscope Reckenholz Tänikon Research Station ART., Zurich, Switzerland. e-mail: [email protected] 1. Devos, Y., Demont, M. & Sanvido, O. Nat. Biotechnol. 26, 1223–1225 (2008). 2. Moschini, G. Eur. Rev. Agric. Econ. 35, 331–355 (2008). 3. Ryffel, G.U. Nat. Biotechnol. 28, 318 (2010). 4. Devos, Y. et al. Agron. Sustain. Dev. 29, 11–30 (2009). 5. Sanvido, O. et al. Transgenic Res. 17, 317–335 (2008). 6. European Commission. Commission Staff Working Document: Report from the Commission to the Council and the European Parliament on the Coexistence of Genetically Modified Crops with Conventional and Organic Farming. Implementation of National Measures on the Coexistence of GM crops with Conventional and Organic Farming. (Commission of the European Communities, Brussels, 2009). 7. Pla, M. et al. Transgenic Res. 15, 219–228 (2006). 8. Goggi, A.S. et al. Field Crops Res. 99, 147–157 (2006). 9. Vogler, A., Eisenbeiss, H., Aulinger-Leipner, I. & Stamp, P. Eur. J. Agron. 31, 99–102 (2009). 10. Demeke, T., Perry, D.J. & Scowcroft, W.R. Can. J. Plant Sci. 86, 1–23 (2006). 11. Messeguer, J. et al. Plant Biotechnol. J. 4, 633–645 (2006). 12. Gustafson, D.I. et al. Crop Sci. 46, 2133–2140 (2006). 13. Gómez-Barbero, M., Berbel, J. & Rodríguez-Cerezo, E. Nat. Biotechnol. 26, 384–386 (2008). 14. Demont, M. & Devos, Y. Trends Biotechnol. 26, 353– 358 (2008). 15. Messéan, A. et al. Oleagineux 16, 37–51 (2009).

volume 28 number 8 AUGUST 2010 nature biotechnology

case study

c o m m e n ta r y

India’s billion dollar biotech

Justin Chakma, Hassan Masum, Kumar Perampaladas, Jennifer Heys & Peter A Singer By focusing on an unmet medical need, providing a cost-efficient solution and reinvesting the resulting revenues into R&D and state-of-the-art manufacturing, Shantha Biotechnics was able to build one of India’s first biotech successes.

© 2010 Nature America, Inc. All rights reserved.

S

hantha Biotechnics, an Indian biotech firm started by K. I. Varaprasad Reddy with $1.2 million of angel funds, was acquired last year by Sanofi-Aventis of Paris for €571 million. Since developing a copy of the hepatitis B surface antigen subunit vaccine—one of the first recombinant products to be ‘home grown’ in India—Shantha has been on a tear, bringing 11 products to market. Much of the company’s success can be attributed to the vision of its management, which brought its first product to market in only four years, reinvested revenues into internal R&D and built a state-of-the art manufacturing capability. This not only enhanced the company’s ability to address local health needs, but also built its global reputation—all of which has subsequently proved good business.1 After attending a conference in 1992, Varaprasad, an electrical engineer by training, recognized the urgent need for an inexpensive Indian hepatitis B vaccine; over 100,000 Indians die every year from the viral infection, with 4% of the population carriers. Prices were as high as $23 a dose with primary suppliers being Merck and SmithKlineBeecham (now part of GlaxoSmithKline). With most Indian families living on $1 a day, with multiple children and three doses required per child, vaccination was simply unaffordable. Varaprasad saw the possibility of a local venture that could supply an affordable version. After recruiting local talent and two expatriate scientists in 1993 (see Supplementary Tables), the company took only four years to develop and register Shanvac-B, a version of the vaccine produced in Pichia pastoris. Shanvac-B was launched at $1 a dose and was an immediate success. Indian consumption of hepatitis B vaccine rose from a few hundred thousand doses in the early 1990s to tens of millions today with prices dropping as low as $0.25. Rapid uptake of the vaccine was partly helped by a confidential partnership with a large pharmaceutical multinational, which provided manufacturing/regulatory acumen and also resold the vaccine. Shantha followed Shanvac-B with Shanferon (interferon alpha 2b), which it also produced in P. pastoris. The company’s development of a purification process compliant with International Conference on Harmonization regulations led it to become the first Indian company to have a hepatitis B vaccine prequalified by the World Health Organization (WHO; Geneva). The initial investment in quality control helped accelerate approval for its other products. The company’s growing reputation for manufacturing excellence and regulatory expertise in recombinant vaccines also helped to secure business from entities in other developing countries, such as the International Vaccine Institute (IVI; South Korea) for low-cost oral cholera vaccine, and the Pediatric Dengue Vaccine Initiative (South Korea). This success led to international attention in 2006 when Mérieux Alliance (Paris, France) acquired a 60% stake in Shantha after its Omani investors sought an exit. The acquisition further bolstered Shantha’s reputation internationally as well as opening new markets. In 2009, the firm was awarded a $340 million United Nations International Children’s The authors are at the McLaughlin-Rotman Centre for Global Health, University Health Network and University of Toronto, Toronto, ON, Canada. e-mail: [email protected]

nature biotechnology volume 28 number 8 august 2010

Emergency Fund (UNICEF) contract for pentavalent vaccines from 2010–2012. Soon after, rumors emerged that multinationals were interested in bidding on Shantha, ultimately culminating in the takeover by Sanofi-Aventis the same year. The case of Shantha shows developing world biotech innovators can maintain a balance between local health impact and financial returns by keeping four principles in mind. First, identify therapeutic areas where cost efficiencies can be achieved locally and combine this with strong leadership skills. Varaprasad leveraged India’s homegrown scientists, lower labor costs, process innovation and a low-margins business strategy to exploit this opportunity. Second, seek investments/partnerships from non-traditional and international sources. Shantha embraced collaborations with research institutes such as the US National Institutes of Health (Bethesda, MD), and with competing multinationals for regulatory guidance. Third, focus on innovation and reinvestment. By plowing back significant profits toward R&D, Shantha has recently released new products every year or two. This initial focus on process and quality innovation may have delayed Shanvac-B’s launch, but it allowed Shantha to become the first WHO-prequalified Indian firm for hepatitis B vaccine, and opened the door to large international contracts, including contract research. However, experience with Shanferon suggested that India’s regulatory environment had challenges in conducting complex clinical trials. Other innovators in developing countries should not insist upon home-grown manufacturing or clinical trials if it entails compromise on quality for the sake of patriotism. Finally, Shantha shows integrated business models are viable in developing countries. Pre-acquisition, Shantha would not invest in any products for which it did not have internal capacity to execute on a significant part of the project. This contrasts with the developed world, where it is becoming increasingly popular to develop a ‘virtual’ business model, whereby clinical trials and even early stage work is outsourced to contract research organizations. Shantha shows the virtual model may not make sense for an innovative biotech in a developing country because the risks of low quality and delays in outsourcing are too great. By maintaining internal development capabilities, Shantha and other developing country firms can also capitalize on earnings generated by contract research work for other companies. By combining cost-efficiency with focused R&D, biotech firms like Shantha are creating a new source of innovation for global health. Funding This work was funded by a grant from the Bill & Melinda Gates Foundation through the Grand Challenges in Global Health Initiative. Note: Supplementary information is available on the Nature Biotechnology website. Competing Financial Interests The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/. 1. Prahalad, CK. The Fortune at the Bottom of the Pyramid: Eradicating Poverty through Profits. (Wharton School Publishing, Philadelphia; 2004).

783

c o m m e n ta r y

DNA patents and diagnostics: not a pretty picture © 2010 Nature America, Inc. All rights reserved.

Julia Carbone, E Richard Gold, Bhaven Sampat, Subhashini Chandrasekharan, Lori Knowles, Misha Angrist & Robert Cook-Deegan Restrictive licensing practices on DNA patents are stymieing clinical access and research on genetic diagnostic testing. Diagnostic companies, university tech transfer offices and their respective associations need to pay more attention.

F

our decades after the US Supreme Court first held that an artificially created bacterium had the potential to be patented in the United States1, biotech patents continue to generate controversy—particularly human gene patents used in diagnostic testing. The persistence of the debate can be attributed to particular business models for genetic testing and university licensing that, despite public pronouncements to the contrary, have failed to acknowledge and appropriately address the real social and economic concerns raised by clinical geneticists, health care professionals, patient groups, politicians and academics. Their failure has led both policymakers and the courts to express increasing concern about broad patent rights over human genes that affect diagnostic testing. The most recent flare-up in the ongoing DNA patent and genetic testing debate is Julia Carbone is at Duke University’s School of Law, Durham, North Carolina, USA; E. Richard Gold is at McGill University’s Faculty of Law and Faculty of Medicine, Montreal, Québec, Canada; Bhaven Sampat is at Columbia University’s Department of Health Policy and Management, New York, NY, USA; Subhashini Chandrasekharan is at Duke University’s Center for Genome Ethics, Law & Policy, Institute for Genome Sciences and Policy, Durham, NC, USA; Lori Knowles is at the University of Alberta’s Health Law Institute, Edmonton, Alberta, Canada; Misha Angrist is at Duke University’s Institute for Genome Sciences & Policy, Durham, NC, USA; and Robert Cook-Deegan is at Duke University’s Center for Genome Ethics, Law & Policy, Institute for Genome Sciences and Policy, Durham, NC, USA. e-mail: Robert Cook-Deegan: [email protected]

784

Myriad Genetics has been the poster child for controversial DNA patent licensing.

the decision of the US District Court for the Southern District of New York in Association for Molecular Pathology et al. v. United States Patent and Trademark Office et al.2. On 29 March, US Federal District Court Judge Robert Sweet ruled that isolated DNA is not patentable in the United States, and also that Myriad Genetics’ (Salt Lake City, UT, USA) method claims relevant to testing for BRCA1 and BRCA2 genes are invalid. Essentially, the District Court held that neither isolated DNA nor cDNA is sufficiently different from DNA as it occurs within host cells to be considered an invention. As for the diagnostic tests, the court held that they simply involved drawing a mental correlation between facts, something that does not fall within the scope of what is patentable. A week earlier, the US Court of Appeals for the Federal Circuit held in Ariad Pharmaceuticals, Inc. et al. v. Eli Lilly and

Company3 that a researcher must do more than identify that a class of compounds has a certain effect: he or she must actually describe what those compounds are. This effectively eliminated the award of patents over basic research, requiring, instead, that the inventor “actually perform the difficult work of ‘invention’—that is, conceive of the complete and final invention with all its claimed limitations—and disclose the fruits of that effort to the public.” One month before that, on 10 February, the Secretary’s Advisory Committee on Genetics, Health and Society (SACGHS; Bethesda, MD, USA) at the US Department of Health and Human Services4, after a careful study of current knowledge on the effects of patenting genes on research and accessibility to genetic tests, found that there is no convincing evidence that patents either facilitate or accelerate the development and accessibility of such tests. What’s more, the committee

volume 28 number 8 august 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

C O M M E N TA R Y found that there was some, albeit limited, evidence that patents had a negative effect on clinical research and on the accessibility of genetic tests to patients. In addition, most gene patents relevant to diagnostics are held by universities on the basis of research funded by public money. In this context, the committee recommended that universities be more cautious in patenting and licensing human genes, that there be more transparency and accountability for university licensing practices and that an existing exception protecting medical practitioners from patent infringement when they undertake surgery or treat a patient’s body be extended to include the provision of genetic diagnostic testing. What all three developments have in common is that they reflect growing disenchantment with the patenting and licensing practices of universities and industry. These concerns have existed for over a decade without resolution5,6. The maturity of microarray technology that allows for multi-allele genotyping and now the prospect of full-genome sequencing deepen these concerns7. A legacy of exclusively licensed gene patents casts a shadow of patent infringement liability over the future of multiallele testing and full-genome analysis. In an attempt to better understand why concerns about DNA patenting persist and what role universities play as patentees and often exclusive licensors, this article outlines university technology transfer practices and business models that have given rise to the concerns. After outlining the practices that have given rise to concerns about the patenting of human genes for diagnostic genetic tests, we review past efforts attempting to address concerns. We then lay out the obstacles to addressing these concerns going forward, including a lack of recognition that diagnostics is a highly unusual market—and that the problem is not so much a legal question or necessarily about what gets patented, so much as how patents are licensed and enforced by both universities and industry. The ability to change these restrictive licensing practices, will, in turn, depend on several factors: first, a sharper definition of what constitutes research that needs to be protected in licensing provisions; second, more coherent university policies that promote broad dissemination, along with incentives for industry compliance with best practices; third, greater recognition of problems and the proposal of constructive solutions by key players; fourth, transparent reporting of DNA patents and diagnostic testing license agreements; and fifth, secure funding for technology transfer offices. Although legislative change may ultimately be necessary to facilitate these changes in practice, many problems can be addressed without statutory change.

A legacy of short-sighted tech transfer and business practices Currently, universities frequently file patents on early-stage inventions9, and license patents exclusively half the time10–13. A study by Mowery et al.10 notes the following: “A relatively high fraction of all inventions that are licensed—as high as 90% for UC [University of California] licenses and no less than 58.8% for Stanford licenses of ‘all technologies’ during this period—is licensed on a relatively exclusive basis, and these shares are similar for biomedical inventions.” Many of those licenses will endure for many years, including licenses on university patents relevant to DNA diagnostics. Universities and academic medical centers that provide diagnostic testing services face private genetic testing companies that enforce patents against university genetic testing services and national reference laboratories5—in contrast to the situation for therapeutics, where universities are often the plaintiffs. The story often begins with publicly funded academic or nonprofit research that is either patented and licensed exclusively to a private company or forms the basis for a spin-off company that attracts further investment and develops an invention that is patented. Whether exclusive licensees or spin-offs, these companies then develop genetic testing services based on a business model that relies not only on patenting sequences and mutations—not objectionable in itself—but also on preventing other institutions, including universities from offering those genetic tests. The case of Myriad patents over BRCA1, BRCA2 and methods for diagnostic testing14, as well as Athena Diagnostics’ exclusive licenses for clinical testing from Duke University (Durham, NC, USA) over three method patents related to diagnostic testing for Alzheimer’s disease15,16, exemplify these practices and business models. Furthermore, other neurological and metabolic conditions, as well as other entities’ screening for Canavan disease, hemochromatosis and other single-gene conditions, has also generated fierce debate. In the case of Canavan testing, litigation resulted from licensing restrictions that inhibited freedom of action among those seeking to get genetic tests. In the case of Myriad, initial research took place at the University of Utah—with public funding from the US National Institutes of Health (NIH; Bethesda, MD, USA). The researchers then spun off Myriad, which attracted investment from Eli Lilly (Indianapolis, IN, USA) and succeeded in patenting BRCA1 and a diagnostic test for breast cancer (patents that were ultimately jointly assigned to the University of Utah, Myriad and

nature biotechnology volume 28 number 8 august 2010

the NIH). Rather than licensing out the test to clinical geneticists and laboratories around the world, Myriad required initial testing in each family to be performed at its laboratories in Salt Lake City. In the United States, the company sent out cease-and-desist letters to laboratories—both academic and commercial—already performing tests when the patent was issued. Threatened patent enforcement resulted in a backlash around the world from public laboratories, clinicians, molecular geneticists and some patient groups—against both the patenting of human genes and what they viewed as Myriad’s strong-arm tactics. These groups feared that by closing down public laboratories, Myriad would thwart research identifying weaknesses in Myriad’s test or distinguishing the effects of different mutations in the genes on disease severity or progression, and prevent the integration of breast and ovarian cancer genetic tests into genetic health services. Although some of these fears were clearly exaggerated, Myriad’s aggressive initial patent enforcement affected practice in the clinical genetics community and stirred long-standing resentment. Furthermore, in countries with public health care systems, health administrators objected to Myriad’s business model because it removed their ability to deploy genetic tests to their citizens in the manner that they viewed as most efficient14. Myriad always permitted what it considered to be basic research on BRCA1 and BRCA2, and also engaged in research collaborations. In fact, until 2004—after which Myriad ceased to do so for unknown reasons—the company contributed data to public databases. To illustrate Myriad’s openness to others performing basic research using BRCA1 and BRCA2, the company’s president, Greg Critchfield, has identified 7,000 papers published by independent authors that mention BRCA1 or BRCA2 (http://docs.justia. com/cases/federal/district-courts/new-york/ nysdce/1:2009cv04515/345544/158/0.pdf ). This indicates that, with the exception of clinical testing at the University of Pennsylvania in 1998, Myriad did not pursue those who conducted research. Myriad also defined the University of Pennsylvania’s testing as ‘commercial’, as later defined under the terms of a 1999 Memorandum of Understanding with the US National Cancer Institute (NCI: Bethesda, MD, USA). Myriad has been successful in arranging for payment agreements with insurers and other payers. However, as a result of Myriad’s enforcement actions coupled with broad patent claims, its fairly narrow conception of what constituted acceptable research and its failure to clearly state that it would not pursue those conducting such research, university and private laboratories ceased to offer the test publicly 785

© 2010 Nature America, Inc. All rights reserved.

C O M M E N TA R Y in the United States. Outside the United States, resistance to Myriad’s model—particularly from health care administrators and government departments—caused the company to lose most of its market. Furthermore, Myriad’s relationship with scientists and policymakers around the world was seriously damaged14. Although the biotech industry tried to portray Myriad as an outlier, a series of detailed case studies conducted by some of us (J.C., S.C., M.A. and R.C.-D.) and others15,18–24 at Duke University’s Center for Genome Ethics Law and Policy reveal that, in fact, Myriad’s business model is not unique. As these studies show, diagnostic companies such as Athena Diagnostics (Worcester, MA) and PGxHealth (New Haven, CT) have adopted similar or even more aggressive business models and have shut out university laboratories from offering genetic testing for diseases such as long-QT syndrome and Alzheimer’s disease. In the case of Alzheimer’s disease, genes and method patents for diagnostic testing were initially patented by Duke University (and other academic institutions) and licensed exclusively to Athena Diagnostics. Athena Diagnostics then used its patents aggressively to prevent others from carrying out the test. These case studies strongly suggest both that universities are often not managing research and patents in a way that promotes dissemination and that companies deploy their patents or exclusive licenses to remove genetic testing laboratories at academic health centers and lowmargin national reference laboratories from the market. This is demonstrably a viable business model, or at least it has proven to be until recently—but is it good national policy, and does it add value to the national health system? As clinicians and laboratory directors react to cease-and-desist letters by withdrawing from those activities, clinical research and genetic testing are impeded. GeneDx (Gaithersburg, MD) and university laboratories ceased testing for the life-threatening long-QT syndrome after patent enforcement in 2002, for example, but no commercial test entered the market until 2004 (ref.  9); neither the University of Utah (which held the patents) nor the NIH (which could have been petitioned to march in, given that ‘health and safety’ needs were not being met) took action. Certain tests may not be offered if the patent holder or exclusive licensee does not provide them; second-opinion and verification testing may be unavailable; and tests are costly to public and private payers, sometimes prohibitively so for those lacking insurance25,26. Although negative effects on price and access to genetic testing are not uniform, consistent or pervasive, one cannot read the case studies as a whole without realizing 786

there are real problems—and also that there are relatively easy solutions modeled on nonexclusive licensing, as used for Huntington’s disease and cystic fibrosis testing. Gene patents over diagnostics are not just like all other patents, and the diagnostic market is not just like markets for therapeutics and instruments. Holders of gene patents need to take care in licensing them for diagnostic use. Hurdles to resolution of concerns The past decade saw a plethora of policy reports about DNA patents, such as those from the Nuffield Council on Bioethics17, the US National Academy of Sciences27, the Ontario Ministry of Health28 and the Australian Law Reform Commission29. Academic articles examined the concerns, the extent to which concerns were founded and the roles of industry, universities and legislative reform in addressing these concerns5,6,26,30–38. Some countries also made statutory changes to their patent and health laws. France expanded compulsory licensing laws39, and Belgium did the same, also carving out a diagnostic-use exemption from patent-infringement liability40. The

In addition to evidence that gene patents covering diagnostics do not necessarily impede research, there is very little evidence of patent litigation in the field. US Patent and Trademark Office (USPTO; Washington, DC) developed guidelines on ‘utility’ and ‘written description’ specifically for examining gene patent applications41. Recognizing that many of the concerns could be addressed through better licensing practices, many institutions also developed licensing guidelines, some aimed at universities and others at industry. These include the NIH’s Best Practices for the Licensing of Genomic Inventions42, the Organisation for Economic Cooperation and Development’s (OECD; Paris) Guidelines for Licensing of Genetic Inventions43 and In the Public Interest: Nine Points to Consider in Licensing University Technology44, a document crafted by 12 institutions and subsequently endorsed by the Board of Trustees of the Association of University Technology Managers (AUTM; Deerfield, IL, USA). Since then, ~50 other institutions and organizations have also endorsed the guidelines. In November 2009, as part of AUTM’s Global Health Initiative to promote licensing practices that facilitate access to essential

medicines in developing countries, AUTM also endorsed a document entitled University Principles on Global Access to Medicines45. Most recently, the SACGHS recommended the implementation of an exception to patentinfringement liability for research use and diagnostic testing4. All of these reports and recommendations focus on broad dissemination through nonexclusive licensing of gene-based inventions, particularly for publicly funded research. They reserve exclusive licensing for situations in which it is needed to induce investment in private-sector development to bring a product or service to fruition—which, as will later be discussed, is rarely the case for genetic diagnostics. Despite the plethora of policy reports, academic articles, guidelines and legislative changes, concerns about DNA patents persist. We must therefore turn our attention to factors that impede changing the system. A question of law or of practice. The first response to concerns is often a call to change patent law39,46,47. As recent research indicates, however, the central problem does not lie with patents over human genes themselves so long as the law incorporates the appropriate checks and balances. The recent suit challenging Myriad’s patents on BRCA genes notwithstanding2, the following discussion indicates that there is little evidence on which to conclude that limiting the ability to patent genes is the only way to solve the problems in the system. A recent study by Huys et al.48 from Belgium suggests that relatively few claims in gene patents block competing laboratories from providing genetic tests. This study of 145 active patent documents (267 independent claims) related to genetic diagnostic testing of 22 inherited diseases (including method claims, gene claims, oligo claims and kit claims) that the European Patent Office (Munich, Germany) and the USPTO issued. It concluded that clinicians could easily get around 36% of claims and could, with work, circumvent another 49% of claims. Only 15% of claims would be difficult or impossible to circumvent. Of the gene claims studied, only 3% were found to be blocking. However, as discussed below, blocking claims were more prevalent among method claims. In addition to evidence that gene patents covering diagnostics do not necessarily impede research, there is very little evidence of patent litigation in the field. A recent study8 on trends in human gene patent litigation notes that there is rarely any litigation over diagnostic tests arising from gene patents. This study identified only 31 examples of litigation over human genes in the United States from 1987 to 2008. Although the low frequency of litigation

volume 28 number 8 august 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

C O M M E N TA R Y could hypothetically support the conclusion that patents successfully exclude others (that is, threatened patent enforcement stops potentially infringing activities), an examination of patent claims suggests that most patents over human genes and related diagnostic tests find themselves in a relatively weak legal position. This weak legal position is further reinforced by the dissent in Laboratory Corp. of America Holdings v. Metabolite Laboratories, Inc.49, which concluded that a natural correlation between two substances in the body was an unpatentable product of nature (the majority decided not to address the issue); by the United States District Court decision in Association for Molecular Pathology et al. v. the United States Patent and Trademark Office et al.; and by the general trajectory of recent decisions on assessing damages, the lack of automatic injunctive relief (eBay Inc v. MercExchange, L.L.C.50), as well as by the increasing ambit for finding an invention to be obvious under patent law. The recent US Supreme Court decision In re Bilski51 only exasperates the uncertainty over method claims on DNA diagnostics. In fact, an eventual appeal from the District Court decision in Association for Molecular Pathology et al. v. the United States Patent and Trademark Office et al. may be required to determine whether these type of claims are valid. Adding to the trend in legal thinking is the Federal Circuit’s decision in Ariad, relating to claims based on DNA patents, where the court writes: “Much university research relates to basic research, including research into scientific principles and mechanisms of action…, and universities may not have the resources or inclination to work out the practical implications of all such research [i.e., finding and identifying compounds able to affect the mechanism discovered]. That is no failure of the law’s interpretation, but its intention. Patents are not awarded for academic theories, no matter how groundbreaking or necessary to the later patentable inventions of others.” That research hypotheses do not qualify for patent protection possibly results in some loss of incentive, although Ariad presents no evidence of any discernable impact on the pace of innovation or the number of patents obtained by universities. But claims to research plans also impose costs on downstream research, discouraging later invention.” Taken together, these studies and cases indicate that gene patents per se have closed off far less of the research landscape than is often supposed, and where expansive claims have been granted, many are vulnerable to challenge. Method claims in patents related to diagnostic testing, however, bear special mention. Although many pharmaceutical patents claim

products as chemical entities, universities and biotech firms also tend to patent ways of using knowledge, including method patents that affect genetic tests. In fact, Huys et al.48 conclude that 30% of method claims relating to genetic testing are difficult, if not impossible, to circumvent. Such claims tend to be broad, often to the point of vagueness, and many cover all conceivable ways to conduct genetic tests on a gene or for a clinical condition. In the 15 of 22 conditions that Huys et al.48 found had at least one blocking claim, most such claims were to methods. In the diagnostic realm, blocking patents thus appear to be common, present in 68% of the clinical conditions studied. Changes in jurisprudence could reduce the number of truly blocking patents in genetic diagnostics. Recent and pending court decisions suggest that some fraction of broad claims in US patents on DNA sequences and methods pertinent to genetic diagnostics would be judged invalid if challenged. Although dealing with a patent claim in the information technology field, the recent US Court of Appeals for the Federal Circuit decision in In re Bilski narrowed criteria for patents on methods to inventions that entail a transformative step or involvement of a particular machine. Depending on how the Federal Circuit deals with the US Supreme Court in Bilski—perhaps in an appeal in the Myriad case—it could signal that broad method claims in DNA diagnostics might be held invalid because the link between a mutation and a probability of contracting a disease may be considered unpatentable. As it stands, many broad method claims pertinent to DNA diagnostics suffer under a cloud of uncertainty and may turn out to be invalid, thus dramatically increasing freedom to operate without fear of patent-infringement liability. Other recent US court decisions have moved in the same direction, increasing the stringency of criteria for nonobviousness52,53 and written description3. Taken as a group, these decisions suggest that some of the potential obstacles to innovation that patents cause in diagnostics may not be as high, nor the amount of intellectual territory enclosed and enforced as expansive, as some had feared. A clear research exemption, a simplified method for challenging patents (for example, opposition proceedings or inter partes re-examination requests) and improved examination procedures to avoid overly broad patent claims could help quell concerns over blocked research and overly broad patents54. Overall, the problem does not lie wholly in patent law but rather concerns how decisions are made about what is patented (methods versus products) and how patents are managed and used. With one or a few successful challenges to broad patents enforced for

nature biotechnology volume 28 number 8 august 2010

diagnostic purposes, the business models of enforcing monopolies on genetic testing for specific conditions would probably give way to more cross-licensing, more competition and faster innovation in testing methods. A need for changes in patent licensing practices at universities. As patent law evolves, it is increasingly apparent that the exclusive licensing strategies of universities and the business models of a few companies doing DNA diagnostics are as much, or even more, of an impediment to DNA diagnostics as any problems with the law. Meanwhile, no evidence suggests that exclusive licensing is as important in the field of diagnostic testing as in therapeutics in creating products that would not otherwise exist. The exclusive licenses over erythropoietin, growth hormone, interferon and other therapeutic proteins are of commercial significance, as illustrated by the fact that eleven legal cases that presume the validity of gene patents have been decided by the US Court of Appeals for the Federal Circuit8. The same cannot be said for diagnostic testing: no exclusive license in this field has been deemed to be of such importance for anyone to take to court. In fact, most cases involving diagnostic testing are settled after initial notification letters or cease and desist letters are sent out. A handful have led to litigation, but settled early. The Federal District Court’s ruling of 29 March in Association for Molecular Pathology et al. v. the United States Trademark and Patent Office is the first diagnostic case to go before a judge for a decision. Furthermore, barriers to entering the market with a new genetic test, at least for the first-generation genetic tests that search for mutations in one or a few genes, are far lower than for therapeutics. This is because for universities and national reference laboratories that already offer other genetic tests, the cost of ‘setting up’ a new genetic test based on data in scientific publications is comparable to the cost of patenting the underlying inventions since they are already laboratories approved by US regulators. Supporting this proposition is the fact that exclusive licensing does not appear to have been necessary to get a test to market in any of the cases15,18–24 studied for SACGHS. In the study of 10 clinical conditions considered by SACGHS, three cases did not involve patent rights (i.e., there were no patents or patents were not licensed or enforced) or patents were nonexclusively licensed to multiple providers. These were cystic fibrosis, hereditary colorectal cancer and Tay-Sachs disease. Such patenting and licensing practices comply with current guidelines. In six cases, however, exclusive licensing led to patent enforcement that 787

© 2010 Nature America, Inc. All rights reserved.

C O M M E N TA R Y reduced availability of genetic tests already being offered: HFE (hemochromatosis), APOE, Alzheimer’s disease and genes associated with Canavan disease, long-QT syndrome, hearing loss and spinocerebellar ataxias. Because tests were already available, exclusive licensing in these cases deviates from the norms that technology licensing offices generally claim to be following. In some cases, but not all, this led, at least transiently, to genetic testing by a single provider, and that exclusive license holder then eliminated other testing services that had beaten it to market. In all cases except hemochromatosis, exclusive licenses from universities were involved. Although the exclusive licensee may ultimately have developed a better test, in no case was the exclusive licensee the first to market. The tenth clinical condition studied by the SACGHS, hearing impairment, is subject to a hybrid of exclusive and nonexclusive licensing, and entails many genes and different means of testing. This case does have some examples of controversial patent enforcement action, but tests are generally widely available from several vendors. Patent incentives may induce investment in genetic diagnostics, but in none of the case studies did this lead to new availability of a test that was not already available, at least in part. This is in stark contrast with the role of patents in therapeutics and scientific-instrument development, where the benefits attributable to private R&D and new products are much clearer. The SACGHS case studies thus reinforce the benefits of licensing nonexclusively for genetic diagnostics, unless an unusual situation arises in which exclusivity is needed to get a product to market for the first time. The cases also highlight deviations from the NIH Best Practices43, OECD Guidelines43 and the AUTM-endorsed Nine Points44. Exclusive licensing practices consistently reduce availability, at least as measured by the number of available laboratories offering a test, and thus reduce competition in genetic diagnostics, but with little evidence of a public benefit from services not otherwise available. Instead of recognizing this reality, some universities continue to seek broad patents regardless of subject matter and then license exclusively, enabling business models that impede competition in genetic testing. Although the real risk of being successfully sued for patent infringement in DNA diagnostics may be low, a 2003 survey33 and recent case studies14,15,18–24 indicate that laboratory directors change their testing practices and clinicians avoid research areas in reaction to cease-and-desist letters. Diagnostics are generally low-margin sources of revenue, and when faced with a threat of patent enforce788

ment, most laboratories simply stop offering a genetic test, or at least no longer advertise a test’s availability publicly (in all the case studies, we learned of ‘research’ testing as an ‘escape valve’ for patients who could not get or could not afford commercial genetic tests). Although part of the problem is that licenses executed over the past decade do not embody the principles of the NIH, OECD or AUTM guidelines and yet remain in force, the reality is that only a minority of universities have endorsed the consensus Nine Points44—with no repercussions for those who do not or those who sign and then violate the norms. Shortsighted licensing practices persist. Potential solutions Changes that could remedy problems with the current strategy of the licensing system include the following: first, a clear definition of research that should be exempt from patent-infringement liability; second, universities’ leadership in promoting the alignment of tech transfer licensing practices with the univeristies’ broader goal of dissemination; third, coupling of the latter with incentives to promote industry compliance and leadership by AUTM and the Biotechnology Industry Organization (BIO; Washington, DC) in recognizing problems and proposing constructive solutions; fourth, adequate funding for tech transfer offices to learn about and implement changing practices; and finally, greater transparency in reporting patent holdings and licensing agreement terms. A more detailed discussion of each of these follows. Defining what qualifies as research. Although most industries tolerate a broad range of research activities and most researchers ignore patents when deciding whether to do research55, such blithe ignorance is not an obvious option in human genetic diagnostics, where threatened enforcement is common, laboratory directors and clinicians tend to respond to threatened enforcement by ceasing the activities under threat and workaround in the case of method patents are not always available48. Norms over what research is to be tolerated are unsettled, despite the existence of research exceptions56 in many national laws (including an exemption in the United States for research into products that may eventually lead to the filing of an application with the US Food and Drug Administration (Rockville, MD)57). One prominent example of disputed norms is the controversy between Myriad and the University of Pennsylvania Genetic Diagnostic Laboratory (GDL; Philadelphia, PA). Although Myriad states that it is generally supportive of research, it nevertheless sent GDL a cease-and-

desist letter because it did not consider GDL’s activities to be research. To Myriad, GDL’s provision of testing services to researchers was commercial, not a research service14. GDL took the position, however, that its activities, which supported others’ research, fell within the norm of tolerated research use, and much of the contested testing was part of clinical trials funded by the NCI, which is clearly clinical research. Much debate ensued, leaving many researchers with the (wrong) impression that Myriad would not tolerate any form of research. In an attempt to establish a clear norm over the question of which activities should be considered ‘research’, Myriad entered into a Memorandum of Understanding with the NCI to provide at-cost or below-cost testing to the NCI and any researcher working under an NCI-funded project. Myriad also similarly offered to provide NIH researchers with at-cost testing, given that the NIH was a co-owner of some of the relevant patents. Importantly, the agreement with the NCI defined the type of research Myriad would tolerate as being “part of the grant supported research of an Investigator, and not in performance of a technical service for the grant supported research of another (as a core facility, for example).” Furthermore, testing services had to be paid for out of grant funds and not by a patient or by insurance. Under this definition, GDL was not conducting research. This agreement was acceptable to both parties (Myriad and the NCI), and given the ‘at-cost’ provisions and the known efficiency of Myriad in testing, perhaps it is a salutary precedent. It is worth noting, however, that the NCI did not seek to delegate its government use rights under the Bayh-Dole Act 35 U.S.C. § 200-212 (“Bayh-Dole Act”) or Stevenson-Wydler Act 15 U.S.C. 3701 (which pertain because Myriad’s patents include inventors covered by both laws). The restricted nature of the Myriad-NCI Memorandum of Understanding limits its value as a precedent. It covered only the provision of services by Myriad; it did not address the general question of which research practices a patent holder should tolerate in the diagnostics field. Some of the conflict surrounding patents and genetics laboratories could be avoided by adopting a clearer definition of ‘research’ for the purposes of incorporating licensing terms that lower the threat of patent-infringement liability. The scope of government use rights under the Bayh-Dole and Stevenson-Wylder Acts is another legal gray zone. In any case, the definition of research should not be left to the individual negotiation between one company and one NIH institute. The NIH could take on a key role in developing this norm by convening a meeting of interested parties to develop

volume 28 number 8 august 2010 nature biotechnology

C O M M E N TA R Y

© 2010 Nature America, Inc. All rights reserved.

the principles by which individual actors can determine how to apply the norm. University leadership. Implementation of licensing guidelines and best practices is difficult when interests and goals are not aligned. Participants at a workshop held at Duke University in April 2009 addressed the role of universities in DNA patents and diagnostic testing and noted that those at the front line of implementing these guidelines, tech transfer offices, face many hurdles to implementation. Many university administrators view patents as a means to secure revenues (to subsequently reinvest in research) and believe that exclusive licenses generate the most revenues. Although the evidence58 is quite clear that most tech transfer offices either break even or lose money and that many of the most lucrative university patents have entailed nonexclusive licensing, this view persists. Compounding this problem, universities expect tech transfer offices to generate sufficient revenues to be sustainable. Despite usually being unrealistic, such expectations can lead these offices toward licensing strategies that promote short-term income over dissemination and broad availability. If there is to be a change of behavior, it must come from two sources: first, university administrators must align tech transfer strategy with the university mission of broad knowledge dissemination; and second, universities should provide more push-back when threatened patent enforcement gets in the way of research and impedes the university’s central mission. Regarding the first point, university presidents and senior management must take seriously the university mission to disseminate knowledge and technology. They must consider technology transfer as one component of their strategy to enable the wider world to access, enjoy and use university-generated knowledge. To achieve change, they need to change the way they fund tech transfer offices so that the latter have the freedom to explore alternatives to the way they currently license out technology. They also need to develop clear goals for dissemination and ensure that they impose measures of success for their technology licensing offices that correspond to those goals. Expecting technology licensing officers to forgo exclusive licenses when companies seek them is unrealistic unless the officers are rewarded for decisions that acknowledge the broad social benefit of avoiding patent thickets in genetic diagnostics. Recognition must also be given to the fact that these offices do not negotiate licenses in a vacuum: they negotiate largely with industry partners. If diagnostic companies are unwilling to accept

nonexclusive licenses, broad research exemptions or other terms that universities propose to support research, tech transfer offices have little room to maneuver. Currently, there is no incentive—whether external or through the threatened use of government march-in rights under the Bayh-Dole Act—to curb industry behavior even when it is problematic. Tech transfer departments with limited funding, limited staff and unreasonable expectations to be sustainable cannot be expected to resist intransigence by licensees. Universities need also to take a lead in encouraging their researchers, clinicians and laboratory directors to push back when threatened with patent enforcement. University administrators need to educate themselves and their staff about the freedom to operate for purposes of research and improving diagnostic testing—that is, the scope of activities allowed that do not infringe on a valid patent. University

Implementation of licensing guidelines and best practices is difficult when interests and goals are not aligned. administrators, researchers, clinicians and laboratory directors can act together by sharing cease and desist letters or other patent enforcement actions to determine whether the activities are, in fact, infringing. They can share expertise about the validity of patent claims that threaten research or clinical testing. Although individual laboratories may lack the resources to conduct these analyses, other institutions may have the requisite resources (for example, the American Society of Human Genetics, the American College of Medical Genetics, the College of American Pathologists and academic units such as the science policy research units at the University of Sussex in Brighton, UK, and the University of Leuven, Belgium). Leadership from AUTM and BIO. The development of a ‘gene patent supermarket’ by Denver firm MPEG-LA is a promising step toward enabling nonexclusive licensing, increasing simplicity and consistency in licensing terms, and reducing transaction costs59. Unfortunately, instead of proposing such constructive solutions, BIO and AUTM have chosen not to acknowledge the real problems that exist in the unusual market for genetic diagnostics and have been quick and vociferous in their opposition to the recommendations of the SACGHS60,61. It is impossible to judge the full extent of the problems, but it is

nature biotechnology volume 28 number 8 august 2010

certainly poor policy to deny that they exist at all. Moreover, BIO and AUTM have expended time and resources opposing SACGHS recommendations while failing to enforce the established norms laid out by the NIH and the OECD, as well as the AUTM-endorsed Nine Points, among their respective constituencies. Companies and universities that violate those norms have faced no action, or even recognition that they have deviated. Indeed, there has been no public statement from either BIO or AUTM that members have been responsible for some of the problems uncovered in licensing practices for genetic diagnostics. It is reasonable to disagree with the SACGHS recommendations, but it is not reasonable to read the SACGHS report and the case studies prepared for it and conclude that the system is working well across the board. BIO and AUTM should recognize the very real problems that have been uncovered, exhort compliance with established norms and—even more importantly if such norms are to be meaningful—criticize deviations from them, rather than following the politically expedient tactic of focusing their fire on SACGHS recommendations intended to prevent these problems. The two most controversial SACGHS recommendations are, first, a proposed exemption from infringement liability for research use, and second, a similar exemption for diagnostic use. As previously noted, university licensing offices opposing a research exemption puts them at odds with their own stated principles, as licensing to ensure freedom to do research appears in every document proposing norms for licensing. Opposition to a diagnostic-use exemption is more understandable because it may be that there are unusual situations in which exclusivity is needed to get a product or service to market, and such situations simply have not been captured in the cases studied to date. Nevertheless, it is quite clear that in many if not most cases of genetic diagnostics, the main use of exclusive licenses from universities has been to reduce competition and reduce the number of laboratories offering tests, without apparent benefits of introducing tests that were not already available. Rather, tests would demonstrably have been available even without the participation of the companies involved. The SACGHS may have judged that tech transfer offices are failing to respect existing norms, and in the absence of any credible compliance measures, the simplest legal solution is to address the problem through exemption from infringement liability. If AUTM and BIO want to preserve the option of exclusive licensing when needed to get genetic tests to market, then compliance with guidelines needs to be credible. Criticizing deviations when 789

© 2010 Nature America, Inc. All rights reserved.

C O M M E N TA R Y they come to light, with the long-term goal of increasing compliance with stated norms, would go a long way toward reducing the need for a diagnostic-use exemption. Moreover, enforcing nonexclusive licensing norms can preserve revenue streams, as seen in the cystic fibrosis and Huntington’s models, whereas a diagnostic-use exemption would eliminate those revenues because the patents would be unenforceable for diagnostic uses. One could object that it is neither the function nor the responsibility of either BIO or AUTM to criticize their members. BIO is an industry lobby group that sees itself as “the champion of biotechnology and the advocate for its member organizations,” whereas AUTM is an association of individuals working in tech transfer that seeks “to support and advance academic technology transfer globally.” Developing and enforcing patenting and licensing policies fall within neither mandate. This argument is, however, disingenuous, given that both AUTM and BIO claim to be working to ensure that tech transfer serves the public good. It is just as important to reduce practices that fall short as to promote practices that achieve the goals of their respective constituencies. Both organizations have endorsed the Nine Points guidelines and actively promote technology transfer “in a manner that is beneficial to the public interest” (http://bio.org/ip/techtransfer/) while “improving quality of life, building social and economic well-being, and enhancing research programs” (http://betterworldproject.org/ tech_transfer.cfm). Having voluntarily taken these positions, both organizations should be held accountable for them. Increasing transparency to permit ‘system learning’. To promote change, universityindustry relationships need to be more transparent; indeed, the current opaqueness over existing university-industry interactions is a major hurdle to improving the intellectual property system for DNA diagnostics11. For example, license agreements between universities and start-up and private companies are unavailable, even in general terms. The only exceptions are universities or companies that voluntarily make such information public. Participants at the workshop on the role of universities in DNA patents and diagnostic testing held at Duke in April 2009 noted that most licensing information is not publicly available, even for inventions arising from public funding. In some cases, but only some, it is possible to reconstruct licensing terms from company annual reports or from press announcements. There is often no way for researchers and institutions to know what practices a license covers, whether there remains scope for others to 790

practice an invention, which regions it covers and whether it applies to any specific fields of use or contains special restrictions. The lack of information makes it difficult to substantiate claims that licensing practices are changing or comply with best practices. As a study11 on university licensing practices notes, simply stating whether a license is exclusive or nonexclusive misses important nuances. Not only would more transparency help researchers better understand the scope and ownership of intellectual property rights, it would also allow policymakers, academics and tech transfer offices to determine in what cases exclusive licensing is justified, as opposed to enforcing a blanket norm of nonexclusive licensing. Although under provisions62 of the BayhDole Act, all recipients of federal grants must report on activities involving the disposition of certain intellectual property rights that result from federally funded research, the information is incomplete and cannot be obtained

Data on patenting and licensing practices are languishing in a government database that is not mined for valuable insights. because of strictures on access to the data. A clause of the legislation was intended to protect proprietary data from public access through the Freedom of Information Act 35 U.S.C. § 202(c) (5). The way the implementing regulations were written, however, went well beyond this, and gave licensees veto power over nongovernment disclosure of information. Tech transfer offices file reports with the interagency Edison (iEdison) database when they license inventions supported by most government funders. The reporting requirements do not require the disclosure of the licensing terms, and what is reported to iEdison is not publicly available. Indeed, access to iEdison is highly restricted; the database is unavailable for study or use outside government, and even government officials wanting to study technology transfer have been denied access unless they get permission from all licensees, a nearly impossible hurdle to overcome. Making licensing terms of publicly funded inventions more transparent would require a rewrite of the implementing regulations to change interpretation of the Bayh-Dole Act’s confidentiality clause. The confidentiality provision in the Bayh-Dole Act was intended to protect agencies from being forced to disclose proprietary data, but its implementing regulation is so broad that, in effect, it restricts the

government’s ability to use data without permission of the relevant licensee. Current nondisclosure practices lead to data being unavailable for research aimed at improving knowledge about patenting and licensing practices. Many studies could be undertaken on aggregated reported data, and there are many precedents for using census data, health statistics and other very private information in government databases. The original rationale for the Bayh-Dole Act was that government-owned inventions were languishing for want of effective patent incentives to grantees and contractors; the current problem is that data on patenting and licensing practices are languishing in a government database that is not mined for valuable insights. On the industry side, there is a somewhat higher standard for disclosure by public companies to protect shareholders. As of 2003, the Securities and Exchange Commission (SEC) requires disclosure of material agreements, including license agreements, as part of SEC filings. Section 401(a) of the Sarbanes-Oxley Act of 2002 (Public Company Accounting Reform and Investor Protection Act of 2002, Pub. L. No. 107-204, 116 Stat. 745) requires the SEC to adopt rules to require each annual and quarterly financial report filed with the commission to disclose “all material off-balance sheet transactions, arrangements, obligations (including contingent obligations), and other relationships of the issuer with unconsolidated entities or other persons, that may have a material current or future effect on financial condition, changes in financial condition, results of operations, liquidity, capital expenditures, capital resources, or significant components of revenues or expenses.” In many cases, however, these disclosures are of little assistance in understanding the licensing landscape. The reporting pertains only when a license underpins a genetic test that is a large enough portion of a publicly traded company’s business that it needs to be disclosed to investors. Even then, which patents have been licensed under what terms may be disclosed vaguely. Many biotech start-up companies are not publicly traded and are not subject to SEC disclosure requirements. By the time a biotech company goes public, its prospectus may contain some, but only limited, information about licensing agreements. In the usual case of a public company acquiring technology by buying another company, disclosure of the original license may not be required. Universities argue that if they are forced to disclose the terms of prior licensing agreements, it will undermine their negotiating position with new potential licensees. If, however, public companies must disclose the contents of their license agreements to protect the interests of those funding them (namely, shareholders) as

volume 28 number 8 august 2010 nature biotechnology

C O M M E N TA R Y

© 2010 Nature America, Inc. All rights reserved.

a matter of public policy, then it is not clear why a university should not be required to disclose the contents of its license agreements to protect those who fund it (namely, the public). The question of human resources needed to ensure transparency is very real and needs to be taken into account, but the principle of public disclosure should be entrenched within public institutions, particularly when the licensed inventions arise from publicly funded research and when data are being collected and reported already. Government and nonprofit research dollars should come with public accountability. Secure funding of tech transfer offices. As noted above, some tech transfer offices are expected to be self-sustaining and suffer from a serious lack of resources. This situation has several consequences. First, the agreements that these offices pursue will not necessarily aim to promote dissemination but instead will focus first on securing revenues. Second, tech transfer offices lack resources to train managers on implementing guidelines and the particular challenges that different technologies raise. The DNA diagnostic market is complex and rapidly evolving. For example, technology licensing officers need to know that the development of genetic testing after the discovery of the gene requires far less investment than the development of therapeutics, suggesting that exclusive licenses are usually not as necessary11. Without a more nuanced and informed understanding of how optimal patenting, dissemination and licensing decisions vary across different types of technologies and uses, these offices cannot fulfill their mandate: transferring technology. Conclusions To address the ongoing failure to achieve the goals of the multiple guidelines, policies and even legislation aimed at ensuring continued research on and access to clinical genetic tests, practices within universities and their industry partners must conform to existing guidelines. Although some changes to patent law—such as clearer research exemptions and an opposition proceeding—could be of use, fundamentally the problem is one of strategy about what to patent (products versus methods), how broadly to make claims to early-stage gene-based inventions and how to deploy those patents (broadly versus exclusively). Patents will be properly deployed only when university constituencies unite in promoting broad dissemination, when technology transfer offices are given the necessary financial support and incentives and when universities and industry have transparent and publicly accountable practices for licensing of DNA diagnostic technologies. Industry groups

such as BIO and university technology transfer organizations such as AUTM have a crucial and constructive role to play in resolving this predicament. Progress toward addressing the problems in genetic diagnostics can begin with less caustic and unhelpful rhetoric and more focus on engagement with their constituencies on seriously implementing guidelines, as well as with federal advisory bodies such as the SACGHS. By acknowledging and engaging with the distinctive problems that patenting and licensing practices raise for DNA diagnostics, both the universities licensing out technology and the companies licensing it in can bring about real improvement without the need for legislation. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. 1. Diamond v. Chakrabarty, 447 U.S. 303 (1980). 2. Association for Molecular Pathology et al. v. United States Patent and Trademark Office et al. (USDC SDNY 09 Civ. 4515, 2010). 3. Ariad Pharmaceuticals, Inc. v. Eli Lilly and Co. (560 F3d 1366 (Fed Cir 2009). 4. Secretary’s Advisory Committee on Genetics Health and Society, National Institutes of Health. Report on Gene Patents and Licensing Practices and Their Impact on Patient Access to Genetic Tests (SACGHS, Washginton, DC, 2010). 5. Merz, J.F. Clin. Chem. 45, 324–330 (1999). 6. Heller, M.A. & Eisenberg, R.A. Science 280, 698–701 (1998). 7. Chandrasekharan, S. & Cook-Deegan, R. Genome Med. 1, 92 (2009). 8. Holman, C.M. Science 322, 198–199 (2008). 9. Nelson, R. J. Technol. Transf. 26, 13–19 (2001). 10. Mowery, D.C. et al. Res. Policy 30, 99–119 (2001). 11. Pressman, L. et al. Nat. Biotechnol. 24, 31–39 (2006). 12. Schissel, A., Merz, J.F. & Cho, M.K. Nature 402, 118 (1999). 13. Henry, M.R., Cho, M.K., Weaver, M.A. & Merz, J.F. Science 297, 1279 (2002). 14. Gold, E.R. & Carbone, J. Genet. Med. 12 Suppl, S39– S70 (2010). 15. Skeehan, K., Heaney, C. & Cook-Deegan, R. Genet. Med. 12 Suppl, S71–S82 (2010). 16. Merz, J.F. in The Penn Center Guide to Bioethics (eds. Ravitsky, F., Feister, A. & Caplan, A.L.) 383–385 (Springer, New York, 2009). 17. Nuffield Council on Bioethics. The Ethics of Patenting DNA (Nuffield Council on Bioethics, London, 2002). 18. Cook-Deegan, R. et al. Genet. Med. 12 Suppl, S15– S38 (2010). 19. Angrist, M., Chandrasekharan, S., Heaney, C. & Cook-Deegan, R. Genet. Med. 12 Suppl, S111–S154 (2010). 20. Chandrasekharan, S. & Fiffer, M. Genet. Med. 12 Suppl, S171–S193 (2010). 21. Chandrasekharan, S., Heaney, C., James, T., Conover, C. & Cook-Deegan, R. Genet. Med. 12 Suppl, S194–S211 (2010). 22. Chandrasekharan, S., Pitlick, E., Heaney, C. & CookDeegan, R. Genet. Med. 12 Suppl, S155–S170 (2010). 23. Colaianni, A., Chandrasekharan, S. & Cook-Deegan, R. Genet. Med. 12 Suppl, S5–S14 (2010). 24. Powell, A., Chandrasekharan, S. & Cook-Deegan, R. Genet. Med. 12 Suppl, S83–S110 (2010). 25. Cook-Deegan, R., Chandrasekharan, S. & Angrist, M. Nature 458, 405–406 (2009). 26. Caulfield, T., Cook-Deegan, R.M., Kieff, F.S. & Walsh, J.P. Nat. Biotechnol. 24, 1091–1094 (2006). 27. National Research Council. Reaping the Benefits of Genomic and Proteomic Research: Intellectual

nature biotechnology volume 28 number 8 august 2010

Property Rights, Innovation and Public Health (National Research Council, Washington, DC, 2006). 28. Ontario Report to the Provinces and Territories. Genetics, Testing and Gene Patenting: Charting New Territory in Healthcare (Government of Ontario, Toronto, Ontario, Canada, 2002). 29. Australian Law Reform Commission. Essentially Yours: The Protection of Human Genetic Information in Australia (ALRC 96) (ALRC, Sydney, New South Wales, Australia, 2003). 30. Gold, E.R., Bubela, T., Miller, F.A., Nicol, D. & Piper, T. Nat. Biotechnol. 25, 388–389 (2007). 31. Gold, E.R. Nat. Biotechnol. 18, 1319–1320 (2000). 32. Nicol, D. & Nielsen, J. Patents and Medical Biotechnology: An Empirical Analysis of Issues Facing the Australian Industry (Occasional Paper no. 6) (Centre for Law & Genetics, Sandy Bay, Tasmania, Australia, 2003). 33. Cho, M.K., Illangasekare, S., Weaver, M.A., Leonard, D.G.B. & Merz, J.F. J. Mol. Diagn. 5, 3–8 (2003). 34. Rai, A. Northwest. Univ. Law Rev. 94, 77–152 (1999). 35. Merz, J.F., Kriss, A.G., Leonard, D.G. & Cho, M.K. Nature 415, 577–579 (2002). 36. Merz, J.F., Cho, M.K., Robertson, M.J. & Leonard, D.G. Mol. Diagn. 2, 299–304 (1997). 37. Merz, J.F. & Cho, M.K. Camb. Q. Healthc. Ethics 7, 425–428 (1998). 38. Andrews, L.B. Nat. Rev. Genet. 3, 803–808 (2002). 39. LOI no 613–16 as amended in 2004. 40. Overwalle, G.V. Int. Rev. Intellect. Property Competition Law 889, 908–918 (2006). 41. Fed. Reg. 66, 1092–1099 (2001). 42. Fed. Reg. 70, 18413–18415 (2005). 43. Organisation for Economic Co-operation and Development. Guidelines for the Licensing of Genetic Inventions (OECD, Paris, 2006). 44. In the Public Interest: Nine Points to Consider in Licensing University Technology (AUTM, Deerfield, Illinois, USA, 2007). 45. Association of University Technology Managers. University Principles on Global Access to Medicines (AUTM, Deerfield, Illinois, USA, 2009). 46. Rimmer, M. Eur. Intellectual Prop. Rev. 25, 20–33 (2003). 47. American Medical Association. Report 9 of the Council on Scientific Affairs (AMA, Chicago, 2000). 48. Huys, I., Berthels, N., Matthijs, G. & Van Overwalle, G. Nat. Biotechnol. 27, 903–909 (2009). 49. Laboratory Corporation of America Holdings, dba Labcorp v. Metabo-Lite Laboratories, Inc. et al., 548 U.S. 124 (2006). 50. eBay Inc. v. MercExchange, LLC, 547 U.S. 388 (2006). 51. Bilski v. Kappos, 561 U.S. ____ 20010 (No. 08–964), affirming F.3d 943 3d 943 (Fed. Cir. 2008). 52. In re Kubin (Fed Cir. 2009). 53. KSR International Co. v. Teleflex, Inc., 550 U.S. 398 (2007). 54. Van Overwalle, G., van Zimmeren, E., Verbeure, B. & Matthijs, G. Nat. Rev. Genet. 7, 143–148 (2006). 55. Walsh, J.P., Ashish, A. & Cohen, W. in Effects Of Research Tool Patents And Licensing On Biomedical Innovation (eds. Cohen, W. & Merrill, S.) 285–336 (National Academies Press, Washington, DC, 2003). 56. Gold, E.R. et al. The Research or Experimental Use Exception: A Comparative Analysis (Centre for Intellectual Property Policy/Health Law Institute, Montreal, Quebec, Canada, 2005). 57. Merck KGaA v. Integra Lifesciences I, Ltd., 545 U.S. 193 (2005). 58. Siegel, D.S. & Wright, M. Oxford Rev. Econ. Policy 23, 529–540 (2007). 59. http://www.mpegla.com/Lists/MPEG%20LA%20 News%20List/Attachments/230/n-10–04–08.pdf, Last Accessed May 4, 2010. 60. (5 February 2010). 61. http://bio.org/ip/genepat/documents/SACGHSsignonletter2–4-2010final_000.pdf 62. Bayh-Doyle Act, 37 C.F.R. Part 401.

791

F E AT U R E

Public biotech 2009—the numbers Brady Huggett, John Hodgson & Riku Lähteenmäki

© 2010 Nature America, Inc. All rights reserved.

The public biotech sector sustained more losses in 2009, but the year ended on a positive note, and the industry has regained its footing.

T

hat whooshing sound at the end of 2009 was the biotech sector letting out its collective breath. The year began as a hard slog, so when it came to a close on an upward swing, the industry rightfully felt a measure of relief. That’s not to say there weren’t casualties: a distressingly large number of companies departed the scene last year. But it was not as bad as some pundits had estimated, and the industry proved itself to be strong and creative. It was helped by a recovering economy in the second half of the year. Overall, counting the vast financial potential of collaborations, the industry recorded one of its best years for fundraising. That has left the sector brightly looking ahead again—a far cry from how things appeared at the end of 2008.

Economic woes The 2009 data from Nature Biotechnology’s annual survey of public biotech firms, which now number 461 (owing to a change in our data-gathering process; see Box 1 and Supplementary Table 1), show little trace of how terribly the year began or how tightly the public markets had been hammered shut at the end of 2008. The reality is that 2009 started bleakly for biotech, and it continued that way for most of the first quarter. Of course, not just biotech suffered—the recession affected all countries and sectors. Along with the other indices, shares on the Nasdaq Biotechnology Index bottomed out on 9 March, resting at 59.05, a low it had not seen since May 2003. The global economy continued to shed jobs last year: the US Central Data retrieval for this article was by Ernst & Young (Boston) with additional reporting by Riku Lähteenmäki. Brady Huggett is business editor at Nature Biotechnology, John Hodgson is editor-at-large at Nature Biotechnology, and Riku Lähteenmäki is a freelance writer in Turku, Finland.

Box 1 The numbers Nature Biotechnology has published an annual report on public biotech companies since 1996. As the industry has grown and changed, so have our definition of what constitutes a biotech company and our methods for gathering the information that serves as the backbone to this piece. We generally include companies built upon applying biological organisms, systems or processes, or the provision of specialist services to facilitate the understanding thereof. We exclude pharmaceutical companies, medical-device firms and contract research organizations to better focus on the unique attributes and situations that make up the biotech sector. This year’s data was provided by Ernst & Young, which has broadened the report’s reach into international exchanges and increased our total number of companies. Additional reporting was done via individual financial reports. The top-ten lists and other aggregate lists are sourced appropriately, with most data supplied by BioCentury. As investors do not stratify the biotech sector as stringently as Nature Biotechnology, we used money figures from across the biotech and biopharmaceutical arena to best highlight trends. In some cases, full-year data were not available and fourth-quarter numbers were extrapolated; this is noted in the company-by-company data table (Supplementary Table 1). Companies delisted in 2009 from major exchanges were excluded.

Intelligence Agency estimates unemployment numbers increased around the world, sometimes drastically—Ireland’s unemployment nearly doubled to 12%, whereas the US went from 5.8% in 2008 to 9.3%. So although biotech wasn’t alone in the dark, as an industry made up mainly of small companies devoid of revenue—and thus more dependent on raising public funds— the sector was hit particularly hard. The fear, expressed by pundits, the Biotechnology Industry Organization (Washington, DC) and even biotech executives themselves, was that the industry would lose up to 25% of its companies to bankruptcy. But the Nasdaq Biotech Index steadily recovered from that March low and closed 2009 at 81.83. Overall funding for the sector jumped in the second half, and although the National Bureau of Economic Research has yet to officially declare the end of the recession in the United States, consensus pegs it around the second quarter of 2009.

nature biotechnology volume 28 number 8 august 2010

Catastrophic shrinkage in the sector has not happened. There were losses (Table 1), but they were not as far-reaching as feared. And among all this detritus, a surprise: the biotech sector was again profitable in 2009. The money trail Financing levels for biotech are a useful gauge of the sector’s overall health, because without repeated investment, the industry shrivels. In this regard, 2009 turned out better than expected. The third quarter saw the first month of positive growth in the US economy since the recession started in December 2007, and as the economy recovered, money again began moving. By year’s end, overall biotech financing was up 84% from the depressed figures seen in 2008. In 2008, as first the United States and then the world slid into recession, overall funding was at its lowest since at least 2002 (Fig. 1), with debt financings, private investments in a public entity (PIPEs), follow-on offerings 793

f eat u re

© 2010 Nature America, Inc. All rights reserved.

Table 1 Casualties in 2009 Company

Reason for status change

Alpha Innotech

Acquired by Cell Biosciences

Altus Pharmaceuticals

Bankruptcy

Arthrokinetics

Delisted

Autoimmune

Inactive

Avalon Pharmaceuticals

Acquired by Clinical Data

Avigen

Acquired by Medicinova

Biopure Corporation

Bankruptcy

BioXell

Acquired by Cosmo

CelSis

Acquired by JM Hambro

Cellegy

Merged with Adamis Pharmaceuticals

Cell Genesys

Acquired by BioSante

Cobra

Merged with Recipharm

Curagen

Acquired by CellDex

Curalogic

Bankruptcy

CV Therapeutics

Acquired by Gilead

EPIX Pharmaceuticals

Liquidated

Evolutec

Transformed into investment company

Genaera

Dissolved

Genentech

Acquired by Roche

Hemacare

Inactive

Hemagen Diagnostics

Inactive

IDM Pharma

Acquired by Takeda

Introgen

Bankruptcy

Isologen

Bankruptcy

Intercytex

Delisted

Liponex

Merged with ImaSight

Medarex

Acquired by BNS

Metabasis Therapeutics

Acquired by Ligand Pharmaceuticals

Monogram

Acquired by LabCorp

Napo Pharma

Inactive

Nastech

Changed name to MDRNA

Neos

Inactive

Neurogen

Inactive

Northfield Laboratories

Inactive

Nucryst

Inactive

Nuvelo

Merged with Arca

Nventa Biopharmaceuticals

Inactive

Phynova

Delisted

Replidyne

Merged with Cardiovascular

Targanta

Acquired by The Medicines Company

ViRexx Medical

Acquired by Paladin

XLT Biopharmaceuticals

Delisted

and initial public offerings (IPOs) all declining substantially from previous years. Only venture capital remained aloft, although venture capitalists were more inclined to put money into companies previously invested in, rather than new ventures. This pattern reversed last year. Debt financings, venture capital and money raised in follow-ons and IPOs all increased, almost achieving the level seen in 2007, before the markets tanked. Only one category went backward, PIPEs­—which was to be expected, 794

as once the general markets (and individual stock prices) improved, the need for private investment faded. The largest follow-on offering of the year ($640 million) was conducted by Qiagen (Venlo, The Netherlands), a profitable provider of sample and assay technologies (Table 2). It had the best year of its existence in 2009, with overall revenues above $1 billion, and is the type of stable company that can easily reach into the secondary-offering market. The sexier story is Human Genome

Sciences (HGS, Rockville, Maryland, USA), which raised about $850 million in two follow-on offerings. As its stock price rocketed after positive pivotal trial results for the lupus drug Benlysta (belimumab), it tapped the public markets in late July for more than $373 million and again in December for about $477 million. The company’s stock, which opened the year at $2.12, ended it at $30.58. This is a similar story to Dendreon’s (Seattle), which in April reported positive phase  3 results for its prostate cancer vaccine Provenge (sipuleucel-T), sending its stock up more than 100% on the day the results were announced. This set the stage for a $427-million public offering in May, followed by another in December. Provenge has now been approved, the company has priced the drug aggressively, and Dendreon’s stock, at the time of publication, sat just above $34; it began 2009 at $4.59. Whereas many of biotech’s established companies completed debt deals last year, returning that funding category to levels seen before a well-below-average 2008, it was hardly a year worth mentioning for IPOs (Table 3). Just ten occurred in 2009, none before August, and none could be considered a typical biotech IPO, either in the type of company or the amount of money raised. For instance, the JSC Human Stem Cell Institute (with sites in Russia, Germany and the Ukraine) raised a mere $4.8 million. The institute doesn’t look much like the usual biotech enterprise preparing to go public: it has a research laboratory and a center for storage of cellular materials, and it publishes the journal Cellular Transplantation and Tissue Engineering. What’s more, an IPO is no longer the cash windfall and viable exit for investors it once was. Consider D-Pharm (Rehovot, Israel), which raised about $7.4 million on the Tel Aviv Stock Exchange to fund clinical testing of its small-molecule stroke drug, DP-b99, a membrane-active derivative of the calcium chelator 1,2-bis-(2-aminophenoxy)ethane- N,N,N′,N′-tetraacetic acid (BAPTA). Alongside the IPO, the company also completed a rights offering (which gives existing shareholders the right to buy shares during a defined period, usually at a discount), raising NIS 57 million ($14.8 million). The existing investors didn’t exit—they instead had the choice to increase their stake. In truth, the average amount raised per IPO is hardly enough to alleviate financial concerns for long. In 2008, our survey showed IPOs raised on average $22.3 million. In the previous two years, it was considerably more, $58 million in 2007 and $41 million in 2006. Figure 2 shows an IPO in 2009 raised,

volume 28 number 8 august 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

f eat u re

70

Financing raised ($ billions)

on average, $92.8 million. On the surface that seems a marked increase, but further inspection shows that the figure is distorted by the unique case of Talecris Biotherapeutics (Research Triangle Park, NC, USA). The company develops nonrecombinant protein therapeutics from plasma and is profitable. It was pegged as an acquisition target by rival CSL (Victoria, Australia) in 2008 for $3.1 billion, but the US government challenged the purchase as anticompetitive, and the deal fell apart. Talecris instead conducted an IPO in 2009 for a whopping $550 million. Toss aside Talecris, and the figure falls more in line with recent years: $42 million. Talecris is again in line for an acquisition, by Grifols (Barcelona, Spain) for $3.4 billion. Overall, the public markets in Europe remain relatively parsimonious. They provided only 15% of all European financing, whereas US public markets provided 33% of the total US fundraising (Table 4). The main shortfall, as in previous years, was in follow-on offerings. Where follow-on financings occurred in Europe, they raised amounts comparable to those raised by US firms—$112 million on average, compared with $107 million for US companies. But in 2009, 48 US biotech companies got followon offerings away, compared with only seven in Europe. For European public companies, secondary offerings are still the exception— leaving them open to acquisition bids and investors open to disillusionment. Two European firms dominated debt financing this year (Table 5), with giant UCB (formed around Celltech, Brussels) taking in more than $2.6 billion in a series of three notes. Elan (Dublin) also raised $625 billion in a bond issue. These two massive chunks of debt financing distort the European fundraising picture, giving it an undue rosy glow. The $3.2 billion raised represents over three-quarters of the ‘Other’ categories

60 50 40 30

0.544 3.883 2.231 4.018 9.075 8.933

2.556 3.335 2.93 5.318 8.833 10.933

1.859 4.838 2.661 5.398 6.112 17.268

2.03 5.578 4.695 5.682 11.853 19.796

2.95 4.377 4.748 6.809 11.68 22.365

0.928 6.041 2.277 5.198 10.335 36.923 0.134 1.867 3.143 5.177 3.232 20.023

IPO Follow-on PIPES Venture capital Debt and other

20 Partnerships

10 0

2003

2004

2005

2006

2007

2008

2009

Year Figure 1 Global biotech industry financing. Biotech funding was up 84% to $62 billion in 2009 from $33 billion in 2008. Partnership figures from Burrill & Co. are for deals involving a US company. BioCentury makes updates to its financing data on an ongoing basis. Sources: BCIQ: BioCentury Online Intelligence; Burrill & Co.

of finance in Europe and nearly half of all finance in Europe during 2009 (Table 4). Without this money, the amount raised in Europe during 2009 would have been only 15% of the global total finance in this survey, rather than 26%. Those IPOs had a small role in the sizable increase in overall funding from 2008, but the biggest factor was headline-grabbing partnering deals: $36.9 billion in 2009, up from $20 billion the previous year. This heightened partnering activity was propelled both by pharma’s need to bolster fading pipelines and biotech’s need for help of any kind during the recession. But here again, that high figure is misleading, because a large portion of it represents milestone payments that may never be paid. The leading deal among our companies (Table 6) was formed between Nektar and AstraZeneca for two programs that use Nektar’s advanced polymer conjugate tech-

nology platform—the program NKTR-118, which had completed phase 2 for opioidinduced constipation, and NKTR-119, an early-stage program intended to deliver products for pain without a constipation side effect. Nektar did receive an up-front payment of $125 million in the deal, but it’s the potential milestones that give the partnership its $1.5 billion high-end value. That was one of six deals in 2009 that had a potential payout of more than $1 billion, making the average potential of our top-ten group worth more than a billion dollars. But the average amount of funds received up front (including equity investments or money for milestones hit at the time of deal signing) was much lower, at about $109 million, meaning nearly 90% of the value in these deals remained unrealized at year’s end. When considering all partnerships between pharma and biotech (public and private), using data from Elsevier’s Strategic

Table 2 Top ten follow-on offerings of 2009 Date completed

Amount raised ($ millions)

Qiagen

9/24

640.4

Deutsche Bank, Goldman Sachs, J.P. Morgan, Barclays Capital, Commerzbank, DZ Bank

Vertex Pharmaceuticals

12/2

500.5

Goldman Sachs, Merrill Lynch, J.P. Morgan, Morgan Stanley

Company name

Human Genome Sciences

Underwriters

12/2

476.8

Goldman Sachs, Citigroup, J.P. Morgan, Morgan Stanley, UBS

12/10

426.9

J.P. Morgan, Deutsche Bank, Citigroup, Morgan Stanley, Lazard, Leerink

Human Genome Sciences

7/28

373.8

Vertex Pharmaceuticals

2/18

320

Dendreon

Goldman Sachs, Citigroup Merrill Lynch, Cowen

Cephalon

5/21

300

Dendreon

5/13

229.9

Deutsche Bank

Deutsche Bank, J.P. Morgan, Barclays Capital Inc., Credit Suisse, Morgan Stanley

Incyte

9/25

139.7

Goldman Sachs, Morgan Stanley, J.P. Morgan

Seattle Genetics

8/11

135.9

J.P. Morgan, Goldman Sachs, Needham, Oppenheimer, RBC Capital Markets

Data are matched to the definition of biotech in Box 1. Source: BCIQ: BioCentury Online Intelligence

nature biotechnology volume 28 number 8 august 2010

795

f eat u re

Table 3 Initial public offerings of 2009 Amount raised Date completed ($ millions) Underwriters

Company name

Location

CanBas

Shizuoka, Japan

9/17

14.8

Mitsubishi UFJ Securities International plc, Mizuho, Ichiyoshi, JPMorgan, Mizuho Investors, Takagi

China Nuokang Bio-Pharmaceutical

Beijing

12/9

40.7

Jefferies, Oppenheimer

Cumberland Pharmaceuticals

Nashville, TN, USA

8/10

85

D. Western Therapeutics Institute

Aichi, Japan

10/13

9.7

Nomura, Mitsubishi UFJ Securities International plc, Takagi, SBI Securities Co. Ltd., Tokai Tokyo, Mizuho

D-Pharm

Rahovot, Israel

8/17

7.3

Clal Finance, Rosario, Meitav

12/10

4.8

CJSC Alor Invest

Movetis N.V.

Turnhout, Belgium

12/3

146

Omeros Corp.

Seattle

10/7

68.2

Talecris Biotherapeutics

Research Triangle Park, NC, USA

9/30

549.9

T-Ray Science Inc.

Vancouver

12/9

1.4

Credit Suisse, KBC, Piper Jaffray Deutsche Bank, Wedbush, Canaccord, Needham, Chicago Investment Group, National Securities Morgan Stanley, Goldman Sachs, JPMorgan, Citigroup, Wells Fargo, Barclays Capital Research Capital Corp.

Source: BCIQ: BioCentury Online Intelligence

Transactions database, we found the average total amount paid up front in 2009 was about $58.9 million. That’s the highest average over the past 10 years (only 2006 came close, at $55.7 million), and a long way from the upfront money paid out in 2000, which was just $12.4 million. Still, it also drives home the reality that a deal with a potential value of $1 billion is just that: potential. 2009 also provided an interesting wrinkle for equity investments around partnerships. Over the past 10 years, the average equity bought as part of a deal in each year was well below $10 million, with the exception of 2001, when it leaped to $32.3 million. Last year, it leaped again, to $20.6 million. In both 2001 and 2009, the public markets had come down from peaks, and thus selling equity as part of partnering deals rose in favor.

Buyouts and climbing sales Mergers and acquisitions fell in 2009, both in total number and in the values assigned to the companies acquired (Table 7). Leading our list is Roche’s buyout of Genentech, but that deal was actually announced in 2008. Although it closed in the spring of last year, the acquisition is old news. But also high on the list is the purchase of Medarex by Bristol-Myers Squibb (BMS, New York), an acquisition that gained a validation of sorts in 2010. The purchase gave BMS access to Medarex’s antibody-drug conjugate technology and UltiMAb human antibody development system, but the main draw was ipilimumab. BMS was already partnered with Medarex on ipilimumab in phase 3 for metastatic melanoma, in phase 2 for lung cancer and in phase 3 for adjuvant melanoma and Number of IPOs

49 41

45 41

50

10 92.8

51 58

Average amt raised ($M) 100 90 80 70

40

60 30 20

10 28

14 39

50

6 22

40 30 20

10

10 0

2002

2003

2004

2005

2006

2007

2008

2009

0

Average amount raised ($ millions)

53 48

60

Number of IPOs

© 2010 Nature America, Inc. All rights reserved.

Human Stem Cell Institute Moscow

UBS, Jefferies, Wells Fargo, Morgan Joseph and Co.

Year Figure 2 Global biotech initial public offerings. IPOs in 2009 seemingly made a recovery in amount raised, if not number of offerings.But the data is skewed by one large offering.

796

hormone-refractory prostate cancer, so it had seen the product up close. Perhaps that’s the reason it offered a greater than 90% premium to the trading price of Medarex shares; the deal went through at $16 apiece, or $2.4 billion. Ipilimumab, a monoclonal antibody designed to block the inhibitory signal of cytotoxic T lymphocyte-associated antigen-4 (CTLA-4), had failed in a phase 3 trial in 2007, and there was uncertainty around the new pivotal program for melanoma. But BMS announced in June 2010 at the American Society of Clinical Oncology’s annual meeting in Chicago that ipilimumab met the primary endpoint of survival in advanced melanoma in a phase 3 doubleblind randomized trial, and BMS said it expects to submit for regulatory approval of ipilimumab this year. Should the drug win approval, the $2.4 billion price tag for Medarex will seem a steal. Also of interest last year was Gilead’s (Foster City, CA, USA) buyout of CV Therapeutics, giving a company typically known for its HIV franchise a presence in the cardiovascular space. The move brought aboard Ranexa (ranolazine extended-release tablets), approved for chronic angina, and Lexiscan (regadenoson) injection for use as a pharmacologic stress agent in radionuclide myocardial perfusion imaging. Gilead remains a leader in HIV drugs—its highest-selling product was Truvada at about $2.5 billion last year, and 90% of Gilead’s product sales came from its antiviral franchise—but through this acquisition it is seeking growth in other areas. Big sellers like Truvada are the beacons in the biotech fog, promising a move into the black after years spent dumping money into R&D and

volume 28 number 8 august 2010 nature biotechnology

f eat u re

Table 4 Comparison of US and EU financing in 2009 Amount raised in US ($ millions) Venture capital IPO Follow-on offering Other Total

Number of US deals

Amount raised in EU ($ millions)

EU as a EU financing EU financing minus UCB percentage Number of UCB and Elan minus UCB (% of US + of US + EU total EU total) EU deals ($ millions) ($ millions)

EU as a percentage of EU total

US as a percentage of US total 22%

3,939

197

1,114

87

–

1,114

22%

22%

18%

703

3

158

3

–

158

18%

18%

3%

4%

5,166

48

785

7

–

785

13%

13%

12%

29%

7,756

236

4,253

108

3,200

1,053

12%

35%

67%

44%

17,564

484

6,310

205

3,200

3,110

15%

26%

100%

100%

© 2010 Nature America, Inc. All rights reserved.

Source: BCIQ: BioCentury Online Intelligence

the clinic. Achieving that level of revenue usually follows this path: drug approval, then a marketing push and physician acceptance, followed by subsequent approvals in other indications to further increase sales. Most of the biologics in our list of the top ten drugs (Table 8) went that route. Enbrel (etanercept), from Amgen (Thousand Oaks, CA, USA), exemplifies this tactic. Originally approved in 1998 for rheumatoid arthritis, Amgen has received approvals in four other indications (ankylosing spondylitis, psoriasis, psoriatic arthritis and juvenile rheumatoid arthritis), and its worldwide revenue has jumped from $2.6 billion in 2005 to an estimated $6.4 billion in 2009, according to BioMedTracker. The drug, which inhibits the tumor necrosis factor (TNF) pathway, is the top-selling biologic in the world. In fact, three of the top five revenue-producing drugs target TNF: Remicade (infliximab, Johnson & Johnson, New Brunswick, NJ, USA) and Humira (adalimumab, Abbott, Abbott Park, IL, USA), are the other two, selling $5.9 billion and $5.5 billion worldwide, respectively. Those numbers, like the revenues for all the drugs in this table, are an improvement over the previous year. Given the lack of generic competition for biologics, it’s almost an anomaly when a drug does not increase sales year on year; it suggests something must have gone wrong. That’s been the case with Amgen’s Aranesp. Peaking at $4.1 billion in worldwide sales in 2006, the drug has lost ground yearly since then, and in 2009 declined 15% to about $2.7 billion, falling off our list of the top ten biotech drugs. Amgen attributes the decline to the negative impact, mostly in supportive cancer care, of a “product label change” that came in August 2008. In fact, Aranesp serves as an example of the downside of product growth: the drug was being used offlabel in various indications until reports of adverse effects caused the US Food and Drug Administration (FDA) to tighten its label. The decline of Aranesp revenue meant Amgen reported lower overall revenues for

2009, although the company’s adjusted net income for the year was more than $5 billion, compared with $4.9 billion in 2008, a 3% increase. Affymetrix (Santa Clara, CA, USA) also saw its revenue decrease in 2009, though the reason has more to do with accounting: the figures had been buoyed in 2008 by a onetime intellectual property payment of $90 million. So while a comparison year-by-year shows the company lost 20% of revenue in 2009, in truth the business ground along smoothly. It had product revenue of $279.2 million and service revenue of $39.6 million last year, both up from the previous year (2008 product revenue was $270.4 million and service revenue was $32.1 million.) Like Amgen and Affymetrix, other established firms fared well. Gilead experienced the largest increase in revenues, posting product sales that increased 27% over 2008 to nearly $6.5 billion, driven mostly by its HIV franchise of Truvada (emtricitabine and tenofovir disoproxil fumarate) and Atripla (efavirenz 600 mg, emtricitabine 200 mg, tenofovir disoproxil fumarate 300 mg). Truvada sales increased 18% to about $2.5 billion, and Atripla brought in $2.4 billion, up 51% over 2008. HGS also reported impressive revenues of $275.7 million for 2009, compared with

revenues of only $48.4 million the previous year. The company logged its first product sales—$180.2 million for delivering to the US Strategic National Stockpile raxibacumab (human monoclonal antibody drug for treatment of inhalation anthrax) under a government contract. That helped HGS earn a net income of $5.7 million for the year, compared with a net loss of $268.9 million in 2008. The company also reported positive results for Benlysta (belimumab) phase 3 trials announced in July and November 2009. The good news drove up HGS’s stock price considerably, and as we noted earlier, it raised public funds twice during the year. End of the line Whereas 2008 saw 34 companies depart from the public biotech landscape—11 because of delisting or bankruptcy—those numbers increased in 2009. The total number of companies departing for any reason (buyout or merger included) climbed to 44, and the number removed owing to financial difficulty also went up, reaching 20. But a 9.5% drop in the number of companies is fewer casualties than was feared. Of those that teetered but survived, some were helped partially by the markets opening back up in the spring; by the ability to conduct debt deals,

Table 5 Top ten debt financings of 2009 Company name

Financing type

Amgen

Sr notes (other)

UCB Group

Bond (other)

Date completed

Amount raised ($ millions)

1/14

2,000

10/27

1,128

UCB Group

Bond (other)

12/3

751.9

UCB Group

Sr convert notes (other)

9/30

730.3

Elan

Sr notes (other)

9/29

625

Cephalon

Sr subord convert notes (other)

5/22

500

Gilead Sciences

Debt (other)

4/20

400

Incyte

Convert notes (other)

9/25

400

Bio-Rad Laboratories

Sr notes (other)

5/19

300

PDL BioPharma

Sr notes (other)

10/28

300

Source: BCIQ: BioCentury Online Intelligence

nature biotechnology volume 28 number 8 august 2010

797

f eat u re

© 2010 Nature America, Inc. All rights reserved.

Table 6 Top ten research partnership and licensing deals of 2009 Date announced

Researcher

Investor

Nektar

AstraZeneca

Incyte

Novartis

Targacept

Deal value ($ millions)

Details

9/21

1,505

Worldwide rights to NKTR-118 for opioid-induced constipation and NKTR-119 for pain

11/25

1,310

Ex-US rights to oral INCB18424, which is in phase 3 for myelofibrosis, and worldwide rights to preclinical cancer compound INCB28060

AstraZeneca

12/3

1,240

Worldwide rights to develop and commercialize major depressive disorder compound TC-5214

Exelixis

Sanofi-aventis

5/29

>1,161

ZymoGenetics

Bristol-Myers Squibb

1/12

1,105

Codevelop and commercialize phase 1 HCV compound PEG-Interferon lambda (IL-29)

Amylin

Takeda

11/1

1,075

Codevelop and commercialize therapeutics for obesity and related indications

Santaris Pharma

Wyeth

1/12

847

Worldwide rights to ALD518 for all indications except cancer

Algeta

Bayer

9/3

800

Codevelop Alpharadin for bone metastases

Medivation

Astellas Pharma

10/27

765

Codevelop MDV3100 for the treatment of prostate cancer

Cytokinetics

Amgen

5/26

650

Exclusive world-wide (except Japan) license for cardiac contractility program

Acorda

Bayer

7/1

510

Exclusive collaboration and license agreement to develop Fampridine-SR for multiple sclerosis

Exclusive, worldwide rights to XL147 and XL765, oral phosphoinositide 3-kinase inhibitors in phase 1b/2 and phase 2 to treat cancer

Data are matched to the definition of biotech in Box 1. Source: BCIQ: BioCentury Online Intelligence

which returned to a more normal level after suffering through the battered credit markets in 2008; and by partners supplying up-front money and other funding. Also, considering that Genentech (and its $3.4 billion of net income in 2008) is no longer in our survey (now part of Roche), it seemed unlikely the sector would be able to repeat its performance from 2008, when it posted a net profit of $3.8 billion. But it did, drawing a collective net income in 2009 of $8 billion—with the heavy lifting, unsurprisingly, done by the large-cap firms (Fig. 3). Three main drivers contributed to the unexpected profitability in 2009. The first is an accounting change by the US Federal Accounting Standards Board, issued in late 2007 but applicable in fiscal year 2009. Called SFAS 141R, the new guidance allows the costs associated with mergers and acquisitions to be expensed over time, rather than all at once

as part of the purchase price. It’s a small factor, and biotech-biotech mergers are less common and of lesser value than those between biotech and pharma, but still noteworthy. The second is that some companies simply had good years, and their revenue growth helped make up for the loss of Genentech. We’ve seen this with companies such as Gilead, which pushed its revenue up 31% and net income up 33% from 2008, and Biogen Idec (Weston, MA, USA), which posted a net income of $970 million, up 24% over the previous year. But the major reason for the collective profit is the same one that kept the number of bankruptcies lower than feared: a cutback on expenses. When the money isn’t there, spending has to decrease, and biotech tightened its belt in 2009. Companies spent less in two notable ways. First, they carried smaller payrolls than previously. In 2008, the

Table 7 Top ten announced mergers and acquisitions of 2009 Target

Acquirer

Genentech

Roche

Medarex

Bristol-Myers Squibb

CV Therapeutics

Gilead Sciences

ESBATech

Alcon

BiPar Sciences

Sanofi-aventis

Noven

Month completed

Deal value ($ millions)

March

46,800

September

2,400

April

1,400

September

589

April

500

Hisamitsu Pharmaceuticals

August

428

ViroChem

Vertex

March

413

Peplin

Leo Pharma

November

288

Dow Pharmaceutical Sciences

Valeant Pharmaceuticals

January

285

Arana Therapeutics

Cephalon

August

276

Data are matched to the definition of biotech in Box 1. Source: BCIQ: BioCentury Online Intelligence

798

companies surveyed had an average of 489 employees per company. In 2009, although our pool of biotech firms surveyed grew to 461 and with it the total number of employees increased, the average number of employees per company actually dropped to 442. Second, the biotech sector collectively reduced its R&D spending. In 2008, even as it faced financial turmoil, biotech increased its spending on R&D, as it had for years, from $22.8 billion in 2007 to $25.5 billion. This pattern came to a halt last year, when the sector’s overall R&D spending fell to $22.3 billion, with the greatest decrease seen in the microcaps, which went from $5.4 billion in 2008 to $4.0 billion (a fall of nearly 30%) in 2009. (Large caps reduced their R&D spending by just under 10%.) This considerable drop helped keep biotech profitable, but it is likely that it penalized the sector’s ability to carry out innovative science. The horizon Compared with other business sectors, biotech will continue to face the challenges of long timelines for product development. The heavy costs of R&D have shaped this industry since its inception, and that’s not about to change. But precisely because biotech remains centered on the provision of medical products, it has had the advantage of being considered ‘recession proof ’—people need drugs no matter how the economy is performing. The bottom lines of biotech’s big producers—Amgen, Gilead, Biogen—in 2008 and 2009 reflect this. Yet the sector’s ability to fund itself ebbs and flows with the global economy, and this

volume 28 number 8 august 2010 nature biotechnology

f eat u re

Table 8 Top ten biologic drugs in terms of sales in 2009 2009 revenue ($ million)

Name

Lead company

Approved indication(s)

Enbrel

Amgen

Rheumatoid arthritis (RA), ankylosing spondylitis, psoriasis, psoriatic arthritis (PA), juvenile rheumatoid arthritis

~6,400

Remicade

Johnson & Johnson

Psoriasis, ulcerative colitis (UC), ankylosing spondylitis, Crohn’s disease, PA, RA

5,892

Avastin

Roche

Colorectal cancer, breast cancer, brain cancer, renal cell cancer, non–small cell lung cancer

5,747

Rituxan

Biogen IDEC

Non-Hodgkin’s lymphoma, RA, chronic lymphocytic leukemia

5,617

Humira

Abbott Laboratories

RA, ankylosing spondylitis, juvenile rheumatoid arthritis, Crohn’s disease, PA, psoriasis

5,488

Herceptin

Roche

Breast cancer

4,833

Lantus

Sanofi-aventis

Diabetes mellitus type II, diabetes mellitus type I

4,295

Gleevec

Novartis

Chronic myelogenous leukemia, hypereosinophilic syndrome, dermatofibrosarcoma protuberans, myeloproliferative disorders, gastrointestinal stromal tumor, acute lymphocytic leukemia, myelodysplastic syndrome, mastocytosis

3,944

Neulasta

Amgen

Neutropenia, leucopenia

Prevnar

Pfizer

Prevention of otitis media, Streptococcus pneumoniae pneumonia

3,355 ~3,100

(sipuleucel-T) for prostate cancer, both of which are expected to be huge sellers. Biotech, with its small firms and entrepreneurial spirit, has long thought of itself as the underdog, made up of fast, nimble companies built to innovate, overachieve, withstand hardship and adapt. This attitude has always been part of the industry’s culture, and these days it’s also a carefully cultivated personality used to distance biotech from the more troubled

pharmaceutical industry. In short, it has often seemed like biotech was built to deal with adversity. After surviving the past two years, it now knows it can. ACKNOWLEDGMENTS The authors would like to acknowledge the insight of G. Giovannetti and G. Jaggi in crafting this article. Note: Supplementary information is available on the Nature Biotechnology website.

Amount ($ billions)

a 60

58.7

Micro cap

50

Small cap

40 30 20

Mid-cap Large cap

21.0 7.1 4.8

10

10.3

12.9 4.3 3.7 4.0

0 –10

Revenue

R and D

1.6 –2.4 –4.0 Net profit/loss

Number of companies

b

Number of employees 350

97,207

334

100,000

300 80,000

250

63,876

200

60,000

150 82

100 50 0

13 Large cap

32 Mid-cap

24,394 Small cap

40,000 22,954 Micro cap

20,000

Number of employees

is especially true for the smaller-cap firms. These companies require investors, they require the support of the public markets, and they require lending, and when the world’s money locks up the way it did over 2008 and the beginning of 2009, they suffer. At times like these, some will break, and R&D expertise and know-how will be dispersed— or worse, will be gone for good. But what biotech showed us in 2008 and 2009 is its ability to hibernate until money flows again. The industry has long had to make do with less—a valuable trait when the tap runs dry. It forces the sector’s executives to look constantly for new ways to trim expenses and to partner. This can be seen through collaborations by Symphony Capital (New York), which invests in clinical programs rather than a company itself, or the low-infrastructure model espoused by groups such as Talaris Advisors (Hopkinton, MA, USA), or the use of contract research organizations to outsource portions of drug development. The economic upswing seen in the second half of 2009 has continued. Overall funding in the first six months of 2010 is on pace to easily surpass 2009 for both private and public biotechs. The FDA approved 16 biologics last year, an increase over both 2008 (11 biologic approvals) and 2007 (9 biologic approvals). The Nasdaq biotech index has held ground for the first six months of 2010. J. Craig Venter and colleagues caught the world’s attention by creating a bacterium with an artificial genome. Biotech made its way to the Supreme Court, winning a decision favorable to Monsanto (St. Louis, MO, USA) and others developing genetically modified seeds. And so far this year, there have been approvals of Amgen’s Prolia (denosumab) for post-menopausal osteoporosis and Provenge

Number of companies

© 2010 Nature America, Inc. All rights reserved.

Source: BioMedTracker

Figure 3 Public biotech company revenue, R&D spending, profits and number of employees by market cap. Large cap, ≥$5 billion; mid-cap, $1 billion to <$5 billion; small cap, $250 million to <$1 billion; microcap, <$250 million.

nature biotechnology volume 28 number 8 august 2010

799

p at e n t s

Bilski v. Kappos: the US Supreme Court broadens patent subject-matter eligibility William J Simmons

© 2010 Nature America, Inc. All rights reserved.

The court narrowly ruled that business methods may be patent eligible, while striking down the primacy of its main test.

W

ith over 60 biopharmaceutical products applied for or expected to be filed at the US Food and Drug Administration this year, joining over 335 currently approved biopharmaceuticals, determining what can or cannot be patented is a threshold question protecting inventions in biotech and pharmaceutical industry1. Up until 2008, the answer to this important question was relatively clear. However, a 2008 decision that set a new single standard for patent eligibility made addressing this inquiry fundamentally uncertain. In a landmark decision issued 28 June, the US Supreme Court issued its holding regarding patent-eligible subject matter in Bilski v. Kappos. The court unanimously agreed that Bilski’s claims recited no more than “abstract ideas” and were therefore not patentable under US law. Importantly, a majority of the court held that the language of the relevant law (35  USC  §§100–101) broadly encompassed vast forms of subject matter as patent eligible. The court unanimously struck down the ‘machine-or-transformation’ test1, a test implemented by the US Court of Appeals for the Federal Circuit in 2008 that was criticized as “unnecessary,” as the sole test for determining whether a process is directed to patentable subject matter and held that the machine-or-transformation test is one test among many that can be used to determine patent eligibility. Justice Kennedy delivered the court’s opinion, with Justices Roberts, Thomas and Alito joining in full and Justice Scalia joining in part. Justice Stevens filed a concurring opinion in which Justices Ginsburg, Breyer and Sotomayor joined. Justice Breyer filed a concurring opinion in which Justice Scalia joined in part. William J. Simmons is at Sughrue Mion, PLLC, Washington, DC, USA. e-mail: [email protected]

The facts of Bilski v. Kappos did not involve biotech or pharmaceutical subject matter but rather a process for hedging risk in commodity markets (that is, an invention regarding instructing buyers and sellers of commodities in the energy market to protect against the risk of price fluctuations)2. For example, the application recited a series of steps instructing how to hedge risk, and in another instance the application of risk hedging was described in the form of a mathematical formula. The US Patent and Trademark Office (USPTO) denied Bilski a patent because, according to the USPTO, the patent application was directed to business methods that were patent-ineligible subject matter. The USPTO reasoned that the invention was too abstract, that it merely manipulated an idea and that it failed to practically apply concepts enough to render them patentable. The administrative appeal board affirmed, concluding that the application involved only mental steps and did not result in the transformation of physical matter. On appeal, the Federal Circuit, sitting en banc, did not rely on any of the several tests used by prior courts, including the Supreme Court, but instead created and applied a new legal standard for patentability: processes are patentable only if they are tied to a particular machine or apparatus, or transform a particular article into a different state or thing— namely, the machine-or-transformation test3. The Federal Circuit reasoned that because Bilski’s claims did not satisfy the new governing test, which the court made clear should be grossly applied to all areas of technology, the USPTO’s decision was correct and Bilski was not entitled to a patent. Judge Rader, now the Chief Judge of the Federal Circuit, in dissent, indicated that the language of 35 USC §101 “contains no hint of an exclusion for certain types of methods” and stated that “ironically the Patent Act itself

nature biotechnology volume 28 number 8 august 2010

specifically defines ‘process’ without any of these judicial innovations.” Rader argued that the only limits on eligibility are inventions that embrace natural laws, natural phenomena and abstract ideas. He wrote, “this court today invents several circuitous and unnecessary tests.” Even so, Rader suggested that the hedging claim on appeal was abstract, and he stated, “Bilski’s method for hedging risk in commodities trading is either a vague economic concept or obvious on its face.” Rader pointed out that US patent law was designed to encourage ingenuity and that the law is focused not on particular subject categories but on the patentability of the specific claimed invention. He maintained that the law distinguishes eligibility from conditions of patentability and generously provides for patent eligibility. His dissent was clear: the court should not create any categorical exclusion. Rader also pointed out that in Diehr4, the Supreme Court indicated that only natural laws, natural phenomena and abstract ideas are patent ineligible. He clarified, however, that if an abstract idea is applied to a practical use, it may be patent eligible. Notably, Rader commented that the earlier Supreme Court opinion of three dissenting justices in Lab. Corp.5 misapprehended the distinction between a natural phenomenon and a patentable process, and in so doing, this opinion did not ask the fundamental question of whether the subject matter at issue is deserving of patent protection. Rader was clear that courts should not avoid this fundamental inquiry nor categorically preclude any form of invention. In response to the Federal Circuit’s decision, Bilski petitioned for and obtained Supreme Court review. The Bilski decision garnered the attention of many, prompting an unprecedented number of submissions of unsolicited briefs expressing the views of nonparties. Among the 66 briefs, 13 were submitted by or on behalf of life science organizations, including biotech and 801

pat e n t s

Bilski Affirmance Neither party 5

5

© 2010 Nature America, Inc. All rights reserved.

3

Figure 1 Number of amicus briefs from biotech and pharma sector vis-à-vis Bilski v. Kappos. Chart compares numbers of briefs arguing for a decision in favor of Bilski, for affirmance of the court’s decision against Bilski or for neither party.

pharmaceutical interests (Fig. 1 and Table 1). Interestingly, there are differing opinions on the desired outcome of the case within the industry, including support for affirmance of the decision. However, among the 13 briefs submitted, only one brief appeared to support the machineor-transformation test (with the caveat that the test be applied correctly; Figs.  1  and  2, and Table 1). During the oral arguments heard at the Supreme Court in November 2009, several justices expressed their concerns that in the absence of unambiguous limitations regarding patent eligibility, the public could be harmed by the grant of patents to inventions directed to unworthy subject matter or commercially useful subject matter that might stifle business or innovation if granted a monopoly. The chief justice and several other justices appeared dissatisfied with the Federal Circuit’s machine-ortransformation test as the sole test for patent eligibility but seemed to be concerned to avoid expanding the scope of patent-eligible subject matter beyond that limited by the court’s precedent5. In defense of its decision, the USPTO argued that the Bilski process did not comply with the machine-or-transformation test, that the claimed process was a method of conducting business that was per se unpatentable and that the claimed process was no more than an abstract idea and therefore unworthy of a patent. The USPTO was clear about the devastating effects of banning entire categories of inventions from patenting and further asserted, “to say that business methods are categorically ineligible for patent protection would eliminate new machines, including programmed computers, that are useful because of their contributions to the operation of business.” 802

The Supreme Court’s decision was supported by all justices but the Court divided 5–4 in holding that under some undefined circumstances, at least some business methods may be patented. The Court did not clarify under which circumstance one could distinguish a patenteligible business method from an unpatentable “abstract idea,” leaving this issue for the Federal Circuit to decide. In reaching its decision, the court looked to the language of the law that describes four categories of patentable subject matter: processes, manufactures, machines and compositions of matter. A problem, however, arises in that the law sets forth a circular definition of ‘process’, making it difficult, at times, to determine whether a process meets the requirements of the statute. According to the court, the machine-ortransformation test, when applied as the sole test of determining a statutory process, violates proper statutory interpretation because “[t]he term ‘process’ means process, art or method, machine, manufacture, composition of matter or material” and the ordinary definition of process does not require that it be tied to a machine or transform an article.”5 Joined by three other justices, Justice Kennedy explained that “[s]ection 101 is a dynamic provision designed to encompass new and unforeseen inventions” and that as new technologies evolve, the statute allows for the development and application of additional tests to assist in determining which processes are patent eligible. Regarding the contention that business methods are per se unpatentable, the court rejected this argument. However, the court reasoned that, in view of specific on-point legislation— namely, 35 USC §273—which creates a defense to alleged infringement of a business method

claim, the legislature intended that claims directed to business methods can be patentable subject matter. Justice Kennedy reiterated that abstract ideas (which he did not define) are not patentable and that the court’s decisions regarding the unpatentability of abstract ideas were useful in determining which business methods may be protected under the patent law. The court held that Bilski’s claims were unpatentable because they were directed to “abstract ideas.” According to the court, Bilski sought a patent on “the use of the abstract idea of hedging risk in the energy market,” which was too abstract to be patent eligible. Even though the court rejected application of an exclusive machineor-transformation test, the court was careful to point out that inventions should be considered as a whole, not analyzed by dissecting the claims into old and new elements. Although it rejected the machine-or-transformation sole standard, the court provided the Federal Circuit with great flexibility in developing and applying “other limiting criteria” useful for determining patent eligibility. This guidance by the court is important and when properly implemented will fundamentally impact method patenting in every act. The court’s reasoning was grounded on precedent, such as that articulated in Benson7, Flook8 and Diehr5. The court held that the claims at issue were unpatentable because allowing Bilski “to patent risk hedging would pre-empt use of this approach in all fields, and would effectively grant a monopoly over an abstract idea.” The court did not formulate a new test but instead held “precedents establish that the machine-ortransformation test is a useful and important clue, an investigative tool” and nothing more. It is therefore clear that the machine-or-transformation test is a nonexclusive option for lower courts, in addition to the tests set forth in the court’s earlier decisions. The guidance set forth in Benson, Flook and Diehr should therefore be carefully considered and revisited. Briefly, in Benson, the patent sought related to an algorithm that converts numbers from binary-coded decimal form into pure binary form, which arguably could be applied to specific computer applications. The Supreme Court held that the recited algorithms were not patentable because they were drawn to abstract ideas, were not tied to a particular machine or apparatus and did not change articles or materials to a “different state or thing.” The court found it important to determine whether, assuming the algorithm to be patentable, patenting of the invention would pre-empt use of the mathematical formula. In Flook, the Supreme Court held that a method for updating alarm limits in catalytic conversion processes, which recited a mathematical

volume 28 number 8 august 2010 nature biotechnology

pat e n t s

© 2010 Nature America, Inc. All rights reserved.

Table 1 Amicus summary (selected) in Bilski v. Kappos Amicus

Industry or group represented

Summary

Novartis Corp.15

Health care solutions; pharmaceutical

Machine-or-transformation test unduly narrows the scope of diagnostic process claims. If upheld, the court should clarify that the test is not the dispositive standard.

Caris Diagnostics, Inc.16

Personalized medicine; tailoring therapeutics for individual patients using biomarkers

Machine-or-transformation test is not the exclusive test for patent eligibility of processes. Many diagnostic tests do not involve a machine or transformation.

Georgia Biomedical Partnership, Inc.17 Life sciences

Machine-or-transformation test is too rigid. Precedent is flexible and permissive.

University of South Florida18

University; research facility

Only presents arguments for the first question presented. Machine-ortransformation test excludes from patent eligibility certain processes that Congress intended to be patent eligible.

Ananda Chakrabarty19

University medical research

Machine-or-transformation test finds no support in the statute and is bad policy.

Prometheus Laboratories20

Manufacturer of pharmaceutical, medical treatment and diagnostic processes

Court’s interpretation of section 101 may have significant ramifications beyond business methods and may adversely affect the field of medical diagnostic and treatment processes.

Monogram Biosciences, Inc. et al.21

Emerging field of personalized medicine, using molecular diagnostic tests to correlate genetic and molecular biomarkers with clinically useful disease characteristics

Federal Circuit erred in holding that a process must be tied to a particular machine or transformation. This should not be the sole test. Nonphysical processes should not be excluded.

Medtronic, Inc.22

R&D of medical technology

Machine-or-transformation test would adversely affect medical technology innovation. Such a test would render significant medical advances patent ineligible.

Pharmaceutical Research and Manufacturers of America23

Pharmaceutical and biotechnology industry

Court should not adopt a new test for the boundaries of section 101. Medical processes have long been protected.

Biotechnology Industry Organization et al.11

Biotech and medical technology industries

Bilski test is not appropriate for determining patent eligibility of biotechnology and medical technology under section 101.

Knowledge Ecology International24

Advocate of new incentive and financing models for biomedical information

It is not necessary to fashion an overly broad definition of patentable subject matter merely to save medical innovations from an imagined and speculative danger.

Adamas Pharmaceuticals et al.25

Biomarkers and pharmaceuticals

Problematic business method patents should be eliminated. Machine-ortransformation test violates NAFTA and the 1994 TRIPS Agreement. This test directly over-rules Congress’s choice (35 USC section 287(c)) to maintain broad subject-matter coverage for health care–related technology.

American Medical Association et al.26

Medical profession; physicians and geneticists

Bilski’s claims are not directed to technology. Machine-or-transformation test must remain secondary and cannot supplant this court’s requirement that claims address a technology or the court’s pre-emption standard. Machine-ortransformation test must be allowed to vary with each particular case.

algorithm for computing an updated alarm limit from measured present values of variables, was not patent eligible. The court held that the identification of a limited category of useful post-solution applications of a formula does not make an otherwise unpatentable formula patentable because a process itself, not merely the mathematical algorithm, must be new and useful in order to meet the requirements for patentability. In Diehr, the court addressed a process for molding uncured synthetic rubber into a cured product. The claims were directed to a method that constantly measured the actual temperature inside a mold. The court held that the process constituted patentable subject matter under 35 USC §101 because the transformation and reduction of an article “to a different state or thing” is one clue to the patentability of a process claim that does not include a specific machine. In this instance, the court determined that the invention manifested the transformation of an article, uncured synthetic rubber, into a different state or thing. Although the invention used a well-known mathematical equation, the court remarked that the applicants did not seek to pre-empt the use of the equation.

Impact on life science technologies In Bilski, four Supreme Court justices unequivocally indicated that nascent technologies, such as biotech and pharmaceutical processes, are patent eligible. This plurality expressed appreciation for technological progress and acknowledged that “unforeseen innovations such as computer programs” are patent eligible9. Justice Kennedy reasoned that the machine-or-transformation test may be an appropriate test for evaluating the patent eligibility of processes of the Industrial Age but should not be the sole test for newer types of inventions, such as medical diagnostic techniques. Interestingly, Justice Kennedy was careful to point out that he was “not commenting on the patentability of any particular invention, let alone holding that any of the above-mentioned technologies from the Information Age should or should not receive patent protection.” Regarding limiting interference with the development of nascent technologies, such as biotechnology and biopharmaceuticals, the court indicated that some types of inventions “raise special problems in terms of vagueness and suspect validity” and could “put a chill on creative endeavor and dynamic change.”

nature biotechnology volume 28 number 8 august 2010

In dramatic contrast, however, Justice Stevens’ concurrence (joined by Justices Breyer, Ginsburg and Sotomayor), in a separate 47-page opinion, “strongly disagree[d] with the court’s disposition of this case.” Justice Stevens expressed great concern that the court “never provides a satisfying account of what constitutes an unpatentable abstract idea” and indicated that business method patents are per se unpatentable even though Bilski’s claims and application materials presented concrete parameters that may have amounted to more than an abstract idea or generalized concept. Justice Stevens cited English and early American patent jurisprudence and legislation as supportive of the opinion, concluding that the scope of patenteligible subject matter is “broad” but not limitless because, according to history, neither the patent statute nor patent law was intended to include business methods. Interestingly, biotech or pharmaceutical processes were not differentiated from business methods in the opinion, and it remains unclear to what extent such inventions could be distinguished, sufficient to survive Stevens’ per se ban. 803

pat e n t s

Against MoT test 1

Support MoT test

© 2010 Nature America, Inc. All rights reserved.

12

Figure 2 Amicus briefs from biotech and pharma sector supporting or against Bilski machine-ortransformation test. MoT, machine-or-transformation.

The dissenting justices agreed with the majority that the machine-or-transformation test was not the exclusive test for method claim patentability, but they went further, indicating that business methods are categorically excluded from patentable subject matter. Justice Stevens indicated that the court should have held “that Petitioners’ claim is not a ‘process’ within the meaning of Section 101 because methods of doing business are not, in themselves, covered by the statute.” Regarding the majority opinion’s holding that the patentability of business methods was clear from a reading of the satute, Stevens asserted that Congress did not explicitly state that it was amending and expanding the patent statute to include business methods; thus, he wrote, it was improper for the court to make such a presumption. Justice Stevens did not indicate how business method patents are categorically distinct from other forms of patent protection (for example, life science processes or therapeutic processes) but rather expressed “serious doubts” about whether business method patents are needed to encourage business innovation. It is unclear to what extent a safe harbor defense to those alleged of infringement of a business method claim applies to biotech or pharmaceutical businesses. The dissent therefore encompasses life science methods and Stevens’ logic applies equally well to biotech and pharmaceutical method patents vis-à-vis therapeutic innovations, making it critical for the industry to consider how each of their process inventions encourage medical innovations. Justice Breyer filed a separate concurring opinion, joined by Justice Scalia, indicating that agreement was reached by all of the 804

justices on at least four points: (i) the statute is broad but has some narrow limits; (ii) the machine-or-transformation is a useful test; (iii) the machine-or-transformation test is not to be misunderstood as the governing test; and (iv) by no means is everything that produces a “useful, concrete and tangible result” a patenteligible process. Regarding the breadth of patent-eligible subject matter, Justice Breyer considered the issue at oral argument wherein he indicated, “…every successful businessman typically has something. His firm wouldn’t be successful if he didn’t have anything others didn’t have…—and it’s new, too, and it’s useful, made him a fortune—anything that helps any businessman succeed is patentable because we reduce it to a number of steps, explain it in general terms, file our application, granted…” to which the attorney answered yes, what was described by Justice Bryer is potentially patentable. The Justice was also concerned that by simply assigning a set of instructions to a computer, and including the computer in the patent, an otherwise unpatentable process would be rendered patentable, asking, “how you are going to later, down the road, deal with the situation of all you do is get somebody who knows computers, and you turn every business patent into a setting of switches on the machine because there are no businesses that don’t use those machines.” This concern was directly addressed by Judge Rader, in the Bilski dissent at the Federal Circuit, wherein he focused the court not on patent ineligibility but rather on the fundamental inquiry of determining if an invention was worthy of patent protection (e.g., if an invention is novel and not obvious).

Important pending life sciences cases Following the Federal Circuit’s decision in Bilski, several cases were decided based solely on the machine or transformation test. Parties whose patent claims were held to be invalid under Bilski will take advantage of the change in law and seek reversal of these decisions. One such case is Association for Molecular Pathology  v.  USPTO (hereinafter, AMP), wherein the patent claims at issue are related to isolated DNA containing all or portions of the BRCA1 and BRCA2 gene sequence and methods for comparing or analyzing BRCA1 and BRCA2 gene sequences to identify the presence of mutations correlating with a predisposition to breast or ovarian cancer10. In a decision that radically changed the law, the court held that the step of isolating or purifying DNA does not sufficiently change the genetic sequence found in nature to make a claim to the gene per se patent eligible and that comparisons of DNA sequences are abstract mental processes, and thus not patent eligible. The court discussed abstract ideas, referring to the Federal Circuit’s opinion in Bilski, and applied the machine-ortransformation test to invalidate the process claims. In deciding AMP, the court discussed and distinguished another critical case, Prometheus Laboratories v. Mayo11. Prometheus Laboratories owns patents covering a method to optimize dosage of two drugs useful for autoimmune diseases, which involves administering a drug at certain dosage, detecting the concentration of certain metabolites and then comparing the value to a preset threshold value and subsequently increasing or decreasing the drug dosage accordingly. The Federal Circuit considered that this diagnostic process based on a correlation between drug metabolites level and drug efficacy and toxicity was patent eligible because, consistent with In re Bilski, a claimed process is patenteligible if the claimed process is transformative (e.g., citing the administering step and various chemical and physical changes of the drug’s metabolites that enable their concentrations to be determined). The court reasoned that determining the levels of drug metabolites was per se transformative because drug metabolite levels cannot be determined by mere inspection. And because these transformations were central to the invention, according to the court, the process was found to be patent eligible and the patent was held valid. The court provided no guidance as to when the interaction of a drug metabolite with the human body is a natural phenomenon. In Bilski, Justice Kennedy discusses the technological aspects of the Industrial Age and the Information Age, suggesting that the differences between the two periods provides insight

volume 28 number 8 august 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

pat e n t s into how inventions are reduced to “physical or tangible form12.” Justice Kennedy seemed to be concerned that adoption of a single test­ —the machine-or-transformation test—could retard innovation by “creating uncertainty as to the patentability of … advanced diagnostic medicine techniques…”. This issue is addressed again later, where Justice Kennedy refers to “the tension, ever present in patent law, between stimulating innovation by protecting investors and impeding progress by granting patents when not justified by the statutory design13.” This tension is most evident in the field of biotechnology and biopharmaceuticals. In deciding AMP, the district court looked to Prometheus Labs10 in determining what constitutes a ‘transformation’ in the biotechnological arts; for example, if an alleged transformation is mere preparatory “data gathering,” it falls outside the “central” focus of the recited method. Myriad’s patents were directed to methods of “analyzing” or “comparing” isolated or purified DNA, not host DNA. Although the district court recognized the great difficulty in isolating the subject DNAs, the court characterized this technical accomplishment as a mere “datagathering step,” thus invalidating the claimed methods as being directed to patent-ineligible subject matter. The district court’s new patent eligibility test is that to be patent eligible, isolated material must be “markedly different” from its naturally occurring counterpart. The court referred to the Supreme Court landmark decision in Diamond v. Chakrabarty14 as precedent but did not define a “markedly different” invention. However, the district court went further and applied a “fundamental qualities” test to invalidate Myriad’s isolated DNA composition claims, indicating that a naturally occurring DNA’s “fundamental quality” is to contain “the physical embodiment of biological information,” which is the same “fundamental quality” as isolated DNA. The court appeared to reason that because both forms of DNA shared this quality, the isolated DNA was not sufficiently different from the naturally occurring DNA to render it patent eligible—a sweeping conclusion that draws into question the validity of thousands of patents susceptible to the application of similar logic. It is also important to remember the questions raised by the court in Lab Corp. v. Metabolite5 in attempting to differentiate patent eligible subject matter from ineligible biotech inventions. In this case, the Supreme Court declined to explicitly consider the issue of the patent eli-

gibility of claims to a method for detecting the deficiency of cobalamin or folate by measuring the level of homocysteine in body fluids. The Federal Circuit held that the claims were valid but did not address the issue of patent eligibility under 35 USC §101. The Supreme Court then declined to review the decision, with Justices Stevens, Breyer and Souter dissenting. The dissenting opinion maintained that the claims were invalid because they recited only natural phenomena, which are not patent eligible. The dissent was compelled by public policy considerations and indicated that if the correlations between metabolite levels and disease were patent eligible, physicians may not be able to exercise their best judgment or might waste time, and the cost of healthcare would increase a result that would outweigh the value of protecting the invention at issue. Conclusion In Bilski, the Supreme Court expanded the forms of biotech and pharmaceutical inventions that are patent eligible in the US, holding that the machine-or-transformation test is not the sole test for patent eligibility in the US and the types of patent-eligibile subject matter are vast. But the Court narrowly avoided a catastrophe for the biotech and pharmaceutical industry. A majority of the court declined to adopt the view that “new technologies may call for new inquiries” directed to patent eligibility, which would adapt patent law to inventions of the Information Age. While the court unanimously held that Bilski’s process claims were not patent eligible, it indicated that the machine-or-transformation test may be useful for determining whether a method claim meets the threshold requirements of eligibility. Thus, universities and companies should consider providing sufficient evidence to satisfy the machine-or-transformation test when seeking to obtain patents. Although a 5–4 majority held that business methods are not categorically unpatentable, the court was a single vote away from denying business methods patent protection. This is chilling in view of the implications of such a ruling for other areas of technology, such as biotech and pharmaceutical method patenting. The court refrained from articulating a generic test that would distinguish a patentable method from an abstract idea. It remains to be seen how the USPTO, district courts and the Federal Circuit will proceed to define a new standard of patent eligibility designed to accommodate future

nature biotechnology volume 28 number 8 august 2010

innovations such as those emerging in the life sciences. The courts must provide guidance to the biotech industry as to what is patentable. What is clear, however, is that based on the court’s determination that Bilski’s claims were unpatentable because they were directed to abstract ideas, it is essential for the pharmaceutical and biotech industry to pursue and obtain method claims of varying scope and pre-emptively evaluate any available evidence to address future attacks on their intellectual property based on Bilski, at least until a medically important “abstract idea,” which could include an otherwise patentable invention under US law, is distinguished by the courts or the legislature. acknowledgments The views expressed are solely the author’s and do not represent those of Sughrue Mion and its clients, and are subject to changes in the art and law. The author thanks An Kang Li and Stuart Levy for their cotributions. COMPETING FINANCIAL INTERESTS The author declares no competing financial interests. 1. Langer, E.S. Realistic expectations likely to prevail in 2010. Gen. Eng. News (March 1, 2010). 2. In re Bilski, 545 F.3d 943 (Fed. Cir. 2008) (en banc). 3. Simmons, W.J. Nat. Biotechnol. 27, 245–248 (2009). 4. In re Bilski, 545 F.3d at 954. 5. Diamond v. Diehr, 450 US 175, 185 (1981). 6. Lab Corp. v. Metabolite (Fed. Cir. 2004). 7. Gottschalk v. Benson, 409 US 63 (1972). 8. Parker v. Flook, 437 US 584 (1978). 9. Bilski v. Kappos, 561 US (2010) at 8. 10. Association for Molecular Pathology et al. v. USPTO et al. 1:09 cv-04515 (SDNY). 11. Prometheus Laboratories v. Mayo Collaborative Servs., 581 F.3d 1336 (Fed. Cir. 2009). 12. 561 US (2010) - Kennedy, p. 9. 13. 561 US (2010) - Kennedy, pp. 12–13. 14. Diamond v. Chakrabarty 447 US 303 (1980). 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 26.

805

patents

© 2010 Nature America, Inc. All rights reserved.

Recent patent applications in proteomics Priority application date

Publication date

Patent number

Description

Assignee

Inventor

US 20100155243

A method for separating a sample, involving introducing the sample into a microchannel formed in a module and separating the sample into sub-samples according to isoelectric point and into protein components based on electrophoresis; useful in, e.g., proteomics.

Baraniuk JN, Schneider TW

Baraniuk JN, Schneider TW

2/26/2003

6/24/2010

US 20100129842

Proteomic analysis of polypeptides for biomarker analysis, involving reacting two polypeptide samples, each having reactive analytes, with different labeling reagents of a set of labeling reagents, mixing, digesting with enzyme and performing mass analysis.

Life Technologies (Carlsbad, CA, USA)

Coull JM, Pappin DJC, Purkayastha S

1/5/2004

5/27/2010

WO 2010052510

A method of diagnosing S-adenyl-l-homocysteine hydrolase deficiency involving determining qualitativequantitative blood plasma proteomic profile and diagnosing S-adenyl-l-homocysteine hydrolase deficiency based on data obtained by the subject method.

Rudjer Boskovic Institute (Zagreb, Croatia)

Cindric M, Hock K, Kraljevic Pavelic S, Sedic M

11/5/2008

5/14/2010

CN 101696238

The total protein extract of a plant, and a method for its preparation, comprising phenol and the reducing agents mercaptoethanol or dithiothreitol; used for proteomics research on plant tissue samples.

Guangdong Academy of Agricultural Sciences Crop Research Institute (Guangdong, China)

Chen X, Liang X, Zhang E

10/27/2009

4/21/2010

JP 2010078455

A method for isolating a peptide, e.g., disease marker protein, from blood, involving performing multidimensional column chromatography using an amphoteric ion column to isolate the peptide and performing protein mass spectrometry.

Japan Science and Technology Agency (Saitama, Japan)

Asajima M, Fukuda H, Into A, Kurisaki A

9/26/2008

4/8/2010

WO 2010035129

An apparatus for separating constituents of a complex protein mixture for proteomic analysis, comprising the separation of elements having chemical-physical features such that they can capture proteins belonging to the determined homogeneous group by adsorption.

National Research Council (Rome)

Boccardi C, Citti L, Mercatanti A, Parodi O, Rocchiccioli S

9/29/2008

4/1/2010

WO 2010026742

A liquid chromatograph for proteomic analysis that injects a sample solution into an injection valve through an injection port that is arranged in the flow path of the injection valve.

GL Sciences (Tokyo)

Uzu H, Zhou X

9/2/2008

3/11/2010

WO 2010011860

A method for determining if a subject of interest has pre-diabetes or diabetes or is at risk for developing pre-diabetes or diabetes, or for monitoring the efficacy of a therapy, comprising comparing a proteomic profile of a test sample with a reference sample.

Diabetomics (Beaverton, OR, USA)

Nagalla SR, Paturi VR, Roberts CT

7/23/2008

1/28/2010

WO 2010010108

A new cell with no or low endogenous dihydrofolate reductase (DHFR) levels comprising at least two heterologous vector constructs; useful as a model cell for production cell proteomics and for manufacturing proteins.

Boehringer Ingelheim Pharma (Ingelheim, Germany)

Becker E, Florin L, Kaufmann H, Studts JM

7/23/2008

1/28/2010

US 7653493

Stanford University A system for automatic mass spectroscopy analysis of a (Palo Alto, CA, USA) group of proteomic samples, e.g., peptides, comprising a unit for detecting ions, ion data processing units to receive the ion data and a material characterization processor.

Brown M, Chungfat N, Dutta S, Mathewson S, Wang EW

2/24/2006

1/26/2010

JP 2010014689

A method for the determination of melanoma, involving detecting or quantifying a melanoma marker gene, e.g., serum amyloid A2 gene, or melanoma marker protein, e.g., serum amyloid A2 protein, in a biological sample extracted from a human.

6/6/2008

1/21/2010

Shizuoka Ken

Akiyama Y, Takigawa M

Source: Thomson Scientific Search Service. The status of each application is slightly different from country to country. For further details, contact Thomson Scientific, 1800 Diagonal Road, Suite 250, Alexandria, Virginia 22314, USA. Tel: 1 (800) 337-9368 (http://www.thomson.com/scientific).

806

volume 28 number 8 august 2010 nature biotechnology

news and views

Can HIV be cured with stem cell therapy? Steven G Deeks & Joseph M McCune

© 2010 Nature America, Inc. All rights reserved.

Transplantation of human hematopoietic stem cells engineered to lack the viral coreceptor CCR5 confers resistance to HIV infection in mice. Antiretroviral therapy has transformed the treatment of HIV infection, but, despite its profound successes, it will not halt the relentless advance of the epidemic. Against this sobering reality, several promising, recent developments in the basic-science arena have led HIV researchers to envision new thera­peutic approaches that would completely eradicate the virus, effectively ‘curing’ HIV disease. In an exciting and impressive display of data published in this issue, Holt et al.1 provide a scientific bellwether for the practical implementation of one such strategy. They show that CCR5, a human gene often required for HIV to enter target cells, can be effectively and permanently disrupted in long-lived, multilineage, human hematopoietic stem cells (HSCs). When introduced into mice, these cells generate an apparently intact human immune system that is resistant to subsequent infection with HIV (Fig. 1a). This result raises the intriguing possibility that HIV-infected individuals might be cured with a one-time infusion of autologous, gene-modified HSCs. The introduction of combination antiretroviral regimens against HIV in the mid-1990s was undoubtedly one of the great triumphs of modern medicine. Almost overnight, those who could receive and adhere to the therapies gained a new lease on life. But the passage of time has revealed the limitations of these regimens. Because HIV DNA persists as an integrated genome in long-lived cellular reservoirs, current antiretroviral drugs are unlikely to prove curative2. In addition, the therapies require life-long adherence, which many find challenging, and are often associated with some short-term and long-term Steven G. Deeks and Joseph M. McCune are in the Department of Medicine, University of California, San Francisco, California, USA. e-mail: [email protected] and [email protected]

t­ oxicity. Moreover, although they suppress HIV replication in a potent, durable manner, they do not restore health; for reasons that remain unknown, treated HIV disease is attended by chronic inflammation, persistent T-cell dysfunction and a shortened life expectancy3,4. Finally, and perhaps most importantly, antiretroviral therapies and their management are expensive and hard to deliver on a worldwide basis. It is now apparent that the number of HIV-infected people will continue to eclipse the number that can be successfully treated. To stop the epidemic and to provide care for all, a fundamentally different approach is needed. Gene therapy with HSCs The concept of HSC-based gene therapy for HIV disease emerged in the epidemic’s first decade, when effective antiretroviral regimens were nonexistent. Multiple advances in delivering and expressing transgenes in eukaryotic cells suggested that therapeutic applications were within reach5. Baltimore coined the term “intracellular immunization” to describe the introduction of HIV resistance genes into HSCs to allow long-term repopulation of the host with progeny cells that would be impervious to HIV6. By the late 1980s, startup biotech companies were isolating and preparing human HSCs for transplantation, devising techniques and vectors to genetically modify the cells, and conducting preclinical testing7. During the same period, studies of HIV pathogenesis were generating data that begged for a therapeutic approach that went beyond antiretroviral drugs. On the one hand, it became clear that CD4+ T-cell depletion, the hallmark of HIV disease, was caused not simply by destruction of late-stage CD4+ T effector cells but also by the host’s inability to maintain progenitor cells, including HSCs, intrathymic T progenitor cells and central memory T cells in the periphery8 (Fig. 1b). On the other hand,

nature biotechnology volume 28 number 8 August 2010

it was recognized that HIV can persist within multiple lineages of long-lived cells, including T cells and cells of the myeloid lineage (some of which appear to be progenitor cells)2,9. Taken together, these observations underscored the need to confer HIV resistance to both progenitor cells and their progeny. Early attempts to engineer HIV resistance into hematopoietic progenitor cells encountered insurmountable hurdles: the scientific and practical constraints of HSC-based therapies were substantial; protocols for genetic modification of HSCs were inefficient and cytotoxic; the preclinical animal models were inadequate; and the choice of anti-HIV genes was driven more by convenience (and/or ­patent considerations) than by data7. Moreover, it proved difficult to devise a business model that could support the introduction of such a dramatically different, untested and potentially toxic form of therapy into the clinic. More recently, however, two important developments have prompted a reevaluation of HSC-based therapy for HIV: a critical target—the cell-surface receptor CCR5—was identified, and an HIV-infected individual was reported to be virus free in the absence of antiretroviral ­medications 20 months after receiving a transplant of CCR5defective allogeneic HSCs10. Targeting the Achilles’ heel of HIV To enter cells, HIV must bind to either CCR5 or CXCR4, chemokine receptors present on many immune cells11. The vast majority of transmitted viruses use CCR5 (R5 variants). As the disease progresses, HIV evolves and often, but not always, expands its co-receptor preference to include CXCR4 (X4 variants). A small fraction of people carry a 32-base-pair deletion in the CCR5 gene, leading to a truncated gene product, CCR5 ∆32. Those who are heterozygous for CCR5 ∆32 have delayed disease progression after they acquire HIV, whereas homozygotes rarely acquire HIV11. Although lack of 807

ne w s and vie w s a

HIV-resistant immune cells

CCR5 knockout with ZFNs

Human HSPCs

Low viremia

High viremia

No CCR5 modification

Transplant HSPCs into NOG mice

Challenge with CCR5-tropic HIV

b © 2010 Nature America, Inc. All rights reserved.

SP4 CLP

ITTP

HIV-mediated destruction of immune cells

CD4M

CD4E

CD4N

CD4E

CD8N

CD8E

CD8M

CD8E

DP SP8

HSC

HSPC/thymusmediated expansion of peripheral CD4+ T cells

Thymus M/E

Bone marrow

Tissue myeloid cells

Peripheral lymphoid system

Figure 1 Reconstitution of an HIV-resistant lymphoid and myeloid system in an experimental model. (a) Holt et al.1 isolated human hematopoietic stem/progenitor cells (HSPCs) and used zinc-finger nucleases (ZFNs) to disrupt the CCR5 gene, which is often required for the entry of HIV into target cells. Mice that were successfully engrafted with CCR5-disrupted HSPCs tolerated infection with HIV, whereas those engrafted with unmodified HSPCs exhibited loss of CD4+ T cells and high-level viremia. (b) Long-lived, multilineage hematopoietic stem cells (HSCs) give rise to common lymphocyte progenitors (CLPs) and progenitors of the myelo-erythroid (M/E) lineages. CLPs move through the thymus and differentiate through a series of stages, from CD3–CD4+CD8– intrathymic T progenitor (ITTP) cells to CD3+/– CD4+CD8+ double positive (DP) thymocytes to CD3+ thymocytes that are single positive for CD4 (SP4) or CD8 (SP8) to circulating naïve (N), effector (E), and memory (M) CD4 + or CD8+ T cells. All of the cell stages colored in red can be directly or indirectly disabled by HIV infection.

CCR5 may be associated with increased risk of developing serious sequela of some uncommon infections12, it does not seem to affect life expectancy and may even be associated with a reduced risk of certain inflammatory diseases. Once the role of CCR5 became clear in the late 1990s, the pharmaceutical industry devoted tremendous resources to the development of small-molecule inhibitors, one of which, maraviroc (Selzentry), is highly effective, well­tolerated and now FDA approved. This important set of observations inspired several groups to pursue CCR5-targeted gene therapy13,14. One highly innovative approach relied on engineered zinc-finger nucleases specific for the CCR5 gene15. Such ‘molecular scissors’ can be delivered to cells ex vivo using methods such as integrase-defective lentiviral vectors, adenoviral vectors and plasmid DNA nucleofection. After specific binding of a pair of zinc-finger nucleases to the CCR5 gene, a double-stranded DNA break is introduced and then repaired by pathways that include 808

e­ rror-prone nonhomologous end-joining. This can create a permanent gene disruption that is passed to daughter cells in the absence of persistent transgene expression. The end result is the functional disruption of the CCR5 gene. Previous work using this approach demonstrated its feasibility in human peripheral blood CD4+ T cells15, and unpublished data from a phase 1 trial suggest that autologous CD4+ T cells modified in this way can be reinfused safely into HIV-infected individuals (P. Tebas, University of Pennsylvania, personal communication). Although of great interest, this strategy does not disrupt CCR5 in HSCs and thus would not enable the long-term generation of both T and myeloid-lineage cells resistant to HIV infection. Evidence supporting such a leap came from another quarter. The Berlin patient: an instructive N of 1 For all of those engaged in the care and treatment of patients with HIV disease, the world changed in 2009 with the remarkable story of

a stably treated, HIV-infected individual—the ‘Berlin patient’—who developed acute myeloid leukemia and was transplanted with HSCs from a human leukocyte antigen–matched, homozygous CCR5 ∆32 donor10. Combination antiretroviral therapy was discontinued the day before the transplant. Twenty months later, HIV could not be detected in any of the patient’s tissues examined, even when very sensitive techniques were used. Given ­disappointing treatment outcomes in the past, the HIV research community is hesitant to use the word ‘cure’, but this single case could very well be the first example to fit the bill. It is important to emphasize that this road to a cure was arduous and will not be available to the vast majority of patients. The Berlin patient underwent fully ablative condi­tion­ing with a potentially lethal regimen that included fludarabine (Fludara), Ara-C, amsacrine (Amerkin, Amsidyl, Amsidine), cyclosporin, mycophenolate mofetil (CellCept), antithymocyte globulin and 4 Gy of total body irradiation. ­Graft-versus-host disease developed during the post-transplant period. Owing to recurrent acute myeloid leukemia, a second stem cell transplantation using cells from the same donor was performed one year after the first transplant, which again required exposure to myeloablative therapy, including irradiation. No one believes that this approach will soon be used beyond the highly unusual indications for which allogeneic transplantations are normally performed. However, the example of the Berlin patient does provide a strong rationale for the development of CCR5targeted stem cell therapy. This case also provides fascinating insights into HIV pathogenesis, some of which may be relevant to future attempts at HIV eradication. For example, it is not entirely clear why HIV did not rebound after combination antiretroviral therapy was discontinued. According to genotypic assays, the patient likely harbored a minority (2.9%) of X4 variant viruses. Also, host-derived CCR5-expressing myeloid cells, which are permissive for HIV infection, persisted for at least five months after the transplant. Given this volatile combination of residual CXCR4-tropic virus and long-lived CCR5-expressing targets, HIV replication and spread should have continued even as the rest of the hematopoietic system was being replaced by homozygous CCR5 ∆32 donor cells. There are at least two possible explanations for this surprising result. First, the low-level X4 variant may have been a poorly fit dualtropic virus that was dependent on CCR5 for replication, whereas the number of residual CCR5-expressing myeloid cells was too low to support systemic replication of the ­CCR5-tropic

volume 28 number 8 August 2010 nature biotechnology

ne w s and vie w s

© 2010 Nature America, Inc. All rights reserved.

v­ ariants. Second, it is possible that the myeloablative preparative regimen itself contributed to the cure by destroying latently infected T and myeloid cells and by reducing the numbers of susceptible activated CD4+ T cells (HIV more readily infects activated rather than resting target cells). It is also possible that the ongoing graft-versus-host disease may have acted to clear residual susceptible target cells. Detailed exploration of these and other mechanisms will surely provide profound insights into almost any possible intervention aimed at HIV eradication in the future, and should be pursued. Disruption of CCR5 in autologous HSCs The strategy of Holt et al.1 is related to the treatment received by the Berlin patient but is potentially relevant to a larger number of patients (Fig. 1a). The authors obtained human CD34+ hematopoietic stem/progenitor cells (HSPCs) (a population enriched in HSCs) from umbilical cord blood and stimulated them to divide with Flt-3 and thrombopoietin. The cells were nucleofected with plasmids expressing CCR5-specific zinc-finger nucleases. A mean of 17% of the cells were successfully modified, 5–7% of which were estimated to be homozygous CCR5–. Modified or unmodified CD34+ cells were then transplanted into nonobese diabetic/severe combined immunodeficient/interleukin 2rγnull (NOD/ SCID/IL2rγnull or NOG) mice, a model known to support multilineage human hematopoiesis. As expected, mice engrafted with unmodified stem cells and subsequently challenged with CCR5-tropic HIV (Bal) showed high levels of viremia and loss of peripheral and tissuebased human T cells. Remarkably, in animals repopulated with CCR5-disrupted HSPCs, the virus levels were lower and CD4+ T cells were not depleted, either in the peripheral blood or in the hemato­lymphoid tissues (e.g., bone marrow, thymus, spleen and small intestine). The preservation of human CD4+ T cells in the experimental group was due to selection for multiple independent clones of successfully gene-modified cells. The frequency of cells containing evidence of CCR5 disruption increased to >80% in the peripheral blood and to >40% in multiple tissues by week 12 of infection. These experiments raise a number of technical issues and derivative questions. For instance, do genetically modified HSCs confer benefit to a mouse that is already infected (the situation most closely approximating the therapeutic need in humans)? Does a CCR5disrupted hematopoietic compartment confer protection against infection by X4 viral variants? Is the CCR5-disrupted immune system normal? Are there long-term toxicities that will

become evident later? Is off-target cleavage by the zinc-finger nucleases a significant concern (e.g., the CCR2 gene may also be targeted by this nuclease15)? These and other issues can be resolved with further work. In the meantime, the data of Holt et al.1 show convincingly that a relatively small number of gene-modified HSCs can be rapidly selected to ultimately confer resistance to HIV in vivo. Next steps If stem cell–based gene therapy for HIV is to become a reality in the clinic, a number of ­nontrivial theoretical and practical concerns must be addressed. First, in the current era, when clinicians are increasingly concerned about the ‘toxicity’ of ongoing viral replication, will patients and their healthcare providers be willing to allow HIV to replicate at high levels in the absence of antiretroviral therapy so that CCR5-deficient cells can be selected? There is now a growing consensus that HIV replication causes significant and perhaps irreversible harm to many organs, including those of the cardiovascular, renal, hepatic and neurologic systems4, so this approach must be assumed to carry some risk. Second, will a partially effective antiviral intervention (which is what the gene-­modified cells represent) select for the outgrowth of a resistant virus population, such as X4 variants? The history of HIV therapeutics is absolutely clear on this issue: if HIV is allowed to replicate in the presence of a selective pressure, it will find a way to survive. This concern is even more pressing as it is widely believed that X4 variants are more virulent than R5 variants. Although X4 variants are only infrequently selected in patients treated with small­molecule CCR5 inhibitors (e.g., maraviroc), this is only true when a fully suppressive regimen is used from day one. It is not likely that transplantation of gene-modified HSCs will be fully suppressive at first, particularly if partially myeloablative therapy is used. Third, will ablative therapy be needed to allow stem cell engraftment and, if so, will short- and long-term toxicity preclude its use in those most likely to be offered this intervention first? Those most in need of aggressive interventions typically have dual-tropic virus and are therefore unlikely to respond to any approach based on disruption of CCR5 (ref. 16). And with advanced disease, they have a paucity of HSCs and damaged hematopoietic microenvironments (such as bone marrow, thymus and lymph node) that would normally support the maturation of modified HSCs. Finally, the mechanism whereby HIV causes CD4+ T-cell depletion remains unclear8.

nature biotechnology volume 28 number 8 AUGUST 2010

Although HIV can clearly kill cells directly, many if not most cells in an HIV-infected individual die as a consequence of indirect viral effects. Generalized activation of the immune system, for example, is harmful to the function of T and myeloid cells and to the regeneration of multiple lineages. These indirect effects may persist even as the virus is driven to extinction by the gradual emergence of an HIV-resistant T-cell population. Reaching for blue sky Although the above concerns are daunting, the epidemic is not going to disappear, the science of stem cells is becoming more tractable, sociopolitical forces are forging new ­perspectives in healthcare, and now is not the time to stop. From our perspective, HSC-based gene therapy for HIV disease may make a significant impact on the worldwide epidemic if two goals can be met. First, it is essential to find a way to deliver these therapies to all in need, in a manner that is safe, affordable and generally available around the world. Many clever approaches to do this have been proposed in the past, and more will surely emerge. Second, the preclinical and clinical development of these strategies requires a sustainable financial model. Such a model may involve reprioritization of governmental efforts, creative plans to incentivize existing pharmaceutical and healthcare delivery systems, and global assistance programs motivated by a common desire for a world free of HIV. This may seem like a formidable exercise, but it is worth noting that if oneshot, modified HSC-based gene therapy can be made efficacious and accessible in the context of HIV disease, similar approaches will likely be applicable to a host of other chronic diseases, infectious and otherwise. If so, the treatment paradigms of the future will look vastly different from today’s. In the same way that problems associated with the reliance on fossil fuels have stimulated the development of alternative ­strategies of energy delivery, so too may the ongoing crisis in the HIV epidemic spark novel approaches to the provision of healthcare in the future. Conclusion The progress in HIV therapeutics over the past 15 years has been tremendous. The life expectancy of most people who present with HIV disease today in resource-rich regions is on the order of decades. Yet antiretroviral drugs have intrinsic limitations that are unlikely to be surmounted. What is needed, therefore, is a ‘game changer’, such as a cure for HIV infection or an effective vaccine. Could a one-shot manipulation of HSCs be the answer? We will 809

ne w s and vie w s

© 2010 Nature America, Inc. All rights reserved.

not know unless we continue to move these new technologies into the clinic. Even if CCR5targeted gene therapy is not the ultimate solution, human studies are certain to be highly informative with regard to HIV pathogenesis and human immunology. ACKNOWLEDGMENTS The authors wish to acknowledge amfAR, Project Inform, TAG and the AIDS Policy Project for supporting and stimulating cross-disciplinary discussion on the issues outlined in this commentary. The authors’ work that contributed to this review was supported by the National Institute of Allergy and Infectious Diseases (RO1 AI087145 and K24AI069994 to S.G.D. and R37 AI40312 and DPI OD00329 to J.M.M.), the University of California, San Francisco (UCSF) Center for AIDS Research (P30 MH59037), the UCSF Clinical and Translational Science Institute (UL1 RR024131), the Harvey V. Berneking Living Trust and amfAR. J.M.M. is a recipient of the National Institutes of Health (NIH) Director’s Pioneer Award Program, part of the NIH Roadmap for Medical Research.

COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/ naturebiotechnology/. Holt, N. et al. Nat. Biotechnol. 28, 839–847 (2010). Siliciano, J.D. et al. Nat. Med. 9, 727–728 (2003). Kuller, L.H. et al. PLoS Med. 5, e203 (2008). Phillips, A.N., Neaton, J. & Lundgren, J.D. AIDS 22, 2409–2418 (2008). 5. Friedman, A.D., Triezenberg, S.J. & McKnight, S.L. Nature 335, 452–454 (1988). 6. Baltimore, D. Nature 335, 395–396 (1988). 7. Rossi, J.J., June, C.H. & Kohn, D.B. Nat. Biotechnol. 25, 1444–1454 (2007). 8. McCune, J.M. Nature 410, 974–979 (2001). 9. McCune, J.M. Cell 82, 183–188 (1995). 10. Hutter, G. et al. N. Engl. J. Med. 360, 692–698 (2009). 11. Moore, J.P., Kitchen, S.G., Pugach, P. & Zack, J.A. AIDS Res. Hum. Retroviruses 20, 111–126 (2004). 12. Glass, W.G. et al. J. Exp. Med. 203, 35–40 (2006). 13. DiGiusto, D.L. et al. Sci. Transl. Med. 2, 36ra43 (2010). 14. Shimizu, S. et al. Blood 115, 1534–1544 (2010). 15. Perez, E.E. et al. Nat. Biotechnol. 26, 808–816 (2008). w 16. Hunt, P.W. et al. J. Infect. Dis. 194, 926–930 (2006).

1. 2. 3. 4.

Microarrays in the clinic Guy W Tillinghast The MicroArray Quality Control (MAQC) consortium has evaluated methods for making clinically useful predictions from large-scale gene expression data. Clinical application of gene expression microarrays1 and other ’omics technologies is widely expected to usher in a new era of personalized medicine. But although DNA microarrays are beginning to be used in patient care2,3, progress has been slow, in part because of analytic challenges and concerns about accuracy and reproducibility. In this issue, the MAQC consortium presents the results of a large study, MAQC-II4, to evaluate methods for building genomic classifiers—software programs that convert microarray profiles of an individual sample into a prediction, such as membership in a clinical class. The results show that microarray algorithms can be reliable enough to justify clinical application, at least within certain contexts. More broadly, the findings of MAQC-II on microarray classifiers may be useful for analyzing data from other highthroughput assays. Existing clinical predictors have well-known limitations, especially with respect to complex diseases such as cancer. Given two individuals who present identical clinical parameters, one Guy Tillinghast is at the Riverside Cancer Care Center, Newport News, Virginia, USA. e-mail: [email protected]

810

may respond to a therapy whereas the other may not. In principle, genome-wide data should be able to discriminate between them. The most common goals of a clinical test are to make a diagnosis or to determine an appropriate therapy. In light of statistical considerations, these goals depend on the prevalence of a disease, suggesting that clinical DNA microarray tests will augment, and not supplant, other clinical information. Thus, a possible strategy would be to first use traditional clinical predictors to broadly identify patients who might benefit from a treatment, and to then use an expensive assay, such as a microarray, to eliminate those for whom the treatment is unlikely to be effective. Despite this promise, DNA microarrays have not been rapidly adopted in clinical practice. One reason is the noise that results from analyzing thousands of genes, which can lead to false predictions. Consequently, microarrays have been criticized because studies of the same clinical groups using different microarray measurements or analytic methods have often yielded dissimilar lists of differentially expressed genes. A second concern is the inherent error in the technology. Error stems from high background at the bottom of the dynamic

range, saturation at the top of the dynamic range, and nonlinearity, at least with measurements of some transcripts. Many statistical methods have been developed to address these challenges, including approaches for grouping samples and genes, data normalization schemes to allow meaningful comparisons across samples, multiple testing procedures to select differentially expressed genes and ‘cross-validation’ methods for using samples to train prediction algorithms while reducing bias. These methods are applied sequentially to transform massive data sets of raw microarray gene expression profiles into clinically useful classifiers (Fig. 1a). As the optimal combination of methods is difficult to determine, MAQC-II sought to evaluate approaches to building classifiers. Clinical use of microarrays is particularly challenging owing to the variability of the arrays themselves and to the variability between patients and between laboratories performing the analyses. These effects fall under the rubric of ‘batch effects’ and cause false positives. Moreover, before MAQC-II, it had not been clear whether classifiers trained on an initial data set would be able to make accurate predictions based on completely independent samples collected at a later date. The five-step process for building a classifier in MAQC-II involved designing the experiment, collecting microarray data, creating a ­predictive model, validating the model internally with the training samples and validating the model externally with new samples obtained independently from the training data. MAQC-II enlisted 36 teams of data analysts within government agencies, academia and industry. The teams were given six microarray data sets and charged with predicting 13 ‘endpoints’ potentially relevant to clinical or preclinical applications. The data sets included toxicological studies of chemicals on rodents and expression profiles of human cancer patients. In total, the teams built >30,000 classifiers using hundreds of combinations of analytic methods. A team of referees comprising biostatisticians and experienced data analysts chose one ‘candidate’ model that was expected to have the best performance for each endpoint from among models nominated by each of the 36 teams. Next, the consortium analyzed how well the models classified samples. Performance was measured using several metrics, but the one most familiar to clinicians is the receiver operating characteristic area under the curve (AUC), a metric that varies between 0 and 1, where 0.5 indicates performance no better than chance and 1 means that all samples are correctly classified and none misclassified. For most of the endpoints, the ­candidate

volume 28 number 8 AUGUST 2010 nature biotechnology

ne w s and vie w s

© 2010 Nature America, Inc. All rights reserved.

not know unless we continue to move these new technologies into the clinic. Even if CCR5targeted gene therapy is not the ultimate solution, human studies are certain to be highly informative with regard to HIV pathogenesis and human immunology. ACKNOWLEDGMENTS The authors wish to acknowledge amfAR, Project Inform, TAG and the AIDS Policy Project for supporting and stimulating cross-disciplinary discussion on the issues outlined in this commentary. The authors’ work that contributed to this review was supported by the National Institute of Allergy and Infectious Diseases (RO1 AI087145 and K24AI069994 to S.G.D. and R37 AI40312 and DPI OD00329 to J.M.M.), the University of California, San Francisco (UCSF) Center for AIDS Research (P30 MH59037), the UCSF Clinical and Translational Science Institute (UL1 RR024131), the Harvey V. Berneking Living Trust and amfAR. J.M.M. is a recipient of the National Institutes of Health (NIH) Director’s Pioneer Award Program, part of the NIH Roadmap for Medical Research.

COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/ naturebiotechnology/. Holt, N. et al. Nat. Biotechnol. 28, 839–847 (2010). Siliciano, J.D. et al. Nat. Med. 9, 727–728 (2003). Kuller, L.H. et al. PLoS Med. 5, e203 (2008). Phillips, A.N., Neaton, J. & Lundgren, J.D. AIDS 22, 2409–2418 (2008). 5. Friedman, A.D., Triezenberg, S.J. & McKnight, S.L. Nature 335, 452–454 (1988). 6. Baltimore, D. Nature 335, 395–396 (1988). 7. Rossi, J.J., June, C.H. & Kohn, D.B. Nat. Biotechnol. 25, 1444–1454 (2007). 8. McCune, J.M. Nature 410, 974–979 (2001). 9. McCune, J.M. Cell 82, 183–188 (1995). 10. Hutter, G. et al. N. Engl. J. Med. 360, 692–698 (2009). 11. Moore, J.P., Kitchen, S.G., Pugach, P. & Zack, J.A. AIDS Res. Hum. Retroviruses 20, 111–126 (2004). 12. Glass, W.G. et al. J. Exp. Med. 203, 35–40 (2006). 13. DiGiusto, D.L. et al. Sci. Transl. Med. 2, 36ra43 (2010). 14. Shimizu, S. et al. Blood 115, 1534–1544 (2010). 15. Perez, E.E. et al. Nat. Biotechnol. 26, 808–816 (2008). w 16. Hunt, P.W. et al. J. Infect. Dis. 194, 926–930 (2006).

1. 2. 3. 4.

Microarrays in the clinic Guy W Tillinghast The MicroArray Quality Control (MAQC) consortium has evaluated methods for making clinically useful predictions from large-scale gene expression data. Clinical application of gene expression microarrays1 and other ’omics technologies is widely expected to usher in a new era of personalized medicine. But although DNA microarrays are beginning to be used in patient care2,3, progress has been slow, in part because of analytic challenges and concerns about accuracy and reproducibility. In this issue, the MAQC consortium presents the results of a large study, MAQC-II4, to evaluate methods for building genomic classifiers—software programs that convert microarray profiles of an individual sample into a prediction, such as membership in a clinical class. The results show that microarray algorithms can be reliable enough to justify clinical application, at least within certain contexts. More broadly, the findings of MAQC-II on microarray classifiers may be useful for analyzing data from other highthroughput assays. Existing clinical predictors have well-known limitations, especially with respect to complex diseases such as cancer. Given two individuals who present identical clinical parameters, one Guy Tillinghast is at the Riverside Cancer Care Center, Newport News, Virginia, USA. e-mail: [email protected]

810

may respond to a therapy whereas the other may not. In principle, genome-wide data should be able to discriminate between them. The most common goals of a clinical test are to make a diagnosis or to determine an appropriate therapy. In light of statistical considerations, these goals depend on the prevalence of a disease, suggesting that clinical DNA microarray tests will augment, and not supplant, other clinical information. Thus, a possible strategy would be to first use traditional clinical predictors to broadly identify patients who might benefit from a treatment, and to then use an expensive assay, such as a microarray, to eliminate those for whom the treatment is unlikely to be effective. Despite this promise, DNA microarrays have not been rapidly adopted in clinical practice. One reason is the noise that results from analyzing thousands of genes, which can lead to false predictions. Consequently, microarrays have been criticized because studies of the same clinical groups using different microarray measurements or analytic methods have often yielded dissimilar lists of differentially expressed genes. A second concern is the inherent error in the technology. Error stems from high background at the bottom of the dynamic

range, saturation at the top of the dynamic range, and nonlinearity, at least with measurements of some transcripts. Many statistical methods have been developed to address these challenges, including approaches for grouping samples and genes, data normalization schemes to allow meaningful comparisons across samples, multiple testing procedures to select differentially expressed genes and ‘cross-validation’ methods for using samples to train prediction algorithms while reducing bias. These methods are applied sequentially to transform massive data sets of raw microarray gene expression profiles into clinically useful classifiers (Fig. 1a). As the optimal combination of methods is difficult to determine, MAQC-II sought to evaluate approaches to building classifiers. Clinical use of microarrays is particularly challenging owing to the variability of the arrays themselves and to the variability between patients and between laboratories performing the analyses. These effects fall under the rubric of ‘batch effects’ and cause false positives. Moreover, before MAQC-II, it had not been clear whether classifiers trained on an initial data set would be able to make accurate predictions based on completely independent samples collected at a later date. The five-step process for building a classifier in MAQC-II involved designing the experiment, collecting microarray data, creating a ­predictive model, validating the model internally with the training samples and validating the model externally with new samples obtained independently from the training data. MAQC-II enlisted 36 teams of data analysts within government agencies, academia and industry. The teams were given six microarray data sets and charged with predicting 13 ‘endpoints’ potentially relevant to clinical or preclinical applications. The data sets included toxicological studies of chemicals on rodents and expression profiles of human cancer patients. In total, the teams built >30,000 classifiers using hundreds of combinations of analytic methods. A team of referees comprising biostatisticians and experienced data analysts chose one ‘candidate’ model that was expected to have the best performance for each endpoint from among models nominated by each of the 36 teams. Next, the consortium analyzed how well the models classified samples. Performance was measured using several metrics, but the one most familiar to clinicians is the receiver operating characteristic area under the curve (AUC), a metric that varies between 0 and 1, where 0.5 indicates performance no better than chance and 1 means that all samples are correctly classified and none misclassified. For most of the endpoints, the ­candidate

volume 28 number 8 AUGUST 2010 nature biotechnology

ne w s and vie w s a

b Classifier

Tissue sample

Prediction

Treatment plan

Microarray

Remove batch effects

Normalize

Select features

Train algorithm

Internal validation

© 2010 Nature America, Inc. All rights reserved.

Process evaluated in MAQC-II

True-positive rate (sensitivity)

AUC = 0.991 AUC = 0.956 AUC = 0.787 AUC = 0.615

False-positive rate (1 – specificity)

Figure 1 Using microarrays to make clinical predictions. (a) Current clinical decision-making processes can be refined by gene expression–based predictions generated by microarray classifiers (top). MAQC-II evaluated methods for constructing classifiers (bottom). Constructing a classifier from raw microarray data requires processing the data using a sequence of analytic steps (colored boxes). Many different approaches have been developed to solve each step (represented as dots above each box). In MAQC-II, >30,000 classifiers were constructed to test different combinations of analytic steps to predict 13 clinical and preclinical ‘endpoints’. (b) Curves showing the range of performance of classifiers developed for different data sets as part of MAQC-II. Performance is quantified using AUC. Data sets are characterized by the ratio of positive to negative samples in the cohort (P/N). Classifiers performed well for some endpoints, such as the sex of patients. The ~400 genes exclusively present on the Y chromosome made this an easy-to-predict positive control (red, training set P/N 1.44). The most difficult-to-predict endpoint was the overall survival of multiple myeloma patients, which has traditionally been difficult for other tests as well (orange, training set P/N 0.34). Classifiers for liver toxicity in rats (blue, training set P/N 0.58) and pathological complete remission in breast cancer (green, training set P/N 0.34) showed intermediate performance.

­ icroarray-based classifiers performed far m better than chance on the independent validation data set, with a range of 0.62–0.99. Moreover, the performance of the referee­selected candidate models was better than that of nominated models, suggesting that expert advice can enhance the modeling outcome. Notably, classifier performance was found to depend heavily on the endpoint being predicted (Fig. 1b). However, it is evident from inspecting the data that there is a linear correlation between the AUC performance and the ratio of positive to negative samples in the cohort (‘training set P/N’). The composition of the training set is known to affect classification performance, and extreme imbalance, such as with the breast cancer and multiple myeloma endpoints (Fig. 1b, orange and green), may have adversely affected performance. Alternatively, the genetics of neuroblastoma and certainly the rodent data sets may be less variable and hence more tractable to modeling (Fig. 1b, blue). Moreover, genetic variation typically accumulates over time, making the genomes of the patients with breast cancer and multiple myeloma more variable than those with neuroblastoma and therefore less consistent with the reference genome from which the microarray platforms were constructed. These substantial differences in endpoints may have affected the validation AUC results. Several findings from MAQC-II may help bring the technology closer to clinical use. Microarray experiments should be designed to minimize batch effects, such as those introduced by different laboratories or material lots. There should be a plan for detecting such

effects (e.g., by testing for unexpected genes that are expressed in different experimental conditions), and the same statistical test used to detect differentially expressed genes should be applied to all samples5. A gene that is differentially expressed in a pattern that matches the grouping of samples into batches should be examined closely and probably not used in a classifier. Related to batch effects, quality control metrics should be used to distinguish variation in gene expression caused by laboratory artifact rather than by clinical phenotype. Quality control metrics are formulated to assess specific aspects of laboratory processing, such as RNA degradation or faulty equipment. These metrics can be used to adjust gene expression measurements or to identify problem microarrays. In the MAQC-II project, rather than adjusting measurements to account for laboratory noise, data analysts did not use samples that appeared to have quality control problems. Several factors were found to influence classifier performance more than the type of algorithm used. One of these is the inherent difficulty of the biological phenomena being predicted. Another is the method for tuning the algorithm. Inexperience in tuning can be a major source of bias in the final classifier, especially if the predictive algorithm is not tuned for the population of interest. For example, in a population with low prevalence of a disease, it may be more desirable to have a test that makes few false predictions. The results of MAQC-II highlight two priorities for future work. First, the field needs rigorous standards for reporting the steps

nature biotechnology volume 28 number 8 AUGUST 2010

used to develop a classifier, its parameters of use and the appropriate quality metrics. Examples in the literature2 may provide useful starting points. A classifier submitted for publication or for regulatory approval should specify how to use it to classify new samples— for example, the normalization and batch effect correction procedures to perform, the essential quality control checks and how to handle quality control flaws. The final report of a prediction algorithm should provide the variance (that is, standard error) of the performance measure as well as an estimation of the bias. A prediction report based on analysis of an individual patient sample should be accompanied by a report of quality metrics and their normal values and a report of batch effect measures that could provide a clinician with a sense of whether a microarray is within the range of the samples for which the test was developed5. Second, methods are needed to combine microarray predictions with existing clinical decision-making tools, such as nomograms (a graphical chart for ­performing calculations). In constructing a nomogram, it will be necessary to determine how to balance the data from a microarray classifier with traditional clinical predictors. In addition, approaches should be developed to handle variability. For instance, the microarray chips used in MAQC-II have already been replaced by newer versions. A key observation of MAQC-II—namely, that some endpoints seem inherently more predictable than others, regardless of the analytic methods used—suggests that gene expression microarrays may not capture a 811

ne w s and vie w s sufficiently rich snapshot of disease physiology. In such cases, complementary technologies, which measure mRNA expression, protein levels, genetic mutation, copy number variation, gene silencing or regulatory RNA expression, could be considered. Alternatively, the best technology may vary by tumor type. High-throughput sequencing, in particular, offers advantages over microarrays in that coverage of the genome is less biased and the dynamic range is larger6. With luck, the results of MAQC-II will be useful for ­shepherding

other high-­throughput technologies toward the clinic as well. COMPETING FINANCIAL INTERESTS The author declares no competing financial interests. 1. DeRisi, J.L., Iyer, V.R. & Brown, P.P. Science 278, 680–686 (1997). 2. Dumur, C.I. et al. J. Mol. Diagn. 10, 67–77 (2008). 3. Buyse, M. et al. J. Natl. Cancer Inst. 98, 1183–1192 (2006). 4. The MicroArray Quality Control (MAQC) consortium. Nat. Biotechnol. 28, 827–838 (2010). 5. Luo, J. et al. Pharmacogenomics J. 10, 278–291 (2010). 6. Schuster, S.C. Nat. Methods 5, 16–18 (2008).

© 2010 Nature America, Inc. All rights reserved.

Shaking up genome engineering KA Tipton & John Dueber A new method generates genome-scale modified bacteria with unprecedented ease. Systematic approaches to mutate and characterize the function of every gene in a microbe have been hampered by the need to manually create thousands of separate strains through tedious genetic manipulation. In this issue, Warner et al.1 describe an approach to create and characterize rationally modified versions of almost every gene in Escherichia coli. Using this strategy, the authors quickly zero in on genes that influence industrially relevant traits, such as tolerance to toxins in a biofuel feedstock. The method enables single genome modifications to be probed rapidly and comprehensively and correlated to a phenotype, yielding information that lays a foundation for gene mapping and for engineering strains with desired phenotypes. Until now, systematic phenotyping of mutants in yeasts2,3 and E. coli4 has been accomplished by Herculean manual efforts to create thousands of mutant strains, each with a different singlegene knockout. Although the resulting strain collections have proven valuable, it remains a challenge to create, on a genome scale, new collections of mutants for targeted applications or to control gene expression levels using a strong promoter, an inducible promoter or a low-­efficiency ribosome binding site. In contrast, the method of Warner et al.1— trackable multiplex recombineering (TRMR), pronounced ‘tremor’ (Fig. 1)—offers a fast and cheap approach for creating collections of mutants. Impressively, the authors were able to KA Tipton and John Dueber are at the University of California Berkeley, Berkeley, California, USA. e-mail: [email protected]

812

construct libraries containing up- and downregulated versions of 96% of the genes in the E. coli genome in one week at a materials cost of ~$1 per targeted gene. The first step in TRMR is to obtain thousands of 189-base-pair oligonucleotides that target and uniquely identify every E. coli gene. Each of these oligos consists of a barcode tag unique to a gene and regions of homology that

flank the targeted gene in the genome. Warner et al.1 purchased the oligos, which were made on a programmable microarray. Next, using a clever cloning strategy, they appended the oligos to DNA elements that modulate gene expression. Attaching the targeting oligos to the strong PLtetO-1 promoter created a DNA cassette that was expected to upregulate the targeted gene after incorporation into the genome. Conversely, attaching the targeting oligo to a weak ribosome binding site produced a DNA cassette that downregulated the targeted gene. An antibiotic resistance gene allowed selection for the genetic modifications. As a result of the DNA synthesis and manipulation steps, Warner et al.1 created two libraries of linear DNA fragments, each with 4,077 DNA cassettes pooled together in a single tube. These libraries of DNA oligonucleotides were used to modify the E. coli genome by means of recombineering, a homologous recombination– based method in E. coli expressing λ phage recombination factors (λgam, bet and exo)5. Growth on antibiotic medium selects for successful recombinants, and the sites of recombination are determined by homology of the targeting oligos to genomic regions flanking each gene. The resulting collections of modified E. coli strains were then challenged by growth in environmental conditions of interest. Warner et al.1 measured the relative fitness of each

E. coli +

Multiplex oligonucleotide library

E. coli strains with modified gene expression levels

Selection in new environmental conditions

Figure 1 TRMR enables genome-scale selection of rational modifications to the expression of single genes. A multiplex library of oligonucleotides is synthesized to encode a unique barcode tag and regions of homology flanking individual target genes in the E. coli genome (left). A series of cloning steps generates linear DNA fragments that contain sequences necessary for up- or downregulating the expression of each target gene. E. coli are transformed with this library of linear fragments to create a collection of genetically modified strains (middle, green cells containing a modified genetic network). The modifications alter the functional linkages between genes. (Lines in the networks represent linkages, with thickness being the strength of the link. Circles represent genes, with translucency and a dashed outline representing attenuated expression). The E. coli strain collection is grown on medium containing an environmental challenge of interest (right). The identities and relative abundances of individual survivors are determined by sequencing colonies using universal primer sequences. Alternatively, survivors are determined in bulk by microarray analysis of the barcode tags. Importantly, the basic TRMR strategy is amenable to rapid iteration such that the most promising gene modifications are used to seed subsequent cycles of mutation and selection (dotted arrow).

volume 28 number 8 AUGUST 2010 nature biotechnology

ne w s and vie w s sufficiently rich snapshot of disease physiology. In such cases, complementary technologies, which measure mRNA expression, protein levels, genetic mutation, copy number variation, gene silencing or regulatory RNA expression, could be considered. Alternatively, the best technology may vary by tumor type. High-throughput sequencing, in particular, offers advantages over microarrays in that coverage of the genome is less biased and the dynamic range is larger6. With luck, the results of MAQC-II will be useful for ­shepherding

other high-­throughput technologies toward the clinic as well. COMPETING FINANCIAL INTERESTS The author declares no competing financial interests. 1. DeRisi, J.L., Iyer, V.R. & Brown, P.P. Science 278, 680–686 (1997). 2. Dumur, C.I. et al. J. Mol. Diagn. 10, 67–77 (2008). 3. Buyse, M. et al. J. Natl. Cancer Inst. 98, 1183–1192 (2006). 4. The MicroArray Quality Control (MAQC) consortium. Nat. Biotechnol. 28, 827–838 (2010). 5. Luo, J. et al. Pharmacogenomics J. 10, 278–291 (2010). 6. Schuster, S.C. Nat. Methods 5, 16–18 (2008).

© 2010 Nature America, Inc. All rights reserved.

Shaking up genome engineering KA Tipton & John Dueber A new method generates genome-scale modified bacteria with unprecedented ease. Systematic approaches to mutate and characterize the function of every gene in a microbe have been hampered by the need to manually create thousands of separate strains through tedious genetic manipulation. In this issue, Warner et al.1 describe an approach to create and characterize rationally modified versions of almost every gene in Escherichia coli. Using this strategy, the authors quickly zero in on genes that influence industrially relevant traits, such as tolerance to toxins in a biofuel feedstock. The method enables single genome modifications to be probed rapidly and comprehensively and correlated to a phenotype, yielding information that lays a foundation for gene mapping and for engineering strains with desired phenotypes. Until now, systematic phenotyping of mutants in yeasts2,3 and E. coli4 has been accomplished by Herculean manual efforts to create thousands of mutant strains, each with a different singlegene knockout. Although the resulting strain collections have proven valuable, it remains a challenge to create, on a genome scale, new collections of mutants for targeted applications or to control gene expression levels using a strong promoter, an inducible promoter or a low-­efficiency ribosome binding site. In contrast, the method of Warner et al.1— trackable multiplex recombineering (TRMR), pronounced ‘tremor’ (Fig. 1)—offers a fast and cheap approach for creating collections of mutants. Impressively, the authors were able to KA Tipton and John Dueber are at the University of California Berkeley, Berkeley, California, USA. e-mail: [email protected]

812

construct libraries containing up- and downregulated versions of 96% of the genes in the E. coli genome in one week at a materials cost of ~$1 per targeted gene. The first step in TRMR is to obtain thousands of 189-base-pair oligonucleotides that target and uniquely identify every E. coli gene. Each of these oligos consists of a barcode tag unique to a gene and regions of homology that

flank the targeted gene in the genome. Warner et al.1 purchased the oligos, which were made on a programmable microarray. Next, using a clever cloning strategy, they appended the oligos to DNA elements that modulate gene expression. Attaching the targeting oligos to the strong PLtetO-1 promoter created a DNA cassette that was expected to upregulate the targeted gene after incorporation into the genome. Conversely, attaching the targeting oligo to a weak ribosome binding site produced a DNA cassette that downregulated the targeted gene. An antibiotic resistance gene allowed selection for the genetic modifications. As a result of the DNA synthesis and manipulation steps, Warner et al.1 created two libraries of linear DNA fragments, each with 4,077 DNA cassettes pooled together in a single tube. These libraries of DNA oligonucleotides were used to modify the E. coli genome by means of recombineering, a homologous recombination– based method in E. coli expressing λ phage recombination factors (λgam, bet and exo)5. Growth on antibiotic medium selects for successful recombinants, and the sites of recombination are determined by homology of the targeting oligos to genomic regions flanking each gene. The resulting collections of modified E. coli strains were then challenged by growth in environmental conditions of interest. Warner et al.1 measured the relative fitness of each

E. coli +

Multiplex oligonucleotide library

E. coli strains with modified gene expression levels

Selection in new environmental conditions

Figure 1 TRMR enables genome-scale selection of rational modifications to the expression of single genes. A multiplex library of oligonucleotides is synthesized to encode a unique barcode tag and regions of homology flanking individual target genes in the E. coli genome (left). A series of cloning steps generates linear DNA fragments that contain sequences necessary for up- or downregulating the expression of each target gene. E. coli are transformed with this library of linear fragments to create a collection of genetically modified strains (middle, green cells containing a modified genetic network). The modifications alter the functional linkages between genes. (Lines in the networks represent linkages, with thickness being the strength of the link. Circles represent genes, with translucency and a dashed outline representing attenuated expression). The E. coli strain collection is grown on medium containing an environmental challenge of interest (right). The identities and relative abundances of individual survivors are determined by sequencing colonies using universal primer sequences. Alternatively, survivors are determined in bulk by microarray analysis of the barcode tags. Importantly, the basic TRMR strategy is amenable to rapid iteration such that the most promising gene modifications are used to seed subsequent cycles of mutation and selection (dotted arrow).

volume 28 number 8 AUGUST 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

ne w s and vie w s modified strain by isolating genomic DNA, amplifying the barcode tags using PCR and hybridizing the amplified DNA to a microarray that contains probes complementary to each tag. A signal on the microarray identifies strains that grew. To demonstrate the approach, the authors selected for growth in media containing salicin, d-fucose, valine or methylglyoxyl. These compounds inhibit cell growth by ­different mechanisms. Salicin is a carbon source that normally cannot be metabolized. d-fucose is an analogue of arabinose that inhibits the ability of E. coli to metabolize this sugar. Valine acts as a feedback inhibitor of growth-limiting leucine and isoleucine biosynthesis. Methylglyoxal presents an oxidative stress  if present in elevated concentrations. These conditions demonstrated the effectiveness of TRMR in identifying gene-trait relationships and in identifying genes that were not expected to be involved in resistance to the given cellular stress, thus supporting the power of a genome-scale, unbiased approach. In a particularly challenging and exciting application of TRMR, Warner et al.1 grew their libraries of strains in lignocellulosic hydrolysate derived from corn stover. Hydrolysates represent a complex potpourri of molecules toxic to E. coli. It has been difficult to predict a priori which genes would best confer resistance to growth inhibitors in the hydrolysates6. This problem is thus well suited to test the authors’ methods. Among the modified genes that conferred improved growth were genes with expected functions as well as several with ­seemingly disparate cellular functions, ­including primary metabolism, RNA metabolism, sugar transporters, secondary metabolism, vitamin processes and antioxidant activities. In one notable result, the authors identified the antioxidant ahpC, a gene not previously linked to growth on hydrolysates, which, when upregulated, considerably improved both growth rate and final biomass levels. TRMR has many potential uses. Warner et al.1 note that it could easily be applied iteratively, with strains selected after one round of TRMR used as the starting strains for a second round, thereby accumulating beneficial genome alterations (Fig. 1, dotted arrow). Such iterative processing can take advantage of the same pool of oligos already synthesized. Parallel microarray analysis of the barcode tags present in the selected survivors should produce additional layers of information about genetic contributors to fitness. For instance, the ability to track combinations of alterations in a stepwise fashion as they accumulate has the potential to provide snapshots of genetic interaction data that, if taken at a high enough frequency, may uncover network connections

in conditions particularly relevant to industrial and biotechnological settings. TRMR is also valuable because it identifies genes and network connections that could form the basis for further strain optimization. For instance, a particularly powerful combination of technologies would be to first use TRMR to identify relevant genes and then apply the recently developed multiplex automated genome engineering (MAGE) method7, which finely tunes the expression levels of a limited number of genes. In ­microbial ­engineering applications, such as the creation of a strain of E. coli that can metabolize lignocellulose sugars, TRMR should complement existing technologies, including directed evolution, genome-scale metabolic modeling and synthetic biology approaches for redox balancing, flux improvement and limiting the production of undesirable and toxic metabolic products. In addition to TRMR, other approaches based on genome-wide modifications are

increasingly providing scientists with the ability to generate large, information-rich data sets from which new genetic information may be extracted2–4,8,9. TRMR heralds an approach to genetic analyses in which phenotypes are rapidly mapped to genetic modifications across the genome, simultaneously producing improved strains for immediate practical use as well as data sets enabling future rational creation of sophisticated strains. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. 1. Warner, J. et al. Nat. Biotechnol. 28, 856–862 (2010). 2. Giaever, G. et al. Nature 418, 387–391 (2002). 3. Kim, D.U. et al. Nat. Biotechnol. 28, 617–623 (2010). 4. Baba, T. et al. Mol. Syst. Biol. 2, 2006.0008 (2006). 5. Datta, S., Costantino, N. & Court, D.L. Gene 379, 109–115 (2006). 6. Mohagheghi, A. & Schell, D.J. Biotechnol. Bioeng. 105, 992–996 (2010). 7. Wang, H.H. et al. Nature 460, 894–898 (2009). 8. Tong, A.H. et al. Science 294, 2364–2368 (2001). 9. Mnaimneh, S. et al. Cell 118, 31–44 (2004).

The expanding family of dendritic cell subsets Hideki Ueno, A Karolina Palucka & Jacques Banchereau The recent identification of human CD141+ dendritic cells as a counterpart of mouse CD8+ dendritic cells may be useful in developing vaccines and immunotherapies. Dendritic cells (DCs) are central players in the control of immunity and tolerance, and investigation of their properties is expected to illuminate many diseases of the immune system and lead to innovative therapies. Four recent reports1–4 in The Journal of Experimental Medicine mark new progress in our understanding of the biology of a particular human DC subset identified by co-expression of CD141 (thrombomodulin, BDCA-3) and the Hideki Ueno, A. Karolina Palucka and Jacques Banchereau are at the Baylor Institute for Immunology Research and INSERM U899, Dallas, Texas, USA; A. Karolina Palucka is at the Sammons Cancer Center, Baylor University Medical Center, Dallas, Texas, USA; and A. Karolina Palucka and Jacques Banchereau are in the Department of Gene and Cell Medicine and Department of Medicine, Immunology Institute, Mount Sinai School of Medicine, New York, New York, USA. e-mail: [email protected]

nature biotechnology volume 28 number 8 AUGUST 2010

C-type lectin CLEC9A (DNGR-1). Collectively, the papers show that CD141+ DCs are the human counterpart of mouse CD8+ DCs. As mouse CD8+ DCs are important for the induction of cytotoxic T-lymphocyte responses through their exceptional capacity to present exogenous antigens in an HLA class I pathway (so-called cross-presentation)5, this discovery could have significant clinical impact if human CD141+ DCs have a similar role. DCs were discovered in 1973 by Ralph Steinman as a novel cell type in the mouse spleen and are now recognized as a group of related cell populations that efficiently present antigens. Both mice and humans have two major types of DC: myeloid DCs (mDCs, also called conventional or classical DCs), and plasmacytoid DCs (pDCs). pDCs are considered the front line in anti-viral immunity as they rapidly produce abundant type I interferon in response to viral infection. In their resting state, pDCs may be important in tolerance, including oral tolerance6,7. pDCs are 813

© 2010 Nature America, Inc. All rights reserved.

ne w s and vie w s modified strain by isolating genomic DNA, amplifying the barcode tags using PCR and hybridizing the amplified DNA to a microarray that contains probes complementary to each tag. A signal on the microarray identifies strains that grew. To demonstrate the approach, the authors selected for growth in media containing salicin, d-fucose, valine or methylglyoxyl. These compounds inhibit cell growth by ­different mechanisms. Salicin is a carbon source that normally cannot be metabolized. d-fucose is an analogue of arabinose that inhibits the ability of E. coli to metabolize this sugar. Valine acts as a feedback inhibitor of growth-limiting leucine and isoleucine biosynthesis. Methylglyoxal presents an oxidative stress  if present in elevated concentrations. These conditions demonstrated the effectiveness of TRMR in identifying gene-trait relationships and in identifying genes that were not expected to be involved in resistance to the given cellular stress, thus supporting the power of a genome-scale, unbiased approach. In a particularly challenging and exciting application of TRMR, Warner et al.1 grew their libraries of strains in lignocellulosic hydrolysate derived from corn stover. Hydrolysates represent a complex potpourri of molecules toxic to E. coli. It has been difficult to predict a priori which genes would best confer resistance to growth inhibitors in the hydrolysates6. This problem is thus well suited to test the authors’ methods. Among the modified genes that conferred improved growth were genes with expected functions as well as several with ­seemingly disparate cellular functions, ­including primary metabolism, RNA metabolism, sugar transporters, secondary metabolism, vitamin processes and antioxidant activities. In one notable result, the authors identified the antioxidant ahpC, a gene not previously linked to growth on hydrolysates, which, when upregulated, considerably improved both growth rate and final biomass levels. TRMR has many potential uses. Warner et al.1 note that it could easily be applied iteratively, with strains selected after one round of TRMR used as the starting strains for a second round, thereby accumulating beneficial genome alterations (Fig. 1, dotted arrow). Such iterative processing can take advantage of the same pool of oligos already synthesized. Parallel microarray analysis of the barcode tags present in the selected survivors should produce additional layers of information about genetic contributors to fitness. For instance, the ability to track combinations of alterations in a stepwise fashion as they accumulate has the potential to provide snapshots of genetic interaction data that, if taken at a high enough frequency, may uncover network connections

in conditions particularly relevant to industrial and biotechnological settings. TRMR is also valuable because it identifies genes and network connections that could form the basis for further strain optimization. For instance, a particularly powerful combination of technologies would be to first use TRMR to identify relevant genes and then apply the recently developed multiplex automated genome engineering (MAGE) method7, which finely tunes the expression levels of a limited number of genes. In ­microbial ­engineering applications, such as the creation of a strain of E. coli that can metabolize lignocellulose sugars, TRMR should complement existing technologies, including directed evolution, genome-scale metabolic modeling and synthetic biology approaches for redox balancing, flux improvement and limiting the production of undesirable and toxic metabolic products. In addition to TRMR, other approaches based on genome-wide modifications are

increasingly providing scientists with the ability to generate large, information-rich data sets from which new genetic information may be extracted2–4,8,9. TRMR heralds an approach to genetic analyses in which phenotypes are rapidly mapped to genetic modifications across the genome, simultaneously producing improved strains for immediate practical use as well as data sets enabling future rational creation of sophisticated strains. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. 1. Warner, J. et al. Nat. Biotechnol. 28, 856–862 (2010). 2. Giaever, G. et al. Nature 418, 387–391 (2002). 3. Kim, D.U. et al. Nat. Biotechnol. 28, 617–623 (2010). 4. Baba, T. et al. Mol. Syst. Biol. 2, 2006.0008 (2006). 5. Datta, S., Costantino, N. & Court, D.L. Gene 379, 109–115 (2006). 6. Mohagheghi, A. & Schell, D.J. Biotechnol. Bioeng. 105, 992–996 (2010). 7. Wang, H.H. et al. Nature 460, 894–898 (2009). 8. Tong, A.H. et al. Science 294, 2364–2368 (2001). 9. Mnaimneh, S. et al. Cell 118, 31–44 (2004).

The expanding family of dendritic cell subsets Hideki Ueno, A Karolina Palucka & Jacques Banchereau The recent identification of human CD141+ dendritic cells as a counterpart of mouse CD8+ dendritic cells may be useful in developing vaccines and immunotherapies. Dendritic cells (DCs) are central players in the control of immunity and tolerance, and investigation of their properties is expected to illuminate many diseases of the immune system and lead to innovative therapies. Four recent reports1–4 in The Journal of Experimental Medicine mark new progress in our understanding of the biology of a particular human DC subset identified by co-expression of CD141 (thrombomodulin, BDCA-3) and the Hideki Ueno, A. Karolina Palucka and Jacques Banchereau are at the Baylor Institute for Immunology Research and INSERM U899, Dallas, Texas, USA; A. Karolina Palucka is at the Sammons Cancer Center, Baylor University Medical Center, Dallas, Texas, USA; and A. Karolina Palucka and Jacques Banchereau are in the Department of Gene and Cell Medicine and Department of Medicine, Immunology Institute, Mount Sinai School of Medicine, New York, New York, USA. e-mail: [email protected]

nature biotechnology volume 28 number 8 AUGUST 2010

C-type lectin CLEC9A (DNGR-1). Collectively, the papers show that CD141+ DCs are the human counterpart of mouse CD8+ DCs. As mouse CD8+ DCs are important for the induction of cytotoxic T-lymphocyte responses through their exceptional capacity to present exogenous antigens in an HLA class I pathway (so-called cross-presentation)5, this discovery could have significant clinical impact if human CD141+ DCs have a similar role. DCs were discovered in 1973 by Ralph Steinman as a novel cell type in the mouse spleen and are now recognized as a group of related cell populations that efficiently present antigens. Both mice and humans have two major types of DC: myeloid DCs (mDCs, also called conventional or classical DCs), and plasmacytoid DCs (pDCs). pDCs are considered the front line in anti-viral immunity as they rapidly produce abundant type I interferon in response to viral infection. In their resting state, pDCs may be important in tolerance, including oral tolerance6,7. pDCs are 813

ne w s and vie w s CD141+ DCs Langerhans cells

Dermal CD14+ DCs

Antigen crosspresentation

© 2010 Nature America, Inc. All rights reserved.

CTL Th cells

IL-15

IL-12? IL-12

CTLs

Plasma cells

Long-lived memory CD8+ T cells

Tfh cells

Long-lived memory B cells

Protection in vivo Figure 1 Contribution of human myeloid DC subsets to the regulation of adaptive immunity. The humoral and cellular arms of adaptive immunity are regulated by different human mDC subsets. Humoral immunity is preferentially regulated by CD14+ dermal DCs by means of IL-12, which acts directly on B cells and promotes the development of T follicular helper cells (Tfh). Cellular immunity is preferentially regulated by Langerhans cells, possibly through IL-15 and a dedicated subset of CD4 + T cells specialized to help CD8+ T cells (CTL Th cells). Given their capacity to cross-present antigens to CD8+ T cells, CD141+ DCs are likely to be involved in the development of cytotoxic T-lymphocyte responses. CD141+ DCs might also be involved in the development of humoral responses through IL-12 secretion. This hypothesis is supported by mouse in vivo antigen-targeting studies showing that CD8+ DCs, the mouse counterpart of human CD141+ DCs, can induce both cytotoxic T-lymphocyte and humoral responses12,13, although the mechanisms may be different. It will be important to determine whether and how CD141+ DCs are related to Langerhans cells and to dermal DCs, and how these DC subsets shape adaptive immunity.

t­ hemselves composed of at least two subsets with different functional properties8. Similarly, mDCs comprise different subsets with unique localization, phenotype and functions (Fig. 1). In human skin, the epidermis hosts Langerhans cells, whereas the dermis contains CD1a+ DCs and CD14+ DCs. The ­latter DC subset is involved in the generation of humoral immunity, partly through secretion of interleukin (IL)-12, which stimulates the differentiation of activated B cells into plasma cells and also promotes the differentiation of naive CD4+ T cells into T follicular helper cells9,10, a CD4+ T-cell subset that promotes antibody responses. In contrast, Langerhans cells efficiently prime antigen-specific CD8+ T  cells, possibly by means of IL-15 (ref. 9). The functions of the predominant CD1a+ dermal DCs are as yet unknown. Human DCs expressing CD141 were originally found in blood as a subset of mDCs distinct from CD1c+ mDCs11. The new reports1–4 argue that CD141+ DCs are the human counterpart of mouse CD8+ DCs on the basis of results from several different experimental 814

approaches, including detailed functional and phenotypic analysis1,3, as well as the discovery of a chemokine receptor expressed on both cell types2,4. First, like mouse CD8+ DCs, human CD141+ DCs are present in secondary lymphoid organs such as tonsils and spleen1,3. Further studies are needed to determine whether they are also present in tissues. Second, although human CD141+ DCs do not express CD8, they share with mouse CD8+ DCs expression of other surface molecules, including CLEC9A1,3,12,13 and the adhesion molecule, NECL2 (refs. 3,14). NECL2 binds to class I–restricted T cell–associated molecule, a cell-surface protein primarily expressed by natural killer cells, natural killer T cells and activated CD8+ T cells14. Third, human CD141+ DCs uniquely express the chemokine receptor XCR1 (refs. 2,4), in line with the unique expression of XCR1 by mouse CD8+ DCs shown previously. XCR1 expressed in both human and mouse DCs is functional, as the cells migrate in response to the ligand XCL1 (refs. 2,4), a secreted protein

known to be produced by natural killer cells and activated CD8+ T cells. These observations suggest a potential for interactions between human CD141+ DCs/mouse CD8+ DCs and natural killer cells or CD8+ T cells, which might be a mechanism involved in the efficient induction of cytotoxic T lymphocyte responses. For example, interferon (IFN)-γ released by natural killer cells and/or CD8+ T cells might stimulate CD141+ DCs/CD8+ DCs to secrete more IL-12 (refs. 2,4). Fourth, all of the new studies1–4 demonstrate that human CD141+ DCs are highly efficient in inducing CD8+ T-cell responses through their capacity to cross-present exogenous antigens. This evidence suggests that human CD141+ DCs participate in the development of cytotoxic lymphocyte responses in vivo. Fifth, human CD141+ DCs and mouse CD8+ DCs express the transcription factors Batf3 and IRF-8 (refs. 1,3), both of which are strictly required for the development of mouse CD8+ DCs5. In contrast, CD141+ DCs do not express IRF4 (refs. 1,3), a transcription factor required for the development of other mouse spleen CD4+ DCs5. Thus, CD141+ DCs and mouse CD8+ DCs might share a common developmental pathway. Finally, two of the studies1,3 show similarities between human CD141+ DCs and mouse CD8+ DCs in the expression of Toll-like receptors (TLRs). TLRs belong to the family of pattern recognition receptors through which DCs sense microbes and dying cells. Engagement of these receptors by pathogen- and danger-associated molecular patterns expressed by microbes and dying cells triggers DC maturation, a complex series of events that includes expression of new surface molecules, secretion of cytokines and a reduction in antigen capture. Different DC subsets express different sets of pattern recognition receptors, particularly in humans, which provides flexibility in responding to different microbes. Similar to mouse CD8+ DCs, human CD141+ DCs are found to express TLR3 and TLR8, and stimulation with their respective ligands (poly I:C and poly U) induces their maturation and cytokine secretion. In contrast to the relatively limited TLR expression by CD141+ DCs, it is known that CD1c+ DCs, another blood mDC subset, express a wide array, including TLR4, 5 and 7. Whether human CD141+ DCs express other pattern recognition receptors, such as NOD-like receptors and RIG-I-like receptors, has yet to be determined. The identification of the human counterpart of mouse CD8+ DCs opens the possibility of translating to humans knowledge

volume 28 number 8 AUGUST 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

ne w s and vie w s generated in the mouse. There are still many infectious diseases for which no efficient vaccines are available, including AIDS, malaria, hepatitis C infection and tuberculosis. Most of these would benefit from the induction of potent cytotoxic T lymphocytes to eliminate the infected cells. Similarly, strong cytotoxic T-lymphocyte responses would be beneficial in the context of cancer immunotherapy. Thus, it may be possible to exploit CD141+ DCs in the ‘DC-targeting’ vaccination strategy, in which vaccines are generated from recombinant anti-DC antibodies fused to selected antigens15. Studies in mice have shown that targeting antigen to DCs in this manner in  vivo results in potent antigen­specific CD4+ and CD8+ T-cell immunity15, provided adjuvants are co-administered to activate the targeted DCs. Indeed, ­antibodies to CLEC9 allowed targeting of antigen to

mouse CD8+ DCs in vivo, inducing potent cytotoxic T-lymphocyte responses when combined with anti-CD40 administration12 and potent antibody responses even without co-administration of adjuvants13. It should be emphasized, however, that translating mouse immunological data to the clinic is fraught with uncertainty, as 65 million years of independent evolution have produced many nuances that distinguish the human and mouse immune systems16. As one example, other human DCs, such as CD1c+ DCs1,3 and epidermal Langerhans cells9, can also cross-present antigens. Thus, it remains to be determined whether and how human CD141+ mDCs are related to other mDCs subsets and how all the mDC subsets cooperate in shaping adaptive immunity. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests.

nature biotechnology volume 28 number 8 AUGUST 2010

1. Jongbloed, S.L. et al. J. Exp. Med. 207, 1247–1260 (2010). 2. Bachem, A. et al. J. Exp. Med. 207, 1273–1281 (2010). 3. Poulin, L.F. et al. J. Exp. Med. 207, 1261–1271 (2010). 4. Crozat, K. et al. J. Exp. Med. 207, 1283–1292 (2010). 5. Shortman, K. & Heath, W.R. Immunol. Rev. 234, 18–31 (2010). 6. Goubier, A. et al. Immunity 29, 464–475 (2008). 7. Liu, Y.J. Annu. Rev. Immunol. 23, 275–306 (2005). 8. Matsui, T. et al. J. Immunol. 182, 6815–6823 (2009). 9. Klechevsky, E. et al. Immunity 29, 497–510 (2008). 10. Schmitt, N. et al. Immunity 31, 158–169 (2009). 11. Dzionek, A. et al. J. Immunol. 165, 6037–6046 (2000). 12. Sancho, D. et al. J. Clin. Invest. 118, 2098–2110 (2008). 13. Caminschi, I. et al. Blood 112, 3264–3273 (2008). 14. Galibert, L. et al. J. Biol. Chem. 280, 21955–21964 (2005). 15. Bonifaz, L.C. et al. J. Exp. Med. 199, 815–824 (2004). 16. Mestas, J. & Hughes, C.C. J. Immunol. 172, 2731– 2738 (2004).

815

r e s ea r c h h i g h l i g ht s

© 2010 Nature America, Inc. All rights reserved.

Lung on a chip Efforts to mimic the alveolar-capillary interface—the fundamental functional unit of the lung—in cell culture have been frustrated primarily by the challenge of replicating the structural and functional properties of the system while simulating the mechanical changes associated with normal breathing. Huh et al. recreate the behavior of lung tissue in a microfluidic device by lining a thin (10 µm), porous and flexible membrane with human alveolar epithelial cells on one side and human pulmonary microvascular endothelial cells on the other. Application and release of a vacuum to two flanking chambers causes the membrane with its adherent tissue layers to stretch and then relax to its original size, thus recreating the dynamic mechanical distortion of the alveolar-capillary interface caused by breathing. The device reproduces organ-level responses to bacterial infection and inflammatory cytokines, and its use suggests that mechanical strain can promote nanoparticleinduced toxicity. These findings underscore the potential of the chip for evaluating the safety and efficacy of new drugs for lung disorders, or the effects of environmental toxins. (Science 328, 1662–1668, 2010) PH

miRNAs, Dicer and metastasis MicroRNAs (miRNAs) play a key role in the pathogenesis of cancer. Although the overexpression of individual miRNAs is important in numerous tumors, a global downregulation of miRNA levels is a hallmark of cancer. Martello et al. now show that members of the miR-103/107 family suppress the expression of Dicer, the enzyme responsible for the maturation of pre-miRNAs into miRNAs. Levels of miR-103/107 are inversely proportional to Dicer abundance in cancer cell lines and high miR-103/107 expression correlates with metastasis and poor prognosis in breast cancer. In mouse models of breast cancer, nonmetastatic cell lines can be converted to an invasive phenotype by miR-103/107 expression. Therapeutic targeting of the miRNAs with a specific antisense molecule reduces the number of lung metastases, making these miRNAs promising targets for antimetastatic drugs, although no effect on the growth of the primary tumor was observed. The miR-103/107 molecules promote an epithelial-to-mesenchymal transition, a developmental program associated with increased mobility and loss of cell adhesion that is frequently observed in metastatic cancer. (Cell 141, 1195–1207, 2010) ME Written by Kathy Aschheim, Laura DeFrancesco, Markus Elsner, Peter Hare & Craig Mak

816

Fungal histone acetylation inhibitors Targeting fungal histone acetylation may provide a new source of drugs against Candida albicans infections, a particular problem for immunocompromised individuals, research by Wurtele et al. suggests. The authors set out to determine whether a fungal histone acetyltransferase enzyme (RTT109) not found in humans would make a good drug target. The particular modification that the enzyme makes—acetylation of lysine 56 on histone 3 (H3 Lys56)—is found on close to 30% of C. albicans histones, whereas only 1% of human histones bear the mark. Knocking out both copies of RTT109 creates strains with greater sensitivity to certain antifungal agents; repressing the activity of the HST3 deacetylase enzyme led to fungal cell death. The effects were also mirrored by nicotinamide, an inhibitor of NAD-dependent deacetylases. A/J mice, a model particularly sensitive to C. albicans infection, which were injected with an HST3repressed strain of the fungus or an RTT109-deleted strain failed to show signs of infection. Once again, nicotinamide treatment mirrored the effects of HST3 repression, but only in strains with wild-type RTT109, suggesting that nicotinamide, which acts as an anti-inflammatory, exerts its effects on infection through its interaction with the histone deacetylase pathway. Finally, the researchers showed that whereas some fungal pathogens are sensitive in various degrees to nicotinamide, all tested clinical isolates of C. albicans, the fungus with the greatest impact on human health, were sensitive. (Nat. Med. 16, 774–780, 2010) LD

iPS cells from blood As researchers contemplate clinical applications of induced pluripotent stem (iPS) cells, one practical consideration is the accessibility of the donor cells used for reprogramming. So far, most human iPS cells have been derived from fibroblasts collected through skin biopsies, a procedure that requires an incision and stitches. Following three 2009 papers on the reprogramming of human hematopoietic stem/progenitor cells from cord blood or from adults after mobilization by granulocyte colony stimulating factor, three new studies describe iPS cells from unmobilized adult blood cells. All three groups rely on the standard ‘Yamanaka’ reprogramming factors (OCT4, SOX2, KLF4, C-MYC), but Loh et al. and Staerk et al. deliver these with retroviruses, whereas Seki et al. use the nonintegrating Sendai virus. The latter method appears more efficient, allowing iPSCs to be generated from samples as small as 1 ml. Like keratinocytes from plucked hair (Nat. Biotechnol. 26, 1276–1284, 2008), peripheral blood cells may provide a convenient source of iPS cells in a clinical context. (Cell Stem Cell 7, 15–19; 20–24; 11–14, 2010) KA

Antibody therapy for thrombosis Small-molecule therapeutics, such as aspirin and clopidogrel (Plavix), reduce the risk for heart attack and stroke by inhibiting platelets but at the cost of increased risk for excessive bleeding. Tucker et al. demonstrate an alternative strategy in baboons based on reducing platelet counts using neutralizing antibodies. This strategy was tested using a vascular graft model that mimics a damaged blood vessel at risk for thrombosis. Animals with fewer circulating platelets showed less potential for thrombosis in the graft model. Notably, the blood of these animals did not take longer to clot after cutting the animals’ forearm, whereas aspirin treatment led to a statistically significant increase in bleeding time. Tucker et al. reduced platelet counts by treating animals with serum containing polyclonal neutralizing antibodies raised in baboons against thrombopoietin, a hormone essential for platelet production. Drugs that can be safely used to inhibit platelet production will be required before this strategy can be tested in humans. (Sci. Transl. Med. 2, 37ra45, 2010) CM volume 28 number 8 august 2010 nature biotechnology

A n a ly s i s

Discovery and characterization of chromatin states for systematic annotation of the human genome

© 2010 Nature America, Inc. All rights reserved.

Jason Ernst1,2 & Manolis Kellis1,2 A plethora of epigenetic modifications have been described in the human genome and shown to play diverse roles in gene regulation, cellular differentiation and the onset of disease. Although individual modifications have been linked to the activity levels of various genetic functional elements, their combinatorial patterns are still unresolved and their potential for systematic de novo genome annotation remains untapped. Here, we use a multivariate Hidden Markov Model to reveal ‘chromatin states’ in human T cells, based on recurrent and spatially coherent combinations of chromatin marks. We define 51 distinct chromatin states, including promoter-associated, transcription-associated, active intergenic, large-scale repressed and repeat-associated states. Each chromatin state shows specific enrichments in functional annotations, sequence motifs and specific experimentally observed characteristics, suggesting distinct biological roles. This approach provides a complementary functional annotation of the human genome that reveals the genome-wide locations of diverse classes of epigenetic function. The primary DNA sequence of the human genome encodes the genetic information of each cell, but numerous epigenetic modifications can modulate the interpretation of the primary sequence. These modifications contribute to the diversity of phenotypes found across different human cell types, play key roles in the establishment and maintenance of cellular identity during ­development and have been associated with DNA repair, replication and human disease. Post-translational modifications in the tails of histone proteins that package DNA into chromatin constitute perhaps the most versatile type of such epigenetic information. More than a dozen positions of multiple histone proteins can undergo a number of modifications, such as acetylation and mono-, di- or tri-methylation1,2. More than 100 distinct histone modifications have been described, leading to the ‘histone code hypothesis’ that specific combinations of chromatin modifications would encode distinct biological functions3. Others, however, have instead proposed that individual epigenetic marks act in additive ways and the multitude of modifications simply contributes to stability and robustness4. The specific combinations of

1MIT

Computer Science and Artificial Intelligence Laboratory, Cambridge, Massachusetts, USA. 2Broad Institute of MIT and Harvard, Cambridge, Massachusetts, USA. Correspondence should be addressed to M.K. ([email protected]). Published online 25 July 2010; doi:10.1038/nbt.1662

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

epigenetic modifications that are biologically meaningful, and their corresponding functional roles, are still largely unknown. To directly address these questions, we introduce an approach for the de novo discovery of ‘chromatin states’ (Fig. 1, Supplementary Table 1 and Supplementary Fig. 1), or biologically meaningful and spatially coherent combinations of chromatin marks, by performing a systematic genome-wide analysis based on a multivariate Hidden Markov Model (HMM). Multivariate HMMs are graphical probabilistic models that model multiple ‘observed’ inputs as generated by unobserved ‘hidden’ states, using transitions between hidden states to model spatial relationships (Online Methods). Our model captures two types of chromatin information. The frequency with which different chromatin mark combinations are found with each other are captured by a vector of ‘emission’ probabilities associated with each chromatin state (Fig. 2 and Supplementary Figs. 2 and 3) and the frequency with which different chromatin states occur in spatial relationships of each other along the genome are encoded in a ‘transition’ probability vector associated with each state. These spatial relationships capture both the spreading of certain chromatin domains across the genome, as well as the functional ordering of different states such as from intergenic regions to ­promoter regions and transcribed regions (Supplementary Notes and Supplementary Figs. 4–6). Biologically the genomic locations associated with a given chromatin state may correspond to specific types of functional elements, such as transcription start sites (TSS), enhancers, active genes, repressed genes, exons or heterochromatin, which can be inferred solely from the corresponding combinations of chromatin marks in their spatial context, even though no information about these annotations is given to the model as input. We applied our model to the largest data set of chromatin mark information available,  consisting of the genome-wide occupancy data for a set of 38 different histone methylation and acetylation marks and for the histone variant H2AZ, RNA polymerase II (PolII) and CTCF in human CD4 T-cells. The maps were previously obtained using chromatin immunoprecipitation followed by next generation sequencing (ChIP-seq) (Online Methods)5,6. To understand the biological importance of the resulting chromatin states, we undertook a large-scale, systematic data-mining effort, bringing to bear dozens of genomewide data sets including gene annotations, expression information, evolutionary conservation, regulatory motif instances, compositional biases, genome-wide association data, transcription-factor binding, DNaseI hypersensitivity and nuclear lamina maps. This work provides an unbiased and systematic chromatin-driven annotation for every region of the genome at a 200 base pair ­resolution, refining previously described epigenetic states and ­ introducing

817

A n a ly s i s a­ dditional ones. Regardless of whether these chromatin states are causal in directing regulatory processes, or simply ­reinforcing independent regulatory decisions, these annotations should provide a resource for interpreting biological and medical data sets, such as genome-wide association studies for diverse phenotypes and could potentially help to identify new classes of functional elements.

Chromatin marks

© 2010 Nature America, Inc. All rights reserved.

Chromatin states

Chr 7: State 3 State 5 State 7 State 8 State 10 State 11 State 13 State 15 State 16 State 17 State 18 State 19 State 24 State 25 State 26 State 36 State 37 State 38 State 39 State 43 State 44 State 51 H3K14ac H3K23ac H4K12ac H2AK9ac H4K16ac H2AK5ac H4K91ac H3K4ac H2BK20ac H3K18ac H2BK120ac H3K27ac H2BK5ac H2BK12ac H3K36ac H4K5ac H4K8ac H3K9ac PolII CTCF H2AZ H3K4me3 H3K4me2 H3K4me1 H3K9me1 H3K79me3 H3K79me2 H3K79me1 H3K27me1 H2BK5me1 H4K20me1 H3K36me3 H3K36me1 H3R2me1 H3R2me2 H3K27me2 H3K27me3 H4R3me2 H3K9me2 H3K9me3 H4K20me3

116,260 kb

116,270 kb

116,280 kb

116,290 kb

116,300 kb

RESULTS Chromatin states model and comparison to previous work Previous analyses have largely focused on characterizing the marks ­predictive of specific classes of genomic elements defined a priori such as transcribed regions, promoters or putative enhancers, and using the characterization to identify new instances of these classes5–12. 116,310 kb

116,320 kb

116,330 kb

116,340 kb

116,350 kb

116,360 kb

Promoter states

Transcribed states

Active intergenic

Repressed Repetitive CAPZA2

50 kb

Figure 1  Example of chromatin state annotation. Input chromatin mark information and resulting chromatin state annotation for a 120-kb region of human chromosome 7 surrounding the CAPZA2 gene. For each 200-bp interval, the input ChIP-Seq sequence tag count (black bars) is processed into a binary presence and/or absence call for each of 18 acetylation marks (light blue), 20 methylation marks (pink) and CTCF/Pol2/H2AZ (brown). The precise combination of these marks in each interval in their spatial context is used to infer the most probable chromatin state assignment (colored boxes). Although chromatin states were learned independently of any prior genome annotation, they correlate strongly with upstream and downstream promoters (red), 5′-proximal and distal transcribed regions (purple), active intergenic regions (yellow), repressed (gray) and repetitive (blue) regions (state descriptions shown in Supplementary Table 1). This example illustrates that even when the signal coming from chromatin marks is noisy, the resulting chromatin state annotation is very robust, directly interpretable and shows a strong correspondence with the gene annotation. Several spatially coherent transitions are seen from large-scale repressed to active intergenic regions near active genes, from upstream to downstream promoter states surrounding the TSS and from 5′-proximal to distal transcribed regions along the body of the gene. The frequent transitions to state 16 correlate with annotated Alu elements (57% overlap versus 4% and 25% for states 13 and 15, respectively). Transitions to state 13 are likely due to enhancer elements in the first intron of CAPZA2, a region where regulatory elements are commonly found and correlate with several enhancer marks. The maximum-probability state assignments are shown here, and the full posterior probability for each state in this region is shown in Supplementary Figure 1.

818

VOLUME 28  NUMBER 8  AUGUST 2010 nature biotechnology

a n a ly s i s frequency of each mark. Specifically, we make a local call of whether a mark was present in each 200-bp interval, and use a Bernoulli random variable to model the probability of detection of each mark in isolation, and a product of independent probabilities to model the probability of each combination of marks (Online Methods). Our approach has the advantage that the model parameters are directly interpretable as the frequencies of each mark and each mark combination, in contrast to previous approaches for which the biological significance of the parameters corresponding to varying signal intensity levels for each mark is often unclear. Moreover, the ­binarization also makes our

An ­unsupervised (without using prior knowledge) local ­chromatin pattern discovery method13 first demonstrated that many of the patterns previously associated with promoters and enhancers could be discovered de novo, but did not discover patterns associated with broader domains and left the vast majority of the genome unannotated (Supplementary Fig. 7). Unsupervised HMM approaches that modeled chromatin mark signal intensity levels using multivariate normals or nonparametric histograms14–18 have been previously used, but in contrast we use a binarization approach that explicitly models the presence/absence

% Repeat

% Lamina

xF TF binding

% GC

xF CpG island

xF DNaseI

xF Conserved

xF TES

xF 3′ UTR

xF All exons

xF Spliced exons

xF 5′ UTR

xF ZNF gene

% RefSeq gene

Expression level

Percent of TSS

xF TSS exact

Percent of genome

% +-2kb TSS

c

Transcribed states

Promoter states

Promoter upstream high expr; potential enh looping Promoter upstream med expr; potential enh looping Promoter upstream low expr; potential enh looping Repressed promoter TSS low-med expr; most GC rich TSS med expr TSS high expr Transcribed promoter; highest expr, TSS for active genes Transcribed promoter; highest expr, downstream Transcribed promoter; high expr, near TSS Transcribed promoter; high expr, downstream Transcribed 5′ proximal, higher expr, open chr, TF binding Transcribed 5′ proximal, higher expr, open chr Transcribed 5′ proximal, high expr, open chr Transcribed 5′ proximal, high expr Transcribed 5′ proximal, med expr; Alu repeats Transcribed less 5′ proximal, med expr, open chr Transcribed less 5′ proximal, med expr Transcribed less 5′ proximal, lower expr; Alu repeats Candidate strong enhancer in transcribed regions Spliced exons/GC rich; open chr, TF binding Spliced exons/GC rich Spliced exons/GC rich; Alu repeats Transcribed 5′ distal; exons Transcribed further 5′ distal; exons Transcribed 5′ distal; Alu repeats End of transcription; exons; high expr ZNF genes; KAP-1 repressed state

Active intergenic

Cand strong distal enh; higher open chr; higher target expr Cand strong distal enh; high open chr; higher target expr Intergenic H2AZ with open chr/TF binding. Cand. distal enh Candidate weak distal enhancer Candidate distal enhancer Proximal to active enhancers; Alu repeats Active intergenic regions not enhancer specific Active intergenic further from enhancers; Alu repeats Non-repressive intergenic domains; Alu repeats H2AZ specific state CTCF island; candidate insulator

Repressive

Unmappable Heterochr; nuclear lamina; most AT rich Heterochr; nuclear lamina; ERVL repeats Heterochr; lower gene depletion Heterochr; ERVL repeats: lower gene/exon depletion Specific repression Simple repeats (CA)n, (TG)n L1/LTR repeats Satellite repeat Satellite repeat; moderate mapping bias Satellite repeat; high mapping bias Satellite repeat/rRNA; extreme mapping bias

Repetitive

© 2010 Nature America, Inc. All rights reserved.

State

b

State H3K14ac H3K23ac H4K12ac H2AK9ac H4K16ac H2AK5ac H4K91ac H3K4ac H2BK20ac H3K18ac H2BK120ac H3K27ac H2BK5ac H2BK12ac H3K36ac H4K5ac H4K8ac H3K9ac PolII CTCF H2AZ H3K4me3 H3K4me2 H3K4me1 H3K9me1 H3K79me3 H3K79me2 H3K79me1 H3K27me1 H2BK5me1 H4K20me1 H3K36me3 H3K36me1 H3R2me1 H3R2me2 H3K27me2 H3K27me3 H4R3me2 H3K9me2 H3K9me3 H4K20me3

a

Chromatin mark frequency 0.01

0.08

Genome total/average

1

Figure 2  Chromatin state definition and functional interpretation. (a) Chromatin mark combinations associated with each state. Each row shows the specific combination of marks associated with each chromatin state and the frequencies between 0 and 1 with which they occur (color scale). These correspond to the emission probability parameters of the Hidden Markov Model (HMM) learned across the genome during model training (values shown in Supplementary Fig. 2). Marks and states colored as in Figure 1. (b) Genomic and functional enrichments of chromatin states. %, percentage; xF, fold enrichment. In order, columns are: percentage of the genome assigned to the state; percentage of state that overlaps a 200-bp interval within 2 kb of an annotated RefSeq TSS; percentage of RefSeq TSS found in the state; fold enrichment for TSS; percentage of state overlapping a RefSeq transcribed region; average expression level of genomic intervals overlapping the state; fold-enrichment for zinc-finger–named gene; fold-enrichment for RefSeq 5′ Untranslated Region (5′-UTR) exon and introns; fold enrichment for RefSeq exons; fold enrichment for spliced exons (2 nd exon or later); fold enrichment for RefSeq 3′ Untranslated Region (3′-UTR) exons and introns; fold enrichment for RefSeq transcription end sites (TES); fold enrichment for PhastCons conserved elements; fold enrichment for DNaseI hypersensitive sites; median fold enrichment for transcription factor binding sites over a set of experiments (expanded in Supplementary Fig. 23); fold-enrichment for CpG islands; percentage of GC nucleotides; percent overlapping experimental nuclear lamina data; percent overlapping a RepeatMasker element (expanded in Supplementary Fig. 31). All enrichments are based on the posterior probability assignments. Genome total indicates the total percentage of 200 bp interval intersecting the feature or the genome average for expression and percent GC. (c) Brief description of biological state function and interpretation (chr, chromatin; enh, enhancer, full descriptions in Supplementary Table 1).

nature biotechnology VOLUME 28  NUMBER 8  AUGUST 2010

819

A n a ly s i s b

Chromatin state at TSS of corresponding gene

80 60

1.07 (1.0)

0.85 (1.0)

0.54 (1.0)

1.00 (1.0)

Chromatin

1.20 (1.0)

0.48 (1.0)

2.17 (10–7)

1.64 (1.0)

0.85 (1.0)

0.85 (1.0)

Response to DNA damage

1.20 (1.0)

0.35 (1.0)

1.55 (0.07)

2.13 (10–11)

1.97 –4 (10 )

0.84 (1.0)

RNA processing

0.49 (1.0)

0.26 (1.0)

1.31 (1.0)

1.91 –11 (10 )

2.64 –24 (10 )

2.46 –4 (10 )

T-cell activation

0.77 (1.0)

0.88 (1.0)

1.27 (1.0)

0.70 (1.0)

0.79 (1.0)

4.72 (10–7)

Fold enrichment

120 80 40 0 80 60 40 20

TSS centered State 4 State 5 State 6 State 7

Downstream State 8 State 9 State 10 State 11

0 ,6 0

–1

–2

,0 0

0

0

0

2.82 (10–22)

0

1.24 (1.0)

0 160

2, 00

Embryonic development

20

0

1.51 (1.0)

1, 60

1.15 (1.0)

0

1.45 (1.0)

1, 20

1.61 –3 (10 )

0

0.57 (1.0)

80

2.70 –7 (10 )

State 1 State 2 State 3

40

0

Cell cycle phase

Dual peaking

40

8

00

7

–4

6

0

5

00

4

–8

3

,2 0

Gene GO category

–1

a

6 State 23

4

18 22

8

13

State 20

2 0

Distance from transcription start site

10 8

State 21 State 23

6 4 2 0

0 80 0 1, 20 0 1, 60 0 2, 00 0

1, 0 60 3, 0 20 4, 0 80 6, 0 40 8, 0 00 9, 0 6 11 00 ,2 12 00 ,8 14 00 ,4 16 00 ,0 17 00 ,6 19 00 ,2 00

0

0

500

23 20

15

State 21

,0

1,000

21

States 13–23 shown

10

–2

1,500

State 19 16

State 22

Distance from spliced exon start

State 12 State 13 State 14 State 15 State 16 State 17 State 18 State 19 State 20 State 21 State 22 State 23 State 24 State 25 State 26 State 27 State 28

0 80 1, 0 60 0 2, 40 3, 0 20 4, 0 00 0

0 2,000

Fold enrichment

State 25 State 24 State 27 State 28

500

State 27

12

12

1,000

States 12–28 shown

,0 –3 00 ,2 –2 00 ,4 0 –1 0 ,6 00 –8 00

1,500

14

Fold enrichment

14

–4

State 26

40

2,000

States 12–23 shown

0 –1 0 ,6 0 –1 0 ,2 00 –8 00 –4 00

c Number of genes

© 2010 Nature America, Inc. All rights reserved.

Distance from transcription start site States 24–28 shown

Distance from transcription end site

Figure 3  Promoter and transcribed chromatin states show distinct functional and positional enrichments. (a) Distinct Gene Ontology (GO) functional enrichments (fold and corrected P-values) found for genes associated with different promoter states at their TSS. For additional states and GO terms, see Supplementary Figure 29. (b) Distinct positional biases of promoter states with respect to nearest RefSeq TSS distinguish states peaking upstream, only downstream and centered at the TSS. (c) Positional biases of transcribed states with respect to TSS, nearest spliced exon start and transcription end sites (TES). These distinguish 5′-proximal states (12–23, left panel), 5′-distal states (24–28), states strongly enriched for spliced exons (middle panel, see also Supplementary Fig. 24 for plot for states 24–28) and TES-associated states (with state 27 being particularly precisely positioned, right panel).

model less prone to forming states overfitting potentially insignificant variations in signal intensity levels. In contrast to models that use a multivariate normal distribution, our method avoids this strong parametric assumption, which is generally violated by the often relatively small discrete counts found in ChIP-seq experiments, enabling more robust models to be inferred. In comparison to the models previously inferred based on a nonparametric histogram strategy18, our binarization approach uses an order of magnitude fewer parameters per state, further increasing model robustness and interpretability. We developed a procedure for learning sets of chromatin states across a range of model complexities. For a given number of states and from a set of initial parameters, standard expectation maximization based procedures enable simultaneous local optimization of the state definitions (emission and transition probabilities) and the corresponding genome annotation consistent with the observed data. However the model inferred and its quality can depend on the initial set of parameters, which can confound comparing models with different number of states learned from independent initializations. We therefore used a two-stage process that first selected a 79-state model which had the highest complexity-penalized likelihood score across a large compendium of randomly-initialized ­models of ­varying ­complexity.

We then pruned and optimized this model down to smaller numbers of states, leading to a model with 51 states that were relatively consistently recovered across the compendium of models, and that sufficiently captured all states found in larger models for which we could give a distinct biological interpretation (see Online Methods). This enabled us to maintain a relatively small number of states while capturing most of the unique biology uncovered across our compendium of randomly-initialized models. Put in other words, this procedure enabled us to maximize biological interpretability, while minimizing model complexity. We further ensured that general properties of the resulting model validated our approach, including robustness to varying thresholds and different background models, and independence of marks given a chromatin state (Supplementary Notes, Supplementary Figs. 8–21 and Supplementary Table 2). We next describe the likely biological functions of the 51 discovered chromatin states, divided into five large groups.

820

VOLUME 28  NUMBER 8  AUGUST 2010 nature biotechnology

Promoter-associated states The first group of states, states 1–11, all had high enrichment for promoter regions: 40–89% of each state was within 2 kb of a RefSeq TSS, compared with 2.7% genome-wide (P < 10−200, for all states).

© 2010 Nature America, Inc. All rights reserved.

b P value

Promoter states Transcribed states

4.6E-04 3.2E-03

Active intergenic states Repressed states Repetitive states Human mRNAs Spliced ESTs Mammal cons

State 33

IKZF2

IKZF2 rs12619285

6 5 4 5.2E-05 5.8E-04 3.6E-06

These states accounted for 59% of all RefSeq TSS although they covered only 1.3% of genome. These states all had a high frequency of H3K4me3 in ­common, as well as significant enrichments for DNaseI hypersensitive sites, CpG islands, evolutionarily conserved motifs and bound transcription ­factors (Fig. 2). They differed however in the presence and levels of other associated marks, primarily H3K79me2/3, H4K20me1, H3K4me1/2 and H3K9me1, and of numerous acetylations leading to varying strength of the aforementioned functional enrichments, and varying expression levels of the downstream genes (Supplementary Figs. 22 and 23). Promoter states differed in the enrichment of Gene Ontology (GO) terms of associated genes including cell cycle, embryonic development, RNA processing and T-cell activation (Fig. 3a). For instance, the term ‘embryonic development’ is specifically enriched in state 4, whereas the term ‘T-cell activation’ is specifically enriched in state 8. Promoter states also differed in their preferentially enriched positions with respect to the TSS of associated genes (Fig. 3b). States 4–7 were most concentrated over the TSS (showing upwards of 100-fold enrichment), states 8–11 peaked between 400 bp and 1,200 bp downstream of the TSS and corresponded to transcribed promoter regions of expressed genes and states 1–3 peaked both upstream and downstream of the TSS.

–log P

HapMap CEU SNP and GWAS

GWAS

a State

Figure 4  SNP and GWAS enrichments for chromatin states. (a) Several chromatin states show enrichments for disease association data sets. For each state is shown: genome percentage; fold enrichment for SNPs from the HapMap CEU population; fold enrichment from a collection of 1,640 GWAS SNPs associated with a variety of diseases and traits from numerous studies25; fold enrichment of GWAS SNPs relative to the HapMap CEU SNP enrichment; significance of GWAS SNPs relative to the underlying SNP frequency (when the corrected P-value < 0.01). (b) Example of intergenic SNP in GWAS-enriched state 33, found 40 kb downstream of the IKZF2 gene and associated with plasma eosinophil count levels26. SNP significance as reported26 is shown for each SNP in the region (blue circles) and associated chromatin state annotation (similar to Fig. 1). Red circle denotes top SNP and its overlap with state 33. In addition to top SNPs, secondary SNPs were also frequently found at or near GWAS-enriched states in several cases.

Percent genome HapMap CEU SNP

a n a ly s i s

3 2 1 0 213.3

213.4

213.5

213.6

213.7

213.8

Position (Mb)

individual marks in spliced exonic states are also frequently detected in several other states that show only a modest 1.3- to 1.6-fold enrichment for spliced exons (e.g., states 12, 13, 14 and 17). This ­suggests that the chromatin signature of spliced exons is not solely defined by the presence of the previously reported H3K36me3, H2BK5me1, H4K20me1 and H3K79me1 marks, but their specific combinations and the absence of H3K4me2, H3K9me1 and H3K79me2/3. State 27 showed a 12.5-fold enrichment for transcription end sites (TES) with its enrichment peaking directly over these locations (Fig. 3c). It was characterized both by the presence of H3K36me3, PolII and H4K20me1 and the absence of H3K4me1, H3K4me2 and H3K4me3, distinguishing it from other transcribed states with higher PolII or H3K36me3 frequencies. This suggests a distinct signature for 3′ ends of genes for which, to our knowledge, no specific chromatin signature had been described before. This was further validated by a 3.4-fold signal enrichment for the elongating form of PolII surveyed in an independent study22 (Supplementary Fig. 25), even though our input data did not distinguish between the elongating and non-elongating form. State 28 showed a 112-fold enrichment in zinc-finger genes, which comprise 58% of the state. This state was characterized by the high frequency for H3K9me3, H4K20me3 and H3K36me3 and relatively low frequency of other marks. This specific combination has been independently reported as marking regions of KAP1 binding, a zinc-finger– specific co-repressor, which also shows a specific 44-fold enrichment for state 28 (refs. 23,24). Although the association of H3K9me3 and H4K20me3 with zinc-finger genes has been previously reported5, the de novo discovery of this highly specific signature of zinc-finger genes illustrates the utility of the methodology and also reveals the additional presence of H3K36me3 and lower frequency of other marks as complementing the signature of zinc-finger genes.

Transcription-associated states The second large group of chromatin states consisted of 17 ­transcription-associated states. These are 70–95% contained within RefSeq-annotated transcribed regions compared to 36% for the rest of the genome (Fig. 2b, P < 10−200, for all states). This group was not predominantly associated with a single mark, but instead defined by combinations of seven marks, H3K79me3, H3K79me2, H3K79me1, H3K27me1, H2BK5me1, H4K20me1 and H3K36me3 (Fig. 2a). Inspection of the transition frequencies between these states revealed subgroups of states that are associated with 5′-proximal or 5′-distal locations and with different expression levels (Fig. 2c, Supplementary Notes, Supplementary Table 1 and Supplementary Fig. 4). We observed several states strongly enriched for spliced exons (states 21–25 and 27–28 with 5.7- to 9.7-fold enrichments) (Figs. 2b and 3c and Supplementary Fig. 24). Spliced exons were previously reported to be enriched in several individual marks19–21. In contrast to these previous studies, the combinatorial approach we have taken here shows that

Active intergenic states The third broad class of chromatin states consisted of 11 active intergenic states (states 29–39), including several classes of candidate enhancer regions, insulator regions and other regions ­proximal to expressed genes (Supplementary Notes). These states were ­associated with higher frequencies for H3K4me1, H2AZ, numerous acetylation marks and/or CTCF and with lower frequencies for other methylation marks (Fig. 2a and Supplementary Figs. 2

nature biotechnology VOLUME 28  NUMBER 8  AUGUST 2010

821

4

0.4

1

H3K4me3

8

Pol2

6

0.3

H3K9ac

7

0.2

21 20 31

mRNA (%) | not RefSeq

3

45

Individual marks (CD4T) Chromatin states ordered (CD4T cells only) CAGE tags (all cell types) H3K4me3 at varying cutoffs

mRNA (%)

0.5

10 11 2 9

CAGE (%)

0.6 True-positive rate

c

RefSeq gene transcription start sites

0.7

State

a

CAGE (%) | not RefSeq TSS

A n a ly s i s

5

0.1 0

b

0

0.005

0.01

0.015

0.02

0.025

0.5

RefSeq gene transcripts

True-positive rate

0.4

© 2010 Nature America, Inc. All rights reserved.

0.03

False-positive rate

14 2621

28 20 7 6 5 23

4

0.3

0.2

19

8 10 9

27

24 25 11 16

0.1

22 12 18 17

13 15

0

H4K20me1 H3K79me3 H3K36me3 H2BK5me1 H3K79me1 H3K79me2

Individual marks (CD4T cells) Chromatin states ordered (CD4T cells only) Expressed sequence tags (all cell types)

Figure 5  Discovery power of chromatin states for genome annotation. (a) Comparison of the power to discover TSS for individual chromatin marks (red), chromatin states (blue) ordered by their TSS enrichment and a directed experimental approach based on CAGE sequence tag data read counts from all available cell types 36 (gold), whereas the chromatin states and marks use only data from CD4 T-cells. Both chromatin states and CAGE tags are compared using a receiver operating characteristic (ROC) curve that shows the false-positive (x axis) and true-positive (y axis) rates at varying prediction thresholds or increasing numbers of states in the task of predicting if a 200-bp interval intersects a RefSeq TSS. Thin red curve compares performance of H3K4me3 mark at varying intensity thresholds. (b) Comparison of the power to detect RefSeq transcribed regions for chromatin states and marks as in a, and directed experimental information coming from EST data (gold) based on sequence counts from all available cell types 37,38. (c) Independent experimental information provides support that a significant fraction of false positives in a and b are genuine unannotated TSS and transcribed regions currently missing from RefSeq. Percentage of each state supported by a CAGE tag (column 1), and the same percentage for locations at least 2 kb away from a RefSeq TSS (column 2), suggests that many promoter-associated state assignments outside RefSeq promoters are supported by CAGE tag evidence. Similarly, percentage of each state overlapping a GenBank mRNA (column 3), and the same percentage specifically outside RefSeq genes (column 4), suggest that transcriptionassociated state assignments outside RefSeq genes are supported by mRNA evidence. Similar support is found by GenBank ESTs and evolutionarily conserved, predicted new exons (Supplementary Fig. 33).

0

0.005

0.01

0.015 0.02 False-positive rate

0.025

0.03 % Overall

2

2

46

16

for GWAS hits. In contrast, the surrounding region of the genome is assigned to other active or repressed intergenic states with no ­significant GWAS association.

and 3). They occurred primarily away from promoter regions (85–97% outside 2 kb of a TSS) and outside of transcribed genes (48–64% outside of RefSeq annotations, Fig. 2b). When they overlapped gene ­annotations, it was mainly in regions that were repressed or not highly expressed (see expression column in Fig. 2b). States 29–33 were notable as they corresponded to smaller fractions of the genome specifically associated with greater DNaseI hypersensitivity, transcription factor binding and regulatory motif instances and are likely to represent enhancer regions (Fig. 2 and Supplementary Fig. 23). Although these candidate enhancer states all shared higher H3K4me1 frequencies, they showed differences in the expression levels of downstream genes associated with subtle differences in their specific mark combinations (Supplementary Fig. 22). For instance, genes downstream of state 30 had a consistently higher average expression level than genes downstream of state 31 (P < 0.001 at 10 kb, two-sided t-test). The two states differed in the frequency of several acetylation marks (state 30 relative to 31 showed higher frequency for H2BK120ac, H3K27ac and H2BK5ac and lower frequency for H4K5ac, H4K8ac) and also in the level of H2AZ (higher in state 31 than 30), suggesting that these marks may be playing a more complex role than previously thought in enhancer regions. Several active intergenic states showed significant enrichments for genome-wide association study (GWAS) hits (e.g., 3.3-fold for candidate enhancer state 33, Fig. 4a), based on a curated database of top-scoring single-nucleotide polymorphisms (SNPs) in a range of diseases and traits25. These states thus provide a likely common functional role and means of refining many intergenic SNPs even in the absence of other annotations. For example (Fig. 4b), a SNP reported to be strongly associated with plasma eosinophil count levels in inflammatory diseases (rs12619285) 26 and located 40 kb downstream of IKZF2 in an intergenic region devoid of annotations is in a section of the genome in the chromatin state 33, which is enriched

Repetitive states The final group of six states (46–51) showed strong and distinct enrichments for specific repetitive elements (Supplementary Fig. 31). State 46 had a strong enrichment of simple repeats, ­ specifically (CA)n, (TG)n or (CATG)n (44, 45 and 302-fold, respectively), possibly due to sequence biases in ChIP-based experiments30. State 47 was ­characterized specifically by H3K9me3 and enriched for L1 and LTR repeats. State 48–51 all had higher frequencies of H4K20me3 and H3K9me3 and were heavily enriched for satellite repeat ­elements.

822

VOLUME 28  NUMBER 8  AUGUST 2010 nature biotechnology

Large-scale repressed states The next group of states (40–45) marked large-scale repressed and heterochromatic regions, representing 64% of the genome. The two most frequently detected modifications in total for all the states in this group were H3K27me3 and H3K9me3. State 40, covering 13% of the genome, was essentially devoid of any detected modifications, states 41–42 (25% of the genome) had a higher frequency for H3K9me3 than H3K27me3, whereas states 43–45 (26% of the genome) had a higher frequency for H3K27me3. States 41–42 as compared to states 43–45 showed a stronger depletion for genes, promoters and conserved elements and stronger association with nuclear lamina regions27 and the darkest-staining chromosomal bands28. It also had a higher frequency of A/T nucleotides (Fig. 2b and Supplementary Figs. 26–28). State 45 likely corresponds to targeted gene repression. It showed the highest frequency for H3K27me3 and was unique among repressed states to show enrichment for TSS. The corresponding genes were enriched for development-related GO categories (Supplementary Fig. 29), similar to the repressed promoter state 4 marked by H3K4me3. However, in contrast to state 4, state 45 showed almost no change in acetylation levels in response to histone deacetylase ­inhibitor (HDACi) treatment (Supplementary Fig. 30), suggesting that state 4 is poised for activation whereas state 45 is stably repressed29.

a n a ly s i s

Ref. 38

First 10 greedy

State

H3K14ac

H2AK9ac

H4K12ac

b H3K23ac

H4R3me2

H4K16ac

H3K36me1

H3K9ac

H2BK12ac

H3K9me2

H3R2me2

H3K36ac

H4K8ac

H2AK5ac

H3R2me1

H3K4ac

PolII

H3K27me2

H3K79me1

H3K27ac

H3K27me1

H2BK20ac

H3K79me2

H4K5ac

H3K9me1

H4K20me3

H3K27me3

H4K91ac

H2BK5ac

CTCF

H2AZ

H4K20me1

H3K4me1

H3K9me3

H2BK120ac

H3K36me3

H2BK5me1

H3K79me3

H3K4me3

H3K18ac

None

Predictive power for genome annotation We next set out to study the predictive power of chromatin states for the discovery of functional elements. We focused on two classes of elements that benefit from ample experimental information independent of chromatin marks, TSS and transcribed regions. We found that chromatin states consistently outperformed predictions based on individual marks (Fig. 5a,b), emphasizing the importance of using

c 50 45 40 35 Squared error

© 2010 Nature America, Inc. All rights reserved.

State

a

H3K4me2

States 49–51 showed seemingly high frequencies for numerous modifications, but also strong enrichments in sequence reads from a nonspecific antibody (IgG) control 31 (Supplementary Fig. 20), suggesting these enrichments are due to a lack of coverage for the additional copies of these repeat elements in the reference genome assembly32, thus illustrating the ability of our model to capture such potential artifacts by considering all marks jointly.

30 25 20 15 10 5 0 0

2

4

6

8

10

12

14

16

18

20

22

24

26

28

30

32

34

36

38

40

Number of marks

Figure 6  Recovery of chromatin states with subsets of marks. (a) The figure shows the ordering of marks (top, from left to right) based on a greedy forward selection algorithm to optimize a squared error penalty on state misassignments (Online Methods). Conditioned on all the marks to the left having already been profiled, the mark listed is the optimal selection for one additional mark to be profiled based on the target optimization function. Below each mark is the percentage of a state with identical assignments using the subset of marks. (b) Comparison of the percentage of each state recovered between the first ten marks based on the greedy method and the ten marks previously used 33 (Supplementary Fig. 39). The two columns after the state IDs are the proportion of the states recovered using the greedy algorithm and the set previously used 33. (c) The figure shows a progressive decrease in squared error for state misassignment as a function of the number of marks selected based on the greedy algorithm.

nature biotechnology VOLUME 28  NUMBER 8  AUGUST 2010

823

© 2010 Nature America, Inc. All rights reserved.

A n a ly s i s mark combinations and spatial genomic information (Supplementary Notes and Supplementary Fig. 32 for a comparison to k-means clustering and a supervised classifier). The prediction performance of chromatin states based on just CD4 T-cells was similar to that of cap analysis of gene expression (CAGE) tags and expressed sequence tags (ESTs) data, even though these were obtained across many diverse cell types. This was possible because active and inactive states together capture the information about genetic elements across cell type boundaries (Fig. 5 and Supplementary Figs. 33–35). Moreover, based on our 51-state model, we could predict TSS and transcribed regions when applied to occupancy data obtained for a subset of ten chromatin marks in CD36 erythrocyte precursors and CD133 hematopoietic stem cells33 (Supplementary Fig. 36). We also found that chromatin states revealed candidate promoter and transcribed regions not in RefSeq, but further supported by independent experimental evidence. Candidate promoters overlapped with CAGE tags (Fig. 5c) and intergenic PolII (Supplementary Fig. 37), and candidate transcribed regions overlapped GenBank mRNAs (Fig. 5c) and EST data (Supplementary Fig. 33). A number of promoter and transcribed states outside known genes were also strongly enriched for not previously described protein-coding exons predicted using evolutionary comparisons of 29 mammals (Lin and M.K., unpublished data) (Supplementary Fig. 33). We note that some candidate promoters may represent distal enhancers, sharing promoter-associated marks potentially due to looping of enhancer to promoter regions7. Recovery of chromatin states using subsets of marks As the large majority of chromatin states were defined by multiple marks, we next sought to specifically study the contribution of each mark in defining chromatin states. First, we found several notable examples of both additive relationships, such as acetylation marks in promoter regions, and combinatorial relationships, such as methylation marks associated with repressive and repetitive elements (Supplementary Notes and Supplementary Fig. 38). We also ­ evaluated varying subsets of chromatin marks in their ability to distinguish between chromatin states (Supplementary Notes and Supplementary Figs. 39–41). More generally, we sought to provide guidelines for selecting subsets of chromatin marks to survey in new cell types that would be maximally informative. As a proof of principle, we evaluated the recovery power of increasing numbers of marks in a greedy way, that is, selecting the best mark given all previous selected marks, weighing each state equally and ­penalizing mismatches uniformly (see Online Methods), which provided an ­initial unbiased recommendation of marks to survey for a new cell type (Fig. 6). We find that increasing subsets of marks rapidly converge to a fairly accurate annotation of chromatin states (Fig. 6c), providing costefficient recommendations for new cell types. In addition to an overall error score, this analysis provides information on the proportion of each state accurately recovered, and specific pairwise state ­misassignments. Such information could be incorporated in a modified scoring function to provide chromatin mark recommendations targeted to the subset of chromatin states that are of particular biological interest, or the ­particular state distinctions that are most important to each study.

genes. The definition of the states themselves revealed numerous insights into the combinatorial and additive roles of chromatin marks, sometimes hinting at combinations of chromatin marks that were not previously described, and the genome-wide annotation of these states exposed many previously unannotated candidate functional elements. We expect the usefulness of the methods presented here will increase as additional genome-wide epigenetic data sets become available, and as additional cell types are surveyed systematically. Chromatin states can be inferred with virtually any type of epigenetic and related information, including histone variants, DNA methylation, DNaseI hypersensitivity and binding of chromatin-associated and sequence-specific transcription factors. Although we focused on a single human cell type, the methods are generally applicable to any species and any number of cell types and even whole embryos, albeit in mixed cell populations mutually exclusive marks found in different subsets of cells could potentially be interpreted as co-occurring. Specifically for understanding epigenomic dynamics, chromatin states can play a central role going forward, as they provide a uniform language for interpreting and comparing diverse epigenetic data sets, for selecting and prioritizing chromatin marks for additional cell types and for summarizing complex relationships of dozens of marks in directly-interpretable chromatin states. As several largescale data production efforts are currently underway to map the epigenomes of many more cell types, exemplified by the ENCODE34, modENCODE35 and Epigenome Roadmap projects (http://www.­ roadmapepigenomics.org/), chromatin states will likely play a key role in the understanding of the human epigenome and its role in development, health and disease. Methods Methods and any associated references are available in the online ­version of the paper at http://www.nature.com/naturebiotechnology/. Note: Supplementary information is available on the Nature Biotechnology website. Acknowledgments We thank P. Kheradpour for regulatory motif instances and M.F. Lin for predicted new exons. We thank M. Garber, A. Siepel, K. Lindblad-Toh, and E. Lander for use of comparative information on 29 mammals. We thank B. Bernstein, N. Shoresh, C. Epstein and T. Mikkelsen for helpful discussions. We thank L. Goff, C. Bristow, R. Sealfon and all members of the MIT CompBio Group for comments, feedback and support. This material is based upon work supported by the National Science Foundation under award no. 0905968 and funding from the US National Human Genome Research Institute (NHGRI) under awards U54-HG004570 and RC1-HG005334. AUTHOR CONTRIBUTIONS J.E. and M.K. developed the method, analyzed results and wrote the paper. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/.

DISCUSSION The discovery and systematic characterization of chromatin states presented here reveals a diverse epigenomic landscape with 51 functionally distinct chromatin states. Although the exact number of chromatin states can vary based on the number of chromatin marks surveyed and the desired resolution at which state differences are studied, our results suggest that the genome annotation resulting from these states can extend the interpretable part of the human genome, especially outside protein-coding

1. Bernstein, B.E., Meissner, A. & Lander, E.S. The mammalian epigenome. Cell 128, 669–681 (2007). 2. Kouzarides, T. Chromatin modifications and their function. Cell 128, 693–705 (2007). 3. Strahl, B.D. & Allis, C.D. The language of covalent histone modifications. Nature 403, 41–45 (2000). 4. Schreiber, S.L. & Bernstein, B.E. Signaling network model of chromatin. Cell 111, 771–778 (2002). 5. Barski, A. et al. High-resolution profiling of histone methylations in the human genome. Cell 129, 823–837 (2007). 6. Wang, Z. et al. Combinatorial patterns of histone acetylations and methylations in the human genome. Nat. Genet. 40, 897–903 (2008).

824

VOLUME 28  NUMBER 8  AUGUST 2010 nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

a n a ly s i s 7. Heintzman, N.D. et al. Distinct and predictive chromatin signatures of transcriptional promoters and enhancers in the human genome. Nat. Genet. 39, 311–318 (2007). 8. Heintzman, N.D. et al. Histone modifications at human enhancers reflect global cell-type-specific gene expression. Nature 459, 108–112 (2009). 9. Guttman, M. et al. Chromatin signature reveals over a thousand highly conserved large non-coding RNAs in mammals. Nature 458, 223–227 (2009). 10. Hon, G., Wang, W. & Ren, B. Discovery and annotation of functional chromatin signatures in the human genome. PLoS Comput. Biol. 5, e1000566 (2009). 11. Wang, X., Xuan, Z., Zhao, X., Li, Y. & Zhang, M.Q. High-resolution human corepromoter prediction with CoreBoost_HM. Genome Res. 19, 266–275 (2009). 12. Won, K.J., Chepelev, I., Ren, B. & Wang, W. Prediction of regulatory elements in mammalian genomes using chromatin signatures. BMC Bioinformatics 9, 547 (2008). 13. Hon, G., Ren, B. & Wang, W. ChromaSig: a probabilistic approach to finding common chromatin signatures in the human genome. PLOS Comput. Biol. 4, e1000201 (2008). 14. Day, N., Hemmaplardh, A., Thurman, R.E., Stamatoyannopoulos, J.A. & Noble, W.S. Unsupervised segmentation of continuous genomic data. Bioinformatics 23, 1424–1426 (2007). 15. Jia, L. et al. Functional enhancers at the gene-poor 8q24 cancer-linked locus. PLoS Genet. 5, e1000597 (2009). 16. Thurman, R.E., Day, N., Noble, W.S. & Stamatoyannopoulos, J.A. Identification of higher-order functional domains in the human ENCODE regions. Genome Res. 17, 917 (2007). 17. Schuettengruber, B. et al. Functional anatomy of polycomb and trithorax chromatin landscapes in Drosophila embryos. PLoS Biol. 7, e13 (2009). 18. Jaschek, R. & Tanay, A. Spatial clustering of multivariate genomic and epigenomic information. in Proceedings of the 13th Annual International Conference on Research in Computational Molecular Biology (ed. Batzoglou, S.) 170–183 (Springer, 2009). 19. Schwartz, S., Meshorer, E. & Ast, G. Chromatin organization marks exon-intron structure. Nat. Struct. Mol. Biol. 16, 990–995 (2009). 20. Kolasinska-Zwierz, P. et al. Differential chromatin marking of introns and expressed exons by H3K36me3. Nat. Genet. 41, 376–381 (2009). 21. Andersson, R., Enroth, S., Rada-Iglesias, A., Wadelius, C. & Komorowski, J. Nucleosomes are well positioned in exons and carry characteristic histone modifications. Genome Res. 19, 1732–1741 (2009). 22. Schones, D.E. et al. Dynamic regulation of nucleosome positioning in the human genome. Cell. 132, 878–898 (2008).

23. Sripathy, S.P., Stevens, J. & Schultz, D.C. The KAP1 corepressor functions to coordinate the assembly of de novo HP1-demarcated microenvironments of heterochromatin required for KRAB zinc finger protein-mediated transcriptional repression. Mol. Cell. Biol. 26, 8623–8638 (2006). 24. O’Geen, H. et al. Genome-wide analysis of KAP1 binding suggests autoregulation of KRAB-ZNFs. PLoS Genet. 3, e89 (2007). 25. Hindorff, L.A., Junkins, H.A., Mehta, J.P. & Manolio, T.A. A catalog of published genome-wide association studies. accessed July 22, 2009. 26. Gudbjartsson, D.F. et al. Sequence variants affecting eosinophil numbers associate with asthma and myocardial infarction. Nat. Genet. 41, 342–347 (2009). 27. Guelen, L. et al. Domain organization of human chromosomes revealed by mapping of nuclear lamina interactions. Nature 453, 948–951 (2008). 28. Furey, T.S. & Haussler, D. Integration of the cytogenetic map with the draft human genome sequence. Hum. Mol. Genet. 12, 1037–1044 (2003). 29. Wang, Z. et al. Genome-wide mapping of HATs and HDACs reveals distinct functions in active and inactive genes. Cell 138, 1019–1031 (2009). 30. Johnson, D.S. et al. Systematic evaluation of variability in ChIP-chip experiments using predefined DNA targets. Genome Res. 18, 393–403 (2008). 31. Zang, C. et al. A clustering approach for identification of enriched domains from histone modification ChIP-Seq data. Bioinformatics 25, 1952–1958 (2009). 32. Zhang, Y., Shin, H., Song, J.S., Lei, Y. & Liu, X.S. Identifying positioned nucleosomes with epigenetic marks in human from ChIP-Seq. BMC Genomics 9, 537 (2008). 33. Cui, K. et al. Chromatin signatures in multipotent human hematopoietic stem cells indicate the fate of bivalent genes during differentiation. Cell Stem Cell 4, 80–93 (2009). 34. ENCODE Project Consortium. Identification and analysis of functional elements in 1% of the human genome by the ENCODE pilot project. Nature 447, 799–816 (2007). 35. Celniker, S.E. et al. Unlocking the secrets of the genome. Nature 459, 927–930 (2009). 36. Carninci, P. et al. Genome-wide analysis of mammalian promoter architecture and evolution. Nat. Genet. 38, 626–635 (2006). 37. Karolchik, D. et al. The UCSC Genome Browser Database: 2008 update. Nucleic Acids Res. 36, D773–D779 (2008). 38. Benson, D.A., Karsch-Mizrachi, I., Lipman, D.J., Ostell, J. & Wheeler, D.L. GenBank: update. Nucleic Acids Res. 32, D23–D26 (2004).

nature biotechnology VOLUME 28  NUMBER 8  AUGUST 2010

825

ONLINE METHODS

© 2010 Nature America, Inc. All rights reserved.

Input data for modeling. The initial unprocessed data were bed files containing the genomic coordinates and strand orientation of mapped sequence reads from ChIP-seq experiments5,6. There was a separate bed file for each of the 18 acetylations, 20 methylations, H2AZ, CTCF and PolII in CD4 T cells. We used the updated version of the H3K79me1/2/3 data, as reported6, which differs from the version first reported5. To apply the model we first divided the genome into 200-base-pair nonoverlapping intervals within which we independently made a call as to whether each of the 41 marks was detected as being present or not based on the count of tags mapping to the interval. Each tag was uniquely assigned to one interval based on the location of the 5′ end of the tag after applying a shift of 100 bases in the 5′ to 3′ direction of the tag. The threshold, t, for each mark was based on the total number of mapped reads for the mark (Supplementary Table 2), and was set to be the smallest integer t such that P(X>t)<10−4 where X is a random variable with a Poisson distribution with mean parameter set to the empirical mean of the number of tags per interval. The probabilistic model. The probabilistic model is based on a multivariate instance of a Hidden Markov Model (HMM)39. The model assumes a fixed number of hidden states K. In each hidden state, the emission distribution, that is the probability distribution over each combination of marks, is modeled with a product of independent Bernoulli random variables. Formally, for each of the K states, and M = 41 input marks, there is an emission parameter pk,m denoting the probability in state k (k = 1,…,K) that input mark m (m = 1,…,M) has a present call. Let c ∈ C denote a chromosome where C is the set of all chromo­ somes. Let ct denote an interval on chromosome c where t = 1,…,Tc corresponds sequentially to the 200 bp intervals on chromosome c. c1 is the interval corresponding to base pairs 1–200 on chromosome c and Tc is the number of nonoverlapping 200 bp intervals on chromosome c. Let vct , m be ‘1’ if there is a present call for input mark m and ‘0’ otherwise at location ct. Denote the specific combination of marks at interval ct as vct = (vct ,1 ,..., vct , m ). Let bij denote the probability of transitioning from state i to j where i = 1,…,K and j = 1,…,K. We also have parameters ai (i = 1,…,K), which denote the probability that the state of the first interval on the chromosome is i. Let sc ∈ SC be an unobserved state sequence through chromosome c and SC be the set of all possible state sequences. Let sct denote the unobserved state on chromosome c at location t for state sequence sc. The full likelihood of all of the observed data v for the parameters a, b and p can then be expressed as: P (v | a, b, p) =

 TC asc  ∏ bsc , sc t −1 t 1  c ∈C sc ∈Sc t =2

∏ ∑

 TC M vc ,m  ∏ ∏ psc t ,m (1 − psc , m  t t t =1 m =1

argmin r

∑

min d(e, c)

e ∈E c ∈Cn \r

where here we used (1 – ρ) where ρ is the standard correlation coefficient as the distance d, though the method is general and could be used with other distance measures. The entire procedure discovered models with comparable or superior likelihood scores to randomly initialized models, while also having sets of parameters that would be more directly comparable (Supplementary Figs. 8 and 13). The number of states for a model to analyze can then be selected by choosing the model trained from a nested initialization with the smallest number of states that sufficiently captures all states of offering distinct biological interpretations, which in our case was a 51-state model based on the (1 − vc , m ) t recovery of the end of transcription state (State 27) (Supplementary Notes). )

 Model learning. We first used an iterative learning expectation-maximization approach to infer state emission and transition parameters that best summarize the observed genome-wide chromatin mark information using a fixed number of randomly-initialized hidden states, varying from 2 up to 80 states. To minimize the number of states and facilitate recovery of a robust and comparable set of states across models of varying complexity, we then used a nested initialization procedure that seeded parameters of lower-complexity models with states of higher-complexity models. From an initial set of parameters we found a local optimum of the parameter values using a variant of the standard expectation-maximization based BaumWelch algorithm for training HMMs39. Our variant after the first full iteration over all the chromosomes, used an incremental expectation-maximization procedure40, which would update the parameters through a maximization step after performing an expectation over any chromosome. This allowed improved parameter estimates from the maximization step to be more quickly incorporated in the more time consuming expectation step. Also for computational considerations, if a transition parameter fell below 10−10 during training we set the parameter value to 0, which allowed faster training with virtually no

nature biotechnology

impact on the final model learned. The transitions were initialized to be fully connected, and except for the 10−10 criterion there was no regularization forcing them closer to 0. We would terminate the training after 300 passes over all the chromosomes, which was sufficient for the likelihood to demonstrate convergence (Supplementary Fig. 8). The procedure for determining the initial parameters used to learn the final set of HMMs was to first learn in parallel for every number of states from 2 to 80 three HMM models based on three different random initializations of the parameters. Each model was scored based on the log likelihood of the model minus a penalization on the model complexity determined by the Bayesian Information Criterion (BIC) of one-half the number of parameters times the natural log of the number of intervals. We then selected the model with the best BIC score among these 237 models, which had 79 states (Supplementary Figs. 8 and 12). We then iteratively removed states from this 79-state model. When removing a state the emission probabilities would be removed entirely, and any state that transitioned to it would have that transition probability ­uniformly redistributed to all the remaining states. This resulting set of ­models was then used as the initial parameters of the HMMs in the final model ­learning. During this final model learning, one HMM was learned for every number of states between 2 and 79 in parallel (Supplementary Fig. 13). The criterion for selecting a state to remove from a model was based on first forming a set E containing all the emission vectors from all the 237 models learned from the random initializations. The procedure would then remove a state such that the elements in E had in total the least distance from their closest emission vector among the remaining states. Formally for a set of emission vectors Cn corresponding to states in a model the method would form a set Cn-1 and corresponding model by removing r defined by

Associating genomic locations with states. After a model is learned, a posterior probability distribution over the state of each interval is computed using a forward-backward algorithm39. Unless otherwise noted, the analysis was based on the ‘soft’ state assignments of the posterior distribution. We also formed hard assignments of states to locations by using the maximum-posterior state assignment at a location. Both the full posterior and hard assignments are available on the website http://compbio.mit.edu/ChromatinStates/. External fold enrichments and percentage overlaps. For a state the sum of the posterior probability over all 200 bp intervals was computed, denoted by a. For an external data source the total number of 200 bp intervals that it intersects in at least one base was computed, denoted by b. For the state and the external data source the total sum of the posterior for the state in intervals intersecting the external data source were computed, denoted by c. Also the total number of 200 bp intervals is denoted by d. The percentage of a state’s overlap with an external data source is defined as (c/a*100) whereas the fold enrichment is (c/a)/(b/d). P-values of the overlap were computed based on the hypergeometric distribution. The gene annotations used were the RefSeq annotations41 as of December 14, 2008 obtained from the UCSC genome browser browser37 and are based on hg18.

doi:10.1038/nbt.1662

© 2010 Nature America, Inc. All rights reserved.

The sequence data for computed nucleotide frequencies, CpG islands, repeats42 and conservation data were also obtained from the UCSC genome browser. The conservation data were based on PhastCon conserved elements using the 44-way vertebrate alignment43,44 (Lindblad-Toh, K. et al., Broad Institute, unpublished data ). Transcription factor binding enrichments were computed for 18 experiments from numerous publications (Supplementary Fig. 23), the median enrichment over all these experiments is reported in Figure 2b. The DNaseI hypersensitivity data was as described45 obtained from the UCSC genome browser. The nuclear lamina data of human fibroblasts was obtained from ref. 27. The zinc-finger genes were defined as those that had ‘ZNF’ at the beginning of the gene symbol in the RefSeq gene table. For published coordinates that were in hg17 we converted them to hg18 using the liftover tool from the UCSC genome browser46. Expression, motif and gene ontology analyses. We obtained the processed CD4 T expression data from ref. 47 for both replicates. We then averaged the two replicates. After averaging the two replicates we performed a natural log transform of the average values. We then standardized all values by subtracting the mean log transformed value and then dividing by the s.d. of the log transform values. The genome coordinates of each probe set were obtained from the UCSC genome browser. Each 200 bp interval that overlapped a probe set obtained the transformed expression score. If multiple probe sets overlapped the same 200 bp then the average of the expression values associated with these were taken. We generated transcription factor motif enrichments as described48, extended for position-weight matrices (PWMs) (Kheradpour, P., MIT, and M.K., unpublished data) based on the hard state assignments. Gene ontology enrichments were based on the hard state assignment of the interval containing the RefSeq annotated TSS of the gene. Enrichments were computed using the STEM software (v.1.3.4) and the Bonferroni corrected P-values are reported49. SNP and GWAS analysis. The HapMap CEU50 data were downloaded from the UCSC genome browser. Significant GWAS hits were taken from ref. 25. SNPs listed as occurring multiple times were only counted once, and for the SNP set listed as a 17-marker haplotype only the first SNP was used ­giving 1,640 SNPs. In computing enrichment for HapMap and GWAS SNPs, if two SNPs mapped to the same interval, we counted them multiple times. To determine if the number of GWAS SNPs in a chromatin state was more significant than would be expected based on the general SNP frequency in the state we used a binomial distribution where n = 1,640 and p is the proportion of HapMap CEU SNPs assigned to the state. We applied a Bonferroni correction for testing multiple states and only reported those P-values significantly enriched with P < 0.01. RefSeq TSS and gene transcripts discovery. The ROC curve for the CAGE data was based on the number of CAGE tags mapping to a 200 bp interval retrieved from the Fantom database and converted from hg17 to hg18 using the UCSC genome browser liftover tool36. The overlap with EST was based on those EST listed in the UCSC genome browser all_est table as of November, 29, 2009 (refs. 37,38). The overlap with GenBank mRNA is based on the overlap with the UCSC genome browser mRNA listed in the table as of October 31, 2009 (refs. 37,38). The novel exon predictions are from (Lin, M.F., MIT, and M.K., unpublished data). Mark subset evaluation and selection. When evaluating the coverage of a specified subset of marks, first a posterior distribution over the states at

doi:10.1038/nbt.1662

each interval is computed using the model learned on the full set of marks, except that the marks not in the subset are omitted when computing emission probabilities. For an interval t we define here st,k and ft,k to be the posterior assignment to state k at interval t based on the subset and full set of marks, respectively. The proportion of state k recovered with a subset of marks is defined as: ck =

∑ t min( ft ,k , st ,k ) ∑ t ft , k

where the sum is over all intervals t in the genome. The ordering of marks presented without any prior biological knowledge was based on a greedy forward selection algorithm designed to select marks that would minimize this function:

∑ (1 − ck )2 k

where the sum is over all states. At each step the algorithm would then choose the one additional mark, conditioned on all the other previously selected marks that would cause this function to be minimized. We note that this target function considers all nonidentical state assignments to have equal loss. An extension of this approach would be to apply target functions that weigh different misassignments differently. The proportion of state k with the full set of marks that is misassigned to state i using a subset of marks, mk,i, as is presented in Supplementary Figures 39 and 40, is defined as: 

∑ t  max( ft , k mk , i =



 max(s − f , 0)   t ,i t ,i  − st , k , 0)   ∑ j max(st , j − ft , j , 0)   

∑ t ft , k

The first term in the sum in the numerator represents for an interval t the amount of posterior probability assigned to state k using the full set of marks not assigned using the subset of marks. The second term represents the portion of this posterior probability that will be credited to state i. The portion credited to state i is the proportion of the surplus posterior state i received with the subset of marks in the interval relative to the total surplus posterior all states received in the interval. 39. Durbin, R., Eddy, S., Krogh, A. & Mitchison, G. Biological Sequence Analysis (Cambridge Univ. Press, 1998). 40. Neal, R.M. & Hinton, G.E. A view of the EM algorithm that justifies incremental, sparse, and other variants. Learn. Graph. Models 89, 355–368 (1998). 41. Pruitt, K.D., Tatusova, T. & Maglott, D.R. NCBI reference sequences (RefSeq): a curated non-redundant sequence database of genomes, transcripts and proteins. Nucleic Acids Res. 35, D61–D65 (2007). 42. Smit, A., Hubley, R. & Green, P. RepeatMasker Open-3.0 1996-2010 . 43. Miller, W. et al. 28-way vertebrate alignment and conservation track in the UCSC Genome Browser. Genome Res. 17, 1797–1808 (2007). 44. Siepel, A. et al. Evolutionarily conserved elements in vertebrate, insect, worm, and yeast genomes. Genome Res. 15, 1034–1050 (2005). 45. Boyle, A.P. et al. High-resolution mapping and characterization of open chromatin across the genome. Cell 132, 311–322 (2008). 46. Kent, W.J. et al. The human genome browser at UCSC. Genome Res. 12, 996–1006 (2002). 47. Su, A.I. et al. A gene atlas of the mouse and human protein-encoding transcriptomes. Proc. Natl. Acad. Sci. USA 101, 6062–6067 (2004). 48. Kheradpour, P., Stark, A., Roy, S. & Kellis, M. Reliable prediction of regulator targets using 12 Drosophila genomes. Genome Res. 17, 1919–1931 (2007). 49. Ernst, J. & Bar-Joseph, Z. STEM: a tool for the analysis of short time series gene expression data. BMC Bioinformatics 7, 191 (2006). 50. International HapMap Consortium. A second generation human haplotype map of over 3.1 million SNPs. Nature 449, 851–861 (2007).

nature biotechnology

Articles

The MicroArray Quality Control (MAQC)-II study of common practices for the development and validation of microarray-based predictive models © 2010 Nature America, Inc. All rights reserved.

MAQC Consortium* Gene expression data from microarrays are being applied to predict preclinical and clinical endpoints, but the reliability of these predictions has not been established. In the MAQC-II project, 36 independent teams analyzed six microarray data sets to generate predictive models for classifying a sample with respect to one of 13 endpoints indicative of lung or liver toxicity in rodents, or of breast cancer, multiple myeloma or neuroblastoma in humans. In total, >30,000 models were built using many combinations of analytical methods. The teams generated predictive models without knowing the biological meaning of some of the endpoints and, to mimic clinical reality, tested the models on data that had not been used for training. We found that model performance depended largely on the endpoint and team proficiency and that different approaches generated models of similar performance. The conclusions and recommendations from MAQC-II should be useful for regulatory agencies, study committees and independent investigators that evaluate methods for global gene expression analysis. As part of the United States Food and Drug Administration’s (FDA’s) Critical Path Initiative to medical product development (http://www. fda.gov/oc/initiatives/criticalpath/), the MAQC consortium began in February 2005 with the goal of addressing various microarray reliability concerns raised in publications1–9 pertaining to reproducibility of gene signatures. The first phase of this project (MAQC-I) extensively evaluated the technical performance of microarray platforms in identifying all differentially expressed genes that would potentially constitute biomarkers. The MAQC-I found high intra-platform reproducibility across test sites, as well as inter-platform concordance of differentially expressed gene lists10–15 and confirmed that microarray technology is able to reliably identify differentially expressed genes between sample classes or populations16,17. Importantly, the MAQC-I helped produce companion guidance regarding genomic data submission to the FDA (http://www.fda.gov/downloads/Drugs/GuidanceCo mplianceRegulatoryInformation/Guidances/ucm079855.pdf). Although the MAQC-I focused on the technical aspects of gene expression measurements, robust technology platforms alone are not sufficient to fully realize the promise of this technology. An additional requirement is the development of accurate and reproducible multivariate gene expression–based prediction models, also referred to as classifiers. Such models take gene expression data from a patient as input and as output produce a prediction of a clinically relevant outcome for that patient. Therefore, the second phase of the project (MAQC-II) has focused on these predictive models18, studying both how they are developed and how they are evaluated. For any given microarray data set, many computational approaches can be followed to develop predictive models and to estimate the future performance of these models. Understanding the strengths and limitations of these various approaches is critical to the ­ formulation *A

of guidelines for safe and effective use of preclinical and clinical genomic data. Although previous studies have compared and benchmarked individual steps in the model development process19, no prior published work has, to our knowledge, extensively evaluated current community practices on the development and validation of microarray-based predictive models. Microarray-based gene expression data and prediction models are increasingly being submitted by the regulated industry to the FDA to support medical product development and testing applications20. For example, gene expression microarray–based assays that have been approved by the FDA as diagnostic tests include the Agendia MammaPrint microarray to assess prognosis of distant ­metastasis in breast cancer patients21,22 and the Pathwork Tissue of Origin Test to assess the degree of similarity of the RNA expression ­pattern in a patient’s tumor to that in a database of tumor samples for which the origin of the tumor is known23. Gene expression data have also been the basis for the development of PCR-based diagnostic assays, including the xDx Allomap test for detection of rejection of heart transplants24. The possible uses of gene expression data are vast and include diagnosis, early detection (screening), monitoring of disease progression, risk assessment, prognosis, complex medical product characterization and prediction of response to treatment (with regard to safety or efficacy) with a drug or device labeling intent. The ability to generate models in a reproducible fashion is an important consideration in predictive model development. A lack of consistency in generating classifiers from publicly available data is problematic and may be due to any number of factors including insufficient annotation, incomplete clinical identifiers, coding errors and/or inappropriate use of methodology25,26. There

full list of authors and affiliations appears at the end of the paper. Correspondence should be addressed to L.S. ([email protected] or [email protected]).

Received 2 March; accepted 30 June; published online 30 July 2010; doi:10.1038/nbt.1665

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

827

© 2010 Nature America, Inc. All rights reserved.

Articles

RESULTS Generating a unique compendium of >30,000 prediction models The MAQC-II consortium was conceived with the primary goal of examining model development practices for generating binary classifiers in two types of data sets, preclinical and clinical (Supplementary Tables 1 and 2). To accomplish this, the project leader distributed six data sets containing 13 preclinical and clinical endpoints coded A through M (Table 1) to 36 voluntary participating data analysis teams representing academia, industry

and government institutions (Supplementary Table 3). Endpoints were coded so as to hide the identities of two negative-control endpoints (endpoints I and M, for which class labels were randomly assigned and are not predictable by the microarray data) and two ­p ositive-control endpoints (endpoints H and L, representing the sex of patients, which is highly predictable by the microarray data). Endpoints A, B and C tested teams’ ability to predict the toxicity of chemical agents in rodent lung and liver models. The remaining endpoints were predicted from microarray data sets from human patients diagnosed with breast cancer (D and E), multiple myeloma (F and G) or neuro­blastoma (J and K). For the multiple myeloma and neuroblastoma data sets, the endpoints represented event free survival (abbreviated EFS), meaning a lack of malignancy or disease recurrence, and overall survival (abbreviated OS) after 730 days (for multiple myeloma) or 900 days (for neuroblastoma) post treatment or diagnosis. For breast cancer, the endpoints represented estrogen receptor status, a common diagnostic marker of this cancer type (abbre­viated ‘erpos’), and the success of treatment involving chemotherapy followed by surgical resection of a tumor (abbreviated ‘pCR’). The biological meaning of the control endpoints was known only to the project leader and not revealed to the project participants until all model development and external validation processes had been completed. To evaluate the reproducibility of the models developed by a data analysis team for a given data set, we asked teams to submit models from two stages of analyses. In the first stage (hereafter referred to as the ‘original’ experiment), each team built prediction models for up to 13 different coded endpoints using six training data sets. Models were ‘frozen’ against further modification, submitted to the consortium and then tested on a blinded validation data set that was not available to the analysis teams during training. In the second stage (referred to as the ‘swap’ experiment), teams repeated the model building and validation process by training models on the original validation set and validating them using the original training set. To simulate the potential decision-making process for evaluating a microarray-based classifier, we established a process for each group to receive training data with coded endpoints, propose a data analysis protocol (DAP) based on exploratory analysis, receive feedback on the protocol and then perform the analysis and validation (Fig. 1). Analysis protocols were reviewed internally by other MAQC-II participants (at least two reviewers per protocol) and by members of the MAQC-II Regulatory Biostatistics Working Group (RBWG), a team from the FDA and industry comprising biostatisticians and others with extensive model building expertise. Teams were encouraged to revise their protocols to incorporate feedback from reviewers, but each team was eventually considered responsible for its own analysis protocol and incorporating reviewers’ feedback was not mandatory (see Online Methods for more details). We assembled two large tables from the original and swap experiments (Supplementary Tables 1 and 2, respectively) containing summary information about the algorithms and analytic steps, or ‘modeling factors’, used to construct each model and the ‘internal’ and ‘external’ performance of each model. Internal performance measures the ability of the model to classify the training samples, based on cross-validation exercises. External performance measures the ability of the model to classify the blinded independent validation data. We considered several performance metrics, including Matthews Correlation Coefficient (MCC), accuracy, sensitivity, speci­ficity, area under the receiver operating characteristic curve (AUC) and root mean squared error (r.m.s.e.). These two tables contain data on >30,000 models. Here we report performance based on MCC because

828

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

are also examples in the literature of classifiers whose performance cannot be reproduced on independent data sets because of poor study design27, poor data quality and/or insufficient cross-validation of all model development steps28,29. Each of these factors may contribute to a certain level of skepticism about claims of performance levels achieved by microarray-based classifiers. Previous evaluations of the reproducibility of microarray-based classifiers, with only very few exceptions30,31, have been limited to simulation studies or reanalysis of previously published results. Frequently, published benchmarking studies have split data sets at random, and used one part for training and the other for validation. This design assumes that the training and validation sets are produced by unbiased sampling of a large, homogeneous population of samples. However, specimens in clinical studies are usually accrued over years and there may be a shift in the participating patient population and also in the methods used to assign disease status owing to changing practice standards. There may also be batch effects owing to time variations in tissue analysis or due to distinct methods of ­ sample ­collection and handling at different medical centers. As a result, ­samples derived from sequentially accrued patient populations, as was done in MAQC-II to mimic clinical reality, where the first cohort is used for developing predictive models and subsequent patients are included in validation, may differ from each other in many ways that could influence the prediction performance. The MAQC-II project was designed to evaluate these sources of bias in study design by constructing training and validation sets at different times, swapping the test and training sets and also using data from diverse preclinical and clinical scenarios. The goals of MAQC-II were to survey approaches in genomic model development in an attempt to understand sources of variability in prediction performance and to assess the influences of endpoint signal strength in data. By providing the same data sets to many organizations for analysis, but not restricting their data analysis protocols, the project has made it possible to evaluate to what extent, if any, results depend on the team that performs the analysis. This contrasts with previous benchmarking studies that have typically been conducted by single laboratories. Enrolling a large number of organizations has also made it feasible to test many more approaches than would be practical for any single team. MAQC-II also strives to develop good modeling practice guidelines, drawing on a large international collaboration of experts and the lessons learned in the perhaps unprecedented effort of developing and evaluating >30,000 genomic classifiers to predict a variety of endpoints from diverse data sets. MAQC-II is a collaborative research project that includes parti­cipants from the FDA, other government agencies, industry and academia. This paper describes the MAQC-II structure and experimental design and summarizes the main findings and key results of the consortium, whose members have learned a great deal during the process. The resulting guidelines are general and should not be construed as specific recommendations by the FDA for regulatory submissions.

Articles The 36 analysis teams applied many different options under each modeling factor for developing models (Supplementary Table 4) including 17 summary and normalization methods, nine batch-effect removal methods, 33 feature selection methods (between 1 and >1,000 features), 24 classification algorithms and six internal validation ­methods. Such diversity suggests the community’s common practices are

it is informative when the distribution of the two classes in a data set is highly skewed and because it is simple to calculate and was available for all models. MCC values range from +1 to −1, with +1 indicating perfect prediction (that is, all samples classified correctly and none incorrectly), 0 indicates random prediction and −1 indicating perfect inverse prediction.

Table 1  Microarray data sets used for model development and validation in the MAQC-II project Training seta

© 2010 Nature America, Inc. All rights reserved.

Date set code

Endpoint Endpoint code description

Microarray platform

Validation seta

Number of samples

Positives (P)

Negatives (N)

P/N ratio

Number of samples

Positives (P)

Negatives (N)

P/N ratio

Hamner

A

Lung tumorigen vs. non-tumorigen (mouse)

Affymetrix Mouse 430 2.0

70

26

44

0.59

88

28

60

0.47

Iconix

B

Non-genotoxic liver carcinogens vs. non-carcinogens (rat)

Amersham Uniset Rat 1 Bioarray

216

73

143

0.51

201

57

144

0.40

NIEHS

C

Liver toxicants vs. non-toxicants based on overall necrosis score (rat)

Affymetrix Rat 230 2.0

214

79

135

0.58

204

78

126

0.62

Breast cancer (BR)

D

Pre-operative treatment response (pCR, pathologic complete response)

Affymetrix Human U133A

130

33

97

0.34

100

15

85

0.18

E

Estrogen receptor status (erpos) Overall survival milestone outcome (OS, 730-d cutoff) Event-free survival milestone outcome (EFS, 730-d cutoff) Clinical parameter S1 (CPS1). The actual class label is the sex of the patient. Used as a “positive” control endpoint Clinical parameter R1 (CPR1). The actual class label is randomly assigned. Used as a “negative” control endpoint Overall survival milestone outcome (OS, 900-d cutoff) Event-free survival milestone outcome (EFS, 900-d cutoff) Newly established parameter S (NEP_S). The actual class label is the sex of the patient. Used as a “positive” control endpoint Newly established parameter R (NEP_R). The actual class label is randomly assigned. Used as a “negative” control endpoint

130

80

50

1.6

100

61

39

1.56

340

51

289

0.18

214

27

187

0.14

340

84

256

0.33

214

34

180

0.19

340

194

146

1.33

214

140

74

1.89

340

200

140

1.43

214

122

92

1.33

238

22

216

0.10

177

39

138

0.28

239

49

190

0.26

193

83

110

0.75

246

145

101

1.44

231

133

98

1.36

246

145

101

1.44

253

143

110

1.30

Multiple myeloma (MM)

F

G

H

I

Neuroblastoma (NB)

J

K

L

M

Affymetrix Human U133Plus 2.0

Different versions of Agilent human microarrays

Comments and references The training set was first published in 2007 (ref. 50) and the validation set was generated for MAQC-II The data set was first published in 2007 (ref. 51). Raw microarray intensity data, instead of ratio data, were provided for MAQC-II data analysis Exploratory visualization of the data set was reported in 2008 (ref. 53). However, the phenotype classification problem was formulated specifically for MAQC-II. A large amount of additional microarray and phenotype data were provided to MAQC-II for cross-platform and cross-tissue comparisons The training set was first published in 2006 (ref. 56) and the validation set was specifically generated for MAQC-II. In addition, two distinct endpoints (D and E) were analyzed in MAQC-II The data set was first published in 2006 (ref. 57) and 2007 (ref. 58). However, patient survival data were updated and the raw microarray data (CEL files) were provided specifically for MAQC-II data analysis. In addition, endpoints H and I were designed and analyzed specifically in MAQC-II

The training data set was first published in 2006 (ref. 63). The validation set (two-color Agilent platform) was generated specifically for MAQC-II. In addition, one-color Agilent platform data were also generated for most samples used in the training and validation sets specifically for MAQC-II to compare the prediction performance of two-color versus one-color platforms. Patient survival data were also updated. In addition, endpoints L and M were designed and analyzed specifically in MAQC-II

The first three data sets (Hamner, Iconix and NIEHS) are from preclinical toxicogenomics studies, whereas the other three data sets are from clinical studies. Endpoints H and L are positive controls (sex of patient) and endpoints I and M are negative controls (randomly assigned class labels). The nature of H, I, L and M was unknown to MAQC-II participants except for the project leader until all calculations were completed.

aNumbers

shown are the actual number of samples used for model development or validation.

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

829

Figure 1  Experimental design and timeline of the MAQC-II project. Numbers (1–11) order the steps of analysis. Step 11 indicates when the original training and validation data sets were swapped to repeat steps 4–10. See main text for description of each step. Every effort was made to ensure the complete independence of the validation data sets from the training sets. Each model is characterized by several modeling factors and seven internal and external validation performance metrics (Supplementary Tables 1 and 2). The modeling factors include: (i) organization code; (ii) data set code; (iii) endpoint code; (iv) summary and normalization; (v) feature selection method; (vi) number of features used; (vii) classification algorithm; (viii) batch-effect removal method; (ix) type of internal validation; and (x) number of iterations of internal validation. The seven performance metrics for internal validation and external validation are: (i) MCC; (ii) accuracy; (iii) sensitivity; (iv) specificity; (v) AUC; (vi) mean of sensitivity and specificity; and (vii) r.m.s.e. s.d. of metrics are also provided for internal validation results.

1/08 9/07 – 10/07 12/07 – 1/08 4. Six training data sets 1. Exploratory 3. Review & approval data analysis of DAP by RBWG (13 endpoints) (36 DATs)

10/07 9/1/2007

11/07 12/07

10/07 – 12/07 2. Data analysis protocol (DAP)

1/08

2/08

3/08 Face-to-face meeting

3/08

4/08

1/08 – 3/08 5. Classifiers are frozen (mark one for validation)

8/08 9/08 7. Validation 9-10. Meta-data 9/08 – 10/08 (blind test) distribution 11. Swap data sets prediction distribution results

5/08

6/08

7/08

3/08 – 8/08 6. MAQC-II’s candidate models

8/08

9/08

10/08 11/08 12/08

8/08 – 9/08 8. Prediction results

1/09

10/08 – 2/09 12. Meta-data analysis & visualization

2/1/2009

10. Table of model information 4. Data sets

1. Exploration

2. DAP

Performance metrics

3. DAP review Modeling factors 1

2

3

Internal validation ...

11. Swap

5. Classifiers

Models

MF1 MF2 MF3 ...

9

6. Models

7. Validation

...

...

...

IV1 IV2 IV3

External validation ... ... ... ... m ... EV1 EV2 EV3 ...

1 2 3

... ... ... ... ...

n

8. Prediction

12. Meta-data analysis 1.0 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1

r = 0.8495, N = 17,092

0. 2 0. 3 0. 4 0. 5 0. 6 0. 7 0. 8 0. 9 1. 0

5′

approaches on a clinically realistic blinded external validation data set. This is especially important in light of the intended clinical or preclinical uses of classifiers that are constructed using initial data sets and validated for regulatory approval and then are expected to accurately predict samples collected under diverse conditions perhaps months or years later. To assess the reliability of performance estimates derived during model training, we compared the performance on the internal training data set with performance on the external validation data set for of each of the 18,060 models in the original experiment (Fig. 2a). Models without complete metadata were not included in the analysis. We selected 13 ‘candidate models’, representing the best model for each endpoint, before external validation was performed. We required that each analysis team nominate one model for each endpoint they analyzed and we then L H selected one candidate from these nomiC E nations for each endpoint. We observed a higher correlation between internal and J B external performance estimates in terms K

well ­represented. For each of the models nominated by a team as being the best model for a particular endpoint, we compiled the list of features used for both the original and swap experiments (see the MAQC Web site at http://edkb.fda.gov/MAQC/). These comprehensive tables represent a unique resource. The results that follow describe data mining efforts to determine the potential and limitations of current practices for developing and validating gene expression–based prediction models. Performance depends on endpoint and can be estimated during training Unlike many previous efforts, the study design of MAQC-II provided the opportunity to assess the performance of many different modeling 1.0

r = 0.840, N = 18,060

Endpoint A B C D E F G H I J K L M

External validation (MCC)

0.8 0.6 0.4 0.2 0 –0.2

b

1.0

r = 0.951, N = 13

0.8 External validation (MCC)

a

0.6 0.4 0.2

M

–0.2

0

0.2

0.4

0.6

0.8

–0.6 –0.6 –0.4 –0.2

1.0

Internal validation (MCC) 1.0

D A F

–0.4

–0.6 –0.6 –0.4 –0.2

c

G

I

0

–0.4

L

C

H

E

0

0.2

0.4

0.6

0.8

1.0

Internal validation (MCC) K

J

B

D

G

F

A

I

M

Internal validation External validation

0.8 0.6 0.4

MCC

© 2010 Nature America, Inc. All rights reserved.

Articles

0.2 0 –0.2 –0.4

1796

970

866

1143

NB- NIEHS MMBRpositive (rat liver positive erpos necrosis)

830

1079

2263

1192

2905

877

863

NBEFS

NBOS

Iconix (rat liver tumor)

BRpCR

MMEFS

MMOS

1569

807

1730

Hamner MMNB(mouse negative negative lung tumor)

Figure 2  Model performance on internal validation compared with external validation. (a) Performance of 18,060 models that were validated with blinded validation data. (b) Performance of 13 candidate models. r, Pearson correlation coefficient; N, number of models. Candidate models with binary and continuous prediction values are marked as circles and squares, respectively, and the standard error estimate was obtained using 500-times resampling with bagging of the prediction results from each model. (c) Distribution of MCC values of all models for each endpoint in internal (left, yellow) and external (right, green) validation performance. Endpoints H and L (sex of the patients) are included as positive controls and endpoints I and M (randomly assigned sample class labels) as negative controls. Boxes indicate the 25% and 75% percentiles, and whiskers indicate the 5% and 95% percentiles.

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

DAT24

0.532

0.982

0.910

0.845

0.748

0.575

0.557

0.311

0.323

0.244

DAT13

0.513

0.973

0.918

0.829

0.792

0.493

0.437

0.322

0.306

0.307

0.202

0.060

0.044 −0.041

DAT25

0.504

0.965

0.801

0.816

0.652

0.514

0.349

0.383

0.360

0.217

0.243

0.247

0.016 −0.051

DAT11

0.500

0.991

0.752

0.750

0.778

0.509

0.483

0.345

0.305

0.295

0.193

0.099

0.029

DAT12

0.495

0.973

0.869

0.825

0.755

0.403

0.413

0.321

0.275

0.193

0.266

0.152 −0.016 −0.117

DAT32

0.489

0.982

0.762

0.823

0.702

0.533

0.557

0.284

0.203

0.143

0.257

0.129

0.043 −0.006

DAT10

0.485

0.982

0.871

0.445

0.728

0.472

0.249

0.429

0.353

0.295

0.293

0.222

0.016 −0.035 0.067 −0.152

0.193

0.168

0.011 −0.059

0.012

DAT20

0.483

0.930

0.838

0.805

0.773

0.542

0.386

0.345

0.289

0.225

0.181

0.000

DAT4

0.473

0.982

0.847

0.835

0.737

0.488

0.344

0.118

0.324

0.110

0.176

0.247 −0.067 −0.112

0.057

0.243

0.090 −0.059 −0.059

DAT18

0.460

0.973

0.860

0.829

0.690

0.371

0.376

0.344

0.229

DAT36

0.457

0.956

0.815

0.847

0.773

0.491

0.202

0.185

0.385 −0.014 0.187

DAT29

0.443

0.982

0.847

0.780

0.755

0.377

0.423

0.313 −0.042 0.198

DAT35

0.427

0.725

0.782

0.824

0.770

0.531

0.344

0.168

0.203

0.002 −0.075

0.241

0.000

0.000 −0.041

0.349 −0.096 0.165

0.140

0.068

0.782

0.466

0.499

0.184

0.271

0.000 −0.062 0.203

0.051

0.013 −0.103

0.761

0.454

0.748

0.247

0.377

0.062

0.324

0.271

0.016 −0.020

DAT33

0.284

0.856

0.054

0.709

0.751

0.455 −0.213 –0.078 0.114

DAT3

0.263

0.982

0.830

0.595

0.544

0.036 −0.090 −0.027 0.336 −0.143 −0.030 −0.142 −0.047 0.019

Median

0.488

0.973

0.830

0.816

0.748

0.491

0.376

0.311

0.306

0.193

0.193

0.129

0.016 −0.041

0.891

0.829

0.732

0.403

0.479

0.429

0.301

0.217

0.162

0.196

0.067 −0.103

H

C

E

K

J

B

D

A

G

F

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

0.051

0.024

0.2 0 −0.2 −0.4 −0.6 −0.8 −1.0

ne B

ne M M

0.4

g

M

g

I

O S

S EF

tu

M M

M

M

m

or

0.479 −0.096 0.091

0.6

N

0.085

0.8

lu

ng

m

BR

O S

tu liv

er

B N

se ou M

R

at

po s

EF

B N

BR

er

cr

s liv at

R

Data analysis teams show different proficiency Next, we summarized the external validation performance of the ­models nominated by the 17 teams that analyzed all 13 endpoints (Fig. 3). Nominated models represent a team’s best assessment of its model-building effort. The mean external validation MCC per team over 11 endpoints, excluding negative controls I and M, varied from 0.532 for data analysis team (DAT)24 to 0.263 for DAT3, indicating appreciable differences in performance of the models developed by different teams for the same data. Similar trends were observed when AUC

ne

po

er

M

M

B

po s

of MCC for the selected candidate models Candidate 0.511 0.982 (r = 0.951, n = 13, Fig. 2b) than for the overall Mean* L set of models (r = 0.840, n = 18,060, Fig. 2a), suggesting that extensive peer review of analysis protocols was able to avoid selecting models that could result in less reliable predictions in external validation. Yet, even for the hand-selected candidate models, there is noticeable bias in the performance estimated from internal validation. That is, the internal validation performance is higher than the external validation performance for most endpoints (Fig. 2b). However, for some endpoints and for some model building methods or teams, internal and external performance correlations were more modest as described in the following sections. To evaluate whether some endpoints might be more predictable than others and to ­calibrate performance against the positive- and negative-control endpoints, we assessed all models generated for each endpoint (Fig. 2c). We observed a clear dependence of prediction performance on endpoint. For example, endpoints C (liver necrosis score of rats treated with hepatotoxicants), E (estrogen receptor status of breast cancer patients), and H and L (sex of the multiple myeloma and neuroblastoma patients, respectively) were the easiest to predict (mean MCC > 0.7). Toxicological endpoints A and B and disease progression endpoints D, F, G, J and K were more difficult to predict (mean MCC ~0.1–0.4). Negative-control endpoints I and M were totally unpredictable (mean MCC ~0), as expected. For 11 endpoints (excluding the negative controls), a large proportion of the submitted models predicted the endpoint significantly better than chance (MCC > 0) and for a given endpoint many models performed similarly well on both internal and external validation (see the distribution of MCC in Fig. 2c). On the other hand, not all the submitted models performed equally well for any given endpoint. Some models performed no better than chance, even for some of the easy-to-predict endpoints, suggesting that additional factors were responsible for differences in model performance.

pC R

0.707

0.636

or

0.982

0.364

S

0.371

.

DAT7

0.043

1.0

0.036

DAT19

N

© 2010 Nature America, Inc. All rights reserved.

Figure 3  Performance, measured using MCC, of the best models nominated by the 17 data analysis teams (DATs) that analyzed all 13 endpoints in the original training-validation experiment. The median MCC value for an endpoint, representative of the level of predicability of the endpoint, was calculated based on values from the 17 data analysis teams. The mean MCC value for a data analysis team, representative of the team’s proficiency in developing predictive models, was calculated based on values from the 11 non-random endpoints (excluding negative controls I and M). Red boxes highlight candidate models. Lack of a red box in an endpoint indicates that the candidate model was developed by a data analysis team that did not analyze all 13 endpoints.

Data analysis team code

Articles

Endpoint

was used as the performance metric (Supplementary Table 5) or when the original training and validation sets were swapped (Supplementary Tables 6 and 7). Table 2 summarizes the modeling approaches that were used by two or more MAQC-II data analysis teams. Many factors may have played a role in the difference of external validation performance between teams. For instance, teams used different modeling factors, criteria for selecting the nominated models, and software packages and code. Moreover, some teams may have been more proficient at microarray data modeling and better at guarding against clerical errors. We noticed substantial variations in ­performance among the many K-nearest neighbor algorithm (KNN)-based models developed by four analysis teams (Supplementary Fig. 1). Follow-up investigations identified a few possible causes leading to the discrepancies in performance32. For example, DAT20 fixed the parameter ‘number of neighbors’ K = 3 in its data analysis protocol for all endpoints, whereas DAT18 varied K from 3 to 15 with a step size of 2. This investigation also revealed that even a detailed but standardized description of model building requested from all groups failed to capture many important tuning variables in the process. The subtle modeling differences not captured may have contributed to the differing performance levels achieved by the data analysis teams. The differences in performance for the models developed by various data analysis teams can also be observed from the changing patterns of internal and external validation performance across the 13 endpoints (Fig. 3, Supplementary Tables 5–7 and Supplementary Figs. 2–4). Our observations highlight the importance of good model­ing practice in developing and validating microarray-based predictive models including reporting of computational details for results to be replicated26. In light of the MAQC-II experience, recording structured information about the steps and parameters of an analysis process seems highly desirable to facilitate peer review and reanalysis of results. Swap and original analyses lead to consistent results To evaluate the reproducibility of the models generated by each team, we correlated the performance of each team’s models on the original training data set to performance on the validation data set and repeated this calculation for the swap experiment (Fig. 4). The correlation varied from 0.698–0.966 on the original experiment and from 831

Articles

Original analysis (training => validation) Modeling factor

Option

Summary and normalization

Loess RMA MAS5 None Mean shift SAM FC+P T-Test RFE 0~9 10~99 ≥1,000 100~999 DA Tree NB

Batch-effect removal Feature selection

Number of features

Classification algorithm

© 2010 Nature America, Inc. All rights reserved.

KNN SVM

Number of teams

Number of endpoints

Number of models

12 3 11 10 3 4 8 5 2 10 13 3 10 4 5 4

3 7 7 11 11 11 11 11 11 11 11 11 11 11 11 11

2,563 46 4,947 2,281 7,279 3,771 4,711 400 647 393 4,445 474 4,298 103 358 924

8 9

11 11

6,904 986

Analytic options used by two or more of the 14 teams that submitted models for all endpoints in both the original and swap experiments. RMA, robust multichip analysis; SAM, significance analysis of microarrays; FC, fold change; RFE, recursive feature elimination; DA, discriminant analysis; Tree, decision tree; NB, naive Bayes; KNN, K-nearest neighbors; SVM, support vector machine.

0.443–0.954 on the swap experiment. For all but three teams (DAT3, DAT10 and DAT11) the original and swap correlations were within ±0.2, and all but three others (DAT4, DAT13 and DAT36) were within ±0.1, suggesting that the model building process was relatively robust, at least with respect to generating models with ­similar performance. For some data analysis teams the internal validation performance drastically overestimated the performance of the same model in predicting the validation data. Examination of some of those models revealed several reasons, including bias in the feature selection and cross-validation process28, findings consistent with what was observed from a recent literature survey33. Previously, reanalysis of a widely cited single study34 found that the results in the original publication were very fragile—that is, not reproducible if the training and validation sets were swapped35. Our observations, except for DAT3, DAT11 and DAT36 with correlation <0.6, mainly resulting from failure of accurately predicting the positive-control endpoint H in the swap analysis (likely owing to operator errors), do not substantiate such fragility in the currently examined data sets. It is important to emphasize that we repeated the entire model building and evaluation processes during the swap analysis and, therefore, stability applies to the model building process for each data analysis team and not to a particular model or approach. Supplementary Figure 5 provides a more detailed look at the correlation of internal and external validation for each data analysis team and each endpoint for both the original (Supplementary Fig. 5a) and swap (Supplementary Fig. 5d) analyses. As expected, individual feature lists differed from analysis group to analysis group and between models developed from the original and the swapped data. However, when feature lists were mapped to biological processes, a greater degree of convergence and concordance was observed. This has been proposed previously but has never been demonstrated in a comprehensive manner over many data sets and thousands of models as was done in MAQC-II36.

itself is by far the dominant source of variability, explaining >65% of the variability in the external validation performance. All other factors explain <8% of the total variance, and the residual variance is ~6%. Among the factors tested, those involving inter­actions with endpoint have a relatively large effect, in particular the interaction between endpoint with organization and classification algorithm, highlighting variations in pro­ficiency between analysis teams. To further investigate the impact of individual levels within each modeling factor, we estimated the empirical best linear unbiased predictors (BLUPs)37. Figure 5b shows the plots of BLUPs of the corresponding ­factors in Figure 5a with proportion of variation >1%. The BLUPs reveal the effect of each level of the factor to the corresponding MCC value. The BLUPs of the main endpoint effect show that rat liver necrosis, breast cancer estrogen receptor status and the sex of the patient (endpoints C, E, H and L) are relatively easier to be predicted with ~0.2–0.4 advantage contributed on the corresponding MCC values. The rest of the endpoints are relatively harder to be predicted with about −0.1 to −0.2 disadvantage contributed to the corresponding MCC values. The main factors of normalization, classification algorithm, the number of selected features and the feature selection method have an impact of −0.1 to 0.1 on the corresponding MCC values. Loess normalization was applied to the endpoints (J, K and L) for the neuro­blastoma data set with the twocolor Agilent platform and has 0.1 advantage to MCC values. Among the Microarray Analysis Suite version 5 (MAS5), Robust Multichip Analysis (RMA) and dChip normalization methods that were applied to all endpoints (A, C, D, E, F, G and H) for Affymetrix data, the dChip method has a lower BLUP than the others. Because ­normalization methods are partially confounded with endpoints, it may not be suitable to compare methods between different confounded groups. Among classification methods, discriminant analysis has the largest positive impact of 0.056 on the MCC values. Regarding the number of selected features, larger bin number has better impact on the average across endpoints. The bin number is assigned by applying the ceiling function to the log base 10 of the number of selected features. All the feature selection methods have a slight impact of −0.025 to 0.025 1.0 Correlation in swap analysis (validation → training)

Table 2  Modeling factor options frequently adopted by MAQC-II data analysis teams

10

24 12 20 29

0.9 4 0.8

18

13

32

7 25

0.7

36

0.6

11 0.5 3 0.4 0.4

0.8 1.0 0.9 0.5 0.6 0.7 Correlation in original analysis (training → validation)

The effect of modeling factors is modest To rigorously identify potential sources of variance that explain the variability in external-validation performance (Fig. 2c), we applied random effect modeling (Fig. 5a). We observed that the endpoint

Figure 4  Correlation between internal and external validation is dependent on data analysis team. Pearson correlation coefficients between internal and external validation performance in terms of MCC are displayed for the 14 teams that submitted models for all 13 endpoints in both the original (x axis) and swap (y axis) analyses. The unusually low correlation in the swap analysis for DAT3, DAT11 and DAT36 is a result of their failure to accurately predict the positive endpoint H, likely due to operator errors (Supplementary Table 6).

832

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Articles

50

60

70

Summary normalization Classification algorithm

A B C D E F GH J K L Tox BR MM NB

Number of features Feature selection

0.05

0

0

–0.05

–0.05

–0.10

–0.10

Endpoint

Validation iterations

BLUP

Organization*classification algorithm Summary normalization*endpoint

Summary normalization 0.10

0.10

0.05

0.05

0.05

0

0

0

–0.05

–0.05

–0.05

–0.10

–0.10

–0.10

Classification algorithm*endpoint

1

Number of features*endpoint

2

3

4

5

Number of features

Feature selection*endpoint

A B C D E F GH J K L Tox BR MM NB

Feature selection method

Validation iterations*endpoint Organization*endpoint

0.10

Batch effect removal*endpoint Organization*classification algorithm*endpoint

0.05 BLUP

© 2010 Nature America, Inc. All rights reserved.

Classification algorithm

0.10

Bscatter FC FC+P Fisher Golub KS P RFE SAM T-Test Vote Welch Wilcoxon

Organization Batch effect removal

ANN DA Forest GLM KNN Logistic ML NB NC PLS RFE SMO SVM Tree

40

Lo es M s AS M 5 e M an ed ia n R M A

30

0.10

0.05

p

20

0.10

A

10

0.40 0.30 0.20 0.10 0 –0.10 –0.20

G

0

BLUP

Endpoint

dC hi

b

a

Residual 0

1

2 3 4 5 6 7 8 Percentage of variation

9

0.20 0.10 0 –0.10 –0.20 –0.30 –0.40 –0.50

0.20 0.15 0.10 0.05 0 –0.05 –0.10

0 –0.05 –0.10 A B CD E F G H J K L Tox

BR

MM

NB

Classification algorithm* endpoint

Summary normalization* endpoint

A B C D E F GH J K L Tox

BR

MM

NB

Number of features* endpoint

A B C D E F GH J K L Tox

BR

MM

NB

Organization*classification* endpoint

Figure 5  Effect of modeling factors on estimates of model performance. (a) Random-effect models of external validation performance (MCC) were developed to estimate a distinct variance component for each modeling factor and several selected interactions. The estimated variance components were then divided by their total in order to compare the proportion of variability explained by each modeling factor. The endpoint code contributes the most to the variability in external validation performance. (b) The BLUP plots of the corresponding factors having proportion of variation larger than 1% in a. Endpoint abbreviations (Tox., preclinical toxicity; BR, breast cancer; MM, multiple myeloma; NB, neuroblastoma). Endpoints H and L are the sex of the patient. Summary normalization abbreviations (GA, genetic algorithm; RMA, robust multichip analysis). Classification algorithm abbreviations (ANN, artificial neural network; DA, discriminant analysis; Forest, random forest; GLM, generalized linear model; KNN, K-nearest neighbors; Logistic, logistic regression; ML, maximum likelihood; NB, Naïve Bayes; NC, nearest centroid; PLS, partial least squares; RFE, recursive feature elimination; SMO, sequential minimal optimization; SVM, support vector machine; Tree, decision tree). Feature selection method abbreviations (Bscatter, betweenclass scatter; FC, fold change; KS, Kolmogorov-Smirnov algorithm; SAM, significance analysis of microarrays).

Feature list stability is correlated with endpoint predictability Prediction performance is the most important criterion for evaluating the performance of a predictive model and its modeling process. However, the robustness and mechanistic relevance of the model and

the corresponding gene signature is also important (Supplementary Fig. 8). That is, given comparable prediction performance between two modeling processes, the one yielding a more robust and reproducible gene signature across similar data sets (e.g., by swapping the ­training and validation sets), which is therefore less susceptible to sporadic fluctuations in the data, or the one that provides new insights to the underlying biology is preferable. Reproducibility or stability of feature sets is best studied by running the same model selection ­protocol on two distinct collections of samples, a scenario only possible, in this case, after the blind validation data were distributed to the data anal­ysis teams that were asked to perform their analysis after swapping their original training and test sets. Supplementary Figures 9 and 10 show that, although the feature space is extremely large for microarray data, different teams and protocols were able to consistently select the best-performing features. Analysis of the lists of features indicated that for endpoints relatively easy to predict, various data analysis teams arrived at models that used more common features and the overlap of the lists from the original and swap analyses is greater than those for more difficult endpoints (Supplementary Figs. 9–11). Therefore, the level of stability of feature lists can be associated to the level of difficulty of the prediction problem (Supplementary Fig. 11), although multiple models with different feature lists and comparable performance can be found from the same data set39. Functional analysis of the most frequently selected genes by all data analysis protocols shows

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

833

on MCC values except for recursive feature elimination (RFE) that has an impact of −0.006. In the plots of the four selected interactions, the estimated BLUPs vary across endpoints. The large variation across endpoints implies the impact of the corresponding modeling factor on different endpoints can be very different. Among the four interaction plots (see Supplementary Fig. 6 for a clear labeling of each inter­ action term), the corresponding BLUPs of the three-way interaction of organization, classification algorithm and endpoint show the highest variation. This may be due to different tuning parameters applied to individual algorithms for different organizations, as was the case for KNN32. We also analyzed the relative importance of modeling factors on external-validation prediction performance using a decision tree model38. The analysis results revealed observations (Supplementary Fig. 7) largely consistent with those above. First, the endpoint code was the most influential modeling factor. Second, feature selection method, normalization and summarization method, classification method and organization code also contributed to prediction performance, but their contribution was relatively small.

Articles

© 2010 Nature America, Inc. All rights reserved.

that many of these genes represent biological processes that are highly relevant to the clinical outcome that is being predicted36. The sexbased endpoints have the best overlap, whereas more difficult survival endpoints (in which disease processes are confounded by many other factors) have only marginally better overlap with biological processes relevant to the disease than that expected by random chance. Summary of MAQC-II observations and recommendations The MAQC-II data analysis teams comprised a diverse group, some of whom were experienced microarray analysts whereas others were graduate students with little experience. In aggregate, the group’s composition likely mimicked the broad scientific community engaged in building and publishing models derived from microarray data. The more than 30,000 models developed by 36 data analysis teams for 13 endpoints from six diverse clinical and preclinical data sets are a rich source from which to highlight several important observations. First, model prediction performance was largely endpoint (bio­ logy) dependent (Figs. 2c and 3). The incorporation of multiple data sets and endpoints (including positive and negative controls) in the MAQC-II study design made this observation possible. Some endpoints are highly predictive based on the nature of the data, which makes it possible to build good models, provided that sound modeling procedures are used. Other endpoints are inherently difficult to predict regardless of the model development protocol. Second, there are clear differences in proficiency between data analysis teams (organizations) and such differences are correlated with the level of experience of the team. For example, the topperforming teams shown in Figure 3 were mainly industrial participants with many years of experience in microarray data analysis, whereas bottom-performing teams were mainly less-experienced graduate students or researchers. Based on results from the positive and negative endpoints, we noticed that simple errors were sometimes made, suggesting rushed efforts due to lack of time or unnoticed implementation flaws. This observation strongly suggests that mechanisms are needed to ensure the reliability of results presented to the regulatory agencies, journal editors and the research community. By examining the practices of teams whose models did not perform well, future studies might be able to identify pitfalls to be avoided. Likewise, practices adopted by top-performing teams can provide the basis for developing good modeling practices. Third, the internal validation performance from well-implemented, unbiased cross-validation shows a high degree of concordance with the external validation performance in a strict blinding process (Fig. 2). This observation was not possible from previously published studies owing to the small number of available endpoints tested in them. Fourth, many models with similar performance can be developed from a given data set (Fig. 2). Similar prediction performance is attainable when using different modeling algorithms and para­meters, and simple data analysis methods often perform as well as more complicated approaches32,40. Although it is not essential to include the same features in these models to achieve comparable prediction performance, endpoints that were easier to predict generally yielded models with more common features, when analyzed by different teams (Supplementary Fig. 11). Finally, applying good modeling practices appeared to be more important than the actual choice of a particular algorithm over the others within the same step in the modeling process. This can be seen in the diverse choices of the modeling factors used by teams that produced models that performed well in the blinded validation (Table 2) where modeling factors did not universally contribute to variations in model performance among good performing teams (Fig. 5). 834

Summarized below are the model building steps recommended to the MAQC-II data analysis teams. These may be applicable to model building practitioners in the general scientific community. Step one (design). There is no exclusive set of steps and procedures, in the form of a checklist, to be followed by any practitioner for all problems. However, normal good practice on the study design and the ratio of sample size to classifier complexity should be followed. The frequently used options for normalization, feature selection and classification are good starting points (Table 2). Step two (pilot study or internal validation). This can be accomplished by bootstrap or cross-validation such as the ten repeats of a fivefold cross-validation procedure adopted by most MAQC-II teams. The samples from the pilot study are not replaced for the pivotal study; rather they are augmented to achieve ‘appropriate’ target size. Step three (pivotal study or external validation). Many investigators assume that the most conservative approach to a pivotal study is to simply obtain a test set completely independent of the training set(s). However, it is good to keep in mind the exchange34,35 regarding the fragility of results when the training and validation sets are swapped. Results from further resampling (including simple swapping as in MAQC-II) across the training and validation sets can provide important information about the reliability of the models and the modeling procedures, but the complete separation of the training and validation sets should be maintained41. Finally, a perennial issue concerns reuse of the independent validation set after modifications to an originally designed and validated data analysis algorithm or protocol. Such a process turns the validation set into part of the design or training set42. Ground rules must be developed for avoiding this approach and penalizing it when it occurs; and practitioners should guard against using it before such ground rules are well established. DISCUSSION MAQC-II conducted a broad observational study of the current community landscape of gene-expression profile–based predictive model development. Microarray gene expression profiling is among the most commonly used analytical tools in biomedical research. Analysis of the high-dimensional data generated by these experiments involves multiple steps and several critical decision points that can profoundly influence the soundness of the results43. An important requirement of a sound internal validation is that it must include feature selection and parameter optimization within each iteration to avoid overly optimistic estimations of prediction performance28,29,44. To what extent this information has been disseminated and followed by the scientific community in current microarray analysis remains unknown33. Concerns have been raised that results published by one group of investigators often cannot be confirmed by others even if the same data set is used26. An inability to confirm results may stem from any of several reasons: (i) insufficient information is provided about the methodology that describes which analysis has actually been done; (ii) data preprocessing (normalization, gene filtering and feature selection) is too complicated and insufficiently documented to be reproduced; or (iii) incorrect or biased complex analytical methods26 are performed. A distinct but related concern is that genomic data may yield prediction models that, even if reproducible on the ­discovery data set, cannot be extrapolated well in independent validation. The MAQC-II project provided a unique opportunity to address some of these concerns. Notably, we did not place restrictions on the model building methods used by the data analysis teams. Accordingly, they adopted numerous different modeling approaches (Table 2 and Supplementary Table 4). VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

Articles For example, feature selection methods varied widely, from statistical significance tests, to machine learning algorithms, to those more reliant on differences in expression amplitude, to those employing knowledge of putative biological mechanisms associated with the endpoint. Prediction algorithms also varied widely. To make internal validation performance results comparable across teams for different models, we recommended that a model’s internal performance was estimated using a ten times repeated fivefold cross-validation, but this recommendation was not strictly followed by all teams, which also allows us to survey internal validation approaches. The diversity of analysis protocols used by the teams is likely to closely resemble that of current research going forward, and in this context mimics reality. In terms of the space of modeling factors explored, MAQC-II is a survey of current practices rather than a randomized, controlled experiment; therefore, care should be taken in interpreting the results. For ­example, some teams did not analyze all endpoints, causing missing data (models) that may be confounded with other modeling factors. Overall, the procedure followed to nominate MAQC-II candidate models was quite effective in selecting models that performed reasonably well during validation using independent data sets, although generally the selected models did not do as well in validation as in training. The drop in performance associated with the validation highlights the importance of not relying solely on internal validation performance, and points to the need to subject every classifier to at least one external validation. The selection of the 13 candidate models from many nominated models was achieved through a peer-review collaborative effort of many experts and could be described as slow, tedious and sometimes subjective (e.g., a data analysis team could only contribute one of the 13 candidate models). Even though they were still subject to over-optimism, the internal and external performance estimates of the candidate models were more concordant than those of the overall set of models. Thus the review was productive in identifying characteristics of reliable models. An important lesson learned through MAQC-II is that it is almost impossible to retrospectively retrieve and document decisions that were made at every step during the feature selection and model development stage. This lack of complete description of the model building process is likely to be a common reason for the inability of different data analysis teams to fully reproduce each other’s results32. Therefore, although meticulously documenting the classifier building procedure can be cumbersome, we recommend that all genomic publications include supplementary materials describing the model building and evaluation process in an electronic format. MAQC-II is making available six data sets with 13 endpoints that can be used in the future as a benchmark to verify that software used to implement new approaches performs as expected. Subjecting new software to benchmarks against these data sets could reassure potential users that the software is mature enough to be used for the development of predictive models in new data sets. It would seem advantageous to develop alternative ways to help determine whether specific implementations of modeling approaches and performance evaluation procedures are sound, and to identify procedures to capture this information in public databases. The findings of the MAQC-II project suggest that when the same data sets are provided to a large number of data analysis teams, many groups can generate similar results even when different model building approaches are followed. This is concordant with studies29,33 that found that given good quality data and an adequate number of informative features, most classification methods, if properly used, will yield similar predictive performance. This also confirms reports6,7,39 on small data sets by individual groups that have suggested that ­several different feature selection methods and prediction algorithms can

yield many models that are distinct, but have statistically similar performance. Taken together, these results provide perspective on the large number of publications in the bioinformatics literature that have examined the various steps of the multivariate prediction model building process and identified elements that are critical for achieving reliable results. An important and previously underappreciated observation from MAQC-II is that different clinical endpoints represent very different levels of classification difficulty. For some endpoints the currently available data are sufficient to generate robust models, whereas for other endpoints currently available data do not seem to be sufficient to yield highly predictive models. An analysis done as part of the MAQC-II project and that focused on the breast cancer data demonstrates these points in more detail40. It is also important to point out that for some clinically meaningful endpoints studied in the MAQC-II project, gene expression data did not seem to significantly outperform models based on clinical covariates alone, highlighting the challenges in predicting the outcome of patients in a heterogeneous population and the potential need to combine gene expression data with clinical covariates (unpublished data). The accuracy of the clinical sample annotation information may also play a role in the difficulty to obtain accurate prediction results on validation samples. For example, some samples were misclassified by almost all models (Supplementary Fig. 12). It is true even for some samples within the positive control endpoints H and L, as shown in Supplementary Table 8. Clinical information of neuroblastoma patients for whom the positive control endpoint L was uniformly misclassified were rechecked and the sex of three out of eight cases (NB412, NB504 and NB522) was found to be incorrectly annotated. The companion MAQC-II papers published elsewhere give more in-depth analyses of specific issues such as the clinical ­benefits of genomic classifiers (unpublished data), the impact of ­ different m­odeling ­factors on prediction performance45, the objective assessment of micro­array cross-platform prediction46, cross-tissue prediction47, one-color ­ versus two-color prediction comparison 48, functional analysis of gene signatures36 and recommendation of a simple yet robust data analysis protocol based on the KNN32. For example, we systematically compared the classification performance resulting from one- and two-color gene-expression profiles of 478 neuroblastoma samples and found that analyses based on either ­platform yielded similar ­classification performance48. This newly gene­ rated one-color data set has been used to evaluate the applicability of the KNN-based simple data analysis protocol to future data sets32. In addition, the MAQC-II Genome-Wide Association Working Group assessed the variabilities in ­genotype calling due to experimental or algorithmic factors49. In summary, MAQC-II has demonstrated that current methods commonly used to develop and assess multivariate gene-expression based predictors of clinical outcome were used appropriately by most of the analysis teams in this consortium. However, ­differences in proficiency emerged and this underscores the importance of proper implementation of otherwise robust analytical methods. Observations based on analysis of the MAQC-II data sets may be applicable to other diseases. The MAQC-II data sets are publicly available and are expected to be used by the scientific community as benchmarks to ensure proper modeling practices. The experience with the MAQC-II clinical data sets also reinforces the notion that clinical classification problems represent several different degrees of prediction difficulty that are likely to be associated with whether mRNA abundances measured in a specific data set are informative for the specific prediction problem. We anticipate that including other

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

835

Articles types of biological data at the DNA, microRNA, protein or meta­ bolite levels will enhance our capability to more accurately ­predict the clinically relevant endpoints. The good modeling practice guidelines established by MAQC-II and lessons learned from this unprecedented collaboration provide a solid foundation from which other high-dimensional biological data could be more reliably used for the purpose of predictive and personalized medicine.

Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/.

1. Marshall, E. Getting the noise out of gene arrays. Science 306, 630–631 (2004). 2. Frantz, S. An array of problems. Nat. Rev. Drug Discov. 4, 362–363 (2005). 3. Michiels, S., Koscielny, S. & Hill, C. Prediction of cancer outcome with microarrays: a multiple random validation strategy. Lancet 365, 488–492 (2005). 4. Ntzani, E.E. & Ioannidis, J.P. Predictive ability of DNA microarrays for cancer outcomes and correlates: an empirical assessment. Lancet 362, 1439–1444 (2003). 5. Ioannidis, J.P. Microarrays and molecular research: noise discovery? Lancet 365, 454–455 (2005). 6. Ein-Dor, L., Kela, I., Getz, G., Givol, D. & Domany, E. Outcome signature genes in breast cancer: is there a unique set? Bioinformatics 21, 171–178 (2005). 7. Ein-Dor, L., Zuk, O. & Domany, E. Thousands of samples are needed to generate a robust gene list for predicting outcome in cancer. Proc. Natl. Acad. Sci. USA 103, 5923–5928 (2006). 8. Shi, L. et al. QA/QC: challenges and pitfalls facing the microarray community and regulatory agencies. Expert Rev. Mol. Diagn. 4, 761–777 (2004). 9. Shi, L. et al. Cross-platform comparability of microarray technology: intra-platform consistency and appropriate data analysis procedures are essential. BMC Bioinformatics 6 Suppl 2, S12 (2005). 10. Shi, L. et al. The MicroArray Quality Control (MAQC) project shows inter- and intraplatform reproducibility of gene expression measurements. Nat. Biotechnol. 24, 1151–1161 (2006). 11. Guo, L. et al. Rat toxicogenomic study reveals analytical consistency across microarray platforms. Nat. Biotechnol. 24, 1162–1169 (2006). 12. Canales, R.D. et al. Evaluation of DNA microarray results with quantitative gene expression platforms. Nat. Biotechnol. 24, 1115–1122 (2006). 13. Patterson, T.A. et al. Performance comparison of one-color and two-color platforms within the MicroArray Quality Control (MAQC) project. Nat. Biotechnol. 24, 1140–1150 (2006). 14. Shippy, R. et al. Using RNA sample titrations to assess microarray platform performance and normalization techniques. Nat. Biotechnol. 24, 1123–1131 (2006). 15. Tong, W. et al. Evaluation of external RNA controls for the assessment of microarray performance. Nat. Biotechnol. 24, 1132–1139 (2006). 16. Irizarry, R.A. et al. Multiple-laboratory comparison of microarray platforms. Nat. Methods 2, 345–350 (2005). 17. Strauss, E. Arrays of hope. Cell 127, 657–659 (2006). 18. Shi, L., Perkins, R.G., Fang, H. & Tong, W. Reproducible and reliable microarray results through quality control: good laboratory proficiency and appropriate data analysis practices are essential. Curr. Opin. Biotechnol. 19, 10–18 (2008). 19. Dudoit, S., Fridlyand, J. & Speed, T.P. Comparison of discrimination methods for the classification of tumors using gene expression data. J. Am. Stat. Assoc. 97, 77–87 (2002). 20. Goodsaid, F.M. et al. Voluntary exploratory data submissions to the US FDA and the EMA: experience and impact. Nat. Rev. Drug Discov. 9, 435–445 (2010). 21. van ‘t Veer, L.J. et al. Gene expression profiling predicts clinical outcome of breast cancer. Nature 415, 530–536 (2002). 22. Buyse, M. et al. Validation and clinical utility of a 70-gene prognostic signature for women with node-negative breast cancer. J. Natl. Cancer Inst. 98, 1183–1192 (2006). 23. Dumur, C.I. et al. Interlaboratory performance of a microarray-based gene expression test to determine tissue of origin in poorly differentiated and undifferentiated cancers. J. Mol. Diagn. 10, 67–77 (2008). 24. Deng, M.C. et al. Noninvasive discrimination of rejection in cardiac allograft recipients using gene expression profiling. Am. J. Transplant. 6, 150–160 (2006). 25. Coombes, K.R., Wang, J. & Baggerly, K.A. Microarrays: retracing steps. Nat. Med. 13, 1276–1277, author reply 1277–1278 (2007). 26. Ioannidis, J.P.A. et al. Repeatability of published microarray gene expression analyses. Nat. Genet. 41, 149–155 (2009). 27. Baggerly, K.A., Edmonson, S.R., Morris, J.S. & Coombes, K.R. High-resolution serum proteomic patterns for ovarian cancer detection. Endocr. Relat. Cancer 11, 583–584, author reply 585–587 (2004). 28. Ambroise, C. & McLachlan, G.J. Selection bias in gene extraction on the basis of microarray gene-expression data. Proc. Natl. Acad. Sci. USA 99, 6562–6566 (2002). 29. Simon, R. Using DNA microarrays for diagnostic and prognostic prediction. Expert Rev. Mol. Diagn. 3, 587–595 (2003). 30. Dobbin, K.K. et al. Interlaboratory comparability study of cancer gene expression analysis using oligonucleotide microarrays. Clin. Cancer Res. 11, 565–572 (2005). 31. Shedden, K. et al. Gene expression-based survival prediction in lung adenocarcinoma: a multi-site, blinded validation study. Nat. Med. 14, 822–827 (2008). 32. Parry, R.M. et al. K-nearest neighbors (KNN) models for microarray gene-expression analysis and reliable clinical outcome prediction. Pharmacogenomics J. 10, 292–309 (2010). 33. Dupuy, A. & Simon, R.M. Critical review of published microarray studies for cancer outcome and guidelines on statistical analysis and reporting. J. Natl. Cancer Inst. 99, 147–157 (2007). 34. Dave, S.S. et al. Prediction of survival in follicular lymphoma based on molecular features of tumor-infiltrating immune cells. N. Engl. J. Med. 351, 2159–2169 (2004). 35. Tibshirani, R. Immune signatures in follicular lymphoma. N. Engl. J. Med. 352, 1496–1497, author reply 1496–1497 (2005).

836

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Methods Methods and any associated references are available in the online version of the paper at http://www.nature.com/naturebiotechnology/.

© 2010 Nature America, Inc. All rights reserved.

Accession codes. All MAQC-II data sets are available through GEO (series accession number: GSE16716), the MAQC Web site (http://www.fda.gov/nctr/science/centers/toxicoinformatics/maqc/), ArrayTrack (http://www.fda.gov/nctr/science/centers/toxicoinformatics/ArrayTrack/) or CEBS (http://cebs.niehs.nih.gov/) accession number: 009-00002-0010-000-3. Note: Supplementary information is available on the Nature Biotechnology website. Acknowledgments The MAQC-II project was funded in part by the FDA’s Office of Critical Path Programs (to L.S.). Participants from the National Institutes of Health (NIH) were supported by the Intramural Research Program of NIH, Bethesda, Maryland or the Intramural Research Program of the NIH, National Institute of Environmental Health Sciences (NIEHS), Research Triangle Park, North Carolina. J.F. was supported by the Division of Intramural Research of the NIEHS under contract HHSN273200700046U. Participants from the Johns Hopkins University were supported by grants from the NIH (1R01GM083084-01 and 1R01RR021967-01A2 to R.A.I. and T32GM074906 to M.M.). Participants from the Weill Medical College of Cornell University were partially supported by the Biomedical Informatics Core of the Institutional Clinical and Translational Science Award RFA-RM-07002. F.C. acknowledges resources from The HRH Prince Alwaleed Bin Talal Bin Abdulaziz Alsaud Institute for Computational Biomedicine and from the David A. Cofrin Center for Biomedical Information at Weill Cornell. The data set from The Hamner Institutes for Health Sciences was supported by a grant from the American Chemistry Council’s Long Range Research Initiative. The breast cancer data set was generated with support of grants from NIH (R-01 to L.P.), The Breast Cancer Research Foundation (to L.P. and W.F.S.) and the Faculty Incentive Funds of the University of Texas MD Anderson Cancer Center (to W.F.S.). The data set from the University of Arkansas for Medical Sciences was supported by National Cancer Institute (NCI) PO1 grant CA55819-01A1, NCI R33 Grant CA97513-01, Donna D. and Donald M. Lambert Lebow Fund to Cure Myeloma and Nancy and Steven Grand Foundation. We are grateful to the individuals whose gene expression data were used in this study. All MAQC-II participants freely donated their time and reagents for the completion and analyses of the MAQC-II project. The MAQC-II consortium also thanks R. O’Neill for his encouragement and coordination among FDA Centers on the formation of the RBWG. The MAQC-II consortium gratefully dedicates this work in memory of R.F. Wagner who enthusiastically worked on the MAQC-II project and inspired many of us until he unexpectedly passed away in June 2008. DISCLAIMER This work includes contributions from, and was reviewed by, individuals at the FDA, the Environmental Protection Agency (EPA) and the NIH. This work has been approved for publication by these agencies, but it does not necessarily reflect official agency policy. Certain commercial materials and equipment are identified in order to adequately specify experimental procedures. In no case does such identification imply recommendation or endorsement by the FDA, the EPA or the NIH, nor does it imply that the items identified are necessarily the best available for the purpose. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/.

Articles

© 2010 Nature America, Inc. All rights reserved.

36. Shi, W. et al. Functional analysis of multiple genomic signatures demonstrates that classification algorithms choose phenotype-related genes. Pharmacogenomics J. 10, 310–323 (2010). 37. Robinson, G.K. That BLUP is a good thing: the estimation of random effects. Stat. Sci. 6, 15–32 (1991). 38. Hothorn, T., Hornik, K. & Zeileis, A. Unbiased recursive partitioning: a conditional inference framework. J. Comput. Graph. Statist. 15, 651–674 (2006). 39. Boutros, P.C. et al. Prognostic gene signatures for non-small-cell lung cancer. Proc. Natl. Acad. Sci. USA 106, 2824–2828 (2009). 40. Popovici, V. et al. Effect of training sample size and classification difficulty on the accuracy of genomic predictors. Breast Cancer Res. 12, R5 (2010). 41. Yousef, W.A., Wagner, R.F. & Loew, M.H. Assessing classifiers from two independent data sets using ROC analysis: a nonparametric approach. IEEE Trans. Pattern Anal. Mach. Intell. 28, 1809–1817 (2006). 42. Gur, D., Wagner, R.F. & Chan, H.P. On the repeated use of databases for testing incremental improvement of computer-aided detection schemes. Acad. Radiol. 11, 103–105 (2004).

43. Allison, D.B., Cui, X., Page, G.P. & Sabripour, M. Microarray data analysis: from disarray to consolidation and consensus. Nat. Rev. Genet. 7, 55–65 (2006). 44. Wood, I.A., Visscher, P.M. & Mengersen, K.L. Classification based upon gene expression data: bias and precision of error rates. Bioinformatics 23, 1363–1370 (2007). 45. Luo, J. et al. A comparison of batch effect removal methods for enhancement of prediction performance using MAQC-II microarray gene expression data. Pharmacogenomics J. 10, 278–291 (2010). 46. Fan, X. et al. Consistency of predictive signature genes and classifiers generated using different microarray platforms. Pharmacogenomics J. 10, 247–257 (2010). 47. Huang, J. et al. Genomic indicators in the blood predict drug-induced liver injury. Pharmacogenomics J. 10, 267–277 (2010). 48. Oberthuer, A. et al. Comparison of performance of one-color and two-color geneexpression analyses in predicting clinical endpoints of neuroblastoma patients. Pharmacogenomics J. 10, 258–266 (2010). 49. Hong, H. et al. Assessing sources of inconsistencies in genotypes and their effects on genome-wide association studies with HapMap samples. Pharmacogenomics J. 10, 364–374 (2010).

Leming Shi1, Gregory Campbell2, Wendell D Jones3, Fabien Campagne4, Zhining Wen1, Stephen J Walker5, Zhenqiang Su6, Tzu-Ming Chu7, Federico M Goodsaid8, Lajos Pusztai9, John D Shaughnessy Jr10, André Oberthuer11, Russell S Thomas12, Richard S Paules13, Mark Fielden14, Bart Barlogie10, Weijie Chen2, Pan Du15, Matthias Fischer11, Cesare Furlanello16, Brandon D Gallas2, Xijin Ge17, Dalila B Megherbi18, W Fraser Symmans19, May D Wang20, John Zhang21, Hans Bitter22, Benedikt Brors23, Pierre R Bushel13, Max Bylesjo24, Minjun Chen1, Jie Cheng25, Jing Cheng26, Jeff Chou13, Timothy S Davison27, Mauro Delorenzi28, Youping Deng29, Viswanath Devanarayan30, David J Dix31, Joaquin Dopazo32, Kevin C Dorff33, Fathi Elloumi31, Jianqing Fan34, Shicai Fan35, Xiaohui Fan36, Hong Fang6, Nina Gonzaludo37, Kenneth R Hess38, Huixiao Hong1, Jun Huan39, Rafael A Irizarry40, Richard Judson31, Dilafruz Juraeva23, Samir Lababidi41, Christophe G Lambert42, Li Li7, Yanen Li43, Zhen Li31, Simon M Lin15, Guozhen Liu44, Edward K Lobenhofer45, Jun Luo21, Wen Luo46, Matthew N McCall40, Yuri Nikolsky47, Gene A Pennello2, Roger G Perkins1, Reena Philip2, Vlad Popovici28, Nathan D Price48, Feng Qian6, Andreas Scherer49, Tieliu Shi50, Weiwei Shi47, Jaeyun Sung48, Danielle Thierry-Mieg51, Jean Thierry-Mieg51, Venkata Thodima52, Johan Trygg24, Lakshmi Vishnuvajjala2, Sue Jane Wang8, Jianping Wu53, Yichao Wu54, Qian Xie55, Waleed A Yousef56, Liang Zhang53, Xuegong Zhang35, Sheng Zhong57, Yiming Zhou10, Sheng Zhu53, Dhivya Arasappan6, Wenjun Bao7, Anne Bergstrom Lucas58, Frank Berthold11, Richard J Brennan47, Andreas Buness59, Jennifer G Catalano41, Chang Chang50, Rong Chen60, Yiyu Cheng36, Jian Cui50, Wendy Czika7, Francesca Demichelis61, Xutao Deng62, Damir Dosymbekov63, Roland Eils23, Yang Feng34, Jennifer Fostel13, Stephanie Fulmer-Smentek58, James C Fuscoe1, Laurent Gatto64, Weigong Ge1, Darlene R Goldstein65, Li Guo66, Donald N Halbert67, Jing Han41, Stephen C Harris1, Christos Hatzis68, Damir Herman69, Jianping Huang36, Roderick V Jensen70, Rui Jiang35, Charles D Johnson71, Giuseppe Jurman16, Yvonne Kahlert11, Sadik A Khuder72, Matthias Kohl73, Jianying Li74, Li Li75, Menglong Li76, Quan-Zhen Li77, Shao Li36, Zhiguang Li1, Jie Liu1, Ying Liu35, Zhichao Liu1, Lu Meng35, Manuel Madera18, Francisco Martinez-Murillo2, Ignacio Medina78, Joseph Meehan6, Kelci Miclaus7, Richard A Moffitt20, David Montaner78, Piali Mukherjee33, George J Mulligan79, Padraic Neville7, Tatiana Nikolskaya47, Baitang Ning1, Grier P Page80, Joel Parker3, R Mitchell Parry20, Xuejun Peng81, Ron L Peterson82, John H Phan20, Brian Quanz39, Yi Ren83, Samantha Riccadonna16, Alan H Roter84, Frank W Samuelson2, Martin M Schumacher85, Joseph D Shambaugh86, Qiang Shi1, Richard Shippy87, Shengzhu Si88, Aaron Smalter39, Christos Sotiriou89, Mat Soukup8, Frank Staedtler85, Guido Steiner90, Todd H Stokes20, Qinglan Sun53, Pei-Yi Tan7, Rong Tang2, Zivana Tezak2, Brett Thorn1, Marina Tsyganova63, Yaron Turpaz91, Silvia C Vega92, Roberto Visintainer16, Juergen von Frese93, Charles Wang62, Eric Wang21, Junwei Wang50, Wei Wang94, Frank Westermann23, James C Willey95, Matthew Woods21, Shujian Wu96, Nianqing Xiao97, Joshua Xu6, Lei Xu1, Lun Yang1, Xiao Zeng44, Jialu Zhang8, Li Zhang8, Min Zhang1, Chen Zhao50, Raj K Puri41, Uwe Scherf2, Weida Tong1 & Russell D Wolfinger7 1National

Center for Toxicological Research, US Food and Drug Administration, Jefferson, Arkansas, USA. 2Center for Devices and Radiological Health, US Food and Drug Administration, Silver Spring, Maryland, USA. 3Expression Analysis Inc., Durham, North Carolina, USA. 4Department of Physiology and Biophysics and HRH Prince Alwaleed Bin Talal Bin Abdulaziz Alsaud Institute for Computational Biomedicine, Weill Medical College of Cornell University, New York, New York, USA. 5Wake Forest Institute for Regenerative Medicine, Wake Forest University, Winston-Salem, North Carolina, USA. 6Z-Tech, an ICF International Company at NCTR/FDA, Jefferson, Arkansas, USA. 7SAS Institute Inc., Cary, North Carolina, USA. 8Center for Drug Evaluation and Research, US Food and Drug Administration, Silver Spring, Maryland, USA. 9Breast Medical Oncology Department, University of Texas (UT) M.D. Anderson Cancer Center, Houston, Texas, USA. 10Myeloma Institute for Research

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

837

© 2010 Nature America, Inc. All rights reserved.

Articles and Therapy, University of Arkansas for Medical Sciences, Little Rock, Arkansas, USA. 11Department of Pediatric Oncology and Hematology and Center for Molecular Medicine (CMMC), University of Cologne, Cologne, Germany. 12The Hamner Institutes for Health Sciences, Research Triangle Park, North Carolina, USA. 13National Institute of Environmental Health Sciences, National Institutes of Health, Research Triangle Park, North Carolina, USA. 14Roche Palo Alto LLC, South San Francisco, California, USA. 15Biomedical Informatics Center, Northwestern University, Chicago, Illinois, USA. 16Fondazione Bruno Kessler, Povo-Trento, Italy. 17Department of Mathematics & Statistics, South Dakota State University, Brookings, South Dakota, USA. 18CMINDS Research Center, Department of Electrical and Computer Engineering, University of Massachusetts Lowell, Lowell, Massachusetts, USA. 19Department of Pathology, UT M.D. Anderson Cancer Center, Houston, Texas, USA. 20Department of Biomedical Engineering, Georgia Institute of Technology and Emory University, Atlanta, Georgia, USA. 21Systems Analytics Inc., Waltham, Massachusetts, USA. 22Hoffmann-LaRoche, Nutley, New Jersey, USA. 23Department of Theoretical Bioinformatics, German Cancer Research Center (DKFZ), Heidelberg, Germany. 24Computational Life Science Cluster (CLiC), Chemical Biology Center (KBC), Umeå University, Umeå, Sweden. 25GlaxoSmithKline, Collegeville, Pennsylvania, USA. 26Medical Systems Biology Research Center, School of Medicine, Tsinghua University, Beijing, China. 27Almac Diagnostics Ltd., Craigavon, UK. 28Swiss Institute of Bioinformatics, Lausanne, Switzerland. 29Department of Biological Sciences, University of Southern Mississippi, Hattiesburg, Mississippi, USA. 30Global Pharmaceutical R&D, Abbott Laboratories, Souderton, Pennsylvania, USA. 31National Center for Computational Toxicology, US Environmental Protection Agency, Research Triangle Park, North Carolina, USA. 32Department of Bioinformatics and Genomics, Centro de Investigación Príncipe Felipe (CIPF), Valencia, Spain. 33HRH Prince Alwaleed Bin Talal Bin Abdulaziz Alsaud Institute for Computational Biomedicine, Weill Medical College of Cornell University, New York, New York, USA. 34Department of Operation Research and Financial Engineering, Princeton University, Princeton, New Jersey, USA. 35MOE Key Laboratory of Bioinformatics and Bioinformatics Division, TNLIST / Department of Automation, Tsinghua University, Beijing, China. 36Institute of Pharmaceutical Informatics, College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, Zhejiang, China. 37Roche Palo Alto LLC, Palo Alto, California, USA. 38Department of Biostatistics, UT M.D. Anderson Cancer Center, Houston, Texas, USA. 39Department of Electrical Engineering & Computer Science, University of Kansas, Lawrence, Kansas, USA. 40Department of Biostatistics, Johns Hopkins University, Baltimore, Maryland, USA. 41Center for Biologics Evaluation and Research, US Food and Drug Administration, Bethesda, Maryland, USA. 42Golden Helix Inc., Bozeman, Montana, USA. 43Department of Computer Science, University of Illinois at UrbanaChampaign, Urbana, Illinois, USA. 44SABiosciences Corp., a Qiagen Company, Frederick, Maryland, USA. 45Cogenics, a Division of Clinical Data Inc., Morrisville, North Carolina, USA. 46Ligand Pharmaceuticals Inc., La Jolla, California, USA. 47GeneGo Inc., Encinitas, California, USA. 48Department of Chemical and Biomolecular Engineering, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA. 49Spheromics, Kontiolahti, Finland. 50The Center for Bioinformatics and The Institute of Biomedical Sciences, School of Life Science, East China Normal University, Shanghai, China. 51National Center for Biotechnology Information, National Institutes of Health, Bethesda, Maryland, USA. 52Rockefeller Research Laboratories, Memorial Sloan-Kettering Cancer Center, New York, New York, USA. 53CapitalBio Corporation, Beijing, China. 54Department of Statistics, North Carolina State University, Raleigh, North Carolina, USA. 55SRA International (EMMES), Rockville, Maryland, USA. 56Helwan University, Helwan, Egypt. 57Department of Bioengineering, University of Illinois at Urbana-Champaign, Urbana, Illinois, USA. 58Agilent Technologies Inc., Santa Clara, California, USA. 59F. Hoffmann-La Roche Ltd., Basel, Switzerland. 60Stanford Center for Biomedical Informatics Research, Stanford University, Stanford, California, USA. 61Department of Pathology and Laboratory Medicine and HRH Prince Alwaleed Bin Talal Bin Abdulaziz Alsaud Institute for Computational Biomedicine, Weill Medical College of Cornell University, New York, New York, USA. 62Cedars-Sinai Medical Center, UCLA David Geffen School of Medicine, Los Angeles, California, USA. 63Vavilov Institute for General Genetics, Russian Academy of Sciences, Moscow, Russia. 64DNAVision SA, Gosselies, Belgium. 65École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland. 66State Key Laboratory of Multi-phase Complex Systems, Institute of Process Engineering, Chinese Academy of Sciences, Beijing, China. 67Abbott Laboratories, Abbott Park, Illinois, USA. 68Nuvera Biosciences Inc., Woburn, Massachusetts, USA. 69Winthrop P. Rockefeller Cancer Institute, University of Arkansas for Medical Sciences, Little Rock, Arkansas, USA. 70VirginiaTech, Blacksburg, Virgina, USA. 71BioMath Solutions, LLC, Austin, Texas, USA. 72Bioinformatic Program, University of Toledo, Toledo, Ohio, USA. 73Department of Mathematics, University of Bayreuth, Bayreuth, Germany. 74Lineberger Comprehensive Cancer Center, University of North Carolina, Chapel Hill, North Carolina, USA. 75Pediatric Department, Stanford University, Stanford, California, USA. 76College of Chemistry, Sichuan University, Chengdu, Sichuan, China. 77University of Texas Southwestern Medical Center (UTSW), Dallas, Texas, USA. 78Centro de Investigación Príncipe Felipe (CIPF), Valencia, Spain. 79Millennium Pharmaceuticals Inc., Cambridge, Massachusetts, USA. 80RTI International, Atlanta, Georgia, USA. 81Takeda Global R & D Center, Inc., Deerfield, Illinois, USA. 82Novartis Institutes of Biomedical Research, Cambridge, Massachusetts, USA. 83W.M. Keck Center for Collaborative Neuroscience, Rutgers, The State University of New Jersey, Piscataway, New Jersey, USA. 84Entelos Inc., Foster City, California, USA. 85Biomarker Development, Novartis Institutes of BioMedical Research, Novartis Pharma AG, Basel, Switzerland. 86Genedata Inc., Lexington, Massachusetts, USA. 87Affymetrix Inc., Santa Clara, California, USA. 88Department of Chemistry and Chemical Engineering, Hefei Teachers College, Hefei, Anhui, China. 89Institut Jules Bordet, Brussels, Belgium. 90Biostatistics, F. Hoffmann-La Roche Ltd., Basel, Switzerland. 91Lilly Singapore Centre for Drug Discovery, Immunos, Singapore. 92Microsoft Corporation, US Health Solutions Group, Redmond, Washington, USA. 93Data Analysis Solutions DA-SOL GmbH, Greifenberg, Germany. 94Cornell University, Ithaca, New York, USA. 95Division of Pulmonary and Critical Care Medicine, Department of Medicine, University of Toledo Health Sciences Campus, Toledo, Ohio, USA. 96Bristol-Myers Squibb, Pennington, New Jersey, USA. 97OpGen Inc., Gaithersburg, Maryland, USA.

838

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

ONLINE METHODS

© 2010 Nature America, Inc. All rights reserved.

MAQC-II participants. MAQC-II participants can be grouped into several ­categories. Data providers are the participants who provided data sets to the consortium. The MAQC-II Regulatory Biostatistics Working Group, whose members included a number of biostatisticians, provided guidance and standard operating procedures for model development and performance estimation. One or more data analysis teams were formed at each organization. Each data analysis team actively analyzed the data sets and produced prediction models. Other participants also contributed to discussion and execution of the project. The 36 data analysis teams listed in Supplementary Table 3 developed data analysis protocols and predictive models for one or more of the 13 endpoints. The teams included more than 100 scientists and engineers with diverse backgrounds in machine learning, statistics, biology, medicine and chemistry, among others. They volunteered tremendous time and effort to conduct the data analysis tasks. Six data sets including 13 prediction endpoints. To increase the chance that MAQC-II would reach generalized conclusions, consortium members strongly believed that they needed to study several data sets, each of high quality and sufficient size, which would collectively represent a diverse set of prediction tasks. Accordingly, significant early effort went toward the selection of appropriate data sets. Over ten nominated data sets were reviewed for quality of sample collection and processing consistency, and quality of microarray and clinical data. Six data sets with 13 endpoints were ultimately selected among those nominated during a face-to-face project meeting with extensive deliberations among many participants (Table 1). Importantly, three preclinical (toxicogenomics) and three clinical data sets were selected to test whether baseline practice conclusions could be generalized across these rather disparate experimental types. An important criterion for data set selection was the anticipated support of MAQC-II by the data provider and the commitment to continue experimentation to provide a large external validation test set of comparable size to the training set. The three toxicogenomics data sets would allow the development of predictive models that predict toxicity of compounds in animal models, a prediction task of interest to the pharmaceutical industry, which could use such models to speed up the evaluation of toxicity for new drug candidates. The three clinical data sets were for endpoints associated with three diseases, breast cancer (BR), multiple myeloma (MM) and neuroblastoma (NB). Each clinical data set had more than one endpoint, and together incorporated several types of clinical applications, including treatment outcome and disease prognosis. The MAQC-II predictive modeling was limited to binary classification problems; therefore, continuous endpoint values such as overall survival (OS) and event-free survival (EFS) times were dichotomized using a ‘milestone’ cutoff of censor data. Prediction endpoints were chosen to span a wide range of prediction difficulty. Two endpoints, H (CPS1) and L (NEP_S), representing the sex of the patients, were used as positive control endpoints, as they are easily predictable by microarrays. Two other endpoints, I (CPR1) and M (NEP_R), representing randomly assigned class labels, were designed to serve as negative control endpoints, as they are not supposed to be predictable. Data analysis teams were not aware of the characteristics of endpoints H, I, L and M until their swap prediction results had been submitted. If a data analysis protocol did not yield models to accurately predict endpoints H and L, or if a data analysis protocol claims to be able to yield models to accurately predict endpoints I and M, something must have gone wrong. The Hamner data set (endpoint A) was provided by The Hamner Institutes for Health Sciences. The study objective was to apply microarray gene expression data from the lung of female B6C3F1 mice exposed to a 13-week treatment of chemicals to predict increased lung tumor incidence in the 2-year rodent cancer bioassays of the National Toxicology Program 50. If successful, the results may form the basis of a more efficient and economical approach for evaluating the carcinogenic activity of chemicals. Microarray analysis was performed using Affymetrix Mouse Genome 430 2.0 arrays on three to four mice per treatment group, and a total of 70 mice were analyzed and used as MAQC-II’s training set. Additional data from another set of 88 mice were collected later and provided as MAQC-II’s external validation set. The Iconix data set (endpoint B) was provided by Iconix Biosciences. The study objective was to assess, upon short-term exposure, hepatic tumor induction by nongenotoxic chemicals51, as there are currently no accurate and

doi:10.1038/nbt.1665

well-validated short-term tests to identify nongenotoxic hepatic tumorigens, thus necessitating an expensive 2-year rodent bioassay before a risk assessment can begin. The training set consists of hepatic gene expression data from 216 male Sprague-Dawley rats treated for 5 d with one of 76 structurally and mechanistically diverse nongenotoxic hepatocarcinogens and nonhepatocarcinogens. The validation set consists of 201 male Sprague-Dawley rats treated for 5 d with one of 68 structurally and mechanistically diverse nongenotoxic hepatocarcinogens and nonhepatocarcinogens. Gene expression data were generated using the Amersham Codelink Uniset Rat 1 Bioarray (GE HealthCare)52. The separation of the training set and validation set was based on the time when the micro­ array data were collected; that is, microarrays processed earlier in the study were used as training and those processed later were used as validation. The NIEHS data set (endpoint C) was provided by the National Institute of Environmental Health Sciences (NIEHS) of the US National Institutes of Health. The study objective was to use microarray gene expression data acquired from the liver of rats exposed to hepatotoxicants to build classifiers for prediction of liver necrosis. The gene expression ‘compendium’ data set was collected from 418 rats exposed to one of eight compounds (1,2-dichloro­ benzene, 1,4-dichlorobenzene, bromobenzene, monocrotaline, N-nitrosomorpholine, thioacetamide, galactosamine and diquat dibromide). All eight compounds were studied using standardized procedures, that is, a common array platform (Affymetrix Rat 230 2.0 microarray), experimental procedures and data retrieving and analysis processes. For details of the experimental design see ref. 53. Briefly, for each compound, four to six male, 12-week-old F344 rats were exposed to a low dose, mid dose(s) and a high dose of the toxicant and sacrificed 6, 24 and 48 h later. At necropsy, liver was harvested for RNA extraction, histopathology and clinical chemistry assessments. Animal use in the studies was approved by the respective Institutional Animal Use and Care Committees of the data providers and was conducted in accordance with the National Institutes of Health (NIH) guidelines for the care and use of laboratory animals. Animals were housed in fully accredited American Association for Accreditation of Laboratory Animal Care facilities. The human breast cancer (BR) data set (endpoints D and E) was contributed by the University of Texas M.D. Anderson Cancer Center. Gene expression data from 230 stage I–III breast cancers were generated from fine needle aspiration specimens of newly diagnosed breast cancers before any therapy. The biopsy specimens were collected sequentially during a prospective pharmacogenomic marker discovery study between 2000 and 2008. These specimens represent 70–90% pure neoplastic cells with minimal stromal contamination54. Patients received 6 months of preoperative (neoadjuvant) chemotherapy including paclitaxel (Taxol), 5-fluorouracil, cyclophosphamide and doxorubicin (Adriamycin) followed by surgical resection of the cancer. Response to preoperative chemotherapy was categorized as a pathological complete response (pCR = no residual invasive cancer in the breast or lymph nodes) or residual invasive cancer (RD), and used as endpoint D for prediction. Endpoint E is the clinical estrogen-receptor status as established by immunohistochemistry55. RNA extraction and gene expression profiling were performed in multiple batches over time using Affymetrix U133A microarrays. Genomic analysis of a subset of this sequentially accrued patient population were reported previously56. For each endpoint, the first 130 cases were used as a training set and the next 100 cases were used as an independent validation set. The multiple myeloma (MM) data set (endpoints F, G, H and I) was contributed by the Myeloma Institute for Research and Therapy at the University of Arkansas for Medical Sciences. Gene expression profiling of highly purified bone marrow plasma cells was performed in newly diagnosed patients with MM57–59. The training set consisted of 340 cases enrolled in total therapy 2 (TT2) and the validation set comprised 214 patients enrolled in total therapy 3 (TT3)59. Plasma cells were enriched by anti-CD138 immunomagnetic bead selection of mononuclear cell fractions of bone marrow aspirates in a central laboratory. All samples applied to the microarray contained >85% plasma cells as determined by two-color flow cytometry (CD38+ and CD45−/dim) performed after selection. Dichotomized overall survival (OS) and event-free survival (EFS) were determined based on a 2-year milestone cutoff. A gene expression model of high-risk multiple myeloma was developed and validated by the data provider58 and later on validated in three additional independent data sets60–62.

nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

The neuroblastoma (NB) data set (endpoints J, K, L and M) was contributed by the Children’s Hospital of the University of Cologne, Germany. Tumor samples were checked by a pathologist before RNA isolation; only samples with ≥60% tumor content were used and total RNA was isolated from ~50 mg of snap-frozen neuroblastoma tissue obtained before chemotherapeutic treatment. First, 502 preexisting 11 K Agilent dye-flipped, dual-color replicate profiles for 251 patients were provided63. Of these, profiles of 246 neuroblastoma samples passed an independent MAQC-II quality assessment by majority decision and formed the MAQC-II training data set. Subsequently, 514 dyeflipped dual-color 11 K replicate profiles for 256 independent neuroblastoma tumor samples were generated and profiles for 253 samples were selected to form the MAQC-II validation set. Of note, for one patient of the validation set, two different tumor samples were analyzed using both versions of the 2 × 11K microarray (see below). All dual-color gene-expression of the MAQC-II training set were generated using a customized 2 × 11K neuroblastoma-related microarray63. Furthermore, 20 patients of the MAQC-II validation set were also profiled using this microarray. Dual-color profiles of the remaining patients of the MAQC-II validation set were performed using a slightly revised version of the 2 × 11K microarray. This version V2.0 of the array comprised 200 novel oligonucleotide probes whereas 100 oligonucleotide probes of the original design were removed due to consistent low expression values (near background) observed in the training set profiles. These minor modifications of the microarray design resulted in a total of 9,986 probes present on both versions of the 2 × 11K microarray. The experimental protocol did not differ between both sets and gene-expression profiles were performed as described63. Furthermore, single-color gene-expression profiles were generated for 478/499 neuroblastoma samples of the MAQC-II dual-color training and validation sets (training set 244/246; validation set 234/253). For the remaining 21 samples no single-color data were available, due to either shortage of tumor material of these patients (n = 15), poor experimental quality of the generated singlecolor profiles (n = 5), or correlation of one single-color profile to two different dual-color profiles for the one patient profiled with both versions of the 2 × 11K microarrays (n = 1). Single-color gene-expression profiles were generated using customized 4 × 44K oligonucleotide microarrays produced by Agilent Technologies. These 4 × 44K microarrays included all probes represented by Agilent’s Whole Human Genome Oligo Microarray and all probes of the version V2.0 of the 2 × 11K customized microarray that were not present in the former probe set. Labeling and hybridization was performed following the manufacturer’s protocol as described48. Sample annotation information along with clinical co-variates of the patient cohorts is available at the MAQC web site (http://edkb.fda.gov/MAQC/). The institutional review boards of the respective providers of the clinical micro­ array data sets had approved the research studies, and all subjects had provided written informed consent to both treatment protocols and sample procurement, in accordance with the Declaration of Helsinki. MAQC-II effort and data analysis procedure. This section provides details about some of the analysis steps presented in Figure 1. Steps 2–4 in a first round of analysis was conducted where each data analysis team analyzed MAQC-II data sets to generate predictive models and associated performance estimates. After this first round of analysis, most participants attended a consortium meeting where approaches were presented and discussed. The meeting helped members decide on a common performance evaluation protocol, which most data analysis teams agreed to follow to render performance statistics comparable across the consortium. It should be noted that some data analysis teams decided not to follow the recommendations for performance evaluation protocol and used instead an approach of their choosing, resulting in various internal validation approaches in the final results. Data analysis teams were given 2 months to implement the revised analysis protocol (the group recommended using fivefold stratified cross-validation with ten repeats across all endpoints for the internal validation strategy) and submit their final models. The amount of metadata to collect for characterizing the modeling approach used to derive each model was also discussed at the meeting. For each endpoint, each team was also required to select one of its submitted models as its nominated model. No specific guideline was given and groups could select nominated models according to any objective or subjective criteria. Because the consortium lacked an agreed upon reference

nature biotechnology

performance measure (Supplementary Fig. 13), it was not clear how the nominated models would be evaluated, and data analysis teams ranked models by different measures or combinations of measures. Data analysis teams were encouraged to report a common set of performance measures for each model so that models could be reranked consistently a posteriori. Models trained with the training set were frozen (step 6). MAQC-II selected for each endpoint one model from the up-to 36 nominations as the MAQC-II candidate for validation (step 6). External validation sets lacking class labels for all endpoints were distributed to the data analysis teams. Each data analysis team used its previously frozen models to make class predictions on the validation data set (step 7). The sample-by-sample prediction results were submitted to MAQC-II by each data analysis team (step 8). Results were used to calculate the external validation performance metrics for each model. Calculations were carried out by three independent groups not involved in developing models, which were provided with validation class labels. Data analysis teams that still had no access to the validation class labels were given an opportunity to correct apparent clerical mistakes in prediction submissions (e.g., inversion of class labels). Class labels were then distributed to enable data analysis teams to check prediction performance metrics and perform in depth analysis of results. A table of performance metrics was assembled from information collected in steps 5 and 8 (step 10, Supplementary Table 1). To check the consistency of modeling approaches, the original validation and training sets were swapped and steps 4–10 were repeated (step 11). Briefly, each team used the validation class labels and the validation data sets as a training set. Prediction models and evaluation performance were collected by internal and external validation (considering the original training set as a validation set). Data analysis teams were asked to apply the same data analysis protocols that they used for the original ‘Blind’ Training → Validation analysis. Swap analysis results are provided in Supplementary Table 2. It should be noted that during the swap experiment, the data analysis teams inevitably already had access to the class label information for samples in the swap validation set, that is, the original training set. Model summary information tables. To enable a systematic comparison of models for each endpoint, a table of information was constructed containing a row for each model from each data analysis team, with columns containing three categories of information: (i) modeling factors that describe the model development process; (ii) performance metrics from internal validation; and (iii) performance metrics from external validation (Fig. 1; step 10). Each data analysis team was requested to report several modeling factors for each model they generated. These modeling factors are organization code, data set code, endpoint code, summary or normalization method, feature selection method, number of features used in final model, classification algorithm, internal validation protocol, validation iterations (number of repeats of crossvalidation or bootstrap sampling) and batch-effect-removal method. A set of valid entries for each modeling factor was distributed to all data analysis teams in advance of model submission, to help consolidate a common vocabulary that would support analysis of the completed information table. It should be noted that since modeling factors are self-reported, two models that share a given modeling factor may still differ in their implementation of the modeling approach described by the modeling factor. The seven performance metrics for internal validation and external validation are MCC (Matthews Correlation Coefficient), accuracy, sensitivity, specificity, AUC (area under the receiver operating characteristic curve), binary AUC (that is, mean of sensitivity and specificity) and r.m.s.e. For internal validation, s.d. for each performance metric is also included in the table. Missing entries indicate that the data analysis team has not submitted the requested information. In addition, the lists of features used in the data analysis team’s nominated models are recorded as part of the model submission for functional analysis and reproducibility assessment of the feature lists (see the MAQC Web site at http://edkb.fda.gov/MAQC/). Selection of nominated models by each data analysis team and selection of MAQC-II candidate and backup models by RBWG and the steering committee. In addition to providing results to generate the model information

doi:10.1038/nbt.1665

© 2010 Nature America, Inc. All rights reserved.

table, each team nominated a single model for each endpoint as its preferred model for validation, resulting in a total of 323 nominated models, 318 of which were applied to the prediction of the validation sets. These nominated models were peer reviewed, debated and ranked for each endpoint by the RBWG before validation set predictions. The rankings were given to the MAQC-II steering committee, and those members not directly involved in developing models selected a single model for each endpoint, forming the 13 MAQC-II candidate models. If there was sufficient evidence through documentation to establish that the data analysis team had followed the guidelines of good classifier principles for model development outlined in the standard operating procedure (Supplementary Data), then their nominated models were considered as potential candidate models. The nomination and selection of candidate models occurred before the validation data were released. Selection of one candidate model for each endpoint across MAQC-II was performed to reduce multiple selection concerns. This selection process turned out to be highly interesting, time consuming, but worthy, as participants had different viewpoints and criteria in ranking the data analysis protocols and selecting the candidate model for an endpoint. One additional criterion was to select the 13 candidate models in such a way that only one of the 13 models would be selected from the same data analysis team to ensure that a variety of approaches to model development were considered. For each endpoint, a backup model was also selected under the same selection process and criteria as for the candidate models. The 13 candidate models selected by MAQC-II indeed performed well in the validation prediction (Figs. 2c and 3). 50. Thomas, R.S., Pluta, L., Yang, L. & Halsey, T.A. Application of genomic biomarkers to predict increased lung tumor incidence in 2-year rodent cancer bioassays. Toxicol. Sci. 97, 55–64 (2007). 51. Fielden, M.R., Brennan, R. & Gollub, J. A gene expression biomarker provides early prediction and mechanistic assessment of hepatic tumor induction by nongenotoxic chemicals. Toxicol. Sci. 99, 90–100 (2007).

doi:10.1038/nbt.1665

52. Ganter, B. et al. Development of a large-scale chemogenomics database to improve drug candidate selection and to understand mechanisms of chemical toxicity and action. J. Biotechnol. 119, 219–244 (2005). 53. Lobenhofer, E.K. et al. Gene expression response in target organ and whole blood varies as a function of target organ injury phenotype. Genome Biol. 9, R100 (2008). 54. Symmans, W.F. et al. Total RNA yield and microarray gene expression profiles from fine-needle aspiration biopsy and core-needle biopsy samples of breast carcinoma. Cancer 97, 2960–2971 (2003). 55. Gong, Y. et al. Determination of oestrogen-receptor status and ERBB2 status of breast carcinoma: a gene-expression profiling study. Lancet Oncol. 8, 203–211 (2007). 56. Hess, K.R. et al. Pharmacogenomic predictor of sensitivity to preoperative chemotherapy with paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide in breast cancer. J. Clin. Oncol. 24, 4236–4244 (2006). 57. Zhan, F. et al. The molecular classification of multiple myeloma. Blood 108, 2020–2028 (2006). 58. Shaughnessy, J.D. Jr. et al. A validated gene expression model of high-risk multiple myeloma is defined by deregulated expression of genes mapping to chromosome 1. Blood 109, 2276–2284 (2007). 59. Barlogie, B. et al. Thalidomide and hematopoietic-cell transplantation for multiple myeloma. N. Engl. J. Med. 354, 1021–1030 (2006). 60. Zhan, F., Barlogie, B., Mulligan, G., Shaughnessy, J.D. Jr. & Bryant, B. High-risk myeloma: a gene expression based risk-stratification model for newly diagnosed multiple myeloma treated with high-dose therapy is predictive of outcome in relapsed disease treated with single-agent bortezomib or high-dose dexamethasone. Blood 111, 968–969 (2008). 61. Chng, W.J., Kuehl, W.M., Bergsagel, P.L. & Fonseca, R. Translocation t(4;14) retains prognostic significance even in the setting of high-risk molecular signature. Leukemia 22, 459–461 (2008). 62. Decaux, O. et al. Prediction of survival in multiple myeloma based on gene expression profiles reveals cell cycle and chromosomal instability signatures in high-risk patients and hyperdiploid signatures in low-risk patients: a study of the Intergroupe Francophone du Myelome. J. Clin. Oncol. 26, 4798–4805 (2008). 63. Oberthuer, A. et al. Customized oligonucleotide microarray gene expression-based classification of neuroblastoma patients outperforms current clinical risk stratification. J. Clin. Oncol. 24, 5070–5078 (2006).

nature biotechnology

ar t icl e s

Human hematopoietic stem/progenitor cells modified by zinc-finger nucleases targeted to CCR5 control HIV-1 in vivo © 2010 Nature America, Inc. All rights reserved.

Nathalia Holt1, Jianbin Wang2, Kenneth Kim2, Geoffrey Friedman2, Xingchao Wang3, Vanessa Taupin3, Gay M Crooks4, Donald B Kohn4, Philip D Gregory2, Michael C Holmes2 & Paula M Cannon1 CCR5 is the major HIV-1 co-receptor, and individuals homozygous for a 32-bp deletion in CCR5 are resistant to infection by CCR5-tropic HIV-1. Using engineered zinc-finger nucleases (ZFNs), we disrupted CCR5 in human CD34+ hematopoietic stem/ progenitor cells (HSPCs) at a mean frequency of 17% of the total alleles in a population. This procedure produces both mono- and bi-allelically disrupted cells. ZFN-treated HSPCs retained the ability to engraft NOD/SCID/IL2rγ null mice and gave rise to polyclonal multi-lineage progeny in which CCR5 was permanently disrupted. Control mice receiving untreated HSPCs and challenged with CCR5-tropic HIV-1 showed profound CD4+ T-cell loss. In contrast, mice transplanted with ZFN-modified HSPCs underwent rapid selection for CCR5−/− cells, had significantly lower HIV-1 levels and preserved human cells throughout their tissues. The demonstration that a minority of CCR5−/− HSPCs can populate an infected animal with HIV-1-resistant, CCR5−/− progeny supports the use of ZFN-modified autologous hematopoietic stem cells as a clinical approach to treating HIV-1.

The entry of HIV-1 into target cells involves sequential binding of the viral gp120 Env protein to the CD4 receptor and a chemokine co-receptor1. CCR5 is the major co-receptor used by HIV-1 and is expressed on key T-cell subsets that are depleted during HIV-1 infection, including memory T cells2. A genetic 32-bp deletion in CCR5 (CCR5Δ32) is relatively common in Western European populations and confers resistance to HIV-1 infection and AIDS in homozygotes3,4. The absence of any other significant phenotype associated with a lack of CCR5 (refs. 5–7) has spurred the development of therapies aimed at blocking the virus–CCR5 interaction, and CCR5 antagonists have proved to be an effective salvage therapy in patients with drug-resistant strains of HIV-1 (ref. 8). Recently, the ability of CCR5−/− mobilized CD34+ peripheral blood cells to generate HIV-resistant progeny that suppress HIV-1 replication in vivo was demonstrated in an HIV-infected patient undergoing transplantation from a homozygous CCR5Δ32 donor during treatment for acute myeloid leukemia9. The donor cells conferred long-term control of HIV-1 replication and restored the patient’s CD4+ T-cell levels in the absence of antiretroviral drug therapy. These clinical data support the potential of gene or stem cell therapies based on the elimination of CCR5. However, the risks associated with allogeneic transplantation and the impracticality of obtaining sufficient numbers of matched CCR5Δ32 donors10 mean that broader application of this approach will require methods for generating autologous CCR5−/− cells. Various gene therapy approaches to block CCR5 expression are being evaluated, including CCR5-specific ribozymes11,12, siRNAs13 and intrabodies14. The targeted cell populations

include both mature T cells and CD34+ HSPCs. Loss of CCR5 in HSPCs appears to have no adverse effects on hematopoiesis12,13,15. An alternative approach is the use of engineered ZFNs to permanently disrupt the CCR5 open reading frame. ZFNs comprise a series of linked zinc fingers engineered to bind specific DNA sequences and fused to an endonuclease domain16. Concerted binding of two juxtaposed ZFNs on DNA, followed by dimerization of the endonuclease domains, generates a double-stranded break at the DNA target. Such double-stranded breaks are rapidly repaired by cellular repair pathways, notably the mutagenic nonhomologous end-joining pathway, which leads to frequent disruption of the gene due to the addition or deletion of nucleotides at the break site17,18. A significant advantage of this approach is that permanent gene disruption can result from only transient ZFN expression. CD4+ T cells modified by CCR5-targeted ZFNs19 are currently being evaluated in a clinical trial. However, disruption of CCR5 in HSPCs is likely to provide a more durable anti-viral effect and to give rise to CCR5−/− cells in both the lymphoid and myeloid compartments that HIV-1 infects. To evaluate this approach, we optimized the delivery of CCR5-specific ZFNs to human CD34+ HSPCs and transplanted the modified cells into nonobese diabetic/severe combined immunodeficient/interleukin 2rγ null (NOD/SCID/IL2rγnull; NSG) mice, which support both human hematopoiesis20 and HIV-1 infection13. Infection of the mice with a CCR5-tropic strain of HIV-1 led to rapid selection for CCR5– human cells, a significant reduction in viral load and protection of human T-cell populations in the key tissues that HIV-1 infects. These

1Keck

School of Medicine of the University of Southern California, Los Angeles, California, USA. 2Sangamo BioSciences, Inc., Richmond, California, USA. 3Childrens Hospital Los Angeles, Los Angeles, California, USA. 4David Geffen School of Medicine at the University of California Los Angeles, Los Angeles, California, USA. Correspondence should be addressed to P.M.C. ([email protected]).

Received 20 October 2009; accepted 24 June 2010; published online 2 July 2010; corrected online 22 July 2010; doi:10.1038/nbt.1663

nature biotechnology volume 28 number 8 august 2010

839

a

% CD45+ in blood

100

Neg. Mock ZFN

Cel 1 digestion products 0%

0%

16%

80 60 40 20 0

Neg.

c Bone marrow 10

1,000

10

103

102

10

102

103

104

100

101

CD19

102

103

0 100

104

101

CD45

Spleen

102

103

101 100

104

100

101

CD14

102

103

104

Thymus 104

1,000

104

21

103

101

CD45

102

103

104

102

33

101 100 100

101

CD4

102

101

CD8

CD8

66

CD45

103

102

0 100

CD45

13

SSC

CD45

SSC

101

23

2

101 100

CD45

ZFN

104

1,000

71

3

74

0 100

Mock

Lung 4

103

100 100

104

101

CD4

Small intestine

CD45

102

103

findings suggest that ZFN engineering of autologous HSPCs may enable long-term control of HIV-1 in infected individuals. RESULTS Efficient disruption of CCR5 in human CD34+ HSPCs Gene delivery methods suitable to express ZFNs include plasmid DNA nucleofection16, integrase-defective lentiviral vectors21 and adenoviral vectors19. Although nonviral methods are attractive, nucleofection can be associated with relatively high toxicity for human CD34+ HSPCs and loss of engraftment potential22, although, more recently, less toxic outcomes have been described 23–25. We evaluated different parameters to identify nucleofection conditions that allowed efficient disruption of CCR5 while limiting toxicity. The extent of CCR5 disruption was quantified using PCR amplification across the CCR5 locus, denaturation and reannealing of products, and digestion with the Cel 1 nuclease, which preferentially cleaves DNA at distorted duplexes caused by mismatches. The Cel 1 nuclease assay detects a linear range of CCR5 disruption between 0.69% and 44% of the total alleles in a population, with an upper limit of sensitivity of 70–80% disruption (ref. 19 and data not shown). We used this assay to monitor CCR5 disruption as only a minority of human CD34+ cells expresses CCR5 (ref. 26), making it difficult to measure CCR5 expression by flow cytometry. Using CD34+ HSPCs harvested from umbilical cord blood and optimized nucleofection conditions, we achieved mean disruption rates of

104

103

103

102

102

CCR5

CD45

6

104

101 100 100

10

CD3

1

10

2

10

3

10

4

CD45

CD4

103

104

104

Lymphoid

42

10

23

CD45

3

102

CD8

46

104

101

102

Blood

1,000

0 100

840

b

CD34+ cells

SSC

Figure 1 ZFN-mediated disruption of CCR5 in CD34+ HSPCs. (a) Representative gel showing extent of CCR5 disruption in CD34+ HSPCs 24 h after nucleofection with ZFN-expressing plasmids (ZFN) or mock nucleofected (mock). Neg. is untreated CD34+ HSPCs. CCR5 disruption was measured by PCR amplification across the ZFN target site, followed by Cel 1 nuclease digestion and quantification of products by PAGE. (b) Graph showing mean ± s.d. percentage of human CD45+ cells in peripheral blood of mice at 8 weeks after transplantation with either untreated, mock nucleofected or ZFN nucleofected CD34 + HSPCs (n = 5 each group). (c) FACS profiles of human cells from various organs of one representative mouse into which ZFN-treated CD34+ HSPCs were transplanted. Cells were gated on FSC/SSC (forward scatter/ side scatter) to remove debris. Staining for human CD45, a pan leukocyte marker, was used to reveal the level of engraftment with human cells in each organ. CD45+-gated populations were further analyzed for subsets, as indicated: CD19 (B cells) in bone marrow, CD14 (monocytes/macrophages) in lung, CD4 and CD8 (T cells) in thymus and spleen and CD3 (T cells) in the small intestine (lamina propria). The CD45+ population from the small intestine was further analyzed for CD4 and CCR5 expression. Peripheral blood cells from CD45 + and lymphoid gates were analyzed for CD4 and CD8 expression. The percentage of cells in each indicated area is shown. No staining was observed with isotype-matched control antibodies (Supplementary Fig. 1) or in animals receiving no human graft (data not shown).

SSC

© 2010 Nature America, Inc. All rights reserved.

ar t icl e s

101 100 100

10

CD4

1

10

2

10

3

10

4

101

38

100 100

10

CD4

1

10

2

103

104

17% ± 10 (n = 21) of the total CCR5 alleles in the population (Fig. 1a). Similar results were also achieved using CD34+ HSPCs isolated from human fetal liver (data not shown). Previous studies in human cell lines16 and primary human T cells19 have shown that the percentage of bi-allelically modified cells in a ZFN-treated population is 30–40% of the total number of disrupted alleles detected by the Cel 1 assay. We therefore estimated that 5–7% of ZFN-treated cells would be CCR5−/−, although this was not directly measured. We evaluated toxicity by measuring induction of apoptosis. Although nucleofection increased toxicity to human CD34+ cells threefold compared to untreated cells, inclusion of the ZFN plasmids had no additional effect compared to mock nucleofected controls (data not shown). Overall, we consider that any adverse effects of nucleofection on cell viability may be offset by the high levels of CCR5 disruption achieved as well as the speed and simplicity of the procedure compared to viral vector systems19,21. ZFN-modified CD34+ HSPCs are capable of multi-lineage engraftment in NSG mice NSG mice can be engrafted with human CD34+ HSPCs20 and thereby provide a rigorous readout of the hematopoietic potential of genetically modified HSPCs. We evaluated the effects of nucleofection and/ or CCR5 disruption by transplanting both untreated and ZFN-treated human CD34+ HSPCs into 1-d-old mice that had received low-dose (150 cGy) radiation. Engraftment of human cells was efficient and rapid, volume 28 number 8 august 2010 nature biotechnology

typically resulting in 40% human CD45+ leu- Table 1 Secondary transplantation of ZFN-treated HSPCs kocytes in the peripheral blood at 8 weeks after Donor animalsa CD45b blood (%) Cel 1c BM (%) Secondary CD45b blood (%) Cel 1c blood (%) recipients transplantation. The animals showed no obviZFN (1) 41 11 ZFN (3) 34 +/- 5 16 +/- 4 ous toxicity or ill health, as reported for higher Neg. (1) 47 0 Neg. (3) 37 +/- 7 0 +/- 0 radiation doses27. ZFN-treated cells engrafted with ZFN-treated HSPCs (ZFN) or untreated HSPCs (Neg.) and NSG mice as efficiently as untreated control aBone marrow (BM) was harvested from donor mice engrafted b + transplanted into three secondary recipients for each BM. Levels of human CD45 cells were measured in blood of both donor cells (Fig. 1b), with no statistically significant and recipient mice at 8 weeks post-transplantation. cCCR5 disruption rates, measured by Cel 1 analysis of donor BM at time of difference between the two groups (Student’s harvest and in blood of recipient mice at 10 weeks post-transplantation. t-test, P = 0.26). Eight to 12 weeks after transplantation, we analyzed engraftment of various mouse tissues with human CD45+ untreated CD34+ HSPCs was transplanted into three additional anileukocytes and with cells from specific hematopoietic lineages (Fig. mals. Analysis of the peripheral blood of the secondary recipients 8 1c). Human cells were detected using human-specific antibodies, and weeks later revealed that all six animals had engrafted and that there specificity was confirmed using both unengrafted animals and isotype- was no significant difference in the percentage of human CD45+ leumatched antibody controls (Supplementary Fig. 1). High levels of kocytes between the ZFN-treated and control groups. Furthermore, human cells were found in both the peripheral blood and tissues, ranging human cells in the blood of the ZFN cohort had levels of CCR5 disfrom 5–15% of the intestine, >50% of blood, spleen and bone marrow, ruption that slightly exceeded the level in the original donor marand >90% of the thymus (Supplementary Table 1). CD4+ and CD8+ row (12–20%) (Table 1). These data demonstrate that ZFN activity T cells were present in multiple organs, including the thymus, spleen, can lead to permanent disruption of CCR5 in SCID-repopulating and both the intraepithelial and lamina propria regions of the small and stem cells and that such modified cells retain their engraftment and large intestines; B-cell progenitors were present in the bone marrow; and differentiation potential. CD14+ macrophage and/or monocytes were detected in the lung. Of particular interest was the large population of human CD4+CCR5+ cells Protection of CD4+ T cells in peripheral blood of NSG mice after in the intestines, as these cells are targeted by both HIV-1 in humans28–31 HIV-1 infection and SIV in primates32–34. Overall, the profile of human cells in mice Engrafted animals at 8–12 weeks after transplantation that had received receiving ZFN-treated CD34+ HSPCs was indistinguishable from that either unmodified or ZFN-treated CD34+ HSPCs were challenged with of mice transplanted with unmodified cells, both with respect to the the CCR5-tropic virus HIV-1BAL. This strain of HIV-1 causes a robust percentage of human cells in each tissue and the frequencies of different infection and significant CD4+ T-cell depletion in humanized mouse subsets (Supplementary Table 1), suggesting that ZFN-modified CD34+ models35,36, mimicking the human infection, in which depletion of HSPCs are functionally normal. CD4+CCR5+ lymphocytes results from a combination of direct infection, systemic immune activation36 and the upregulation of CCR5 on thymic ZFN-treated CD34+ HSPCs produce CCR5-disrupted progeny precursors37,38. After infection, blood samples were collected from the after secondary transplantation mice every 2 weeks and analyzed for HIV-1 RNA levels, T-cell subsets and To evaluate whether ZFN treatment of the bulk CD34 + popu- the extent of CCR5 disruption. At 8–12 weeks after infection, animals were lation modified true SCID-repopulating stem cells, we har- euthanized and multiple tissues analyzed (Supplementary Fig. 2). Changes in the ratio of CD4+ to CD8+ T cells in the peripheral blood vested bone marrow from an animal 18 weeks after engraftment + with ZFN-treated CD34 HSPCs, in which the extent of CCR5 are characteristic of progressive infection in individuals with AIDS39,40. disruption in the bone marrow was 11% (Table 1). This mar- We therefore examined the CD4/CD8 ratio in blood samples from indirow was transplanted into three 8-week-old recipients. vidual mice both before and after infection and found that the mean At the same time, bone marrow from a control animal engrafted with ratio before infection was similar for both the untreated and ZFN-treated

a

b

HIV-1 infected Uninf. (3) Lymphoid 104

42

Blood

103

103

104

67

103

102

102

102

101

101

101

38 10

10

39

0

101

102

103

32

0

0

100

P = 0.8892

ZFN (9)

104

100

CD4+/CD8+ ratio

CD45

Neg. (3)

104

CD8

© 2010 Nature America, Inc. All rights reserved.

ar t icl e s

10

0

101

102

103

104

100

101

CD4

102

103

104

P = 0.0001

2.5

Neg.

2.0

ZFN

1.5 1.0 0.5 0.0

Pre-infection

Post-infection

Figure 2 Protection of human CD4+ T cells in peripheral blood of HIV-infected mice previously engrafted with ZFN-modified CD34 + HSPCs. (a) FACS plots showing human CD4+ and CD8+ T cells in peripheral blood of representative animals from each of three cohorts: uninfected mice previously engrafted with either untreated or ZFN-treated CD34+ HSPCs (Uninf.), and HIV-1 infected animals previously engrafted with either untreated (Neg.) or ZFN-treated (ZFN) CD34+ HSPCs, at 4 weeks post-infection. The total number of animals analyzed in each cohort is indicated. Cells were gated on FSC/SSC to remove debris, on human CD45, and a lymphoid gate applied. Percentage of cells in indicated compartments is shown. (b) Ratio of human CD4+ to CD8+ lymphocytes in peripheral blood of individual mice into which untreated (Neg.) or ZFN-modified CD34+ HSPCs were transplanted, measured pre-infection and at 6–8 weeks post-infection. Statistical analysis comparing Neg. and ZFN cohorts at each time point is shown.

nature biotechnology volume 28 number 8 august 2010

841

ar t icl e s groups. After HIV-1 challenge, the ratios became highly skewed in the control group owing to the pronounced loss of CD4+ cells, whereas the ZFN-treated animals maintained normal ratios (Fig. 2a,b). Protection of human cells in mouse tissues after HIV-1 infection We next analyzed the human cells present in various mouse tissues 12 weeks after infection with HIV-1BAL. NSG mice into which unmodified cells were transplanted displayed a characteristic loss of certain human cell populations, whereas the ZFN-treated cohort retained normal human cell profiles throughout their tissues despite HIV-1 challenge (Fig. 3a). In the intestines and spleen, which are the organs harboring the highest percentage of human CD4+CCR5+ cells in this model (Supplementary Fig. 3), we observed specific depletion

a

HIV-1 infected

Bone marrow

Neg. (3)

1,000

HIV-1 infected

Spleen

ZFN (9)

Uninf. (3)

1,000

104

CD45

ZFN (9)

1,000

45

11

102

41

33

CD8

SSC

Neg. (3)

1,000

21

103

47

65

101

0 0 0 100 101 102 103 104 100 101 102 103 104 100 101 102 103 104

0

34

0 0 100 0 10 101 102 103 104 100 101 102 103 104 100 101 102 103 104

CD45

Thymus

CD4

Small intestine CD45

CD45 104

83

103 102 101

104

0

103

103

102

102

101

101

94

104

104

103

103

79 10

102

2

CD45

104

CD8

© 2010 Nature America, Inc. All rights reserved.

Uninf. (3) 1,000

of CD4+ T cells from the spleen and the complete loss of all human lymphocytes from the intestines of untreated animals, whereas these populations were fully preserved in the ZFN-treated cohort (Fig. 3b). In the bone marrow, which is not a major target organ of HIV-1 infection, levels of human CD45+ cells were similar in all three groups. Notably, HIV-1BAL infection resulted in the loss of virtually all human cells from the thymus of mice receiving untreated CD34+ HSPCs by 12 weeks after infection (Fig. 3a). Depletion of thymocytes has been proposed to occur as a consequence of the upregulation of CCR5 on these cells during HIV-1 infection37,38, and likely contributed both to the observed depletion in the thymus and to the reduction in the numbers of mature CD4+ and CD8+ T cells observed in other tissues.

101

100 100 100 100 101 102 103 104 100 101 102 103 104 100 101 102 103 104

CD4

b

101

104 103

0 10

2

84

101

100 100 100 100 101 102 103 104 100 101 102 103 104 100 101 102 103 104

CD3

HIV-1 infected No graft (2)

Neg. (2)

ZFN (2)

Neg. (3)

ZFN (9)

Small intestine, anti-CD3

Spleen, anti-CD4 Figure 3 Effects of HIV-1 infection on human cells in HSPC-engrafted NSG mice. (a) FACS analysis of human cells in tissues of representative NSG mice from three cohorts: uninfected mice previously engrafted with either untreated or ZFN-treated CD34 + HSPCs (Uninf.), and HIV-1 infected animals previously engrafted with either untreated (Neg.) or ZFN-treated (ZFN) CD34+ HSPCs. Mice were necropsied at 12 weeks post-infection or at the equivalent time point for uninfected animals. The total number of animals analyzed in each cohort is indicated. FACS analysis was performed as described in Figure 1. Small intestine sample is lamina propria, and similar results were obtained when samples from the large intestine were analyzed. Percentage of cells in indicated compartments is shown. (b) Immunohistochemical analysis of human CD3 expression in small intestine, and CD4 expression in spleen of representative NSG mice, into which untreated (Neg.) or ZFN-treated (ZFN) CD34 + HSPCs were transplanted, with and without HIV-1 infection. Animals were necropsied at 12 weeks after infection or at the same time point for uninfected animals. Control animals receiving no human CD34 + HSPCs (no graft) were also analyzed. The number of animals analyzed in each cohort is shown. Scale bars, 50 µM.

842

HIV-1 infection rapidly selects for CCR5– T cells We examined whether the survival of T cells in the mice receiving ZFN-treated CD34+ HSPCs was the result of selection for ZFN-modified progeny. We measured the percentage of disrupted CCR5 alleles in the blood of mice at sequential time points after HIV-1 challenge, using both the Cel 1 assay and a specific PCR amplification that detects a common 5-bp duplication at the ZFN target site that typically accounts for 10–30% of total modifications19. Both assays revealed a rapid increase in the frequency of ZFN-disrupted alleles, reaching the upper limit of the Cel 1 assay by 4 weeks after infection (Fig. 4a). We also examined levels of CCR5 disruption in multiple tissues from ZFN-treated animals, either uninfected or 12 weeks after HIV-1BAL challenge, and observed a sharp increase in CCR5 disruption after HIV-1 infection (Fig. 4b). FACS analysis of the spleen and intestine revealed that, in contrast to uninfected animals, in which ~25% of CD4+ cells were also CCR5+, very little or no CCR5 expression was detected in the CD4+ T cells that persisted in the ZFN-treated animals (Fig. 4c,d). Together, these data suggest that the protection of CD4+ lymphocytes in ZFN-treated mice was a consequence of selection for CCR5–, HIV-1-resistant cells derived from ZFN-edited cells. Heterogeneity of CCR5 modifications suggests polyclonal origins ZFN-induced double-stranded breaks repaired by nonhomologous end-joining result in highly heterogeneous changes at the targeted locus19. We used this property to investigate whether the CCR5 – cells that developed in mice that received ZFN-treated CD34+ HSPCs were polyclonal in origin. Sequencing of 60 individual CCR5 alleles amplified from the large intestine of an HIV1-infected mouse into which ZFN-treated CD34+ HSPCs were previously transplanted revealed that 59 alleles harbored mutations at the ZFN target site (Fig. 5). As previously

volume 28 number 8 august 2010 nature biotechnology

ar t icl e s significant difference (P = 0.02) in antigenemia between the two groups observed by the 6-week time point (data not shown). These differences between the two cohorts are more striking when the levels of human CD4+ T cells are also considered (Fig. 6a), as the loss of CD4+ T cells in the untreated mice probably contributed to the lowering of overall viral levels seen as the infection progressed. The continued presence of virus in the blood, despite acute loss of CD4+ cells, also occurs during progression to AIDS, where high viral load measurements in serum are typically observed when T-cell death is rapidly occurring41. In contrast, CD4+ T-cell levels in the ZFN-treated mice rebounded after the 2-week nadir and recovered to normal levels by 4 weeks after infection. In contrast to these findings with HIV1BAL, ZFN-treated mice challenged with a CXCR4-tropic HIV-1 strain did not control viral levels or preserve CD4+ T cells, confirming that the mechanism is CCR5 specific (Supplementary Fig. 4). We also measured HIV-1 levels in intestinal samples. In tissues harvested at 8 and 9 weeks after infection, viral levels in the ZFNtreated mice were 4 orders of magnitude lower than in the untreated controls. By the 10- and 12-week time points, HIV-1 RNA was undetectable in the ZFN-treated mice (Fig.  6b). This drop in viral load occurred despite the maintenance of normal numbers of human

Presence of ZFN-modified cells controls HIV-1 replication in vivo Quantitative PCR analysis of HIV-1 RNA levels in the peripheral blood of animals revealed that peak viremia occurred at 6 weeks after infection for animals that received transplants of either untreated or ZFN-treated CD34+ HSPCs (Fig. 6a), although the levels were significantly lower (P = 0.03) in the ZFN cohort. By 8 weeks after infection, viral loads in both cohorts were dropping but there continued to be a statistically significant difference between the two groups (P = 0.001). Measurements of p24 levels in the blood by enzyme-linked immunosorbent assay (ELISA) corroborated these findings, with a

CCR5 disruption in peripheral blood (%)

a

b

100 3

5

80

CCR5 disruption (%)

Total disuptions

4

2

5 bp duplication

60 2

40 20

5 3 2 5

0

0

4

1 1

2

4

6

8

10

3

3

Weeks post-infection

Uninf.

80

3

3

60

Large intestine

2

3

3

HIV-1

Uninf.

HIV-1

8

56

3

40 20

2

2

2

2

2

2

2

2

Cel 1 products (%)

c

BM

LP

LE

SP

SE

Lung

Spleen

Thymus

0

HIV-1 Uninf. (3)

CD45 10

33

102 10

1

Uninf. (3)

CD4 103

10

4

0

103

102

102

101

101

0

100 100 0 1 2 3 4 100 0 1 2 3 4 100 101 102 103 104 10 10 10 10 10 10 10 10 10 10

d

10

30

102 101

ZFN (9) 10

4

103

CD4

Neg. (3)

CD4

4

10

4

103

HIV-1

ZFN (9)

Spleen

CD45

Neg. (3)

CCR5

104

CCR5

103 102 101

4

0

103 102

0

101

100 100 100 100 101 102 103 104 100 101 102 103 104 100 101 102 103 104

CD4

HIV-1 infected 3,000

Uninf. (3)

Neg. (3)

ZFN (9)

CD4+ CCR5+ CD4+

2,000

1,000

LP

SE SP LE

SE

nature biotechnology volume 28 number 8 august 2010

SP LE LP

0 SE SP LE LP

Small intestine

Figure 4 HIV-1 infection selects for disrupted CCR5 alleles. (a) Mean ± s.d. levels of CCR5 disruption (Cel 1 assay, black bars) in sequential peripheral blood samples taken from mice into which ZFN-treated CD34+ HSPCs were transplanted and which were subsequently infected with HIV-1. Upper limit of linearity of Cel 1 assay is 44% (ref. 19) and is indicated by the dotted line; upper limit of sensitivity of assay is 70–80%. White bars show the frequency of a common 5-bp duplication at the ZFN target site that typically comprises 10–30% of total CCR5 mutations19. Numbers of mice analyzed at each time point, and in each assay, are shown above the appropriate bar. (b) Mean ± s.d. levels of CCR5 disruption (Cel 1 assay) in indicated tissues from mice into which ZFN-treated CD34+ HSPCs were transplanted; mice were necropsied at 12 weeks after infection (black bars) or at an equivalent time point for uninfected ZFNtreated animals (white bars). Numbers analyzed in each group are shown above the appropriate bar. One representative Cel 1 analysis from the large intestine (lamina propria) of uninfected and infected mice is shown. Animals receiving untreated cells gave no Cel 1 digestion products at any time point analyzed (data not shown). Asterisk indicates levels too low to quantify. (c) Contour FACS analyses of human CD4+ cells in the small intestine (lamina propria) and spleen of one representative animal from each indicated cohort are shown. Cells were gated on FSC/SSC to remove debris and gated on human CD45 and CD4. Numbers indicate the percentage of cells that are CCR5+. (d) Mean ± s.d. numbers of human CD4+ cells (gray bars) and CD4+CCR5+ cells (white bars) per 5,000 human CD45+ cells analyzed from different sections of the intestine and from the indicated cohorts. Asterisk indicates levels too low to quantify. Number of animals analyzed in each cohort is indicated. Abbr. S, small intestine; L, large intestine; E, intraepithelial lymphocytes; P, lamina propria lymphocytes; BM, bone marrow.

Number cells per 5,000 human CD45+ cells

© 2010 Nature America, Inc. All rights reserved.

reported for this ZFN pair19, a high proportion (13 out of 59) of the mutated loci contained a characteristic 5-bp duplication, with the remaining 46 clones bearing 36 unique sequences. In contrast, all alleles sequenced from a mouse receiving untreated CD34 + HSPCs contained the wild-type sequence (data not shown). The high degree of sequence diversity observed strongly suggests that multiple stem or progenitor cells were modified by the ZFNs. These findings also predict that the overwhelming majority of cells selected by HIV-1BAL infection would be CCR5−/−, which is in agreement with the data from flow cytometry analysis (Fig. 4c).

843

ar t icl e s

© 2010 Nature America, Inc. All rights reserved.

Wild-type (1) gttttgtgggcaacatgctggtcatcctcatcctgataaactgcaaaaggctgaagagcatgactgaca wt Deletions (43) gttttgtgggcaacatgctggtcatcctcat-ctgataaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgat--actgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctg--aaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcc---tcctgataaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc----taaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc-----aaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggAcatcctcatcctgat------caaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc------aaTtgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgat-------aaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcat-------ctgataaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc--------ctgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgat--------aaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctc--------aaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcc--------ataaactgcaaaaggctAaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcat---------ctgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgat----------aggctgaagagcatgactgaca gttttgtgggcaacatgctggt----------ctgataaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc-----------caaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctca-----------tgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc------------aaaaggctgaaAagGatgactgaca gttttgtgggcaacatgctg------------ctgGtaaactgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcct--------------gcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcat---------------ctgcaaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcct---------------caaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgataa----------------gagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgat-----------------Cgagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctga-------------------gagcatgactgaca gttttgtgggcaacatgctggtcatcctcatc-------------------tgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctgat--------------------gcatgactgaca gttttgtgggcaacatgctggtcatcctcatc----------------------agagcatgactgaca gttttgtgggcaacatgc--------------------------aaaaggctgaagagcatgactgaca gttttgtgggcaa------------------------------caaaaggctgaagagcatgactgaca gttttgtgggcaacatgctggtcatcctcatcctg--------------------------------ca gttttgtgggcaacatgctggt---------------------------------------------ca

-1 -2 -2 2X -3 -4 -5 3X -6 -6 -7 -7 -8 -8 -8 -8 -9 -10 -10 -11 -11 2X -12 -12 -14 5X -15 -15 2X -16 -17 -19 -19 -20 -22 -26 -30 -32 -45

Figure 5 ZFN activity produces heterogeneous mutations in CCR5. Sequence analysis was performed on 60 cloned human CCR5 alleles, PCR amplified from intraepithelial cells from the large intestine of an HIV-infected mouse into which ZFN-treated CD34+ HSPCs were previously transplanted, and at 12 weeks post-infection. The number of nucleotides deleted or inserted at the ZFN target site (underlined) in each clone is indicated on the right of each sequence, together with the number of times the sequence was found. Dashes (–) indicate deleted bases compared to the wild-type sequence; uppercase letters are point mutations; underlined upper case letters are inserted bases. Some specific mutations of CCR5 occurred more frequently, in particular a 5-bp duplication at the ZFN target site that was identified 13 times (bottom sequence). No mutations in CCR5 were observed in a similar analysis performed on control samples from a mouse receiving unmodified CD34+ HSPCs (data not shown).

and effect permanent knockout of the targeted gene19,45–47. Only transient expression of the ZFNs is required during a brief period of ex vivo culture, and the genetic mutation is present for the life of the cell and its progeny. Thus, a major shortcoming of other gene therapy technologies—the need for continued Insertions (16) expression of a foreign transgene—is avoided. gttttgtgggcaacatgctggtcatcctcatcctCTgataaactgcaaaaggctgaagagcatgactga +2 Moreover, unlike approaches based on small gttttgtgggcaacatgctggtcatcctcatcctgataTAaactgcaaaaggctgaagagcatgactga +2 molecules, antibodies or RNA interference44, gttttgtgggcaacatgctggtcatcctcatcctgatCTGATaaactgcaaaaggctgaagagcatgac +5 13X ZFN-mediated gene disruption can completely eliminate CCR5 from the surface of T lymphocytes in the intestines and other tissues (Fig.  3). These cells through bi-allelic modification. By using an optimized nucleofecobservations are consistent with a strong selective pressure for HIV- tion procedure, we were able to overcome the technical challenges to resistant CCR5−/− cells to replace CCR5-expressing cells, leading to ZFN-induced genome editing in CD34+ cells previously reported21 and control of viral replication. achieve, on average, disruption at 17% of the loci, which we estimate will produce 5–7% bi-allelically modified cells. DISCUSSION The safety and efficacy of T lymphocytes modified with CCR5Despite major advances in anti-retroviral therapy, HIV-1 infection targeted ZFNs are currently being evaluated in a phase 1 clinical trial. remains an epidemic cause of morbidity and mortality. Effective anti- In a preclinical study, investigation of the specificity of the same CCR5retroviral therapy often involves costly, multi-drug regimens that are targeted ZFNs as used in this study revealed off-target cleavage events in not well tolerated by a significant percentage of patients42, and even T cells at significant levels only at the homologous CCR2 locus19. Studies successful adherence to the therapy does not eradicate the virus, and a in mice have not detected any deleterious phenotype associated with rapid rebound in HIV-1 levels can occur if therapy is discontinued43. loss of CCR2 (ref. 48), and human genetic studies have even suggested An alternative approach to controlling HIV-1 replication is engineering a beneficial phenotype from the loss of this gene in HIV-infected indiof the body’s immune cells to be resistant to infection44. In this regard, viduals49. Although not analyzed here, modification of CD34+ HSPCs the CCR5 co-receptor is an attractive target because of the HIV-resistant with these same CCR5 ZFN reagents is likely to result in similar, low phenotype of homozygous CCR5Δ32 individuals3. In the present study, levels of off-target cleavage events. Any safety concerns associated with we identified conditions that allow efficient disruption of CCR5 in nonspecific cleavage must be evaluated in larger, future studies. Although T lymphocytes are the primary target of HIV-1 infection, human CD34+ HSPCs and demonstrated that such modified cells generate CCR5−/−, HIV-resistant progeny in a mouse model of human ZFN modification of HSPCs may allow longer-term production of hematopoiesis and HIV-1 infection, leading to control of HIV-1 replica- CCR5−/− cells in patients. The scientific rationale for CCR5 modification tion. These findings suggest that transplantation of autologous HSPCs of HSPCs is supported by the recent finding that an HIV+ leukemia patient modified by CCR5-specific ZFNs may provide a permanent supply of receiving a transplant from a CCR5−/− donor was effectively cured of his HIV-resistant progeny that could replace cells killed by HIV-1, recon- infection, despite discontinuing antiretroviral therapy9. As shown by our stitute the immune system and control viral replication long term in the data, ZFN-modified HSPCs retained full functionality and gave rise to absence of anti-retroviral therapy. CCR5– cells in lineages relevant to HIV-1 pathogenesis. ZFNs delivered to The high levels of CCR5 disruption that we achieved were pos- purified CD34+ cell populations by nucleofection were capable of modifysible because of an efficient gene editing technology based on ZFNs. ing true SCID-repopulating stem cells, and the high levels of CCR5 editing ZFNs can be designed to bind to a specific genomic DNA sequence were maintained after secondary transplantation. 844

volume 28 number 8 august 2010 nature biotechnology

The experimental mouse model of HIV-1 infection used in these studies revealed a strong selection for CCR5– progeny during acute infection with a CCR5-tropic strain of HIV-1. This suggests that CCR5−/− stem cells, even if the minority, produced sufficient numbers of CCR5−/− progeny to support immune reconstitution and inhibit HIV-1 replication. Such selection is consistent with clinical observations from genetic diseases such as adenosine deaminase deficiency (ADA)-SCID, X-linked SCID and Wiskott-Aldrich syndrome, in which normal hematopoietic cells have a selective advantage, so that spontaneous monoclonal reversions can lead to selective outgrowth of such cells and amelioration of symptoms50–53. The observation of almost complete replacement of human T cells in the intestines of the infected mice with CCR5– cells is consistent with this tissue harboring the majority of the body’s CD4+CCR5+ effector memory cells. A characteristic feature of HIV-1 replication in mucosal tissues is an ongoing cycle of T-cell death and the recruitment of replacement T cells, which, in an activated state, are highly permissive for HIV-1 infection37. This is especially true in the gut mucosa, a key battleground in HIV-1 infection54–56. We also observed a strong selection for CCR5– cells in the thymus, suggesting that CCR5– cells would be selected at both a precursor stage in the thymus and at an effector stage in the mucosa. Ultimately, the presence of HIV-resistant CCR5– cells in mucosal tissues should both protect individual cells from infection and help to break the cycle of immune hyperactivation that may underlie much of the pathology of AIDS57.

a Neg. (3) ZFN (9)

107

80

CD4+ in blood (%)

HIV-1 RNA copies/ml blood

106 105 104 103 102 101 2

4

6

60 40 20 0

8

Weeks post-infection

0

2

4

6

8

Weeks post-infection

b HIV-1 RNA copies/106 cells

© 2010 Nature America, Inc. All rights reserved.

ar t icl e s

108

2

2

2

Neg. ZFN

2

10

6

2

3

3

2

10

4

2

2

2

2

10

2

100

2

8

9

10

9

12

Small intestine

8

9

2

9

10

12

Weeks post-infection

Large intestine

Figure 6 Control of HIV-1 replication in mice receiving ZFN-treated CD34+ HSPCs . (a) Mean +/− s.d. levels of HIV-1 RNA (left) and percent CD4+ human T cells (right) in peripheral blood of mice into which untreated (Neg.) or ZFN-treated CD34+ HSPCs were transplanted, at indicated times postinfection. Dashed line is limit of detection of assay. Asterisk indicates a statistically significant difference between two groups (P < 0.05). (b) Mean ± s.d. HIV-1 RNA levels in small and large intestine lamina propria from Neg. or ZFN mice, from animals necropsied between 8 and 12 weeks postinfection. Numbers of mice analyzed at each time point are shown above the appropriate bar. Dashed line indicates limits of detection of assay. Asterisk indicates undetectable levels.

nature biotechnology volume 28 number 8 august 2010

Although antiretroviral therapy is highly effective in many patients, the associated costs and potential for side effects can be considerable when extrapolated over a lifetime. In contrast, our approach may provide a one-shot treatment that would be most suited to the setting of autologous HSPC transplantation. Procedures for isolating and processing HSPCs for autologous or allogeneic transplantation are well established. The use of a patient’s own stem cells may remove the requirement for full ablation of the marrow hematopoietic compartment and the immune suppression that is necessary in allogeneic transplantation. Indeed, the toxicity of such regimens is one reason that allogeneic stem cell transplantation from CCR5Δ32 donors is not a realistic treatment option for HIV+ patients in the absence of other conditions that necessitate the transplant. Of note, certain HIV-infected individuals, such as AIDS lymphoma patients, already undergo full ablation and autologous HSPC rescue as part of their therapy58 and may be suitable candidates for HSPCbased gene therapies44. In addition, the experience of autologous HSPC transplantation in gene therapy treatments for ADA-SCID59,60, chronic granulomatous disease61 and X-linked adrenoleukodystrophy62 is that nonmyeloablative conditioning can facilitate engraftment of gene-modified autologous HSPCs with minimal associated toxicity. It is possible that the use of nonmyeloablative regimens, together with the selective advantage conferred on CCR5−/− progeny, could prove an effective combination for HIV+ patients receiving ZFN-treated autologous HSPCs. Targeting CCR5 is not expected to provide protection against viruses that use alternate co-receptors such as CXCR4. Although only a handful of cases of HIV-1 infection of CCR5Δ32 homozygotes have been reported63,64, CXCR4-tropic viruses have been associated with accelerated disease progression65, so that selection for such strains could be an undesirable consequence of targeting CCR5. However, this outcome is not generally observed in patients treated with CCR5 inhibitors unless CXCR4-tropic viruses were present before therapy, and resistance to these drugs occurs by viral adaptation to the drug-bound form of CCR5 (refs. 66,67). Notably, although the patient who received the CCR5Δ32 transplant harbored CXCR4-tropic virus before the procedure, his HIV-1 infection was still controlled long term9,10. Similar to the recommendations for CCR5 inhibitors, it may be prudent to restrict CCR5 ZFN treatment of HSPCs to individuals with no detectable CXCR4-tropic virus. In contrast to the acute HIV-1 infection modeled in this study, HIV-1 patients usually present in a chronic phase of the disease, and their viral levels can be effectively controlled by antiretroviral therapy. The requirement for the selective pressure of active HIV-1 replication in the success of this, or other, anti-HIV gene therapies is at present unknown. It has been suggested that low-level viral replication continues in certain sanctuary sites, even in well-controlled patients on antiretroviral therapy43,68, which could provide a low level of selection, although drug intensification trials have not provided evidence of ongoing replication69. It is also possible that the high levels of CCR5 disruption we achieved without selection, if extrapolated to HIV+ patients, could be sufficient to provide a therapeutic effect even in the absence of a strong selective pressure. Alternatively, ZFN knockout of CCR5 in HSPCs could be viewed as a backup strategy in the event that antiretroviral therapy fails or is withdrawn. It may also be possible to incorporate antiretroviral therapy interruptions into an overall therapeutic strategy, as recently described for HIV-infected individuals receiving autologous HSPCs engineered with anti-HIV ribozymes, where gene-marked progeny were found at higher levels after treatment interruptions70. In summary, our data demonstrate that transient ZFN treatment of human CD34+ HSPCs can efficiently disrupt CCR5 while yielding cells that remain competent to engraft and support hematopoiesis. In the presence of CCR5-tropic HIV-1, CCR5−/− progeny rapidly replaced cells depleted by the virus, leading to a polyclonal population that ultimately 845

ar t icl e s preserved human immune cells in multiple tissues. Our findings indicate that the modification of only a minority of human CD34+ HSPCs may provide the same strong anti-viral benefit as was conferred by a complete CCR5Δ32 stem cell transplantation in a patient9. And they further suggest that a partially modified autologous transplant, administered under only mildly ablative transplantation regimens may also be effective, opening up the treatment to many more HIV-infected individuals. Finally, the identification of conditions that allow the efficient use of ZFNs in human CD34+ HSPCs suggests the use of this technology in other diseases for which HSPC modification may be curative. METHODS Methods and any associated references are available in the online version of the paper at http://www.nature.com/naturebiotechnology/.

© 2010 Nature America, Inc. All rights reserved.

Note: Supplementary information is available on the Nature Biotechnology website. ACKNOWLEDGMENTS We would like to thank A. Cuddihy, S. Ge, R. Hollis and N. Smiley for expert technical assistance; C. Lutzko, V. Garcia, R. Akkina, B. Torbett and M. McCune for advice regarding humanized mice; and M. McCune for communicating unpublished data. This work was supported by funding from the California HIV/AIDS Research Project (P.M.C.), The Saban Research Institute (V.T.), and the National Heart, Lung, and Blood Institute P01 HL73104 (G.M.C., D.B.K. and P.M.C.). AUTHOR CONTRIBUTIONS N.H. performed most of the experiments; J.W., K.K., G.F. and X.W. developed assays and analyzed samples; V.T. contributed to discussions; N.H., G.M.C., D.B.K., P.D.G., M.C.H. and P.M.C. designed the experiments and analyzed data; N.H. and P.M.C. wrote the manuscript. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/. Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/reprintsandpermissions/. 1. Wu, L. et al. CD4-induced interaction of primary HIV-1 gp120 glycoproteins with the chemokine receptor CCR-5. Nature 384, 179–183 (1996). 2. deRoda Husman, A.M., Blaak, H., Brouwer, M. & Schuitemaker, H. CC chemokine receptor 5 cell-surface expression in relation to CC chemokine receptor 5 genotype and the clinical course of HIV-1 infection. J. Immunol. 163, 84597–84603 (1999). 3. Samson, M. et al. Resistance to HIV-1 infection in Caucasian individuals bearing mutant alleles of the CCR-5 chemokine receptor gene. Nature 382, 722–725 (1996). 4. Novembre, J. et al. The geographic spread of the CCR5 Delta32 HIV-resistance allele. PLoS Biol. 3, e339 (2005). 5. Glass, W.G. et al. CCR5 deficiency increases risk of symptomatic West Nile virus infection. J. Exp. Med. 203, 35–40 (2006). 6. Kantarci, O.H. et al. CCR5∆32 polymorphism effects on CCR5 expression, patterns of immunopathology and disease course in multiple sclerosis. J. Neuroimmunol. 169, 137–143 (2005). 7. Rossol, M. et al. Negative association of the chemokine receptor CCR5 d32 polymorphism with systemic inflammatory response, extra-articular symptoms and joint erosion in rheumatoid arthritis. Arthritis Res. Ther. 11, R91–98 (2009). 8. Dau, B. & Holodiny, M. Novel targets for antiretroviral therapy: clinical progress to date. Drugs 69, 31–50 (2009). 9. Hutter, G. et al. Long-term control of HIV by CCR5 Delta32/Delta32 stem-cell transplantation. N. Engl. J. Med. 360, 692–698 (2009). 10. Hutter, G., Schneider, T. & Thiel, E. Transplantation of selected or transgenic blood stem cells—a future treatment for HIV/AIDS? J. Int. AIDS Soc. 12, 10–14 (2009). 11. Anderson, J. et al. Safety and efficacy of a lentiviral vector containing three anti-HIV genes–CCR5 ribozyme, tat-rev siRNA, and TAR decoy–in SCID-hu mouse-derived T cells. Mol. Ther. 15, 1182–1188 (2007). 12. Bai, J. et al. Characterization of anti-CCR5 ribozyme-transduced CD34+ hematopoietic progenitor cells in vitro and in a SCID-hu mouse model in vivo. Mol. Ther. 1, 244–254 (2000). 13. Kumar, P. et al. T cell-specific siRNA delivery suppresses HIV-1 infection in humanized mice. Cell 134, 577–586 (2008). 14. Swan, C.H. et al. T-cell protection and enrichment through lentiviral CCR5 intrabody gene delivery. Gene Ther. 13, 1480–1492 (2006). 15. Swan, C.H. & Torbett, B.E. Can gene delivery close the door to HIV-1 entry after escape? J. Med. Primatol. 35, 236–247 (2006).

846

16. Urnov, F.D. et al. Highly efficient endogenous human gene correction using designed zinc-finger nucleases. Nature 435, 646–651 (2005). 17. Jasin, M. et al. Genetic manipulation of genomes with rare-cutting endonucleases. Trends Genet. 12, 224–228 (1996). 18. Sonoda, E. et al. Differential usage of non-homologous end-joining and homologous recombination in double strand break repair. DNA Repair (Amst.) 5, 1021–1029 (2006). 19. Perez, E.E. et al. Establishment of HIV-1 resistance in CD4+ T cells by genome editing using zinc-finger nucleases. Nat. Biotechnol. 26, 808–816 (2008). 20. Ishikawa, F. et al. Development of functional human blood and immune systems in NOD/ SCID/IL2 receptor {gamma} chain(null) mice. Blood 106, 1565–1573 (2005). 21. Lombardo, A. et al. Gene editing in human stem cells using zinc finger nucleases and integrase-defective lentiviral vector delivery. Nat. Biotechnol. 25, 1298–1306 (2007). 22. Hollis, R.P. et al. Stable gene transfer to human CD34(+) hematopoietic cells using the Sleeping Beauty transposon. Exp. Hematol. 34, 1333–1343 (2006). 23. Sumiyoshi, T. et al. Stable transgene expression in primitive human CD34+ hematopoietic stem/progenitor cells, using the Sleeping Beauty transposon system. Hum. Gene Ther. 20, 1607–1626 (2009). 24. Mátés, L. et al. Molecular evolution of a novel hyperactive Sleeping Beauty transposase enables robust stable gene transfer in vertebrates. Nat. Genet. 41, 753–761 (2009). 25. Xue, X. et al. Stable gene transfer and expression in cord blood-derived CD34+ hematopoietic stem and progenitor cells by a hyperactive Sleeping Beauty transposon system. Blood 114, 1319–1330 (2009). 26. Basu, S. & Broxmeyer, H.E. CCR5 ligands modulate CXCL12-induced chemotaxis, adhesion, and Akt phosphorylation of human cord blood CD34+ cells. J. Immunol. 183, 7478–7488 (2009). 27. Watanabe, S. et al. Hematopoietic stem cell-engrafted NOD/SCID/IL2Rgamma null mice develop human lymphoid systems and induce long-lasting HIV-1 infection with specific humoral immune responses. Blood 109, 212–218 (2007). 28. Brenchley, J.M. et al. CD4+ T cell depletion during all stages of HIV disease occurs predominantly in the gastrointestinal tract. J. Exp. Med. 200, 749–759 (2004). 29. Brenchley, J.M. et al. HIV disease: fallout from a mucosal catastrophe? Nat. Immunol. 7, 235–239 (2006). 30. Guadalupe, M. et al. Severe CD4+ T-cell depletion in gut lymphoid tissue during primary human immunodeficiency virus type 1 infection and substantial delay in restoration following highly active antiretroviral therapy. J. Virol. 77, 11708–11717 (2003). 31. Talal, A.H. et al. Effect of HIV-1 infection on lymphocyte proliferation in gut-associated lymphoid tissue. J. Acquir. Immune Defic. Syndr. 26, 208–217 (2001). 32. Li, Q. et al. Peak SIV replication in resting memory CD4+ T cells depletes gut lamina propria CD4+ T cells. Nature 434, 1148–1152 (2005). 33. Mattapallil, J.J. et al. Massive infection and loss of memory CD4+ T cells in multiple tissues during acute SIV infection. Nature 434, 1093–1097 (2005). 34. Veazey, R.S. et al. Gastrointestinal tract as a major site of CD4+ T cell depletion and viral replication in SIV infection. Science 280, 427–431 (1998). 35. Berges, B.K. et al. HIV-1 infection and CD4 T cell depletion in the humanized Rag2−/−gamma c−/− (RAG-hu) mouse model. Retrovirology 3, 76–90 (2006). 36. Appay, V. & Sauce, D. Immune activation and inflammation in HIV-1 infection: causes and consequences. J. Pathol. 214, 231–241 (2008). 37. Stoddart, C.A. et al. IFN-alpha-induced upregulation of CCR5 leads to expanded HIV tropism in vivo. PLoS Pathog. 6, e1000766 (2010). 38. Choudhary, S.K. et al. R5 human immunodeficiency virus type 1 infection of fetal thymic organ culture induces cytokine and CCR5 expression. J. Virol. 79, 458–471 (2005). 39. Kahn, J.O. & Walker, B.D. Acute human immunodeficiency virus type 1 infection. N. Engl. J. Med. 339, 33–39 (1998). 40. Margolick, J.B. et al. Impact of inversion of the CD4/CD8 ratio on the natural history of HIV-1 infection. J. Acquir. Immune Defic. Syndr. 42, 620–626 (2007). 41. Henrard, D.R. et al. Natural History of HIV-1 cell-free viremia. J. Am. Med. Assoc. 274, 554–558 (1995). 42. Chen, R.Y. et al. Distribution of health care expenditures for HIV-infected patients. Clin. Infect. Dis. 42, 1003–1010 (2006). 43. Richman, D.D. et al. The challenge of finding a cure for HIV infection. Science 323, 1304–1307 (2009). 44. Rossi, J.J., June, C.H. & Kohn, D.B. Genetic therapies against HIV. Nat. Biotechnol. 25, 1444–1454 (2007). 45. Bibikova, M. et al. Targeted chromosomal cleavage and mutagenesis in Drosophila using zinc-finger nucleases. Genetics 161, 1169–1175 (2002). 46. Doyon, Y. et al. Heritable targeted gene disruption in zebrafish using designed zincfinger nucleases. Nat. Biotechnol. 26, 702–708 (2008). 47. Santiago, Y. et al. Targeted gene knockout in mammalian cells by using engineered zinc-finger nucleases. Proc. Natl. Acad. Sci. USA 105, 5809–5814 (2008). 48. Peters, W., Dupuis, M. & Charo, I.F. A mechanism for the impaired IFN-gamma production in C–C chemokine receptor 2 (CCR2) knockout mice: Role of CCR2 in linking the innate and adaptive immune responses. J. Immunol. 165, 7072–7077 (2000). 49. Smith, M.W. et al. CCR2 chemokine receptor and AIDS progression. Nat. Med. 3, 1052–1053 (1997). 50. Davis, B.R. & Candotti, F. Revertant somatic mosaicism in the Wiskott-Aldrich syndrome. Immunol. Res. 44, 127–131 (2009). 51. Hirschhorn, R. et al. Spontaneous in vivo reversion to normal of an inherited mutation in a patient with adenosine deaminase deficiency. Nat. Genet. 3, 290–295 (1996).

volume 28 number 8 august 2010 nature biotechnology

ar t icl e s 62. Cartier, N. et al. Hematopoietic stem cell gene therapy with a lentiviral vector in X-linked adrenoleukodystrophy. Science 326, 818–823 (2009). 63. Biti, R. et al. HIV-1 infection in an individual homozygous for the CCR5 deletion allele. Nat. Med. 3, 252–253 (1997). 64. Oh, D.Y. et al. CCR5Delta32 genotypes in a German HIV-1 seroconverter cohort and report of HIV-1 infection in a CCR5Delta32 homozygous individual. PLoS ONE 3, e2747–2753 (2008). 65. Weiser, B. et al. HIV-1 coreceptor usage and CXCR4-specific viral load predict clinical disease progression during combination antiretroviral therapy. AIDS 22, 469–479 (2008). 66. Ogert, R.A. et al. Mapping Resistance to the CCR5 co-receptor antagonist vicriviroc using heterologous chimeric HIV-1 envelope genes reveals key determinants in the C2–V5 domain of gp120. Virology 373, 387–399 (2008). 67. Soulie, C. et al. Primary genotypic resistance of HIV-1 to CCR5 antagonist treatmentnaïve patients. AIDS 22, 2212–2214 (2008). 68. Palmer, S. et al. Low-level viremia persists for at least 7 years in patients on suppressive antiretroviral therapy. Proc. Natl. Acad. Sci. USA 105, 3879–3884 (2008). 69. Dinoso, J.B. et al. Treatment intensification does not reduce residual HIV-1 viremia in patients on highly active antiretroviral therapy. Proc. Natl. Acad. Sci. USA 106, 9403–9408 (2009). 70. Mitsuyasu, R.T. et al. Phase 2 gene therapy trial of an anti-HIV ribozyme in autologous CD34+ cells. Nat. Med. 15, 285–292 (2009).

© 2010 Nature America, Inc. All rights reserved.

52. Hirschhorn, R. et al. In vivo reversion to normal of inherited mutations in humans. J. Med. Genet. 40, 721–728 (2003). 53. Stephan, V. et al. Atypical X-linked severe combined immunodeficiency due to possible spontaneous reversion of the genetic defect in T cells. N. Engl. J. Med. 335, 1563–1567 (1996). 54. Chun, T.W. et al. Persistence of HIV in gut-associated lymphoid tissue despite longterm antiretroviral therapy. J. Infect. Dis. 197, 714–720 (2008). 55. Lackner, A.A. et al. The gastrointestinal tract and AIDS pathogenesis. Gastroenterology 136, 1965–1978 (2009). 56. Picker, L.J. Immunopathogenesis of acute AIDS virus infection. Curr. Opin. Immunol. 18, 399–405 (2006). 57. Veazey, R.S., Marx, P.A. & Lackner, A.A. The mucosal immune system: primary target for HIV infection and AIDS. Trends Immunol. 22, 626–633 (2001). 58. Krishnan, A. et al. Autologous stem cell transplantation for HIV associated lymphoma. Blood 98, 3857–3859 (2001). 59. Aiuti, A. et al. Correction of ADA-SCID by stem cell gene therapy combined with nonmyeloablative conditioning. Science 296, 2410–2413 (2002). 60. Aiuti, A. et al. Gene therapy for immunodeficiency due to adenosine deaminase deficiency. N. Engl. J. Med. 360, 447–458 (2009). 61. Ott, M.G. et al. Correction of X-linked chronic granulomatous disease by gene therapy, augmented by insertional activation of MDS1–EVI1, PRDM16 or SETBP1. Nat. Med. 12, 401–409 (2006).

nature biotechnology volume 28 number 8 august 2010

847

ONLINE METHODS

© 2010 Nature America, Inc. All rights reserved.

Hematopoietic stem/progenitor cell isolation. Human CD3+ HSPCs were isolated from umbilical cord blood collected from normal deliveries at local hospitals, according to guidelines approved by the Children’s Hospital Los Angeles Committee on Clinical Investigation, or as waste cord blood material from StemCyte Corp. Immunomagnetic enrichment for CD34 + cells was performed using the magnetic-activated cell sorting (MACS) system (Miltenyi Biotec), per the manufacturer’s instructions, with the modification that the initial purified CD34+ population was put through a second column and washed three times with 3 ml of the supplied buffer per wash before the final elution. This additional step gave a > 99% pure CD34+ population, as measured by FACS analysis using the anti-CD34 antibody, 8G12 (BD Biosciences). Nucleofection of CD34+ HSPCs with ZFN expression plasmids. Freshly isolated CD34+ cells were stimulated for 5–12 h in X-VIVO 10 media (Lonza) containing 2 nM l-glutamine, 50 ng/ml SCF, 50 ng/ml Flt-3 and 50 ng/ml TPO (R&D Systems). 1 × 106 cells were nucleofected with 2.5 µg each of a plasmid pair expressing ZFNs binding upstream (ZFN-L) or downstream (ZFN-R) of codon Leu55 within TM1 of human CCR5 (ref. 19). The CD34+ cell/DNA mix was processed in an X series Amaxa Nucleofector (Lonza) using the U-01 setting and the human CD34+ nucleofector solution, according to the manufacturer’s instructions. Following nucleofection, cells were immediately placed in pre-warmed IMDM media (Lonza) containing 26% FBS (Mediatech), 0.35% BSA, 2nM l-glutamine, 0.5% 10−3 mol/l hydrocortisone (Stem Cell Technologies), 5 ng/ml IL-3, 10 ng/ml IL-6 and 25 ng/ml SCF (R&D Systems). Cells were allowed to recover in this media for 2–12 h before injection into mice. Apoptosis assay. CD34+ HSPCs were collected at 24 h post-nucleofection and analyzed for the percent of viable cells marked for apoptosis using the PE apoptosis detection kit (BD Biosciences) according to the manufacturer’s instructions. Cells were stained with 7-AAD (detects viable cells) and annexin V (detects apoptotic cells) and analyzed using a FACScan flow cytometer (BD Biosciences). This double staining allowed the identification of cells in the early stages of apoptosis.

(CCTTCTTACTGTCCCCTTCTGGGCTCAC) BHQ-1-3′ (Biosearch Technologies), and analyzed using a 7,900HT real-time PCR machine (Applied Biosystems). At the same time, 5 µl of a 1:50,000 dilution of the PCR product were used in a Taqman qPCR reaction using primers (5′- CCAAAAAATCAATGTGAAGCAAATC-3′ and 5′- TGCCCACAAAAC CAAAGATG -3′) and probe 5′- FAM d(CAGCCCGCCTCCTGCCTCC) BHQ-1-3′ to detect total copies of human CCR5. Data were analyzed using software supplied by the manufacturer and the frequency of pentamer insertions in CCR5 calculated. The assay is sensitive enough to detect a single pentamer insertion event in 100,000 cells (data not shown). ZFN-induced modifications of CCR5 were analyzed by directly sequencing cloned CCR5 alleles, isolated by PCR amplification as described above, and TOPO-TA cloning (Invitrogen). Plasmid DNA was isolated from 60 individual bacterial colonies for each tissue analyzed. HIV-1 infection and analysis. A cell-free virus stock of HIV-1BaL and a molecular clone of HIV-1NL4-3 were obtained from the AIDS Research and Reference Reagent Program (ARRRP), Division of AIDS, NIAID, NIH from material deposited by Suzanne Gartner, Mikulas Popovic, Robert Gallo and Malcolm Martin. HIV-1BaL virus was propagated in PM1 cells, obtained from the ARRRP and deposited by Marvin Reitz and harvested 10 d post-infection. HIV-1NL4-3 viruses were generated by transient transfection of 293T cells (ATCC). Viruses were titrated using the Alliance HIV-1 p24 ELISA kit (PerkinElmer) and by TCID50 analysis on U373-MAGI cells (ARRRP, deposited by Michael Emerman and Adam Geballe). Mice to be infected with HIV-1 were anesthetized with inhalant 2.5% isoflourane and injected intraperitoneally with virus stocks containing 200 ng p24, 7 × 104 TCID50 units, in 100 µl total volume. HIV-1 levels in peripheral blood or tissues harvested at necropsy were determined by extracting RNA from 5 × 105 cells using the master pure complete DNA and RNA purification kit (Epicentre Biotechnologies) and performing Taqman qPCR using a primer and probe set targeting the HIV-1 LTR region, as previously described72. In addition, p24 levels were measured in blood samples by ELISA.

NSG mouse transplantation. NOD.Cg-Prkdc scid Il2rg tm1Wj/SzJ (NOD/ SCID/IL2rγnull, NSG) mice71 were obtained from Jackson Laboratories. Neonatal mice within 48 h of birth received 150 cGy radiation, then 2–4 h later 1 × 106 ZFN-modified or mock-treated human CD34+ HSPCs in 50 µl PBS containing 1% heparin were injected through the facial vein. For secondary transplantations, bone marrow was harvested by needle aspiration from the upper and lower limbs of 18-week-old animals previously engrafted with human CD34+ HSPCs, filtered through a 70 µm nylon mesh screen (Fisher Scientific) and washed in PBS. The cells were transplanted into three 8-week-old mice that had previously received 350 cGy radiation, using retroorbital injection of 2 × 107 bone marrow cells per mouse. Mouse cohorts are described in Supplementary Table 2.

Mouse blood and tissue collection. Peripheral blood samples were collected every 2 weeks starting at 8 weeks of age, using retro-orbital sampling. Whole blood was blocked in FBS (Mediatech) for 30 min., the red blood cells were lysed using Pharmlyse solution (BD Biosciences) and cells were washed with PBS. Tissue samples were collected at necropsy and processed immediately for cell isolation and FACS analysis, or kept in freezing media (IMDM plus 20% DMSO) in liquid nitrogen, for later analysis and DNA extraction. Tissue samples were manually agitated in PBS before filtering through a sterile 70 µm nylon mesh screen (Fisher Scientific) and suspension cell preparations produced as previously described19. Intestinal samples were processed as previously described73, with the modification that the mononuclear cell population was isolated after incubation in citrate buffer and collagenase enzyme for 2 h, followed by nylon wool filtration (Amersham Biosciences) and ficoll-hypaque gradient isolation (GE Healthcare).

Analysis of CCR5 disruption. The percentage of CCR5 alleles disrupted by ZFN treatment was measured by performing PCR across the ZFN target site followed by digestion with the Surveyor (Cel 1) nuclease (Transgenomic), which detects heteroduplex formation, as previously described19. Briefly, genomic DNA was extracted from mouse tissues and subject to nested PCR amplification using human CCR5-specific primers, with the resulting radiolabeled products digested with Cel 1 nuclease and resolved by PAGE. The ratio of cleaved to uncleaved products was calculated to give a measure of the frequency of gene disruption. The assay is sensitive enough to detect single-nucleotide changes and has a linear detection range between 0.69 and 44%19. In addition, a common 5-bp (pentamer) duplication that occurs after nonhomologous end-joining repair of ZFN-cleaved CCR5 (ref. 19) was detected by PCR. The first-round PCR product generated during Cel 1 analysis was diluted 1:5,000 and 5 µl used in a Taqman qPCR reaction using primers (5′-GGTCATCCTCATCCTGATCTGA-3′ and 5′-GATGATGAAGAAGATTCCAGAGAAGAAG-3′) and probe 5′-FAM d

Analysis of human cells in mouse tissues. FACS analysis of human cells was performed using a FACSCalibur instrument (BD Biosciences) with either BD CellQuest Pro version 5.2 (BD Biosciences) or FlowJo software version 8.8.6 for Macintosh (Treestar). The gating strategy performed was an initial forward scatter versus side scatter (FSC/SSC) gate to exclude debris, followed by a human CD45 gate. For analysis of lymphocyte populations in peripheral blood, a further lymphoid gate (low side scatter) was also applied to exclude cells of monocytic origin74. All antibodies used were fluorochrome conjugated and human specific, and obtained from BD Biosciences: CD45 (clone 2D1), CD19 (clone HIB19), CD14 (clone MϕP9), CD3 (clone SK7), CD4 (clone SK3), CD8 (clone HIT8a), CCR5 (2D7). Gates were set using fluorescence minus one controls, where cells were stained with all antibodies except the one of interest. Specificity was also confirmed using isotype-matched nonspecific antibodies (BD Biosciences) (Supplementary Fig. 1) and with tissues from animals that had not been engrafted with human cells. Immunohistochemical analysis of human CD3 and CD4 expression, respectively, in the small intestine and spleen tissue from HSPC-engrafted

nature biotechnology

doi:10.1038/nbt.1663

mice was performed on fixed paraffin-embedded tissue sections, as previously described73. Controls included isotype-matched nonspecific antibodies and unengrafted NSG mice.

© 2010 Nature America, Inc. All rights reserved.

Statistical analysis. All statistical analysis was performed using GraphPad Prism version 5.0b for Mac OSX (GraphPad Software). Unpaired two-tailed t-tests were performed assuming equal variance to calculate P-values. A 95% confidence interval was used to determine significance. A minimum of three data points was used for each analysis.

71. Shultz, L.D. et al. Human lymphoid and myeloid cell development in NOD/LtSz-scid IL2R gamma null mice engrafted with mobilized human hematopoietic stem cells. J. Immunol. 174, 6477–6489 (2005). 72. Rouet, F. et al. Transfer and evaluation of an automated, low-cost real-time reverse transcription-PCR test for diagnosis and monitoring of human immunodeficiency virus type 1 infection in a West African resource-limited setting. J. Clin. Microbiol. 43, 2709–2717 (2005). 73. Sun, Z. et al. Intrarectal transmission, systemic infection, and CD4+ T cell depletion in humanized mice infected with HIV-1. J. Exp. Med. 204, 705–714 (2007). 74. Loken, M.R. et al. Establishing lymphocyte gates for immunophenotyping by flow cytometry. Cytometry 11, 453–459 (1990).

doi:10.1038/nbt.1663

nature biotechnology

ar t ic l e s

© 2010 Nature America, Inc. All rights reserved.

Cell type of origin influences the molecular and functional properties of mouse induced pluripotent stem cells Jose M Polo1–4, Susanna Liu5, Maria Eugenia Figueroa6, Warakorn Kulalert1–4, Sarah Eminli1–4, Kah Yong Tan1,4,7, Effie Apostolou1–4, Matthias Stadtfeld1–4, Yushan Li6, Toshi Shioda2, Sridaran Natesan8, Amy J Wagers1,4,7, Ari Melnick6, Todd Evans5 & Konrad Hochedlinger1–4 Induced pluripotent stem cells (iPSCs) have been derived from various somatic cell populations through ectopic expression of defined factors. It remains unclear whether iPSCs generated from different cell types are molecularly and functionally similar. Here we show that iPSCs obtained from mouse fibroblasts, hematopoietic and myogenic cells exhibit distinct transcriptional and epigenetic patterns. Moreover, we demonstrate that cellular origin influences the in vitro differentiation potentials of iPSCs into embryoid bodies and different hematopoietic cell types. Notably, continuous passaging of iPSCs largely attenuates these differences. Our results suggest that early-passage iPSCs retain a transient epigenetic memory of their somatic cells of origin, which manifests as differential gene expression and altered differentiation capacity. These observations may influence ongoing attempts to use iPSCs for disease modeling and could also be exploited in potential therapeutic applications to enhance differentiation into desired cell lineages.

IPSCs are usually obtained from fibroblasts after infection with viral constructs expressing the four transcription factors Oct4, Sox2, Klf4 and c-Myc1–10. In addition, other cell types, including blood2,4,11, stomach and liver cells1, keratinocytes12,13, melanocytes14, pancreatic β cells7 and neural progenitors3,15–17 have been reprogrammed into iPSCs. Although these iPSC lines have been shown to express pluripotency genes and support the differentiation into cell types of all three germ layers, recent studies detected substantial molecular and functional differences among iPSCs derived from distinctive cell types. For example, iPSCs produced from various fibroblasts, stomach and liver cells showed different propensities to form tumors in mice, although the underlying molecular mechanisms remain elusive18. Another study identified persistent donor cell–specific gene expression patterns in human iPSCs produced from different cell types, suggesting an influence of the somatic cell of origin on the molecular properties of resultant iPSCs19. Whether cellular origin also affected the functional properties of iPSCs remained unexplored in that report. Of note, the findings of some of these studies may be confounded by the presence of different viral insertions in individual iPSC lines and by the fact that the analyzed iPSC lines were of different genetic background, which can affect both gene expression patterns20 and the functionality9,21 of cells. Indeed, we have recently shown that many mouse iPSC lines derived from different somatic cell types show aberrant silencing of a surprisingly small set of transcripts compared with embryonic stem cells (ESCs)22. However, our study did not investigate

whether additional cell-of-origin–specific differences may exist in iPSC lines derived from different cell types. Patient-specific iPSCs are a valuable tool for the study of disease and possibly for the development of therapies20,23–26. Thus, resolving the question of whether iPSCs produced from different cell types are molecularly and functionally equivalent is crucial for using these cells to model disease, which entails detecting subtle differences in the differentiation potential of patient-derived iPSCs24,27. Furthermore, the identification of somatic cells that influence the differentiation capacities of resultant iPSCs into desired cell lineages could be useful in a therapeutic setting. To assess whether iPSCs derived from different somatic cell types are distinguishable, we compared here the transcriptional and epigenetic patterns, as well as the in vitro differentiation potentials, of iPSCs produced from four genetically identical adult mouse cell types that differed only in the lineage from which they were derived. RESULTS Genetically matched iPSCs derived from different cell types Because the genetic background of ESCs can influence their transcriptional and functional behaviors, we used a previously described ‘secondary system’ to generate genetically identical iPSCs2,28 (Fig. 1a). Briefly, iPSCs were generated from somatic cells using doxycyclineinducible lentiviruses expressing Oct4, Sox2, Klf4 and c-Myc 29, and then injected into blastocysts to produce isogenic chimeric mice.

1Howard Hughes Medical Institute and Department of Stem Cell and Regenerative Biology, Harvard University and Harvard Medical School, Cambridge, Massachusetts, USA. 2Massachusetts General Hospital Cancer Center, Charlestown, Massachusetts, USA. 3Massachusetts General Hospital Center for Regenerative Medicine, Boston, Massachusetts, USA. 4Harvard Stem Cell Institute, Cambridge, Massachusetts, USA. 5Department of Surgery, Weill Cornell Medical College, New York, New York, USA. 6Department of Medicine, Hematology Oncology Division, Weill Cornell Medical College, New York, New York, USA. 7Joslin Diabetes Center, Boston, Massachusetts, USA. 8Sanofi-Aventis Cambridge Genomics Center, Cambridge, Massachusetts, USA. Correspondence should be addressed to K.H. ([email protected]).

Received 26 March; accepted 9 July; published online 19 July 2010; doi:10.1038/nbt1667

848

volume 28 number 8 august 2010 nature biotechnology

ar t ic l e s a Blast

injection

Chi no. 1

Granulocytes

Gra-iPSC

SMP cells

SMP-iPSC

• Gene expression • DNA methylation

+ dox

Secondary iPSC clone (carry dox-inducible copies of Oct4, Sox2, Klf4, c-Myc)

• ChIP for histone modifications

Chi no. 2

B cells

B-iPSC

TTFs

TTF-iPSC

• In vitro differentiation

b Gr-1

Itgb1

Lysozyme 0.05

0.10

0.80

0.04

0.04

0.08 0.06 0.04 0.02

0.60 0.40 0.20

0.00

0.00

GraiPSC

c

SMPiPSC

Chi no. 1

1 2 3 1 2 3

GraiPSC

SMPiPSC

0.03 0.02 0.01 0.00

GraiPSC

Chi no. 2

1 2 3 1 2 3

BiPSC

Fold GAPDH

0.05

Fold GAPDH

1.00

Fold GAPDH

Fold GAPDH

© 2010 Nature America, Inc. All rights reserved.

Cxcr4 0.12

TTFiPSC

0.02 0.01 0.00

GraiPSC

SMPiPSC

0.03

SMPiPSC

GraiPSC

SMPiPSC

d SMP-iPSC1

B-iPSC3

SMP-iPSC2

B-iPSC1

SMP-iPSC3

B-iPSC2

Gra-iPSC3

TTF-iPSC3

Gra-iPSC1

TTF-iPSC1

Gra-iPSC2

TTF-iPSC2

Chi no. 1

Chi no. 2

Figure 1 iPSCs derived from different cell types are transcriptionally distinguishable. (a) Flow chart explaining the derivation and analysis of genetically matched iPSCs from different cell types. Secondary iPSCs were first injected into blastocysts to generate chimeric mice, from which the indicated somatic cell types were isolated. Exposure of these cells to doxycycline (dox) then gave rise to iPSCs. ChIP, chromatin immunoprecipitation. (b) Quantification of the expression levels of Cxcr4, Itgb1, Gr-1 and Lysozyme by quantitative PCR in SMP-iPSCs, in red, and Gra-iPSCs, in gray. The values were normalized to GAPDH expression; the error bars depict the s.e.m. (n = 3). (c) Heat map showing top 104 probes with highest variance in their expression levels. Left panel, SMP-iPSCs and Gra-iPSCs derived from chimera no. 1. Right panel, TTF-iPSCs and B-iPSCs derived from chimera no. 2. (d) Hierarchical, unsupervised clustering of iPSC expression profiles using the correlation distance and the Ward method. SMP-iPSCs and Gra-iPSCs were derived from chimera no. 1 (left panel), TTF-iPSCs and B-iPSCs originate from chimera no. 2 (right panel). Chi no. 1, chimera no. 1; chi no. 2, chimera no. 2.

Thus, isolation of different cell types from these chimeras and their subsequent exposure to doxycycline gave rise to iPSCs with the same genetic makeup. In this study, we focused on iPSCs derived from tail tip–derived fibroblasts (TTFs), splenic B cells (B), bone marrow– derived granulocytes and skeletal muscle precursors (SMPs)30, which were continuously cultured for 2–3 weeks (passage 4 to 6) after picking. The pluripotency of some of these cell lines has been previously documented2, or was analyzed in this study (Supplementary Table 1 and Supplementary Fig. 1). All cell lines grew at similar rates and independently of viral transgene expression (Supplementary Fig. 2) and upregulated the endogenous pluripotency genes Nanog, Sox2 and Oct4, indicating successful molecular reprogramming (Supplementary Table 1). Moreover, all lines gave rise to differentiated teratomas, and all tested lines supported the development of nature biotechnology volume 28 number 8 august 2010

chimeric animals upon blastocyst injection, demonstrating their pluripotency (Supplementary Table 1). We therefore concluded that the cell lines analyzed here qualify as bona fide iPSC lines. iPSCs produced from different cell types are transcriptionally distinguishable We first evaluated whether iPSCs derived from defined somatic cell types retain gene expression patterns indicative of their cells of origin. Specifically, we assessed the expression of cell lineage–specific candidate genes in iPSCs derived from granulocytes (Gra-iPSCs) and SMPs (SMP-iPSCs). As expected, the SMP markers Cxcr4 and Integrin B1 and the granulocyte markers Lysozyme (also known as Lyz1 and Lyz2) and Gr-1 (also known as Ly6g) were expressed at considerably higher levels in the somatic cells of origin than in resultant 849

ar t ic l e s

a Chi no. 1

b

Chi no. 2 B-iPSC1

Gra-iPSC2 Gra-iPSC1

d = 0.02

B-iPSC2

Gra-iPSC3

TTF-iPSC1

SMP-iPSC2 SMP-iPSC3

TTF-iPSC3

SMP-iPSC1

TTF-iPSC2

c

Chi no. 1

Chi no. 2

Percent of methylation 0%

100%

Not analyzed

SMP-iPSC1 SMP-iPSC2 SMP-iPSC3 Gra-iPSC1 Gra-iPSC2 Gra-iPSC3 ESC

Gr-1

Lysozyme

Cxcr4

Itgb1

d Gr-1

0.12

H3Ac

0.5

Percent input

Percent input

Cxcr4

0.6

0.1 0.08 0.06 0.04 0.02

H3K4me3

0.4

H3K27me3

0.3

IgG

0.2 0.1

0

Gra

SMP

GraiPSC

0

SMPiPSC

Lysozyme

Gra

SMP

0.3

0.2

GraiPSC

SMPiPSC

Itgb1

0.8

Percent input

0.4

Percent input

© 2010 Nature America, Inc. All rights reserved.

d = 0.02

B-iPSC3

0.6

0.4

0.2

0.1

0

0

Gra

SMP

GraiPSC

SMPiPSC

Gra

SMP

GraiPSC

SMPiPSC

Figure 2 iPSCs derived from different cell types exhibit distinguishable epigenetic signatures. (a) Hierarchical unsupervised clustering analysis of HELP genome-wide methylation data from indicated iPSC lines. (b) Correspondence analysis of SMP-iPSCs and Gra-iPSCs (left panel) from chimera no. 1, TTF-iPSCs and B-iPSCs (right panel) from chimera no. 2. (c) Graphic representation of DNA methylation quantification of specific CpGs (circles) in the promoter regions of the indicated candidate genes using EpiTYPER DNA methylation analyses. Yellow indicates 0% methylation and blue 100% methylation. (d) Chromatin immunoprecipitation (ChIP) for H3 pan-acetylated (H3Ac, in blue), H3K4 trimethylated (H3K4me3, in green), H3K27 trimethylated (H3K27me3, in red) and isotype control (IgG, in light blue) of granulocytes (Gra), SMPs, Gra-iPSCs and SMP-iPSCs. Chi no. 1, chimera no. 1; chi no. 2, chimera no. 2. The error bars depict the s.e.m. (n = 3).

850

volume 28 number 8 august 2010 nature biotechnology

ar t ic l e s between different experiments and individual animals. All iPSC lines analyzed were between passage (p) EPO B-iPSC eryPs 4 and 6. There were 1,388 genes dif4 days Dissociate and ferentially expressed (twofold, corTTF-iPSC 5,000 cells/ml plate 100,000/ml IL-3/M-CSF rected P = 0.05) between SMP-iPSCs Macrophages 6 days 7 days and Gra-iPSCs, and 1,090 genes Gra-iPSC cytokines between B-iPSCs and TTF-iPSCs Mixed colonies EBs 8 days (Supplementary Table 2). An analySMP-iPSC sis of the 100 genes with the greatest range of expression levels across all samples indicated that iPSCs Gra-iPSC B-iPSC TTF-iPSC SMP-iPSC b with the same cell of origin clustered together (Fig. 1c). Consistent with this observation, unsupervised hierarchical clustering (Fig. 1d) as well as principal component analysis (Supplementary Fig. 4) of all genes placed SMP-iPSCs and c d Gra-iPSCs, as well as B-iPSCs and 8 7 TTF-iPSCs, into different groups P < 0.001 6 according to their cells of origin. 5 4 Notably, Gene Ontology (GO) 3 analysis of the 100 genes with the 2 greatest range of expression between 1 0 SMP-iPSCs and Gra-iPSCs indicated BTTFGraSMPeryPs Macrophages Mixed colonies iPSC iPSC iPSC iPSC an enrichment for genes belonging Chi no. 2 Chi no. 1 to the categories ‘myofibril’ (7.6fold enrichment), ‘contractile fiber’ e f g (7.3-fold enrichment) and ‘muscle 2,000 2,500 250 250 3 4 development’ (5.9-fold enrichP < 0.001 P < 0.05 P < 0.05 P < 0.07 1,600 2,000 200 200 ment) as well as ‘B-cell activation’ 3 2 1,200 1,500 150 150 (6.8-fold enrichment) and ‘leuko2 800 100 100 1,000 cyte activation’ (3.7-fold enrich1 1 50 400 50 500 ment) (when compared with the 0 0 0 0 0 0 expected background). Together, BTTFBTTFGraSMPBTTFGra- SMPGraSMPiPSC iPSC iPSC iPSC iPSC iPSC iPSC iPSC iPSC iPSC iPSC iPSC these results show that genetically Chi no. 2 Chi no. 1 Chi no. 2 Chi no. 1 Chi no. 2 Chi no. 1 identical iPSCs obtained from four Figure 3 iPSCs derived from different cell types have distinctive in vitro differentiation potentials. (a) Experimental different somatic cell types are disoutline. iPSCs were first differentiated into embryoid bodies. At day 6, embryoid bodies were dissociated and tinguishable from each other using plated in conditions to favor differentiation into erythrocyte progenitors (eryP) and macrophage and mixed genome-wide transcriptional analyhematopoietic colonies. (b) Phase contrast images showing embryoid bodies derived from B-iPSCs, TTF-iPSCs, ses, further supporting the notion Gra-iPSCs and SMP-iPSCs at same magnification. (c) Quantification of embryoid body sizes derived from B-iPSCs, that the donor cell type influences TTF-iPSCs, Gra-iPSCs and SMP-iPSCs; the diameter of the embryoid bodies was measured using arbitrary units the overall gene expression pattern (AU). The error bars depict the s.e.m. (n = 30) (d) Representative images of erythrocyte progenitors (eryPs), of resultant iPSCs. macrophage colonies and mixed hematopoietic colonies. (e–g) Quantification of in vitro differentiation potentials of the different iPSCs into EryPs (e), macrophage colonies (f) and mixed hematopoietic colonies (g). Chi no. 1, To determine the effect on gene chimera no. 1; chi no. 2, chimera no. 2. The error bars depict the s.e.m. (n = 12). expression patterns of deriving iPSCs from different animals in iPSCs (Supplementary Fig. 3). Moreover, SMP-iPSCs expressed sub- independent experiments, we compared the expression profiles of stantially higher levels of Cxcr4 and Itgb1 than did Gra-iPSCs (Fig. Gra-iPSCs derived from chimera no. 1 (n = 3) with Gra-iPSCs from 1b), and Gra-iPSCs showed higher expression levels of Lysozyme chimera no. 2 (n = 3) as well as with SMP-iPSCs from chimera no. 1 and Gr-1 compared with SMP-iPSCs (Fig. 1b). Together, these data and TTF-iPSCs from chimera no. 2 (Fig. 1a). Hierarchical clustering suggest that iPSCs retain a transcriptional memory of their somatic separated Gra-iPSCs according to their origin from different animals, cell of origin. suggesting a significant contribution of this experimental variable to To test this notion globally, we compared the transcriptional profiles gene expression patterns (Supplementary Fig. 5). However, when the of iPSC lines originating from SMPs (n = 3) with those derived from expression data from TTF-iPSCs and SMP-iPSCs were included in the granulocytes (n = 3), as well as expression profiles of iPSC lines origi- analysis, we found that differences due to cell of origin were stronger nating from B cells (n = 3) with those produced from TTFs (n = 3). than those arising from variations in experimental conditions or aniNote that iPSCs were compared with each other only if they originated mals. These data reinforce the observation that iPSCs derived from from the same chimeric mouse (SMP-iPSCs versus Gra-iPSCs and different somatic cell types are transcriptionally distinguishable, even B-iPSCs versus TTF-iPSCs) (Fig. 1a) to eliminate potential variability when they originate from different animals. nature biotechnology volume 28 number 8 august 2010

Mixed colonies

Mixed colonies

Macrophage colonies

Macrophage colonies

EryP colonies

EB diameter (in arbitrary units)

EryP colonies

© 2010 Nature America, Inc. All rights reserved.

a

851

b

a

2,500

T-iPSC2

TTF iPSC1

B-iPSC3

TTF iPSC2 Gra-iPSC3

TTF-iPSC1

Gra-iPSC1 Gra-iPSC2

B-iPSC2

T-iPSC2 TTF-iPSC3 B-iPSC2 Gra-iPSC2 TTF-iPSC2

TTF-iPSC2

Gra-iPSC3

B-iPSC1 TTF-iPSC3

B-iPSC3 B-iPSC1 B-iPSC2

p4

T-iPSC3 Gra-iPSC1

T-iPSC1

B-iPSC3

T-iPSC3

TTF-iPSC1

p10

c Gra-iPSC p4 vs. p16

1,000

Organ development:

Functional cluster

EGLN1, EN2, AA409316, GYS1, IQGAP2, LOXL3, MGP, NDRG1, NOPE, PHF21A, BC021588, CYB5R3, SNRPD3, NM_008681, NM_030247

Tube morphogenesis

2.75

Positive regulation of cellular process

2.09

Morphogenesis of a branching structure

2.04

508

15

500

T-iPSC T-iPSC B-iPSC T-iPSC B-iPSC Gra-iPSC vs. vs. vs. vs. vs. vs. B-iPSC TTF-iPSC TTF-iPSC Gra-iPSC Gra-iPSC TTF-iPSC

125

56 TTF-iPSC p4 vs. p16

1,500

0

685

68

p16

2,000

p16

Enrichment score

Response to heat

1.98

Organ development

1.96

mRNA metabolic process

1.95

Cellular component assembly

1.79

Cartilage and skeletal development

1.75

Regulation of cell cycle

1.69

Tissue development

1.41

Spermatogenesis

1.40

474 B-iPSC p4 vs. p16

d

f 900 800 700 600 500 400 300 200 100 0

Mixed colonies

TTF-iPSC3

B-iPSC2

TTF iPSC1

140 120 100 80 60 40 20 0

B-iPSC

g

B-iPSC1

Macrophage colonies

EryP colonies

e

To exclude the possibility that the observed gene expression differences were due to the specific secondary system used, we derived iPSCs from SMPs, granulocytes, B cells and peritoneal fibroblasts from reprogrammable mice31, which carry dox-inducible copies of all four reprogramming factors in a defined genomic locus. All iPSC lines grew independently of dox and gave rise to differentiated teratomas (Supplementary Fig. 6a). Analysis of gene expression profiles of these lines at p4 showed clustering according to their cells of origin, with the exception of peritoneal fibroblast–derived iPSCs, which may be a consequence of the heterogeneity of the starting population. Collectively, these results corroborate the notion that iPSCs 852

p10

T-iPSC1

Gra-iPSC2 Gra-iPSC3

TTF-iPSC3

p4

B-iPSC1

Gra-iPSC1

Differentially expressed probes

T-iPSC2 T-iPSC1 T-iPSC3

B-iPSC3

Figure 4 Continuous passaging of iPSCs abrogates transcriptional, epigenetic and functional differences. (a) Hierarchical unsupervised clustering of expression profiles from B-iPSCs, T-iPSCs, TTFiPSCs and Gra-iPSCs from chimera no. 2. Left panel shows clustering analysis of all iPSC samples at passage p4, the middle panel at p10 and the right panel at p16. (b) Number of differentially expressed probes between pairs of iPSC samples used in a; iPSCs at p4 are shown in blue bars, iPSCs at p10 are shown in orange bars and iPSCs at p16 are shown in red bars. The number of differently expressed probes between iPSCs was calculated using a pairwise analysis (twofold), with t-test P = 0.05, with Bejamini and Hochberg correction (n = 3). (c) Venn diagram and GO analysis showing overlap of genes that change from p4 to p16 in Gra-IPSCs, TTFiPSCs and B-iPSCs. Red line marks functional GO cluster of genes shared between all three iPSC groups. Black line marks functional GO cluster of genes shared by at least two of the iPSC groups. Functional ontology cluster analysis was performed using the DAVIS algorithm. (d) Hierarchical unsupervised clustering using HELP genome-wide methylation profiles of B-iPSCs and TTF-iPSCs at p16. (e–g) Quantification of in vitro differentiation potentials of B-iPSCs and TTF-iPSCs at p16 into EryPs (e), macrophage colonies (f) and mixed hematopoietic colonies (g). The error bars depict the s.e.m. (n = 9).

TTF iPSC2

© 2010 Nature America, Inc. All rights reserved.

ar t ic l e s

TTF-iPSC

B-iPSC

TTF-iPSC

1.5 1 0.5 0

B-iPSC

TTF-iPSC

generated from different cell types exhibit distinct transcriptional patterns (Supplementary Fig. 6b). iPSCs derived from different cell types exhibit distinguishable epigenetic patterns We next asked whether the differential gene expression patterns we observed correlated with differences in epigenetic marks. To this end, we performed a genome-wide, restriction enzyme–based methylation analysis of promoters termed ‘HpaII tiny fragment enrichment by ligationmediated PCR’ (HELP) on the same cell lines we used for expression analysis. Unsupervised hierarchical clustering showed that Gra-iPSCs volume 28 number 8 august 2010 nature biotechnology

ar t ic l e s

© 2010 Nature America, Inc. All rights reserved.

Reprogramming (transgene-dependent phase)

Reprogramming (transgene-independent phase)

two genes. A similar pattern was observed for the granulocyte-specific genes in Gra-iPSCs compared with SMP-iPSCs, with Gr-1 and Lysozyme being elevated for H3K4me3 (Fig. 2d). These data show that the observed expression differences among iPSCs derived from different cell types may be predominantly the consequence of differences in histone marks, further suggesting that iPSCs retain an epigenetic memory of their cells of origin.

iPSCs derived from different cell types have distinctive in vitro differentiation potentials Because the gene expression differences we observed among different iPSC lines affected genes known to Cell of origin be involved in the lineage-specific differentiation and function of the somatic cell types from which they were derived, we reasoned that these differEarly passage iPSC Late passage iPSC Partially reprogrammed cells ences might affect their capacity to differentiate • No endogenous pluripotent • Activation of endogenous • Activation of endogenous gene expression pluripotency genes pluripotency genes into defined cell lineages. Thus, we evaluated the • No contribution to chimeras • Promoter demethylation • Promoter demethylation autonomous differentiation potential of the four • Teratoma formation • Teratoma formation • Teratoma formation types of iPSC lines by assessing their abilities to • Chimera contribution • Chimera contribution • Transcriptionally distinguishable produce embryoid bodies, erythrocyte progenitors, • Transient epigenetic memory macrophages and mixed hematopoietic colonies • Altered differentiation using established semiquantitative differentiation protocols (Fig. 3a). Most notably, TTF-iPSCs proFigure 5 Model summarizing the presented data. iPSCs derived from different somatic cell duced significantly smaller and fewer embryoid types retain a transient epigenetic and transcriptional memory of their cell type of origin at early bodies compared with all the other iPSC lines (P passage, despite acquiring pluripotent gene expression, transgene-independent growth and the < 0.001; Fig. 3b,c). Moreover, the embryoid bodability to contribute to tissues in chimeras. Continuous passaging resolves these differences, giving rise to iPSCs that are molecularly and functionally indistinguishable. Note the difference ies derived from TTF-iPSC generated relatively between early passage iPSCs and partially reprogrammed cells, which require continuous few erythrocyte, macrophage and mixed colony viral transgene expression and fail to activate endogenous pluripotency genes or support the progenitors compared with B-iPSCs derived from development of viable mice. the same animal despite equal numbers of input cells, indicating striking differences in the differand SMP-iPSCs, as well as B-iPSCs and TTF-iPSCs, which clustered entiation potentials of these iPSCs (Fig. 3d–g). In contrast, SMP-iPSCs separately in the transcriptional assays, were also distinguishable based and Gra-iPSCs showed equivalent abilities to produce embryoid bodies on their methylation patterns (Fig. 2a). Correspondence analysis of the (Fig. 3d–g). However, Gra-iPSCs gave rise to erythrocyte, macrophages same samples corroborated this finding (Fig. 2b), indicating that the and mixed colonies at higher efficiencies than SMP-iPSCs, suggesting donor cell type affects not only the overall transcriptional pattern but a pattern of differentiation that reflects their cells of origin. Together, these data show that the cell type of origin may bias the differentiation also the promoter methylation pattern of resultant iPSCs. Despite the separation of Gra-iPSCs from SMP-iPSCs and of potential of resultant iPSC lines. TTF-iPSCs from B-iPSCs (Fig. 2a,b) by hierarchical clustering, we detected few loci that were differentially methylated with statisti- Continuous passaging of iPSCs abrogates transcriptional, cal significance using supervised analysis (69 genes between Gra- epigenetic and functional differences iPSCs and SMP-iPSCs and 0 genes between B-iPSCs and TTF-iPSCs; Previously published data suggest that early-passage, human iPSCs Supplementary Table 3). To complement these results, we interro- derived from fibroblasts are transcriptionally distinct from late-passage gated the DNA methylation status at the promoter regions of the iPSCs32. However, that study did not examine the effect of passaging on previously analyzed markers Cxcr4, Itgb1, Lysozyme and Gr-1 (Fig. the iPSC functionality. We therefore wondered whether continuous pas1b) using EpiTYPER DNA methylation analysis, which quantifies saging of the various iPSC lines would eliminate the observed differences gene-specific CpG methylation. We failed to detect differences in the in gene expression and differentiation potential. For this analysis, we methylation levels of these candidate genes between SMP-iPSCs and added to the B-iPSC/TTF-iPSC group, studied before (Figs. 1 and 2a,b), a Gra-iPSCs (Fig. 2c), further indicating that methylation differences new set of T cell– and granulocyte-derived iPSCs, which were all derived are more subtle than the observed gene expression differences and from chimera no. 2. These 12 iPSC lines were subjected to several addiraising the possibility that other chromatin marks may be responsible tional rounds of passaging under identical culture conditions, and RNA for the observed expression differences. was harvested at p10 and p16 for expression profiling. Whereas unsuIndeed, we observed high levels of the activating marks H3Ac and pervised hierarchical clustering of these cell lines at early passage (p4) H3K4me3 and low levels of the repressive marks H3K27me3 at the pro- clearly separated each of the different iPSC lines according to their cells moters of Cxcr4 and Itgb1 in SMPs and at the promoters of Lysozyme and of origin (Fig. 4a, left panel), unsupervised clustering of these lines at p10 Gr-1 in granulocytes, respectively, consistent with their abundant expres- showed that B-iPSCs, TTF-iPSCs and T-iPSCs were indistinguishable sion in these cell types (Fig. 2d). Notably, SMP-iPSCs, which showed from each other, whereas the Gra-iPSCs still clustered together (Fig. 4a, higher expression levels of Cxcr4 and Itgb1 than did Gra-iPSCs (Fig. middle panel). Further passaging of these cells until p16 entirely elimi1b), were enriched for H3K4me3 compared with Gra-iPSCs at these nated these differences (Fig. 4a, right panel). Together, these data indinature biotechnology volume 28 number 8 august 2010

853

© 2010 Nature America, Inc. All rights reserved.

ar t ic l e s cate that continuous cell division resolves transcriptional differences among iPSC lines. Consistent with this observation, the total number of differentially expressed genes between various pairs of iPSC lines derived from different cellular origins was reduced from ~500–2,000 in early-passage cultures to only ~50 or even 0 in late-passage cultures, further demonstrating that after extensive in vitro propagation, these iPSC lines have become very similar to each other (Fig. 4b). Analysis of the genes whose expression changed between p4 and p16 in Gra-iPSCs, B-iPSCs and TTF-iPSCs showed 25% overlap with at least one of the other two groups of iPSC lines, suggesting that iPSCs undergo some common changes during passaging, irrespective of their cell of origin (Fig. 4c). GO analysis of these changes indicated a strong enrichment for developmental regulators. Moreover, the only GO cluster common to all three groups was ‘organ development’, indicating that the passaging of iPSCs results in a change of differentiation-associated gene expression patterns (Fig. 4c). The expression levels of the pluripotency genes Sox2 and Oct4, which are high already at early passage (Supplementary Table 1), increased even further during the passaging process, supporting the notion that the pluripotency network becomes increasingly solidified during culture (Supplementary Fig. 7), consistent with a previous report showing gradual upregulation of pluripotency-associated genes upon passaging of human iPSC lines32. To evaluate whether the passaging of iPSCs attenuates the observed epigenetic differences, we performed HELP analysis on B-iPSCs and TTF-iPSCs at late passage. In contrast to early-passage iPSCs, the latepassage iPSCs could not be separated by hierarchical unsupervised clustering analysis based on their cells of origin (Fig. 4d). Accordingly, the methylation levels of histones at candidate genes in Gra-iPSCs and SMP-iPSCs became indistinguishable (Supplementary Fig. 8). Notably, several of the analyzed loci showed an enrichment for both H3K4me3 and H3K27me3, indicative of bivalent domains that are characteristic of pluripotent stem cells33. Thus, continuous passaging leads to an equilibration of the epigenetic differences detected in early-passage iPSCs. Two possible mechanisms could account for the observed loss of epigenetic and transcriptional memory with increased passage number: (i) passive replication-dependent loss of somatic marks in the majority of iPSCs and (ii) selection of rare, preexisting, fully reprogrammed cells over time. Because the selection model predicts that such rare clones would have a growth or survival advantage, we would expect to see impaired growth rates of bulk iPSC cultures at early passage compared with late passage, which we did not observe (Supplementary Fig. 9a). We also did not detect significant differences when the growth rates of single-cell clones established from early and late passage iPSC lines were examined using a colorimetric assay (XTT assay) that detects metabolic activity (Supplementary Fig. 10) or by measuring the increase in cell numbers on three consecutive days (Supplementary Figs. 11 and 12). Similarly, an analysis of the colony formation efficiency of single cell-sorted iPSC from early- and late-passage cultures did not yield detectable differences (Supplementary Fig. 9b). Collectively, these data argue against the presence of rare subclones that become selected over time and are consistent with the notion that all iPSC lines gradually resolve transcriptional and epigenetic differences with increased passaging. However, our results do not exclude a combined model involving passive resolution of epigenetic marks as well as selection of multiple clones. Finally, we asked whether the similar transcriptional and epigenetic patterns of late-passage iPSCs derived from distinct cells of origin would translate into an equalization of their differentiation potentials. We first performed an embryoid-body formation assay at different passages for TTF-iPSCs and B-iPSCs, which showed a strong difference at early passage. TTF-iPSCs gave rise to similarly-sized embryoid bodies as B-iPSCs 854

around p10–p12 (Supplementary Fig. 13a,b) and were indistinguishable at p16 (Supplementary Fig. 13c,d). Moreover, embryoid bodies derived from TTF-iPSCs and B-iPSCs at p16 differentiated into similar numbers of erythrocyte (Fig. 4e), macrophage (Fig. 4f) and mixed-colony progenitors (Fig. 4g), thus proving that extensive cellular passaging eliminates differences in the differentiation potentials of these iPSCs. DISCUSSION Our study shows that genetically matched iPSCs retain a transient transcriptional and epigenetic memory of their cell of origin at early passage, which can substantially affect their potential to differentiate into embryoid bodies and different hematopoietic cell types (Fig. 5). These molecular and functional differences are lost upon continuous passaging, however, indicating that complete reprogramming is a gradual process that continues beyond the acquisition of a bona fide iPSC state as measured by the activation of endogenous pluripotency genes, viral transgene–independent growth and the ability to differentiate into cell types of all three germ layers. Notably, the previously seen silencing of the Dlk1-Dio3 locus in many iPSC lines22 is not affected by the passaging of cells (data not shown). Of note, the early-passage iPSCs described here are different from “partially reprogrammed iPSCs”34,35, which depend on the continuous expression of viral transgenes and do not activate and demethylate pluripotency genes or contribute to the formation of viable chimeras (Fig. 5). The mechanism by which passaging eliminates the molecular and functional differences between iPSCs of different origins remains to be determined. Three key observations argue against the possibility of selective expansion of a rare subset of completely reprogrammed iPSCs: (i) both early- and late-passage iPSCs had similar proliferation rates; (ii) there was little variability in the growth rate of single-cell iPSC clones from early- and late-passage lines; and (iii) the number of passages required to resolve cell-of-origin differences was dependent upon the starting cell type. These observations suggest that the consolidation of the pluripotent transcriptional network upon passaging is a slow process, potentially facilitated by a positive feedback mechanism that gradually resolves the residual cell-of-origin–specific epigenetic marks and transcriptional patterns. In accordance with this idea is the finding that telomeres become gradually elongated with increased passage number of iPSCs36. Our results are also consistent with the previous observation that cloned embryos often retain donor cell–specific transcriptional patterns and do not efficiently activate embryonic genes over many cell divisions37–40, suggesting possible similarities in the mechanisms of reprogramming by nuclear transfer and induced pluripotency. Because of the lack of ESC lines genetically matched to the secondary iPSC lines used here, we did not include ESC lines in our comparative analysis. Nevertheless, the present results may help to explain some of the previously reported differences between ESCs and iPSCs41,42. Some of these studies compared late-passage ESC lines with iPSC lines of undefined, but presumably earlier, passage that may not yet have reached an ESC-equivalent ground state. It should be informative to revisit these studies with genetically matched, transgene-free late-passage iPSCs to determine whether this abrogates such gene expression and differentiation differences. The observed tendency of early-passage iPSC lines to differentiate preferentially into the cell lineage of origin could potentially be exploited in clinical settings to produce certain somatic cell types that have been difficult to obtain from ESCs thus far. However, these data also serve as a cautionary note for ongoing attempts to recapitulate disease phenotypes in vitro using patient-specific, early-passage iPSC lines, as the epigenetic, transcriptional and functional ‘immaturity’ of these cells might confound volume 28 number 8 august 2010 nature biotechnology

ar t ic l e s the data obtained from them. Further elucidation of the molecular indicators of fully reprogrammed iPSCs should help in the establishment of standardized iPSC lines that can be compared with confidence in basic biological and drug discovery studies. METHODS Methods and any associated references are available in the online version of the paper at http://www.nature.com/naturebiotechnology/. Accession code. GEO: GSE22043, GSE22827, GSE22908.

© 2010 Nature America, Inc. All rights reserved.

Note: Supplementary information is available on the Nature Biotechnology website. ACKNOWLEDGMENTS We thank N. Maherali and R. Walsh for helpful suggestions and critical reading of the manuscript, B. Wittner for statistical advice, J. LaVecchio, G. Buruzula, K. Folz-Donahue and L. Prickett for expert cell sorting and K. Coser for technical assistance. J.M.P. was supported by an MGH ECOR fellowship, E.A. by a Jane Coffin Childs fellowship, M.S. by a Schering fellowship and K.Y.T. by the Agency of Science, Technology and Research Singapore. Support to A.M. was from the Lymphoma Society, SCOR no. 7132-08; to T.E. from National Institutes of Health (NIH) grant HL056182 and NYSTEM; to A.J.W. in part from the Burroughs Wellcome Fund, Harvard Stem Cell Institute, Peabody Foundation, and NIH 1 DP2 OD004345-01, and the Joslin Diabetes Center DERC (P30DK036836); to K.H. from Howard Hughes Medical Institute, the NIH Director’s Innovator Award and the Harvard Stem Cell Institute. The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. AUTHOR CONTRIBUTIONS J.M.P. and K.H. conceived the study, interpreted results and wrote the manuscript; J.M.P. performed most of the experiments with help from W.K.; S.L. and T.E. performed and interpreted in vitro differentiation assays; M.E.F and A.M. performed and analyzed HELP methylation experiments; K.Y.T. and A.J.W. isolated SMPs and derived most SMP-iPSCs; T.S. and S.N. performed expression arrays; and S.E., E.A. and M.S. provided essential study material. All authors gave critical input to the manuscript draft. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/. Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/.

Note added in proof: We thank George Daley for sharing unpublished results, which show similar differences in DNA methylation patterns and differentiation propensity of iPSCs derived from distinctive cell types. Of note, this report43 also suggests that somatic cell nuclear transfer more faithfully reprograms cells to a pluripotent state than transcription factor overexpression. 1. Aoi, T. et al. Generation of pluripotent stem cells from adult mouse liver and stomach cells. Science 321, 699–702 (2008). 2. Eminli, S. et al. Differentiation stage determines potential of hematopoietic cells for reprogramming into induced pluripotent stem cells. Nat. Genet. 41, 968–976 (2009). 3. Eminli, S., Utikal, J., Arnold, K., Jaenisch, R. & Hochedlinger, K. Reprogramming of neural progenitor cells into induced pluripotent stem cells in the absence of exogenous Sox2 expression. Stem Cells 26, 2467–2474 (2008). 4. Hanna, J. et al. Direct reprogramming of terminally differentiated mature B lymphocytes to pluripotency. Cell 133, 250–264 (2008). 5. Lowry, W.E. et al. Generation of human induced pluripotent stem cells from dermal fibroblasts. Proc. Natl. Acad. Sci. USA 105, 2883–2888 (2008). 6. Park, I.H. et al. Reprogramming of human somatic cells to pluripotency with defined factors. Nature 451, 141–146 (2008). 7. Stadtfeld, M., Brennand, K. & Hochedlinger, K. Reprogramming of pancreatic beta cells into induced pluripotent stem cells. Curr. Biol. 18, 890–894 (2008). 8. Takahashi, K. et al. Induction of pluripotent stem cells from adult human fibroblasts by defined factors. Cell 131, 861–872 (2007). 9. Takahashi, K. & Yamanaka, S. Induction of pluripotent stem cells from mouse embryonic and adult fibroblast cultures by defined factors. Cell 126, 663–676 (2006).

nature biotechnology volume 28 number 8 august 2010

10. Yu, J. et al. Induced pluripotent stem cell lines derived from human somatic cells. Science 318, 1917–1920 (2007). 11. Loh, Y.H. et al. Generation of induced pluripotent stem cells from human blood. Blood 113, 5476–5479 (2009). 12. Aasen, T. et al. Efficient and rapid generation of induced pluripotent stem cells from human keratinocytes. Nat. Biotechnol. 26, 1276–1284 (2008). 13. Maherali, N. et al. A high-efficiency system for the generation and study of human induced pluripotent stem cells. Cell Stem Cell 3, 340–345 (2008). 14. Utikal, J., Maherali, N., Kulalert, W. & Hochedlinger, K. Sox2 is dispensable for the reprogramming of melanocytes and melanoma cells into induced pluripotent stem cells. J. Cell Sci. 122, 3502–3510 (2009). 15. Kim, J.B. et al. Pluripotent stem cells induced from adult neural stem cells by reprogramming with two factors. Nature 454, 646–650 (2008). 16. Shi, Y. et al. A combined chemical and genetic approach for the generation of induced pluripotent stem cells. Cell Stem Cell 2, 525–528 (2008). 17. Silva, J. et al. Promotion of reprogramming to ground state pluripotency by signal inhibition. PLoS Biol. 6, e253 (2008). 18. Miura, K. et al. Variation in the safety of induced pluripotent stem cell lines. Nat. Biotechnol. 27, 743–745 (2009). 19. Ghosh, Z. et al. Persistent donor cell gene expression among human induced pluripotent stem cells contributes to differences with human embryonic stem cells. PLoS One 5, e8975 (2010). 20. Soldner, F. et al. Parkinson’s disease patient-derived induced pluripotent stem cells free of viral reprogramming factors. Cell 136, 964–977 (2009). 21. Okita, K., Ichisaka, T. & Yamanaka, S. Generation of germline-competent induced pluripotent stem cells. Nature 448, 313–317 (2007). 22. Stadtfeld, M. et al. Aberrant silencing of imprinted genes on chromosome 12qF1 in mouse induced pluripotent stem cells. Nature 465, 175–181 (2010). 23. Dimos, J.T. et al. Induced pluripotent stem cells generated from patients with ALS can be differentiated into motor neurons. Science 321, 1218–1221 (2008). 24. Ebert, A.D. et al. Induced pluripotent stem cells from a spinal muscular atrophy patient. Nature 457, 277–280 (2009). 25. Park, I.H. et al. Disease-specific induced pluripotent stem cells. Cell 134, 877–886 (2008). 26. Saha, K. & Jaenisch, R. Technical challenges in using human induced pluripotent stem cells to model disease. Cell Stem Cell 5, 584–595 (2009). 27. Lee, G. et al. Modelling pathogenesis and treatment of familial dysautonomia using patient-specific iPSCs. Nature 461, 402–406 (2009). 28. Wernig, M. et al. A drug-inducible transgenic system for direct reprogramming of multiple somatic cell types. Nat. Biotechnol. 26, 916–924 (2008). 29. Stadtfeld, M., Maherali, N., Breault, D.T. & Hochedlinger, K. Defining molecular cornerstones during fibroblast to iPS cell reprogramming in mouse. Cell Stem Cell 2, 230–240 (2008). 30. Cerletti, M. et al. Highly efficient, functional engraftment of skeletal muscle stem cells in dystrophic muscles. Cell 134, 37–47 (2008). 31. Stadtfeld, M., Maherali, N., Borkent, M. & Hochedlinger, K. A reprogrammable mouse strain from gene-targeted embryonic stem cells. Nat. Methods 7, 53–55 (2010). 32. Chin, M.H. et al. Induced pluripotent stem cells and embryonic stem cells are distinguished by gene expression signatures. Cell Stem Cell 5, 111–123 (2009). 33. Bernstein, B.E. et al. A bivalent chromatin structure marks key developmental genes in embryonic stem cells. Cell 125, 315–326 (2006). 34. Mikkelsen, T.S. et al. Dissecting direct reprogramming through integrative genomic analysis. Nature 454, 49–55 (2008). 35. Sridharan, R. et al. Role of the murine reprogramming factors in the induction of pluripotency. Cell 136, 364–377 (2009). 36. Marion, R.M. et al. Telomeres acquire embryonic stem cell characteristics in induced pluripotent stem cells. Cell Stem Cell 4, 141–154 (2009). 37. Boiani, M., Eckardt, S., Scholer, H.R. & McLaughlin, K.J. Oct4 distribution and level in mouse clones: consequences for pluripotency. Genes Dev. 16, 1209–1219 (2002). 38. Bortvin, A. et al. Incomplete reactivation of Oct4-related genes in mouse embryos cloned from somatic nuclei. Development 130, 1673–1680 (2003). 39. Ng, R.K. & Gurdon, J.B. Epigenetic memory of active gene transcription is inherited through somatic cell nuclear transfer. Proc. Natl. Acad. Sci. USA 102, 1957–1962 (2005). 40. Ng, R.K. & Gurdon, J.B. Epigenetic memory of an active gene state depends on histone H3.3 incorporation into chromatin in the absence of transcription. Nat. Cell Biol. 10, 102–109 (2008). 41. Feng, Q. et al. Hemangioblastic derivatives from human induced pluripotent stem cells exhibit limited expansion and early senescence. Stem Cells 28, 704–712 (2010). 42. Hu, B.Y. et al. Neural differentiation of human induced pluripotent stem cells follows developmental principles but with variable potency. Proc. Natl. Acad. Sci. USA 107, 4335–4340 (2010). 43. Kim, K. et al. Epigenetic memory in induced pluripotent stem cells. Nature doi:10.1038/nature09342 (19 July 2010).

855

© 2010 Nature America, Inc. All rights reserved.

ONLINE METHODS Generation of iPSC lines. iPSC lines were generated as described previously2. Briefly, iPSC-derived somatic cells were isolated from chimeras by fluorescence-activated cell sorting (FACS), plated on feeders in the presence of cytokines in ESC culture conditions. Resultant iPSC colonies were picked and expanded in the absence of doxycycline and used for subsequent analyses. SMP isolation. Myofiber-associated cells were prepared from intact limb muscles (extensor digitorum longus, gastrocnemius, quadriceps, soleus, traverus abdominis and triceps brachii) as described previously44,45. Briefly, intact mouse limb muscles were digested with collagenase II to dissociate individual myofibers. These were triturated and digested with collagenase II and dispase to release myofiber-associated cells. The myofiber-associated cells were next unfractionated by FACS, using the following marker profiles for each population: (i) SMPs: CD45−Sca-1−Mac-1−CXCR4+β1-integrin+ ; (ii) Myoblast-containing population: CD45−Sca-1−Mac-1−CXCR4− ; (iii) Sca1+ mesenchymal cells: D45−Sca-1+Mac-1−. After the initial sort, cells were resorted by FACS using the same gating profile to increase the purity of the obtained population46. Blastocyst injections. For blastocyst injections, female BDF1 mice were superovulated by intraperitoneal injection of PMS and hCG and mated to BDF1 stud males. Zygotes were isolated from females with a vaginal plug 24 h after hCG injection. Zygotes for 2n injections were cultured for 3 d in vitro in KSOM media, blastocysts were identified, injected with ESCs or iPSCs and transferred into pseudopregnant recipient females. Teratoma formation. iPSCs were harvested by trypsinization, preplated onto untreated culture plates to remove feeders as well as differentiating cells and injected into flanks of nonobese diabetic/severe combined immunodeficient NOD/SCID mice, using ~5 million cells per injection. The mice were euthanized 3–5 weeks after injection, teratomas dissected out and processed for histological analysis.

pH7.5 (10 mM Tris-HCl and 0.1 mM EDTA) and quantified using a Nanodrop (Nanodrop Technologies). Quantitative PCR. cDNA was produced with the First Strand cDNA Synthesis Kit (Roche) using 1 mg of total RNA input. Real-time quantitative PCR reactions were set up in triplicate using 5 ml of cDNA (1:100 dilution) with the Brilliant II SYBR Green QPCR Master Mix (Stratagene) and run on a Mx3000P QPCR System (Stratagene). Primer sequences are listed in Supplementary Table 4. mRNA profiling. Total RNA samples (RIN (RNA integrity number) > 9) were subjected to transcriptomal analyses using Affymetrix HTMG- 430A mRNA expression microarray as previously described. Statistical analyses. Hierarchical clustering was performed using the GeneSifter software (Geospiza). Correlation distance and subsequent clustering were done using Ward’s method. The differentially expressed genes (twofold) were calculated using a t-test (P = 0.05) with Benjamini and Hochberg correction. Principal component analysis was performed using the GeneSifter software. Gene ontology analysis was performed using the DAVID software47, with the classification stringency set to ‘high’. Embryoid body formation. Before plating embryoid bodies, the iPSCs were depleted of mouse embryonic fibroblasts by splitting the cells 1:3 onto gelatin-coated plates on each day, for 2 consecutive days. On the 3rd day (designated day 0), iPSCs were trypsinized and plated at a density of 5,000 cells/ml in Isocove’s Modified Dulbecco’s Medium (IMDM) with 15% FCS (Atlanta Biologicals), 10% protein-free hybridoma medium (PFHM-II; Gibco), 2 mM l-glutamine (Gibco), 200 µg/ml transferrin (Roche), 0.5 mM ascorbic acid (Sigma) and 4.5 × 10–4 M monothioglycerol (MTG; Sigma). Differentiation was carried out in 60-mm ethylene oxide–treated Petri grade dishes (Parter Medical). The embryoid bodies were left to differentiate until day 6, when the cells were harvested to assay for hematopoietic colonies.

Cell culture. ESCs and iPSCs were cultured in ESC medium (DMEM with 15% FBS, l-glutamin, penicillin-streptomycin, nonessential amino acids, β-mercaptoethanol and 1,000 U/ml leukemia inhibitor factor) on irradiated feeder cells. TTF cultures were established by trypsin digestion of tail-tip biopsies taken from newborn (3–8 d of age) chimeric mice produced by blastocyst injection of iPSCs.

Hematopoietic colony formation assays. Day 6 embryoid bodies were collected by gravity, dissociated with trypsin and then passed several times through a 20 gauge needle to ensure dissociation. For the growth of hematopoietic progenitors, the cells were then seeded at a density of 100,000 cells/ml in IMDM containing 1% methylcellulose (Fluka Biochemika), 15% plasma-derived serum (PDS; Animal Technologies), 5% PFHM-II and specific cytokines as follows: primitive erythrocytes (erythropoietin (EPO, 2 U/ml)); macrophages (IL-3 (10ng/ml), M-CSF (5 ng/ml)); megakaryocytes (IL-3 (10 ng/ml), IL-11 (5 ng/ml), thrombopoietin (TPO, 5 ng/ml)); mixed colonies (SCF (5ng/ml), IL-3 (10 ng/ ml), G-CSF (30 ng/ml), GM-CSF (10 ng/ml), IL-11 (5 ng/ml), IL-6 (5 ng/ ml), TPO (5 ng/ml), and M-CSF (5 ng/ml)). All cytokines were purchased from R&D Systems. Primitive erythroctye colonies (eryPs) were counted on day 10 (4 d after embryoid body harvest). Macrophage colonies were counted on day 13 (7 d after embryoid body harvest). Mixed colonies were counted on day 14 (8 d after embryoid body harvest) and consist of a layer of macrophages, a layer of granulocytes, and a central core of red erythroid cells. Statistical analysis was performed using the Krward software. P values were calculated using the nonparametric Wilkinson test.

RNA isolation. ESCs and iPSCs grown on 35-mm dishes were harvested when they reached about 50% confluency and preplated on nongelatinized T25 flasks for 45 min to remove feeder cells. Cells were spun down and the pellet used for isolation of total RNA using the miRNeasy Mini Kit (Qiagen) without DNase digestion. RNA was eluted from the columns using 50 ml RNAse-free water or TE buffer,

HELP DNA methylation analysis. High molecular weight DNA was isolated from iPSCs using the PureGene kit from Qiagen and the HELP (HpaII tiny fragment enrichment by ligation-mediated PCR) assay was carried out as previously described1,2. Briefly, 1 µg of genomic DNA was digested overnight with either HpaII or MspI (New England Biolabs). On the following day, the reactions were

Cellular growth assays. To measure the clonal growth potential of iPSCs, SSEA1-positive cells from the different iPSC lines were sorted into 96-well plates by FACS (BD). After 7 d, the presence of iPSC colonies was scored based on morphology. To establish growth rates, the different bulk iPSCs lines or derivative subclones were plated in six gelatinized wells of a 12-well plates and each day the number of cells was counted in duplicate using a Countess cell counter (Invitrogen). For colorimetric measurement of growth, iPSCs lines were subcloned into 96-well plates and after 7 d, the cells were exposed to XTT (TOX-2) (Sigma) reagent overnight and the absorbance at 450 nm measured with a multiwell plate reader (Molecular Devices).

nature biotechnology 

doi:10.1038/nbt.1667

© 2010 Nature America, Inc. All rights reserved.

extracted once with phenol-chloroform and resuspended in 11 µl of 10 mM Tris-HCl pH 8.0 and the digested DNA was used to set up an overnight ligation of the JHpaII adaptor using T4 DNA ligase. The adaptor-ligated DNA was used to carry out the PCR amplification of the HpaII- and MspI-digested DNA as previously described 48. All samples for microarray hybridization were processed at the Roche-NimbleGen Service Laboratory. Samples were labeled using Cy-labeled random primers (9 mers) and then hybridized onto a mouse custom-designed oligonucleotide array (50-mers) covering 25,720 HpaII amplifiable fragments (HAF) (>50,000 CpGs), annotated to 15,465 unique gene symbols (Roche NimbleGen, Design name: 2006-10-26_MM5_HELP_Promoter Design ID = 4803). HpaII-amplifiable fragments are defined as genomic sequences contained between two flanking HpaII sites found within 200–2,000 bp from each other and is represented on the array by 15 individual probes, randomly distributed across the microarray slide. HAF were first realigned to the MM9 July 2007 build of the mouse genome and then annotated to the nearest transcription start site (TSS), allowing for a maximum distance of 5 kb from the TSS. Scanning was performed using a GenePix 4000B scanner (Axon Instruments) as previously described49. Quality control and data analysis of HELP microarrays was performed as described50. Signal intensities at each HpaII-amplifiable fragment were calculated as a robust (25% trimmed) mean of their component probe-level signal intensities. Any fragments found within the level of background MspI signal intensity, measured as 2.5 mean-absolute-differences (MAD) above the median of random probe signals, were categorized as ‘failed’. These failed loci therefore represent the population of fragments that did not amplify by PCR, whatever the biological (e.g., genomic deletions and other sequence errors) or experimental cause. On the other hand, ‘methylated’ loci were so designated when the level of HpaII signal intensity was similarly indistinguishable from background. PCR-amplifying fragments (those not flagged as either methylated or failed) were normalized using an intra-array quantile approach wherein HpaII/MspI ratios are aligned across density-dependent sliding windows of fragment size–sorted data. DNA methylation was therefore measured as the log2(HpaII/MspI) ratio, where HpaII reflects the hypomethylated fraction of the genome and MspI represents the whole genome reference. Analysis of normalized data revealed the presence of a bimodal distribution. For each sample, a cutoff was selected at the point that more clearly separated these two populations and the data were centered around this point. Each fragment was then categorized as either methylated, if the centered log HpaII/MspI ratio < 0, or hypomethylated if on the other hand the log ratio > 0. HELP data analysis. Statistical analysis was performed using R 2.9 and BioConductor51. Unsupervised hierarchical clustering of HELP data was performed using the subset of probe sets (n = 3745) with s.d. > 1 across all cases. We used 1– Pearson correlation distance, followed by a Lingoes

doi:10.1038/nbt.1667

transformation of the distance matrix to a Euclidean one and subsequent clustering using Ward’s method. Correspondence analysis was performed using the BioConductor package MADE4. The top 100 genes whose methylation status varied the most across the different groups were identified as those with the greatest s.d. across all samples. Quantitative DNA methylation analysis by MassARRAY EpiTyping. Validation of HELP findings was performed by matrix-assisted laser desorption ionization/time-of-flight (MALDI-TOF) mass spectrometry using EpiTyper by MassARRAY (Sequenom) on bisulfite-converted DNA following manufacturer’s instructions52 but using the Fast Start High Fidelity Taq polymerase from Roche for the PCR amplification of the bisulfite-converted DNA. MassArray primers were designed to cover the promoter regions of the indicated genes. (Primer sequences available as Supplementary Table 5). Chromatin immunoprecipitation (ChIP). Cells were fixed in 1% formaldehyde for 10 min, quenched with glycine and washed three times with PBS. Cells were then resuspended in lysis buffer and sonicated 10 × 30 s in a Bioruptor (Diagenode) to shear the chromatin to an average length of 600 bp. Supernatants were precleared using protein-A agarose beads (Roche) and 10% input was collected. Immunoprecipitations were performed using polyclonal antibodies to H3K4trimethylated, H3K27trimethylated, H3 pan-acetylation and normal rabbit serum (Upstate). DNA-protein complexes were pulled down using protein-A agarose beads and washed. DNA was recovered by overnight incubation at 65 °C to reverse cross-links and purified using QIAquick PCR purification columns (Qiagen). Enrichment of the modified histones in different genes was detected by quantitative real-time PCR using the primers in the Supplementary Table 4. 44. Conboy, I.M., Conboy, M.J., Smythe, G.M. & Rando, T.A. Notch-mediated restoration of regenerative potential to aged muscle. Science 302, 1575–1577 (2003). 45. Sherwood, R.I. et al. Isolation of adult mouse myogenic progenitors: functional heterogeneity of cells within and engrafting skeletal muscle. Cell 119, 543–554 (2004). 46. Cheshier, S.H., Morrison, S.J., Liao, X. & Weissman, I.L. In vivo proliferation and cell cycle kinetics of long-term self-renewing hematopoietic stem cells. Proc. Natl. Acad. Sci. USA 96, 3120–3125 (1999). 47. Huang, D.W. et al. Systematic and integrative analysis of large gene lists using DAVID Bioinformatics Resources. Nat. Protoc. 4, 44–57 (2009). 48. Figueroa, M.E., Melnick, A. & Greally, J.M. Genome-wide determination of DNA methylation by Hpa II tiny fragment enrichment by ligation-mediated PCR (HELP) for the study of acute leukemias. Methods Mol. Biol. 538, 395–407 (2009). 49. Selzer, R.R. et al. Analysis of chromosome breakpoints in neuroblastoma at sub-kilobase resolution using fine-tiling oligonucleotide array CGH. Genes Chromosom. Cancer 44, 305–319 (2005). 50. Thompson, R.F. et al. An analytical pipeline for genomic representations used for cytosine methylation studies. Bioinformatics 24, 1161–1167 (2008). 51. Culhane, A.C., Thioulouse, J., Perriere, G. & Higgins, D.G. MADE4: an R package for multivariate analysis of gene expression data. Bioinformatics 21, 2789–2790 (2005). 52. Ehrich, M. et al. Quantitative high-throughput analysis of DNA methylation patterns by base-specific cleavage and mass spectrometry. Proc. Natl. Acad. Sci. USA 102, 15785–15790 (2005).

nature biotechnology

Articles

Rapid profiling of a microbial genome using mixtures of barcoded oligonucleotides

© 2010 Nature America, Inc. All rights reserved.

Joseph R Warner1, Philippa J Reeder1, Anis Karimpour-Fard2, Lauren B A Woodruff1 & Ryan T Gill1 A fundamental goal in biotechnology and biology is the development of approaches to better understand the genetic basis of traits. Here we report a versatile method, trackable multiplex recombineering (TRMR), whereby thousands of specific genetic modifications are created and evaluated simultaneously. To demonstrate TRMR, in a single day we modified the expression of >95% of the genes in Escherichia coli by inserting synthetic DNA cassettes and molecular barcodes upstream of each gene. Barcode sequences and microarrays were then used to quantify population dynamics. Within a week we mapped thousands of genes that affect E. coli growth in various media (rich, minimal and cellulosic hydrolysate) and in the presence of several growth inhibitors (b-glucoside, d-fucose, valine and methylglyoxal). This approach can be applied to a broad range of traits to identify targets for future genome-engineering endeavors. Microbial genomes hold the potential for tremendous combinatorial diversity, comprising a sequence space of 44,600,000. Researchers’ ability to search this diversity for genetic features that affect pertinent traits remains limited by the number of individuals that can be tested, which is a small fraction of all possibilities. Thus, there is a demand for strategies for first defining relevant genetic variation and then thoroughly searching that space. This issue has been studied in great depth at the level of individual genes1,2, where high-throughput protein engineering methods are available for introducing specific mutations and then mapping the effects of such mutations onto protein activity. Advances in genomics3, and more recently multiplex DNA synthesis4–8 and homologous recombination (or recombineering)9–11, now enable the extension of such a strategy to the genome scale. Advances in genomics have resulted in several methods for highly parallel mapping of genes to traits, such as profiling of gene-knockout and plasmid-based libraries12–20. In some instances, microarray ­technology has been used to enable parallel tracking of genetically distinct individuals throughout growth in selective environments. One such tool, molecular barcoding12,17, involves the replacement of every gene in Saccharomyces cerevisiae with a specific DNA sequence that could be tracked via microarray. Although these tools are a power­ful way to profile the effect of mutation, the difficulty of specifically creating new mutations limits these studies to one of two types of mutations that have previously been introduced (insertions or increases in copy number). These limitations have challenged efforts to apply these methods for dissecting phenotypes and reengineering phenotypes that rely upon the coordinated action of multiple genes and mutations. Research over the past decade has resulted in recombination-based methods (recombineering) that make it easier to specifically modify the E. coli genome using synthetic DNA (synDNA)9–11,21–23. Recently, a recombineering-based method, called MAGE, was reported24,

whereby the expression levels of 24 genes were optimized in parallel to improve lycopene production more than all previously reported efforts, in considerably less time. This demonstration was enabled by a priori knowledge of what genes to modify, which is not known in many genome-engineering efforts, such as engineering growth and tolerance. Here we describe TRMR, a complementary method for simultaneously mapping genetic modifications that affect a trait of interest. The method combines parallel DNA synthesis, recombineering and molecular barcode technology to enable rapid modification of all E. coli genes (Fig. 1 and Supplementary Fig. 1). We demonstrate this general approach through the construction of two comprehensive E. coli genomic libraries comprising 8,000 distinct mutations and gene-trait mapping of these cells in seven environments. Results Synthetic DNA cassettes for promoter replacement We designed a comprehensive library of synDNA cassettes that have predictable effects when inserted into the genome of E. coli. Although various genetic features could have been incorporated into the cassettes (such as point mutations or sequences affecting mRNA stability, translational efficiency and other processes), we chose to demonstrate TRMR using functional modifications that either generally increase the expression of a target gene, called ‘up’, or generally decrease the gene’s expression, called ‘down’. The up cassette contains a strong and repressible PLtetO-1 promoter25 and ribosome binding site (RBS)26 sequences, which in general will increase downstream gene transcription and translation (Fig. 2). The down cassette was designed to replace the native RBS with an inert sequence that will generally cause a decrease in translation initiation. Both cassette designs include a blasticidin-S resistance gene27, allowing for selection of recombinant alleles. Molecular barcodes12 (also called ‘tags’) were incorporated to track the presence of each synDNA oligo and to

1Department of Chemical and Biological Engineering, University of Colorado, Boulder, Colorado, USA. 2School of Medicine, University of Colorado at Health Science Center, Denver, Colorado, USA. Correspondence should be addressed to R.T.G. ([email protected]).

Received 4 February; accepted 8 June; published online 18 July 2010; doi:10.1038/nbt.1653

856

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Articles

© 2010 Nature America, Inc. All rights reserved.

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

geneF

geneE

geneD

geneC

geneF

geneE

geneD

geneC

(iii) Multiplex recombineering (iv) Enrichment of improved cells track each allele (engineered cell) within the Bacterial cells Engineered Improved mixed population on a barcode microarray28 wild-type genome genomes genomes (Supplementary Notes). Because the length of the synDNA cassettes used here is beyond the current capabilities of commercially available oligo library synthesis, we developed a strategy (ii) Multiplex (v) Multiplex identification (vi) Genome mapping for multiplex cassette construction that Mixture of ≈ 8,000 synthesis Frequency of designed Frequency of designed Fitness conferred by unique oligomers involves the ligation of sequences shared by mutation (Fx = Cx/Ctot) mutation (Fx = Cx /Ctot) mutation (W′x = Fx,f /Fx,i) all cassettes to a mixture of shorter oligos Genome specific to each targeted gene. Construction (i) Design plot of this library was complicated by the fact microarray Targeting Tracking Functional Targeting microarray that each synDNA cassette must contain unique sequences in the flanking positions Figure 1  TRMR method. (i) Design DNA cassettes encoding the suite of mutations of interest. that are homologous to the chromosome (ii) Synthesize those cassettes, along with associated molecular barcodes, in a single pool. where the cassette is to be inserted. This is (iii) Introduce cassettes into recombination-proficient E. coli46 and produce thousands of variants, traditionally accomplished by using PCR to each with a distinct region of the chromosome that is engineered. (iv) Perform selections or screens amplify a DNA cassette with primers that on the mixture of variants to enrich for those possessing a desired trait. (v) Quantify changes in 47 contain the flanking homology regions21,29. allele frequency using molecular barcode technology . (vi) Use these frequency measurements to map specific genetic changes onto the trait of interest. Cx, concentration of allele x; Ctot, Using such a method to construct thoutotal concentration; Fx,f and Fx,i, final and initial allele frequencies (see equations in Results). sands of alleles is resource- and time­i ntensive 3,11,19, thus limiting the number recombineering experiments, separately generating thousands of and type of allelic libraries that can be investigated. To address these issues, we developed a procedure to generate thou- up and down recombinant colonies. Colonies were scraped from sands of synDNAs containing multiple desirable sequence features plates and frozen in aliquots for subsequent experiments. To confirm that desired mutant alleles were generated, we PCR (such as homology regions and expression modulators) that can be carried out in a complex mixture. Briefly, ‘targeting oligos’ were first amplified and sequenced barcode tags from 390 colonies. Sequencing synthesized on a microarray. Then, we ligated these to the cassette of the cassette and neighboring chromosome DNA indicated that in that modifies gene function, amplified the resulting product with 34 of 34 distinct alleles, the cassettes had inserted into the correct rolling-circle amplification and then cleaved the long amplified DNA location of the genome. Sequencing also provided an estimate of the number of alleles containing an error in DNA sequence. Outside molecule into the synDNAs (Fig. 2a–c). Targeting oligos were designed for every protein-coding gene in the of the barcode sequences, DNA errors were observed in only three E. coli MG1655 genome (Supplementary Table 1 and Supplementary of 34 alleles, two of which had errors in regions of the ­cassette that Notes). In all, 8,154 targeting oligos were designed to create two pos- should not affect allele identification or function. The barcode tag sible expression alleles for 4,077 genes. Targeting regions were chosen sequences provide an estimate of DNA errors present in the initial such that DNA cassettes would insert upstream of genes, replace the oligo libraries because barcodes are not subject to the experi­mental translation start codon and account for gene overlap. Once designed, bias (bias includes selection for correct sequences during PCR the set of targeting oligos, each 189 nucleotides long, was purchased amplification and during homologous recombination) that would through limited access at a cost of roughly $1 per unique oligo filter out incorrect sequences. High fidelity of the molecular barcode (Oligonucleotide Library Synthesis, Agilent). sequences is also required to accurately detect the presence of each To test cassette design and construction and to optimize the pro- allele in cell mixtures. Only 5% of the 390 sequenced tags showed cedure for allele production, we attempted promoter replacement an error, usually substitution or loss of a single nucleotide. The for the lacZ and galK genes. After optimizing design, we were able high percentage of correct alleles observed here is a first indication to efficiently generate these alleles using the procedures outlined in that complex oligonucleotide mixtures may be used to engineer and Figure 2. Alleles were isolated as colonies and all showed the expected identify thousands of distinct genomic loci with high fidelity. change in regulation and expression of the lacZ gene (Fig. 2d,e) or To assess our ability to make complete and uniform libraries in the galK gene. Furthermore, in PCR confirmations and sequencing, multiplex, we used Affymetrix Geneflex TAG4 arrays28 to measure 30 of 30 alleles tested showed the correct site of insertion. By count- the concentration of each barcode tag in the synDNA mixture (before ing colonies we estimated that we were able to routinely generate at recombineering) and in genomic DNA from cell mixtures (after least 75 alleles per microliter of cells transformed and determined recombineering). We observed microarray signals from hybridizathat yields increased linearly with transformation volumes from tion of each of the 8,154 library tags, ten positive-control tags that 40 ml up to 400 ml tested. With increases in scale, it is conceivable that we spiked into the samples to calculate tag concentrations (see one could generate 105–107 alleles in a single day, enough to profile Supplementary Fig. 2), and 1,642 negative-control tags used to provide a measure of background hybridization and noise. The barcode several modifications of every E. coli gene. signals from the synDNA mixtures indicated that 8,016 of the oligos were present (detected above background). Therefore, we successEfficient construction of genome-scale allele libraries Using a library of 8,154 targeting oligos, we attempted to con- fully generated nearly complete (98%) up and down oligo libraries. struct 4,077 up synDNA oligos and 4,077 down synDNA oligos Microarray analysis of the cell mixtures indicated successful generain separate pools. Both oligo pools were constructed in 1 week tion of at least 7,829 unique alleles (96% of designed alleles; Fig. 3a and resulted in enough material for several rounds of multiplex and Supplementary Table 2). We found that the concentration of recombineering. The synDNA oligos were then used in a day of each unique allele depended on the concentration of synDNA used

857

Articles a

Two mixtures of target oligos

Target oligos

g

Shared DNA

Y

geneX down geneY down geneZ down

ii

iii

c

Target oligo (189 nucleotides) H2x

P1

Cut site

H1x

Z

ared DN A Sh

b

up

X

geneX up geneY up geneZ up

i

e

Two mixtures of 4,077 synDNA oligos

neX up

P3 Tagx

4,077 genes targeted simultaneously

geneZ

geneY

geneX

Chromosome

Y Z

down X Y down Y Z down Z

up up up

P2

X

X

iv

E. coli cell

up up

synDNA oligo (~ 800 base pairs)

© 2010 Nature America, Inc. All rights reserved.

H1x Up

P3

Tagx P2

R

antibiotic

Recombineering enzymes

H2x

Up/Down

Chromosome

(PLtetO-1 & RBS)

up

Chromosome Down

d

(no RBS)

Chromosome

Chromosome

blasticidin

R

PLtetO-1RBS

up

geneX

TaglacZ–

blasticidin

R

No RBS

geneZ geneZ

geneY geneY

up

geneZ

lacZ up

lacZ down

Glucose + X-gal

IPTG + X-gal

Wild type

Wild type

lacZ

Down allele (703 bp insertion) Chromosome

geneX

e

Up allele (761 bp insertion) TaglacZ+

geneY

geneX

lacZ

Figure 2  Multiplex strategy to rapidly generate cell mixtures with defined genetic modifications. (a) Construction of synDNA library. (i) ‘Target’ oligos that contain chromosome homology and barcodes are synthesized on a chip, cleaved from the chip, amplified by two rounds of PCR and modified with (ligation) sequences by uracil excision48. (ii) This pool of target oligos is ligated with oligos containing a selectable marker and promoter and RBS variants (Shared DNA), resulting in a pool of DNA circles. (iii) DNA circles are copied into a pool of linear concatemers by rolling-circle amplification49. (iv) Concatemers are cleaved at a repeating site linking the homology regions to provide a pool of synDNA ready for multiplex recombineering. (b) Schematic of target oligos and synDNA oligos for gene x. Red, unique regions; black, shared regions; P, PCR priming site; H, chromosome targeting region; Tag, barcode tag sequence; Up/Down, functional region. Sequence is shown for amplifying barcode tags and for functional regions (promoter sequence in italic, RBS in bold, start codon underlined). (c) Pool of synDNA oligos is inserted into electrocompetent E. coli cells. Recombineering enzymes catalyze the insertion of the synDNA oligos at thousands of unique loci in the genome. (d) Schematic of lacZ alleles used to test the method. Up allele is designed to increase gene transcription and translation. Down allele is designed to decrease translation. (e) LacZ up and down alleles yield the intended phenotypes. Up mutation of the lacZ gene causes cells to turn blue on the surface of agar containing glucose and X-gal. Down mutation of the lacZ gene causes cells to remain colorless on the surface of agar containing IPTG and X-gal.

to construct that allele (Supplementary Fig. 3). After normalization of the concentration of each allele for differences in synDNA concentrations used in recombineering, the s.d. for generating each mutant was ± 65% of the average, distributed uniformly around the genome (Fig. 3b). We also observed a modest dependence of recombineering frequency on the hybridization free energy30 of the homology regions (Supplementary Fig. 4). A small percentage of the alleles were not detected (4%), and in all these cases the preceding synDNA was either absent or found in low concentrations. In subsequent attempts to create allele libraries, most of these missing alleles were detected, suggesting that the alleles were initially not detected because of low concentrations of the synDNA oligos. These results indicate that the uniformity of cell mixtures in future multiplex recombineering experiments may easily be improved by supplementation with synDNA oligos that are initially present in low concentrations. Improvements in the uniformity of the initial mixture should enable the more efficient identification of cells with improved traits.

Notably, a single researcher was able to create these two genomescale up and down allele libraries in a single day, demonstrating that multiplex recombineering is a rapid strategy for reprogramming thousands of genes.

858

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Genome-scale mapping of alleles to selectable traits To illustrate the potential of TRMR to rapidly generate and identify cells with new traits, we plated the cell mixtures on agar medium supplemented to create four different conditions (salicin, d-fucose, methylglyoxal and valine) in which wild-type E. coli typically do not grow. Colonies representing resistant mutants arose from our allele mixtures at frequencies >100-fold greater than from unmodified control cells that relied on spontaneous mutation to generate resistance (Supplementary Table 3). We characterized individual colonies (83 total) by sequencing the barcode tags. Additionally, we used TAG4 microarrays to characterize the populations obtained by scraping all colonies off of the surfaces of selection plates. Using microarray data, we ranked each allele in each condition according to fitness (fitness of

Articles

2,000 Number of probes

1,500 Number of probes

Figure 3  Analysis of synDNA and cell library. (a) Histogram showing the distribution of barcode signals of the up and down allele libraries detected by the TAG4 microarray. The unassigned tag signals (shown in gray) provide a measure of the background signal for each probe on the microarray. Probes that are assigned to unique alleles are shown in green. The unassigned tag signals have a low signal distribution (inset), and the threshold is shown for signals that are significantly above the background signal. The threshold for detection was such that the rate of false positives would be less than 2.2%. (b) TAG4 microarray results showing the distribution of synDNA oligos and alleles plotted by genomic location on the circular E. coli genome. Blue, up library; red, down library; inner circles, the concentration of each unique synDNA oligo before recombineering; outer circles, efficiency of generating each allele, calculated by dividing allele concentration by synDNA concentration.

1,200

1,000

900

Threshold

a

600 300 0

20 80 140 200 260 Unassigned tag signals

500

0

0

1,200

2,400

3,600 4,800 6,000 Allele tag signals

7,200 8,400

© 2010 Nature America, Inc. All rights reserved.

b

Up library pop. 3,869

results and sequencing, as having the greatest effect on fitness in medium supplemented with salicin. Mutations in the hns (histonelike nucleoid structuring protein) regulator31 are known to confer improved growth on salicin. Its identification here confirms that the TRMR method can effectively uncover gene-trait relationships. d-fucose is a nonmetabolizable analog of arabinose that inhibits the ability of E. coli to use arabinose as a carbon source by inhibiting induction of the l-arabinose operon. We identified the xylA up allele, which causes overexpression of xylA and xylB, as conferring the ability to grow in the presence of d-fucose. Notably, these results suggest that E. coli xylose isomerase (XylA) may have in vivo l-arabinose isomerase activity. This discovery is corroborated by the observation that overexpression of E. coli xylAB in Pseudomonas putida confers the ability to metabolize both xylose and l-arabinose32. Such a trait is of potential value for the efficient use of cellulosic biomass as a renewable feedstock. Methylglyoxal is an important intracellular metabolite because it can be used as an intermediate for production of ­commodity ­chemicals and because, when metabolism is disrupted, it can ­ accumulate,

Down library pop. 3,960

allele x = W ′x = Fx,f / Fx,i, which is the ratio of the final allele frequency (Fx = concentration of x/total concentration) after growth to the initial allele frequency). The allele fitness determined by microarray agreed well with the results from picking and sequencing individual colonies (Fig. 4 and Supplementary Table 4). Constructing mutants with beneficial traits and identifying the genetic cause has traditionally been a slow and laborious process. Using TRMR, we were able to rapidly identify traits present in our cell mixtures that are consistent with previous studies and identify unexpected genetic cell Recovered a Frozen modifications that could be used in future Growth on selective agars mixture cells metabolic engineering. The allele(s) that conferred the highest frequency or fitness from these selections were reconstructed separately to confirm that improved growth is Microarray analysis, allele sequencing & reconstruction, phenotype validation due to the insertion of the identified ­cassette. These alleles are summarized in Figure 4 and b Salicin Methylglyoxal Valine D-fucose described in detail below. xyIA Salicin is a carbon source that E. coli norilvN mally cannot metabolize owing to repression of the enzymes BglF and BglB. We identified hns sodC the hns down mutation, using both array

c

hns down

xyIA up

sodC down

IeuL down

ilvN down

Methylglyoxal

Overexpression LeuABCD

Pyruvate + 2-ketobutyrate

D-fucose

Salicin

L-arabinose

D-xylose

2-KIV

BgIF H-NS XylA BgIB

Toxic oxygen radicals Valine

D-fucose

LeuABCD

Figure 4  Trait-conferring genotypes identified in four selective environments. (a) Up and down alleles were recovered from frozen cultures and spread on agar medium in conditions where wild-type cells would not grow (indicated as column headings). (b) Fitness (W′) calculated by microarray detection of barcode tags was plotted for each allele by genomic location. Blue, up allele; red, down allele. (c) Known or hypothesized mechanisms whereby the identified genomic modifications confer the ability to grow. High-fitness alleles were detected on microarrays, except for leuL down, which was identified by sequencing of barcode tags within colonies.

IIvB ilvN

IIvB

Valine

Acetohydroxybutyrate

H-NS XyIB Glycolysis

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

Pentose phosphate pathway

SodC

Leucine & isoleucine

Isoleucine

859

a

Freezer stock

Growth in minimal nutrients

Growth in 15–17% cellulosic hydrolysate

Growth in rich nutrients

Growth in 18–20% cellulosic hydrolysate

b

e

cyaA

ybaB

ahpC

cyaA moeA

ilvM

yciV ygjQ

eutL ptsI

IsrA ydjG

eutL

Number of cells (A600)

2.0 1.5 1.0 0.5 1.0 0.5 1.5 Rich medium allele fitness (W′)

2.5

2

Fitness

3

2.0

1 0

1.5

2.0

1.0 0.5 0

0.5

1.0

1.5

0.5 0 0

2.0

Up allele fitness (W′)

2.5

3.0

2

4

6

8

10

12

14

3.5 3.0 2.5 2.0 1.5 1.0 0.5 0

viii

vii

0

g % change in biomass productivity and growth rate relative to wild-type

0

3.0

0

i ii iii iv & v vi vii

1.5 1.0

2.5

0

d

f

3.0

Loss Gain

c Minimal medium allele fitness (W′)

Figure 5  Alleles identified during pooled growth in media and cellulosic hydrolysate. (a) TRMR alleles were recovered from frozen cultures and allowed to grow in a rich medium, minimal medium or cellulosic hydrolysate. (b) Allele frequencies after growth in media plotted by genomic location. Inner circle, rich; outer circle, minimal; blue, up allele; red, down allele; black, control allele frequency × 10. (c) Allele fitness in minimal medium plotted against fitness of the same allele in rich medium. Shapes describe the affected gene function as determined by clusters of orthologous groups: ◊, information storage and processing; , cellular processes; , metabolism; ×, poorly characterized; blue, up allele; red, down allele; black, control allele. Fitness trend was fit to a line shown in black (R2 = 0.748). (d) The fitness of down alleles compared with the corresponding up alleles. Brown , rich medium; green , minimal medium. For alleles that cluster toward either the x or the y axis, the up allele and the down allele report opposite effects. Inset shows fitness benefits (W′ > 1) of top 40 alleles for growth in minimal medium, and the fitness effects (usually detrimental, W′ < 1) of the orthogonal alleles. (e) Fitness (lnW′) plotted by genomic location of alleles isolated after growth in hydrolysate. Inner circle, 15–17% hydrolysate; outer circle, 18–20% hydrolysate; blue, up allele; red, down allele. Some alleles conferring high fitness are labeled. (f) Growth curves of isolated variants in cellulosic hydrolysate. Each growth curve is the average of three replicates. Curves are fit with a Gompertz function50 (black). Alleles are denoted with roman numerals, as follows: (i) puuE down (pale blue), (ii) yciV down (purple), (iii) ygaZ up (green), (iv) lpp down (pink), (v) ugpE down (blue), (vi) ptsI down (pale green), (vii) wild-type MG1655 (red), (viii) ahpC up (blue). Error bars are minimal and are not shown for clarity. A600, absorbance at 600 nm. (g) Percent change in biomass productivity and maximum growth rate for isolated variants grown in hydrolysate relative to E. coli MG1655 grown in hydrolysate. Biomass productivity (gray bars) is the area under each growth curve. Growth rate (red bars) is the maximum growth rate as calculated from the Gompertz function. Values are the average of three replicates; error bars denote s.d.

Down allele fitness (W′)

© 2010 Nature America, Inc. All rights reserved.

Articles

2

4

6

8 Time (h)

10

12

14

248 ± 18% 233 ± 22% 80 60 40 20 0 –20 ahpC puuE ptsI lpp yciV ygaZ ugpE up down down down down up down

r­ esulting in oxidative damage and eventual cell death33. We used TRMR to ­ discover a previously unknown phenotype: decreased expression of sodC, which produces a superoxide-­mediating enzyme34, confers resistance to exogenous methylglyoxal, possibly by affecting ­superoxide concentrations in the periplasm. Excess valine causes feedback inhibition of leucine and isoleucine biosynthesis, leading to inhibition of cell growth as these amino acids become scarce. Microarray results identified ilvN down as the allele conferring the best growth, and this genomic region has been indicated in several previous studies35,36. Unexpectedly, sequencing showed that the leuL down allele also could grow well on valine plates. The leuL down mutation would cause increased expression of the leucine biosynthesis operon leuABCD by circumventing the alleged transcription attenuation caused by leuL37. Mutations of this operon have not previously been associated with valine resistance. However, a recent attempt to increase production of noncanonical amino acids in engineered E. coli cells demonstrated that overexpression of leuABCD shifts metabolite pools from valine toward isoleucine and leucine38.

Genome-scale quantitative growth phenotypes To further demonstrate that TRMR performs well at the genome scale, we combined the up and down allele libraries and measured fitness in liquid cultures that contained rich or minimal nutrients (Fig. 5a). The liquid cultures were allowed to grow for an average of eight generations, before and after which aliquots of cells were plated for analysis of individuals or frozen for microarray analysis. Additionally, an aliquot of control cells (barcoded and kanamycin resistant; Supplementary Notes) was spiked into the culture at the start of selections. A known concentration of these control cells was used to assess the ability of barcode technology to measure allele concentrations during pooled growth. The control cells also serve as a wild-type standard with which the fitness of alleles can be compared. Using barcode microarrays, we simultaneously tracked all of the alleles, which were reduced to approximately 2,500 alleles after growth selections (Fig. 5b). The numbers of control cells in the populations determined by microarray was not substantially different from estimates of control-cell numbers obtained from counting

860

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

© 2010 Nature America, Inc. All rights reserved.

Articles kanamycin-resistant colonies. Microarrays revealed that the majority of alleles had similar growth phenotypes in both rich and minimal media (Fig. 5c, x-y diagonal). Noteworthy alleles that do not fit this trend are those that allow growth in the rich medium but are no longer observed in the minimal medium (Fig. 5c, alleles along x axis). Consistent with previous observations19, many of these alleles consist of changes in the expression of genes involved in metabolism. Also of interest are those alleles that confer faster growth than that of the control cells in the minimal medium (a list of fitness values can be found in Supplementary Table 5). These experiments also offer the first genome-wide glimpse of generally orthogonal expression alleles grown competitively in the same culture. We anticipated that if a particular up allele shows a fitness benefit, then the down allele is likely to show a negative effect on fitness, possibly being lost from the culture, and vice versa. This is often the case (see Fig. 5d, allele clustering toward the axes), providing further evidence that our synthetic cassettes are generally causing the intended effects at genome-wide loci. Exceptions such as improved growth resulting from both up and down expression alleles in the same environment may be due to secondary effects (such as increased transcription of multiple downstream genes) and require further investigation. Mapping tolerance to lignocellulosic hydrolysate We next applied TRMR to identify genes that improve tolerance to lignocellulosic hydrolysate derived from corn stover (provided by the US National Renewable Energy Laboratory). This class of feedstocks contains a variable array of growth inhibitors (known inhibitors include organic acids, aldehydes and phenolic-based compounds)39,40. To take hydrolysate variability into account, we measured growth of variants bearing our alleles in several mixtures of hydrolysate and minimal medium. Microarray analysis of the alleles indicated that only a small ­subset of the population remained after each selection (Fig. 5e; see Supplementary Table 6 for fitness values and gene ontology ­analysis). Many of the modifications that improved growth in lower concentrations of corn stover hydrolysate affected genes known to be involved in primary metabolism (pgi up, eno up and tdcG up), RNA metabolism (rlmG down, rimM up, rsmE down and rrmA down) and transport of sugars (ptsI down, ptsI up and directly downstream crr). Growth in higher concentrations of hydrolysate selected alleles related to ­secondary metabolism (ispF up and dxs up), vitamin metabolic ­processes (nadD up, menD up, apbE up, pabC up, dxs up and ribB up) and antioxidant activity (ahpC up, tpx up and bcp up). The down mutation of the adenylate cyclase gene (cyaA) conferred a growth advantage in every selection. To confirm that the mutations conferred fitness advantages, we isolated seven alleles after the selections and characterized growth in hydrolysate relative to unmutated E. coli. (Fig. 5f,g). All seven alleles (ahpC up, ugpE down, puuE down, ptsI down, ygaZ up, yciV down and lpp down) yielded improvement in either growth rate or biomass productivity relative to the wild-type strain. Notably, the up allele of ahpC resulted in a large improvement. The ahpC gene and its downstream counterpart ahpF have not previously been identified as important for growth in hydrolysate. However, they have been implicated in resistance to organic solvents41 and various oxidants42,43, possibly indicating that during growth in cellulosic hydrolysate, reactive oxygen species in the form of peroxides and other oxidants are present or forming as a result of imbalances in metabolism44. In addition to identifying several important targets for future genome-engineering endeavors, many of which would have nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

been difficult to predict a priori, these profiling studies shed light on general mechanisms of hydrolysate toxicity (such as the presence of oxidants) and growth advantage in hydrolysate (such as metabolism of preferred carbon sources). Discussion We have described a new method for the genome-scale mapping of genes to traits and have shown that this method can increase the throughput of genetic studies by several orders of magnitude. Although some of the trait-conferring modifications we identified correspond to previously identified genomic regions, the majority would have been difficult to predict. Such unanticipated outcomes provide insight into many uncharacterized genes and, in some cases, into known genes with uncharacterized functions. We have already begun applying this method toward understanding a range of traits of importance in biotechnology, including improved growth in industrially relevant conditions and enhanced product formation. We have designed TRMR to be easy to use and versatile. The molecular cloning procedures were accomplished within a week by a single researcher, with two additional days providing enough cells for 60 genome-wide selection and screening studies. Notably, data acquisition and analysis from TRMR is similar to genomics methods currently used by the yeast community and is amenable to a range of freely and commercially available software packages. The primary challenge to the broad dissemination of this method is the acquisition of oligonucleotide libraries, which will be overcome as DNA synthesis technologies continue to improve. We envision that a broad range of additional studies could be performed using the basic TRMR platform described here by changing the targeting, functional or tracking design. For example, although the functional regions we used were promoters and translation sites, one might conceivably use sites associated with additional functions such as switches, oscillators or sensors45. Moreover, the TRMR approach is not limited to engineering or examining the E. coli genome. The design could be adapted for rapidly engineering yeast and a range of Gram-negative bacteria23, provided the host has sufficient transformation and recombination capabilities. Additionally, TRMR may be carried out recursively, allowing for the accumulation of multiple beneficial mutations within a genome. Researchers could produce second- and third-generation recombinant cells by removing the antibiotic cassette between rounds of recombineering to allow isolation of cells containing an additional mutation, by using different anti­ biotic cassettes in the modular construction of the synDNA oligos so that different antibiotics could be used to isolate recombinants after each round of TRMR, or by eliminating altogether the need to isolate recombinants by relying on the increased efficiency of recombineering strategies such as those used in MAGE24. Integration of TRMR into directed-evolution programs would provide genome-scale construction and tracking of combinations of mutations, which would improve both the understanding and engineering of complex traits. Methods Methods and any associated references are available in the online ­version of the paper at http://www.nature.com/naturebiotechnology/. Note: Supplementary information is available on the Nature Biotechnology website. Acknowledgments We thank D. Court (Center for Cancer Research, National Cancer Institute at Frederick, Maryland) for sharing plasmid pSIM5, C. Nislow and G. Giaever (University of Toronto, Ontario) for help with microarray analysis, A. Mohagheghi and M. Zhang (US National Renewable Energy Laboratories) for hydrolysate samples, M. O’Donnell for help in preparation of selective agar plates, Agilent for

861

Articles access to the Oligonucleotide Library Synthesis product, and H. Marshall and the University of Colorado Microarray Facility for molecular barcode genotyping. The authors appreciate financial support provided by Shell, the Colorado Center for Biorefining and Biofuels (http://www.C2B2web.org) and the Colorado Energy Initiative (http://rasei.colorado.edu).

1. Fox, R.J. et al. Improving catalytic function by ProSAR-driven enzyme evolution. Nat. Biotechnol. 25, 338–344 (2007). 2. Turner, N.J. Directed evolution drives the next generation of biocatalysts. Nat. Chem. Biol. 5, 567–573 (2009). 3. Winzeler, E.A. et al. Functional characterization of the S. cervisiase genome by gene geletion and parallel analysis. Science 285, 901–906 (1999). 4. Fodor, S. et al. Light-directed, spatially addressable parallel chemical synthesis. Science 251, 767–773 (1991). 5. Blanchard, A.P., Kaiser, R.J. & Hood, L.E. High-density oligonucleotide arrays. Biosens. Bioelectron. 11, 687–690 (1996). 6. Singh-Gasson, S. et al. Maskless fabrication of light-directed oligonucleotide microarrays using a digital micromirror array. Nat. Biotechnol. 17, 974–978 (1999). 7. Cleary, M.A. et al. Production of complex nucleic acid libraries using highly parallel in situ oligonucleotide synthesis. Nat. Methods 1, 241–248 (2004). 8. Ghindilis, A. et al. CombiMatrix oligonucleotide arrays: genotyping and gene expression assays employing electrochemical detection. Biosens. Bioelectron. 22, 1853–1860 (2007). 9. Yu, D. et al. An efficient recombination system for chromosome engineering in Escherichia coli. Proc. Natl. Acad. Sci. USA 97, 5978–5983 (2000). 10. Murphy, K. Use of bacteriophage lambda recombination functions to promote gene replacement in Escherichia coli. J. Bacteriol. 180, 2063–2071 (1998). 11. Zhang, Y., Buchholz, F., Muyrers, J. & Stewart, A.F. A new logic for DNA engineering using recombination in Escherichia coli. Nat. Genet. 20, 123–128 (1998). 12. Shoemaker, D.D., Lashkari, D.A., Morris, D., Mittmann, M. & Davis, R.W. Quantitative phenotypic analysis of yeast deletion mutants using a highly parallel molecular bar-coding strategy. Nat. Genet. 14, 450–456 (1996). 13. Cho, R.J. et al. Parallel analysis of genetic selections using whole genome oligonucleotide arrays. Proc. Natl. Acad. Sci. USA 95, 3752–3757 (1998). 14. Gill, R.T. et al. Genome wide screening for trait conferring genes using DNA microarrays. Proc. Natl. Acad. Sci. USA 99, 7033–7038 (2002). 15. Lynch, M.D., Warnecke, T. & Gill, R.T. SCALEs: multiscale analysis of library enrichment. Nat. Methods 4, 87–93 (2007). 16. Badarinarayana, V. et al. Selection analyses of insertional mutants using subgenic resolution arrays. Nat. Biotechnol. 19, 1060–1065 (2001). 17. Giaever, G. et al. Functional profiling of the Saccharomyces cerevisiae genome. Nature 418, 387–391 (2002). 18. Ho, C.H. et al. A molecular barcoded yeast ORF library enables mode-of-action analysis of bioactive compounds. Nat. Biotechnol. 27, 369–377 (2009). 19. Baba, T. et al. Construction of Escherichia coli K-12 in-frame, single-gene knockout mutants: the Keio collection. Mol. Syst. Biol. 2, 1–11 (2006). 20. Kitagawa, M. et al. Complete set of ORF clones of Escherichia coli ASKA library (a complete set of E. coli K-12 ORF archive): unique resources for biological research. DNA Res. 12, 291–299 (2006). 21. Datsenko, K. & Wanner, B. One-step inactivation of chromosomal genes in E. coli K12 using PCR products. Proc. Natl. Acad. Sci. USA 97, 6640–6645 (2000). 22. Ellis, H.M., Yu, D., DiTizio, T. & Court, D.L. High efficiency mutagenesis, repair, and engineering of chromosomal DNA using single-stranded oligonucleotides. Proc. Natl. Acad. Sci. USA 98, 6742–6746 (2001).

23. Datta, S., Costantino, N., Zhou, X. & Court, D.L. Identification and analysis of recombineering functions from Gram-negative and Gram-positive bacteria and their phages. Proc. Natl. Acad. Sci. USA 105, 1626–1631 (2008). 24. Wang, H. et al. Programming cells by multiplex genome engineering and accelerate evolution. Nature 460, 894–898 (2009). 25. Lutz, R. & Bujard, H. Independent and tight regulation of transcriptional units in Escherichis coli via the LacR/O, the TetR/O and AraC/I1–I2 regulatory elements. Nucleic Acids Res. 25, 1203–1210 (1997). 26. Shine, J. & Dalgarno, L. The 3′-terminal sequence of Escherichia coli 16S ribosomal RNA: complementarity to nonsense triplets and ribosome binding sites. Proc. Natl. Acad. Sci. USA 71, 1342–1346 (1974). 27. Kimura, M., Takatsuki, A., Yamaguchi, I. & Blasticidin, S. Deaminase gene from Aspergillus terreus (BSD): a new drug resistance gene for transfection of mammalian cells. Biochim. Biophys. Acta 1219, 653–659 (1994). 28. Pierce, S.E. et al. A unique and universal molecular barcode array. Nat. Methods 3, 601–603 (2006). 29. Baudin, A., Ozier-Kalogeropoulos, O., Denouel, A., Lacroute, F. & Culin, C. A simple and efficient method for direct gene deletion in Saccharomyces cerevisiae. Nucleic Acids Res. 21, 3329–3330 (1993). 30. Markham, N.R. & Zuker, M. DINAMelt web server for nucleic acid melting prediction. Nucleic Acids Res. 33, W577–W581 (2005). 31. Defez, R. & de Felice, M. Cryptic operon for beta-glucoside metabolism in Escherichia coli K12: genetic evidence for a regulatory protein. Genetics 97, 11–25 (1981). 32. Meijnen, J.P., de Winde, J.H. & Ruijssenaars, H.J. Engineering Pseudomonas putida S12 for efficient utilization of d-xylose and l-arabinose. Appl. Environ. Microbiol. 74, 5031–5037 (2008). 33. Zhu, M.M., Skraly, F.A. & Cameron, D.C. Accumulation of methylglyoxal in anaerobically grown Escherichia coli and its detoxification by expression of the Pseudomonas putida glyoxalase i gene. Metab. Eng. 3, 218–225 (2001). 34. Gort, A.S., Ferber, D.M. & Imlay, J.A. The regulation and role of the periplasmic copper, zinc superoxide dismutase of Escherichia coli. Mol. Microbiol. 32, 179–191 (1999). 35. Sutton, A., Newman, T., Francis, M. & Freundlich, M. Valine-resistant Escherichia coli K-12 strains with mutations in the ilvB operon. J. Bacteriol. 148, 998–1001 (1981). 36. Weinstock, O., Sella, C., Chipman, D.M. & Barak, Z. Properties of subcloned subunits of bacterial acetohydroxy acid synthases. J. Bacteriol. 174, 5560–5566 (1992). 37. Wessler, S.R. & Calvo, J.M. Control of leu operon expression in Escherichia coli by a transcription attenuation mechanism. J. Mol. Biol. 149, 579–597 (1981). 38. Sycheva, E.V. et al. Overproduction of noncanonical amino acids by Escherichia coli cells. Microbiology 76, 712–718 (2007). 39. Chen, S.F., Mowery, R.A., Castleberry, V.A., van Walsum, G.P. & Chambliss, C.K. High-performance liquid chromatography method for simultaneous determination of aliphatic acid, aromatic acid and neutral degradation products in biomass pretreatment hydrolysates. J. Chromatogr. A 1104, 54–61 (2006). 40. Mohagheghi, A. & Schell, D.J. Impact of recycling stillage on conversion of dilute sulfuric acid pretreated corn stover to ethanol. Biotechnol. Bioeng. 105, 992–996 (2010). 41. Ferrante, A.A., Augliera, J., Lewis, K. & Klibanov, A.M. Cloning of an organic solvent-resistance gene in Escherichia coli: the unexpected role of alkylhydroperoxide reductase. Proc. Natl. Acad. Sci. USA 92, 7617–7621 (1995). 42. Poole, L.B. Bacterial defenses against oxidants: mechanistic features of cysteinebased peroxidases and their flavoprotein reductases. Arch. Biochem. Biophys. 433, 240–254 (2005). 43. Seaver, L.C. & Imlay, J.A. Alkyl hydroperoxide reductase is the primary scavenger of endogenous hydrogen peroxide in Escherichia coli. J. Bacteriol. 183, 7173–7181 (2001). 44. Kohanski, M.A., Dwyer, D.J., Hayete, B., Lawrence, C.A. & Collins, J.J. A common mechanism of cellular death induced by bactericidal antibiotics. Cell 130, 797–810 (2007). 45. Lu, T.K., Khalil, A.S. & Collins, J.J. Next-generation synthetic gene networks. Nat. Biotechnol. 27, 1139–1150 (2009). 46. Datta, S., Constantino, N. & Court, D.L. A set of recombineering plasmids for gram-negative bacteria. Gene 379, 109–115 (2006). 47. Pierce, S.E., Davis, R.W., Nislow, C. & Giaever, G. Genome-wide analysis of barcoded Saccharomyces cerevisiae gene-deletion mutants in pooled cultures. Nat. Protoc. 2, 2958–2974 (2007). 48. Nour-Eldin, H.H., Hansen, B.G., Norholm, M.H.H., Jensen, J.K. & Halkier, B.A. Advancing uracil-excision based cloning towards an ideal technique for cloning PCR fragments. Nucleic Acids Res. 34, e122 (2006). 49. Dean, F.B., Nelson, J.R., Giesler, T.L. & Lasken, R.S. Rapid amplification of plasmid and phage DNA using Phi29 DNA polymerase and multiply-primed rolling circle amplification. Genome Res. 11, 1095–1099 (2001). 50. Perni, S., Andrew, P.W. & Shama, G. Estimating the maximum growth rate from microbial growth curves: definition is everything. Food Microbiol. 22, 491–495 (2005).

862

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Author Contributions J.R.W. and R.T.G. conceived the study; J.R.W. designed and performed all experiments except for growth selections and allele confirmations in hydrolysate, which were conducted by P.J.R.; A.K.-F. aided J.R.W. in selection of targeting sequences and selection of barcode tags; A.K.-F. and P.J.R. assigned gene ontology terms; L.B.A.W. aided J.R.W. in selection design and microarray analysis; L.B.A.W. constructed circle plots; P.J.R., A.K.-F. and L.B.A.W. helped in manuscript preparation; J.R.W. and R.T.G. wrote the manuscript; R.T.G. supervised all aspects of the study. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests.

© 2010 Nature America, Inc. All rights reserved.

Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/.

ONLINE METHODS

© 2010 Nature America, Inc. All rights reserved.

Strains, DNA and reagents. Escherichia coli MG1655 (wild type) was obtained from ATCC 700926. Genomic sequences were obtained from GenBank U00096.2, and gene annotation was from the Ecogene database version 2.20 (http://www.ecogene.org/). Pseudogenes and insertion elements were excluded from the protein-coding genes that were targeted. The kanamycin-resistant control strain (also called JWKAN) was constructed from E. coli ATCC 700926, with nucleotide 3,909,796 replaced with a barcoded kanamycin cassette21 (Supplementary Notes). Up and down DNA cassettes were constructed using PCR and cloned into the pEM7/BSD plasmid (Invitrogen, Supplementary Notes). Oligonucleotide libraries were purchased from Agilent; all other oligonucleotides were purchased from Integrated DNA Technologies with standard desalting except where noted. The pSIM5 plasmid46 was a gift from D. Court. All reagents were obtained from common commercial sources. All enzymes were from New England Biolabs except where noted. All sequencing was performed by Macrogen USA or Eurofins MGW Operon. Recipes and additional information can be found in Supplementary Notes. Preparation of synthetic DNA and recombineering. A portion of the oligonucleotide library provided by Agilent (8,154 unique 189-mers) was amplified by two rounds of PCR. Products were treated with the USER enzymes (New England Biolabs), purified and ligated to the up cassette. Rolling-circle amplification, nuclease treatment and purification resulted in 8–10 μg synDNA. This procedure was also carried out in parallel to separately generate TRMR down synDNA. More details are available in Supplementary Notes. E. coli cells containing the recombineering plasmid pSIM5 were grown in 800 ml SOB cultures at 30 °C and made recombineering proficient with minor modifications to reported methods46. Briefly, when cells reached an optical density at 600 nm of 0.7, flasks were transferred to water baths at 42 °C to induce the λRed enzymes for 15 min. Flasks were then transferred to an ice-water bath and cells were kept close to 4 °C for the remaining steps. Cells were collected by centrifugation and suspended with cold deionized water. Cell collection and washing was repeated once more, then cells were suspended to a final volume of 6.4 ml in water. Aliquots of cells (400 μl) were transformed in a 0.2-cm electrocuvette with approximately 1 μg of up or down synDNA and a pulse of 12.5 kV cm−1. Transformation was carried out eight times to generate the up allele library and eight times to generate the down allele library. The cells from each transformation were recovered in 12 ml SOC medium for 1 h at 37 °C. Cells were collected by centrifugation and resuspended in 30 ml MA salts (Supplementary Notes). Centrifugation and resuspension was repeated twice more, with the final resuspension to a volume of 2 ml in MA salts. The up and down allele libraries were separately spread onto a total of 40 low-salt LB agar plates containing blasticidin-S (90 μg ml−1) and allowed to grow at 37 °C for 22 h. Colonies were scraped from the agar plates and up and down allele libraries were each suspended in a total of 35 ml LB. Cells were collected by centrifugation and suspended to 3 × 109 cells per milliliter in LB medium containing 16% (vol/vol) glycerol and blasticidin-S (90 μg ml−1). Aliquots of the up or down cell mixtures were stored at −80 °C. Screens and selections. Freezer stocks were used to inoculate 50 ml low-salt LB medium containing 80 μg ml−1 blasticidin-S with 5 × 108 TRMR up cells and 5 × 108 TRMR down cells. This culture was allowed to grow with shaking at 37 °C to an optical density at 600 nm of 0.8. The cells were centrifuged at 4,500g for 6 min, decanted and suspended in 30 ml of MA salts. The cells were collected once more by centrifugation and suspended in MA salts to a

doi:10.1038/nbt.1653

concentration of 5 × 108 cells per milliliter. The JWKAN cells were added to a final concentration of 7.7 × 104 cells per milliliter. A 1.7 ml aliquot of the cell library (called the recovery culture) was frozen for microarray analysis, and the remainder was used for various growth selections. Liquid selections were carried out with shaking at 37 °C in 600 ml of MOPS minimal medium containing 2 mM phosphate and 4% (wt/vol) glucose or in 600 ml LB medium. Each medium was inoculated with 2.4 × 108 cells from a recovery culture and allowed to grow to an optical density at 600 nm of 1.0–1.2. Cells were collected from each culture by centrifugation of 10-ml aliquots at 4,500g for 6 min, decanted and stored at −80 °C for microarray analysis. Growth results are the average of three array hybridizations. Hydrolysate growth selections were carried out in various dilutions of hydrolysate in minimal media (15%, 16%, 17%, 18%, 19% and 20%). During selections, cell samples were taken for microarray analysis of populations, and cells were plated to isolate and identify individual alleles growing as colonies. Unique alleles from selections were identified and confirmed by PCR and studied for growth characteristics in hydrolysate. All growth curves were done in complete triplicate. More details are available in Supplementary Notes. Growth on various selective agars was carried out by spreading a total of 0.7 × 108 cells of the allele mixtures recovered from freezer aliquots on five plates for each selective condition (salicin, d-fucose plus l-arabinose, valine, and methylglyoxal; plate recipes in Supplementary Notes). Plates were incubated at 37 °C until colonies were visible (1–3 d). Selection for galK down alleles was carried out on plates containing 2-deoxygalactose9, and screens were carried out on MacConkey agar containing 1% (wt/vol) d-galactose. Screens of lacZ up alleles were carried out on LB agar plates containing 0.2% (wt/vol) glucose and 40 μg ml−1 X-gal. Screens of lacZ down alleles were carried out on LB agar plates containing 0.05% (wt/vol) IPTG and 40 μg ml−1 X-gal. Selections for control cells were carried out on LB agar plates containing kanamycin (30 μg ml−1). Microarray tracking. Genomic DNA was extracted from ~109 E. coli cells using Purelink Genomic Mini kit (Invitrogen). Barcode tags are amplified in 300 μl PCR reactions (final concentrations: 1× PCR buffer, 2.5 mM MgCl 2, 0.2 mM each dNTP, 1 μM each primer 5′-GTAGCACACGAGGTCTCT-3′ and Biotin-5′-TACGACTCACTATAGGGAGA-3′, 0.6 U μl−1 Taq polymerase and 0.5 μg genomic DNA or 30 pg synDNA). Reactions were cycled 25 times with an annealing temperature of 55 °C. Barcode tags were purified by agarose gel electrophoresis and extraction using the QIAquick gel extraction protocols (Qiagen, substitute buffer QX1 for QG). Tag purification was shown to reduce background hybridization. Microarray hybridizations to the Geneflex Tag4 16K V2 array (Affymetrix) were carried out according to published procedures47 with the following modifications: 600 ng of purified tags (combined up tags and down tags) were hybridized along with ten tags (amplified and purified as above) included at known concentrations (0.5 pM to 10 nM). Intensity values are calculated for each tag after removal of replicate outliers and averaging of unmasked replicates using software (raw_file_maker.pl) that can be downloaded from http://chemogenomics.stanford.edu/supplements/ 04tag/download.html. Background hybridization was calculated from the average intensity of 1,642 unused tag probes; threshold intensity was set to background hybridization plus 2 s.d. The intensities of the ten spiked tags were used to calculate allele concentrations from array signals and correct for array saturation (Supplementary Fig. 2). Barcode frequencies were calculated by dividing barcode concentrations by the total concentration of all barcodes detected on the array.

nature biotechnology

letters

Implications of the presence of N-glycolylneuraminic acid in recombinant therapeutic glycoproteins

© 2010 Nature America, Inc. All rights reserved.

Darius Ghaderi1,2, Rachel E Taylor1, Vered Padler-Karavani1, Sandra Diaz1 & Ajit Varki1 Recombinant glycoprotein therapeutics produced in nonhuman mammalian cell lines and/or with animal serum are often modified with the nonhuman sialic acid N-glycolylneuraminic acid (Neu5Gc; refs. 1,2). This documented contamination has generally been ignored in drug development because healthy individuals were not thought to react to Neu5Gc (ref. 2). However, recent findings indicate that all humans have Neu5Gc-specific antibodies, sometimes at high levels3,4. Working with two monoclonal antibodies in clinical use, we demonstrate the presence of covalently bound Neu5Gc in cetuximab (Erbitux) but not panitumumab (Vectibix). Anti-Neu5Gc antibodies from healthy humans interact with cetuximab in a Neu5Gc-specific manner and generate immune complexes in vitro. Mice with a human-like defect in Neu5Gc synthesis generate antibodies to Neu5Gc after injection with cetuximab, and circulating anti-Neu5Gc antibodies can promote drug clearance. Finally, we show that the Neu5Gc content of cultured human and nonhuman cell lines and their secreted glycoproteins can be reduced by adding a human sialic acid to the culture medium. Our findings may be relevant to improving the half-life, efficacy and immunogenicity of glycoprotein therapeutics. Therapeutic glycoproteins, including antibodies, growth factors, cytokines, hormones and clotting factors, generate sales with annual double-digit growth rates5. They must often be produced in mammalian expression systems because of the crucial influence of the location, number and structure of N-glycans on their yields, bio­activity, solubi­ lity, stability against proteolysis, immunogenicity and rate of clearance from the bloodstream6–8. Two differences between the protein glycosylation apparatus of humans and rodents account for major potential differences between the N-glycans on glycoproteins made in cultured human cells and those made using rodent cell lines. First, humans cannot synthesize a terminal Galα1-3Gal motif (known as alpha-Gal) on N-glycans. As a consequence, they express antibodies against this structure9. Second, unlike other mammals, humans cannot biosynthesize the sialic acid Neu5Gc because the human gene CMAH, encoding CMP-N-acetylneuraminic acid hydroxylase, the enzyme responsible for producing CMP-Neu5Gc from CMP-N-acetylneuraminic acid (CMP-Neu5Ac), is irreversibly mutated10. The use of cultured human cells to address

this issue is not a solution, as Neu5Gc can be taken up from animal products present in the culture medium and then metabolically incorporated into secreted glycoproteins11. Owing largely to limitations of the assays originally used to detect anti-Neu5Gc antibodies, including the fact that only a small number of possible Neu5Gc-containing epitopes were tested, healthy humans were long believed to show no immune reaction to Neu5Gc (ref. 2). Subsequent reports that all humans possess anti-Neu5Gc antibodies3, sometimes at high levels, approaching 0.1–0.2% of circulating IgG 3,4, have led to re-evaluation of the potential significance of Neu5Gc contamination7,8. Especially in light of trends toward administering increasingly higher amounts of certain biotherapeutics over longer periods of time, some biopharmaceutical companies are exploring steps to reduce levels of Neu5Gc in their products12. Given that they are produced using nonhuman cell lines, animal serum or serum-derived factors, or a combination of these, it is likely that most recombinant therapeutic glycoproteins carry some Neu5Gc. However, given the diversity of products and production protocols, it is difficult to make generalizations. Thus, we chose to compare two US Food and Drug Administration (FDA)-approved monoclonal antibodies with the same therapeutic target, the EGF receptor. The first, Erbitux (cetuximab, obtained from the University of California, San Diego Pharmacy), is a chimeric antibody produced in mouse myeloma cells13,14. The second, Vectibix (panitumumab, obtained from Amgen), is a fully human antibody produced in Chinese hamster ovary (CHO) cells15. The samples studied were preparations that would normally be administered to patients. We first performed enzyme-linked immunosorbent assays (ELISAs) using an affinity-purified polyclonal chicken Neu5Gc-specific antibody preparation that is highly monospecific for Neu5Gc (ref. 16, alongside a nonreactive control IgY). Bound Neu5Gc was easily detectable on cetuximab but not on panitumumab (Fig. 1a). Sialidase pretreatment abolished binding, confirming specificity. Western blot analysis also showed sialidase-sensitive anti-Neu5Gc IgY reactivity on the heavy chains of cetuximab but not those of panitumumab (Fig. 1b). The specificity of anti-Neu5Gc IgY binding was reaffirmed by ­pretreatment with mild sodium periodate under conditions that selectively cleave sialic acid side chains (Fig. 1c) and abolish ­reactivity of such ­antibodies3,16. Finally, we quantified sialic acids on the therapeutic antibodies, as described in Online Methods. Panitumumab ­carries 0.22 mol of sialic acids per mole of protein, with <0.1% Neu5Gc.

1Glycobiology 2Present

Research and Training Center, Departments of Medicine and Cellular & Molecular Medicine, University of California, San Diego, La Jolla, California, USA. address: Sialix, Inc., Vista, California, USA. Correspondence should be addressed to A.V. ([email protected]).

Received 7 August 2009; accepted 24 May 2010; published online 25 July 2010; doi:10.1038/nbt.1651

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

863

letters

© 2010 Nature America, Inc. All rights reserved.

Active Heat-inactivated sialidase sialidase

e

Cet Pan

–

**

–1

Concentration (ng µl )

f

0.2

0.4

Periodate treatment

Mock treatment

0

Periodate treatment

Control human IgG

0

Pan +

–

+

Anti-Neu5Gc Human IgG

***

0.8

A495

Cet

Sialidase:

Human antiNeu5Gc IgG

Control IgY

d

***

Control human IgG

Control IgY

Anti-Neu5Gc IgY

Control IgY

Anti-Neu5Gc IgY

0

Cet Pan

+

Human antiNeu5Gc IgG

Anti-Neu5Gc IgY

0.2

–

Control IgY

0.4

c

Pan +

Anti-Neu5Gc IgY

Coomassie staining

Cet –

Control IgY

***

0.6

A495

Sialidase:

Anti-Neu5Gc IgY

b

Cet Pan

A495

a

Mock treatment

S34 S30

**

2

1

0 Cet

Pan

No Ab

Figure 1  ELISA and western blot detection of Neu5Gc on biotherapeutic antibodies by Neu5Gc IgY antibodies from chickens or IgG antibodies from normal human serum. Cetuximab (Cet) and panitumumab (Pan) were treated with active sialidase to eliminate sialic acid epitopes or with heatinactivated sialidase as control. (a,b) Samples were used for ELISA (a) or western blot (b), in which Neu5Gc was detected using an affinity-purified chicken anti-Neu5Gc IgY or control IgY. A495, absorbance at 495 nm. ***P < 0.001, paired two-tailed t-test. (c) In an additional ELISA, Cet and Pan were used for coating, then blocked, and sialic acid epitopes were modified chemically using mild sodium metaperiodate pretreatment. The reaction was stopped using sodium borohydride. As a control, periodate and borohydride were mixed and then added to the wells (the borohydride inactivates the periodate). ELISA samples were studied at least in triplicate and data shown are means ± s.d. ***P < 0.001, paired two-tailed t-test. (d) Cet and Pan were pretreated with mild periodate as in c and used to coat ELISA wells before blocking and incubation with human anti-Neu5Gc IgG that had been purified from the serum of healthy humans and biotinylated as previously described 4. Samples were studied in triplicate and data shown are means ± s.d. ***P < 0.001, paired two-tailed t-test. (e) Cet and Pan (1 μg each) were treated with sialidase or heat-inactivated sialidase as in a separated by SDS-PAGE, Coomassie-stained, blotted (see b), and Neu5Gc detected using biotinylated human anti-Neu5Gc IgG. (f) Immune complex formation with Cet or Pan in whole human serum was detected using the CIC (C1Q) ELISA Kit (Buehlmann) as described in the manufacturer’s guidelines. Absorbance was measured at 405 nm. Samples were studied in triplicate and data shown are means ± s.d. **P < 0.01, paired two-tailed t-test. Gels in b and e were cropped for clarity of presentation. Full-length blots and gels are presented in Supplementary Figures 1–4.

In contrast, cetuximab carries 1.84 mol of sialic acids per mole of ­protein, mostly as Neu5Gc (see Supplementary Table 1). The differences probably reflect different cell-expression systems. For example, in contrast to CHO cells, murine myeloma cell lines express a greater proportion of sialic acids as Neu5Gc (ref. 17; see Supplementary Tables 2 and 3 for a listing of other potential examples). Pull-down assays of cetuximab with SNA-agarose (modified with the lectin Sambucus nigra agglutinin, which recognizes α2-6-linked sialic acids), followed by ELISAs of unbound proteins, showed that only about half of cetuximab mole­cules actually carry bound sialic acids and Neu5Gc (data not shown). Such heterogeneity is typical for glycoproteins. To address the potential significance of the high levels of antiNeu5Gc antibodies found in certain humans, we affinity purified antiNeu5Gc antibodies from normal human sera and biotinylated them exactly as previously described4, before their analysis using ELISA and western blotting assays (Fig. 1d,e). As with Neu5Gc-specific chicken IgY, these affinity-purified human Neu5Gc-specific antibodies reacted with cetuximab but not with panitumumab. Again, reactivity was abrogated by pretreatment with mild sodium periodate (Fig. 1d) or sialidase (Fig. 1e). To further address potential clinical relevance, we studied whether addition of cetuximab to normal human sera is capable of promoting the formation of immune complexes (Fig. 1f). Cetuximab formed immune complexes in a human serum with high levels of anti-Neu5Gc antibodies (serum S34; ref. 4) but not in a low-titer serum (serum S30; ref. 4). In contrast, we detected no formation of immune complexes with either serum in the presence of panitumumab. Assuming that similar interactions occur between cetuximab and circulating antiNeu5Gc antibodies in humans, these complexes could potentially fix complement and cause untoward reactions in some patients and/or affect half-life, possibly explaining some reported clinical differences between cetuximab and panitumumab13,15, We next evaluated whether Neu5Gc affects clearance rate when circulating anti-Neu5Gc antibodies are present. To mimic the situation

in humans, we used mice with a human-like defect in the Cmah gene, which encodes the enzyme that generates activated Neu5Gc (CMP-Neu5Gc)18. Such mice can make anti-Neu5Gc antibodies upon immunization with glycosidically bound, but not free, Neu5Gc19–21. However, the previous studies reporting these mouse anti-Neu5Gc antibodies used whole rodent or chimpanzee cells for immunization19,20, an artificial approach. In contrast, feeding of Neu5Gc (which is present in mouse chow) does not induce a human-like immune response in the mutant mice21. We could not immunize the mice with cetuximab itself, as other antibodies directed against the partly human IgG protein backbone would confound any results. To most closely mimic the situation in humans, we therefore immunized with Neu5Gcloaded Haemophilus influenzae (see Online Methods and ref. 21; this is very similar to the mechanism by which human Neu5Gc-specific antibodies appear to be generated naturally21). Given the great variability in isotypes and affinities of the naturally occurring human anti-Neu5Gc antibodies, as well as their different relative reactivities against various Neu5Gc-containing antigens4, it is impractical to model all possible human conditions. We therefore chose to mimic a situation in a human with relatively high levels of the IgG antibodies against the kind of Neu5Gc epitope (Neu5Gcα2-6Galβ1-4Glc-) found in cetuximab22. It also happens that this epitope is commonly recognized by human anti-Neu5Gc antibodies4. Each of the therapeutic antibodies, cetuximab and panitumumab, was injected intravenously at levels estimated to ensure a concentration of 1 μg ml−1 in the extracellular fluid volume according to mouse body weight23. Next, sera pooled from naïve, control-immunized or Neu5Gc-immunized syngeneic mice were passively transferred via intraperitoneal injection, ensuring equal starting concentrations of circulating Neu5Gc-specific antibodies. Anti-Neu5Gc IgG levels in the pooled sera from Neu5Gc-immunized mice were quantified using ELISA with a Neu5Gcα2-6Galβ1-4Glc-conjugate as a target, as pre­ viously described4 (97.5 μg ml−1, data not shown). The amount of pooled antibody injected was then calculated to achieve an approximate

864

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

a

b

Circulating Cet or Pan (normalized)

Change in A495

letters Figure 2  Effects of Neu5Gc-specific antibodies on the kinetics of mlgG Pan Cet 0.8 100 therapeutic antibodies in mice with a human-like Neu5Gc deficiency, 0.6 levels of anti-Neu5Gc IgG in mice after injections of the therapeutic 0.4 0.2 antibodies, and binding of IgG Neu5Gc-specific antibodies from whole 0 human serum to Neu5Gc on the Fab fragment of cetuximab. (a) Cmah–0.2 null mice were first injected intravenously with either of the therapeutic 0.3 80 antibodies, cetuximab (Cet) or panitumumab (Pan). Serum from Cmah-null 0.2 mice containing anti-Neu5Gc antibodies (or serum from naïve mice or 0.1 control-immunized mice) was then passively transferred by intraperitoneal 0 *** *** injection. Mice were bled periodically after the passive transfer of serum. –0.1 60 Concentrations of Cet or Pan in the isolated sera were determined by 0 50 100 0.6 sandwich ELISA. Absorbance was measured at 495 nm. The y axis starts S34 * Hours S30 at 60% to better display the difference in kinetics. Error bars, s.d.; 0.4 ***P < 0.001, unpaired two-tailed t-test. (b) Cmah-null mice were injected Serum of Serum of NaÏve intravenously with Cet, Pan or mouse IgG weekly and were bled initially controlNeu5Gcmouse serum immunized immunized 0.2 and after the third intravenous injection. To detect Neu5Gc-specific mice mice antibodies by ELISA, we coated wells with human (Neu5Gc-deficient) Cet 0 or chimpanzee (Neu5Gc-positive) serum glycoproteins (upper chart), or Pan Periodate Mock Periodate Mock alternatively with human or bovine fibrinogen (lower chart). Data were Fab Cet Fab Pan obtained in triplicate. (c) Fab fragments of Cet and Pan were isolated using the Pierce Fab Preparation Kit according to the manufacturer’s manual for use as target molecules in ELISA (1 μg per well). Sialic acid–specific binding was determined using mild sodium metaperiodate pretreatment. Wells were then blocked and incubated with human sera (S30 and S34, with low and high anti-Neu5Gc IgG titers, respectively4). Binding of human IgG was detected using anti-human IgG-Fc. Absorbance was measured at 490 nm and ELISA samples were studied in triplicate. Error bars, s.d.; *P < 0.05, paired two-tailed t-test.

© 2010 Nature America, Inc. All rights reserved.

A495

c

starting concentration of 4 μg ml−1 IgG in the extracellular fluid volumes of the mice, which is about a four fold excess of anti-Neu5Gc antibodies compared to the injected drug in the mice, and similar to levels found in some humans4. Clearance was monitored by a sandwich ELISA specific for human IgG-Fc. Although both drugs had a similar clearance rate in mice pre-injected with serum from naïve or control-immunized mice, circulating levels of cetuximab decreased significantly (P < 0.001) when Neu5Gc-specific antibodies were pre-injected (Fig. 2a). Assuming that a similar interaction between cetuximab and circulating anti-Neu5Gc antibodies occurs in patients, there could be relevant effects on clearance rate and efficacy. This might help to explain the wide range of half-life values reported for such antibodies in clinical studies14,15. To further simulate the clinical situation, we injected equal amounts of cetuximab or panitumumab intravenously into Neu5Gcdeficient Cmah−/− mice in typical human dosages (4 μg per gram of body weight) at weekly intervals. To exclude any effect of the human portion of the protein (cetuximab) or of the fully human protein (panitumumab) in mice, we also injected murine IgG as a positive control, as it happens to carry Neu5Gc as the predominant sialic acid (Supplementary Table 1). Notably, cetuximab and murine IgG (but never panitumumab) induced a Neu5Gc-specific IgG immune response (Fig. 2b). As with humans, responses of individual mice varied greatly, and more positive signals were obtained with the Neu5Gc epitope mixture found in chimp serum than that in bovine fibrinogen. Thus, even patients without pre-existing high levels of anti-Neu5Gc antibodies may be at risk of developing them after injection of Neu5Gc-carrying agents, potentially affecting the outcome of subsequent injections. Moreover, repeated injections of Neu5Gccarrying agents could result in the accumulation of this nonhuman sugar in human tissues. Together with Neu5Gc-specific antibodies, accumulation of Neu5Gc in tissues can mediate chronic inflammation and potentially facilitate progression of diseases such as cancer19 and atherosclerosis24. Thus, chronic use of Neu5Gc-bearing therapeutics might increase future risk of such diseases. Finally, we studied direct binding of anti-Neu5Gc antibodies from whole human sera to both cetuximab and panitumumab. To avoid

excessive cross-reactivity involving the secondary reagent, we prepared Fab fragments of both of the agents, used them to coat ELISA plate wells, exposed them to human sera and then detected serum antibody binding with a human IgG-Fc–specific secondary antibody (note that cetuximab is known to have an additional glycosylation site in the V-region21). We detected mild periodate–sensitive binding of serum IgG from a high–anti-Neu5Gc titer serum (S34, ref. 4), which had >15 μg ml−1 IgG antibodies against Neu5Gcα2-6Galβ14Glc-) to the Fab fragments of cetuximab and not to those of panitumumab (Fig. 2c). In contrast, incubation with another human serum containing very low Neu5Gc-antibodies (serum S30, ref. 4, which had <2 μg ml−1 IgG antibodies against Neu5Gcα2-6Galβ1-4Glc-) did not show much periodate-sensitive binding (Fig. 2c). Thus, whole human sera with high (but not low) titers of anti-Neu5Gc antibodies showed sialic acid–dependent (mild periodate–sensitive) binding of serum IgG to cetuximab but not to panitumumab. Assuming that our in vitro and in vivo data translate into clinically relevant difficulties for some patients given Neu5Gc-containing biotherapeutics, how can one prevent this problem? Even biotherapeutics produced in human cell lines can be contaminated with Neu5Gc incorporated from animal-derived culture medium materials. Furthermore, it is difficult to change from the well-established and FDA-approved cell lines currently in use. Also, Neu5Gc in cells is ‘recycled’ in lysosomes and reused11. We hypothesized that free N-acetylneuraminic acid (Neu5Ac) in the medium would be taken up and compete with pre-existing Neu5Gc, preventing such recycling. Also, excess Neu5Ac would compete with any further incorporation of medium Neu5Gc (ref. 11). We therefore preloaded human 293T cells with Neu5Gc and then chased this with a culture medium with or without 5 mM Neu5Ac. Exposure to Neu5Ac indeed resulted in more rapid disappearance of both ethanol-precipitable (glycosidically bound) and ethanol-soluble (free) Neu5Gc from the human cells (Fig. 3a,b). Thus, addition of Neu5Ac to the medium is a simple and nontoxic way to eliminate or reduce Neu5Gc contamination of human cells, which are now being used in the production of biotherapeutic agents—for example, Xigris from Eli Lilly and Elaprase from Shire Pharmaceuticals. Given the considerable number of recombinant biotherapeutic glycoproteins currently produced in nonhuman cells and the presence of

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

865

letters

10

5 mM Neu5Ac – +

Size 197 (kDa) 125 83

5 37

7 D ay

5

0

3

7 D ay

D ay

5

0

e

Membrane fraction Control 5 mM Neu5Ac

D ay

5

Neu5Gc (% of total sialic acids)

10

d

D ay

15

3

Neu5Gc (% of total sialic acids)

Secreted protein Control 5 mM Neu5Ac

D ay

4

5 D ay

D ay

3

0

2

5

3

2

4

D ay

D ay

D ay

1

D ay

D ay

0

0

20

D ay

20

40

1

40

60

D ay

60

80

0

80

c

Ethanol-precipitable fraction Control 5 mM Neu5Ac

D ay

5 mM Neu5Ac

D ay

Control

Neu5Gc (% of total sialic acids)

b

Ethanol-soluble fraction

D ay

Neu5Gc (% of total sialic acids)

a

© 2010 Nature America, Inc. All rights reserved.

Figure 3  An approach to reducing Neu5Gc contamination in biotherapeutic products. (a,b) Human 293T cells were grown in the presence of 5 mM Neu5Gc for 3 d. The cells were then washed with PBS and split into two identical cultures, and 5 mM Neu5Ac was added to one of the cultures. Cells were harvested as described in Online Methods, and the Neu5Gc and Neu5Ac content of both the ethanol-soluble (a) and ethanol-precipitable proteins (b) was analyzed by HPLC. (c–e) Feeding of CHO cells with free Neu5Ac reduced Neu5Gc in the whole-cell membranes and in secreted glycoproteins. Stably transfected CHO-KI cells expressing a recombinant soluble IgG-Fc fusion protein were grown in the absence or presence of 5 mM Neu5Ac. The individually collected medium was centrifuged to remove cell debris and adjusted to 5 mM Tris-HCl pH 8. The fusion protein was purified using protein A–Sepharose, and sialic acid content was determined by DMB-HPLC analysis as described in Online Methods (c). Total cell membranes from the same CHO cells were prepared and used for DMB-HPLC analysis (d). CHO membrane proteins from d were separated by SDS-PAGE and transferred onto nitrocellulose membranes. Expression of Neu5Gc (e) was detected by incubating with polyclonal affinity-purified chicken anti-Neu5Gc IgY, as described in Online Methods.

small amounts of Neu5Gc in recombinant glycoproteins produced in CHO cells1,16, we next asked whether feeding Neu5Ac could reduce total glycoprotein Neu5Gc levels in CHO cells. This was successful for all membrane glycoproteins and for a secreted recombinant protein (Fig. 3c–e). Similar feeding of murine myeloma cells with Neu5Ac did not substantially reduce the higher initial Neu5Gc content (~70–80% of sialic acids), most probably because of the higher baseline levels of Cmah in these cells. Regardless, given that the CHO cell expresses its own Cmah enzyme, these data suggest a novel mechanism, in addition to Neu5Ac competing out recycled Neu5Gc. Whatever that mechanism, reduction of Neu5Gc content of a recombinant glycoprotein can be achieved even in a nonhuman Cmah-positive cell line that starts with low levels of Neu5Gc. Despite their successful use for a variety of indications, infusionrelated reactions, immunogenicity and accelerated clearance remain important concerns for many therapeutic glycoproteins7,25. The incidence and severity of an immune reaction depends on the interplay of infused agents with the immune system and can vary greatly from patient to patient. Understanding the underlying nature of these events will help to identify patients at risk with the use of specific markers. Humanized and fully human antibodies have been developed to reduce immunogenicity due to peptide epitopes5. However, the potential immunogenicity of the glycans they carry has not been as well considered. It is known that immune reactions can be mediated by binding of pre-existing IgEs against the nonhuman alpha-Gal epitope carried by some agents, such as cetuximab13. However, in our studies alpha-Gal residues are not an issue, as Cmah-null mice already express this sequence and do not have antibodies against it. A further concern arises here because pre-existing antibodies against a glycan on a glycoprotein can secondarily enhance antibody reactivity against the underlying protein backbone26, perhaps because immune complexes are cleared efficiently by Fc receptors into dendritic cells and other antigen-presenting cells27,28. Such a mechanism might help explain why patients’ immunogenicity to some glyco­ protein therapeutics sometimes increases over time26,29,30. If this were true, it would likely have a further impact in long-term replacement therapy with recombinant therapeutic glycoproteins. Our findings suggest that the potential significance of the presence of Neu5Gc on glycoprotein biotherapeutics should be revisited. Despite a natural tendency to downplay potential new problems involving currently useful drugs, it is worthwhile to consider lessons from other fields, where initial enthusiasm was not balanced by full

appreciation of immunological implications31. With this in mind, we have also suggested that Neu5Gc contamination of stem cells and other cell types intended for human therapy could pose risks32,33. In addition, others have recently reported that Cmah-null mice can reject Neu5Gc-positive wild-type organ transplants via complement-fixing Neu5Gc-specific antibodies20. For new drugs, it may be possible to avoid Neu5Gc contamination from the outset by using Neu5Gc-deficient cells and media. Meanwhile, as an immediate practical solution, we have also demon­ strated a nontoxic way to reduce the Neu5Gc content of some currently used expression systems and their secreted glycoproteins, by simply adding Neu5Ac to the culture media. This could bypass the need to establish new Neu5Gc-deficient cell lines for already approved drugs. The addition of Neu5Ac to the media could also potentially increase total sialylation of a glycoprotein biotherapeutic agent. But if anything, such an increase would only be beneficial—for example, leading to a longer half-life of the agent in vivo.

866

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Methods Methods and any associated references are available in the online version of the paper at http://www.nature.com/naturebiotechnology/. Note: Supplementary information is available on the Nature Biotechnology website. Acknowledgments This work was supported by US National Institutes of Health grants R01-GM32373 and R01-CA38701 to A.V. and The International Sephardic Education Foundation for V.P.-K. Haemophilus influenzae strain 2019 was a generous gift from M. Apicella, Department of Microbiology, University of Iowa. AUTHOR CONTRIBUTIONS All authors helped design the studies; D.G. and S.D. performed the research; R.E.T. and V.P.-K. generated crucial reagents; D.G. and A.V. wrote the paper; and all authors read the paper. COMPETING FINANCIAL INTERESTS The authors declare competing financial interests: details accompany the full-text HTML version of the paper at http://www.nature.com/naturebiotechnology/. Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/. 1. Hokke, C.H. et al. Sialylated carbohydrate chains of recombinant human glycoproteins expressed in Chinese hamster ovary cells contain traces of N-glycolylneuraminic acid. FEBS Lett. 275, 9–14 (1990).

© 2010 Nature America, Inc. All rights reserved.

letters 2. Noguchi, A., Mukuria, C.J., Suzuki, E. & Naiki, M. Failure of human immunoresponse to N-glycolylneuraminic acid epitope contained in recombinant human erythropoietin. Nephron 72, 599–603 (1996). 3. Tangvoranuntakul, P. et al. Human uptake and incorporation of an immunogenic nonhuman dietary sialic acid. Proc. Natl. Acad. Sci. USA 100, 12045–12050 (2003). 4. Padler-Karavani, V. et al. Diversity in specificity, abundance, and composition of anti-Neu5Gc antibodies in normal humans: potential implications for disease. Glycobiology 18, 818–830 (2008). 5. Aggarwal, S. What′s fueling the biotech engine—2007. Nat. Biotechnol. 26, 1227–1233 (2008). 6. Arnold, J.N., Wormald, M.R., Sim, R.B., Rudd, P.M. & Dwek, R.A. The impact of glycosylation on the biological function and structure of human immunoglobulins. Annu. Rev. Immunol. 25, 21–50 (2007). 7. Durocher, Y. & Butler, M. Expression systems for therapeutic glycoprotein production. Curr. Opin. Biotechnol. 20, 700–707 (2009). 8. Higgins, E. Carbohydrate analysis throughout the development of a protein therapeutic. Glycoconj. J. 27, 211–225 (2009). 9. Galili, U. Immune response, accommodation, and tolerance to transplantation carbohydrate antigens. Transplantation 78, 1093–1098 (2004). 10. Varki, A. Glycan-based interactions involving vertebrate sialic-acid-recognizing proteins. Nature 446, 1023–1029 (2007). 11. Bardor, M., Nguyen, D.H., Diaz, S. & Varki, A. Mechanism of uptake and incorporation of the non-human sialic acid N-glycolylneuraminic acid into human cells. J. Biol. Chem. 280, 4228–4237 (2005). 12. Borys, M.C. et al. Effects of culture conditions on N-glycolylneuraminic acid (Neu5Gc) content of a recombinant fusion protein produced in CHO cells. Biotechnol. Bioeng. 105, 1048–1057 (2009). 13. Chung, C.H. et al. Cetuximab-induced anaphylaxis and IgE specific for galactosealpha-1,3-galactose. N. Engl. J. Med. 358, 1109–1117 (2008). 14. Delbaldo, C. et al. Pharmacokinetic profile of cetuximab (Erbitux) alone and in combination with irinotecan in patients with advanced EGFR-positive adenocarcinoma. Eur. J. Cancer 41, 1739–1745 (2005). 15. Saadeh, C.E. & Lee, H.S. Panitumumab: a fully human monoclonal antibody with activity in metastatic colorectal cancer. Ann. Pharmacother. 41, 606–613 (2007). 16. Diaz, S.L. et al. Sensitive and specific detection of the non-human sialic acid N-glycolylneuraminic acid in human tissues and biotherapeutic products. PLoS ONE 4, e4241 (2009). 17. Muchmore, E.A., Milewski, M., Varki, A. & Diaz, S. Biosynthesis of N-glycolyneuraminic acid. The primary site of hydroxylation of N-acetylneuraminic acid is the cytosolic sugar nucleotide pool. J. Biol. Chem. 264, 20216–20223 (1989). 18. Hedlund, M. et al. N-glycolylneuraminic acid deficiency in mice: implications for human biology and evolution. Mol. Cell. Biol. 27, 4340–4346 (2007). 19. Hedlund, M., Padler-Karavani, V., Varki, N.M. & Varki, A. Evidence for a humanspecific mechanism for diet and antibody-mediated inflammation in carcinoma progression. Proc. Natl. Acad. Sci. USA 105, 18936–18941 (2008). 20. Tahara, H. et al. Immunological property of antibodies against N-glycolylneuraminic acid epitopes in cytidine monophospho-n-acetylneuraminic acid hydroxylasedeficient mice. J. Immunol. 184, 3269–3275 (2010).

21. Taylor, R.E. et al. Novel mechanism for the generation of human xeno-autoantibodies against the non-human sialic acid N-glycolylneuraminic acid. J. Exp. Med. published online, doi: 10.1084/jem.20100575 (12 July 2010). 22. Qian, J. et al. Structural characterization of N-linked oligosaccharides on monoclonal antibody cetuximab by the combination of orthogonal matrix-assisted laser desorption/ionization hybrid quadrupole-quadrupole time-of-flight tandem mass spectrometry and sequential enzymatic digestion. Anal. Biochem. 364, 8–18 (2007). 23. Axworthy, D.B. et al. Cure of human carcinoma xenografts by a single dose of pretargeted yttrium-90 with negligible toxicity. Proc. Natl. Acad. Sci. USA 97, 1802–1807 (2000). 24. Pham, T. et al. Evidence for a novel human-specific xeno-auto-antibody response against vascular endothelium. Blood 114, 5225–5235 (2009). 25. Jahn, E.M. & Schneider, C.K. How to systematically evaluate immunogenicity of therapeutic proteins—regulatory considerations. New Biotechnol. 25, 280–286 (2009). 26. Galili, U. et al. Enhancement of antigen presentation of influenza virus hemagglutinin by the natural human anti-Gal antibody. Vaccine 14, 321–328 (1996). 27. Benatuil, L. et al. The influence of natural antibody specificity on antigen immunogenicity. Eur. J. Immunol. 35, 2638–2647 (2005). 28. Abdel-Motal, U.M., Wigglesworth, K. & Galili, U. Mechanism for increased immunogenicity of vaccines that form in vivo immune complexes with the natural anti-Gal antibody. Vaccine 27, 3072–3082 (2009). 29. Koren, E. et al. Recommendations on risk-based strategies for detection and characterization of antibodies against biotechnology products. J. Immunol. Methods 333, 1–9 (2008). 30. Shankar, G., Pendley, C. & Stein, K.E. A risk-based bioanalytical strategy for the assessment of antibody immune responses against biological drugs. Nat. Biotechnol. 25, 555–561 (2007). 31. Wilson, J.M. Medicine. A history lesson for stem cells. Science 324, 727–728 (2009). 32. Martin, M.J., Muotri, A., Gage, F. & Varki, A. Human embryonic stem cells express an immunogenic nonhuman sialic acid. Nat. Med. 11, 228–232 (2005). 33. Martin, M.J., Muotri, A., Gage, F. & Varki, A. Response to Cerdan et al.: Complement targeting of nonhuman sialic acid does not mediate cell death of human embryonic stem cells. Nat. Med. 12, 1115 (2006). 34. Van Hoeyveld, E. & Bossuyt, X. Evaluation of seven commercial ELISA kits compared with the C1q solid-phase binding RIA for detection of circulating immune complexes. Clin. Chem. 46, 283–285 (2000). 35. Campagnari, A.A., Gupta, M.R., Dudas, K.C., Murphy, T.F. & Apicella, M.A. Antigenic diversity of lipooligosaccharides of nontypable Haemophilus influenzae. Infect. Immun. 55, 882–887 (1987). 36. Greiner, L.L. et al. Nontypeable Haemophilus influenzae strain 2019 produces a biofilm containing N-acetylneuraminic acid that may mimic sialylated O-linked glycans. Infect. Immun. 72, 4249–4260 (2004). 37. Gagneux, P. et al. Proteomic comparison of human and great ape blood plasma reveals conserved glycosylation and differences in thyroid hormone metabolism. Am. J. Phys. Anthropol. 115, 99–109 (2001). 38. Debeire, P., Montreuil, J., Moczar, E., van Halbeek, H. & Vliegenthart, J.F.G. Primary structure of two major glycans of bovine fibrinogen. Eur. J. Biochem. 151, 607–611 (1985).

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

867

ONLINE METHODS

Mice. The Cmah-null mice used for this study have been described previously18 and were backcrossed to C57Bl/6 mice for over ten generations. All experiments were approved by the University of California, San Diego Institutional Review Board committee responsible for approving animal experiments.

© 2010 Nature America, Inc. All rights reserved.

Sialidase treatment of therapeutic antibodies. One milligram each of cetuximab or panitumumab (obtained from the University of California, San Diego pharmacy or the manufacturer) were treated with 50 mU of active or heatinactivated Arthrobacter ureafaciens sialidase (EY Laboratories) in 100 mM sodium acetate pH 5.5, at 37 °C for 24 h. Samples were used for ELISA or western blots. Periodate treatment of therapeutic antibodies on ELISA plate. Untreated cetuximab and panitumumab (1 μg per well) were used for coating, then blocked with PBST for 2 h and incubated with freshly made 2 mM sodium metaperiodate in PBS for 20 min at 4 °C in the dark. The reaction was stopped by addition of 200 mM sodium borohydride to a final concentration of 20 mM. As a control, periodate and borohydride were premixed and then added to the wells (the borohydride inactivates the periodate). To remove resulting borates, wells were then washed three times with 100 mM sodium acetate with 100 mM NaCl pH 5.5 before further analysis. ELISA detection of Neu5Gc on therapeutic antibodies. For the ELISA, wells were coated with 1 μg of cetuximab or panitumumab (either before sialidase treatment or after periodate treatment), blocked with TBST for 2 h and then incubated with affinity-purified chicken anti-Neu5Gc IgY or control IgY for 1 h (1:20,000 in TBST). Binding of IgY was detected using horseradish peroxidase (HRP)-conjugated donkey anti-chicken IgY (1:50,000 in TBST) and developed with O-phenylenediamine in citrate-phosphate buffer, pH 5.5, with absorbance measured at 495 nm. ELISA samples were studied at least in triplicate. Similar to the ELISA with the anti-Neu5Gc chicken IgY, human anti-Neu5Gc IgG that had been purified from the serum of healthy humans and biotinylated (exactly as described in ref. 4) was also used as the primary antibody (1:100 in TBST). Binding of the human antibodies to the therapeutic antibodies was detected using HRP-conjugated streptavidin (1:10,000) followed by development as described above. Samples were studied in triplicate. Western blot detection of Neu5Gc on therapeutic antibodies. For ­western blot detection, cetuximab or panitumumab (1 μg per lane) was separated by 12.5% SDS-PAGE and Coomassie-stained or blotted on nitrocellulose membranes. Blotted membranes were blocked with TBST containing 0.5% cold-water fish-skin gelatin overnight at 4 °C and subsequently incubated with affinity-purified chicken anti-Neu5Gc IgY for 4 h at room temperature (1:100,000 in TBST). Binding of the chicken anti-Neu5Gc IgY was detected using HRP-conjugated donkey anti-chicken IgY for 1 h (1:50,000 in TBST), followed by incubation with SuperSignal West Pico Substrate (Pierce) as per the manufacturer’s recommendation, exposure to X-ray film and development of the film. Similar to the western blot with the chicken anti-Neu5Gc IgY, purified biotinylated human anti-Neu5Gc IgG was also used as the primary antibody (1:100 in TBST). Binding of the human antibodies to the therapeutic antibodies was detected using HRP-conjugated streptavidin (1:10,000 in TBST) followed by development as described above. CIC-C1q binding assay. Immune complex formation was detected using the CIC (C1Q) ELISA Kit (Buehlmann) as described in the manufacturer’s guidelines34. Briefly, 100 μl of human serum with low or high anti-Neu5Gc (S30 and S34, respectively4) was incubated with 40 μg of cetuximab or panitumumab for 14 h at 4 °C. We applied 1:50 dilutions of the mix to human C1q–coated ELISA wells and incubated for 1 h at 25 °C. Binding was detected using alkaline phosphatase–conjugated protein A. After another washing step, the enzyme substrate (para-nitrophenylphosphate) was added, followed by a stopping step. The absorbance was measured at 405 nm. Samples were studied in triplicate. Generation of murine Neu5Gc-specific antibodies. Haemophilus influenzae strain 2019 (ref. 35) was a generous gift from M. Apicella, Department of

nature biotechnology

Microbiology, University of Iowa. Bacteria were grown to mid log phase in sialic acid–free media36 with or without addition of 1 mM Neu5Gc21, heatkilled and injected intraperitoneally (200 μl of culture at an absorbance of 600 nm of 0.4) into Cmah-null mice. Effects of anti-Neu5Gc antibodies on in vivo kinetics of therapeutic antibodies. Cetuximab or panitumumab in PBS (0.24 μg per gram mouse body weight) were injected intravenously, and 14 h later, mouse serum pooled from syngeneic Cmah-null mice containing anti-Neu5Gc antibodies (or pooled serum from syngeneic naïve or control-immunized mice) was passively transferred via intraperitoneal injection into syngeneic Cmah-null mice that were prescreened for the absence of pre-existing antibodies against human IgG. Mice were bled 0, 2, 8, 32, 56 and 80 h after the passive transfer of mouse serum. For quantification of therapeutic antibody concentrations in the sera, wells of ELISA plates were coated with 1 μg of anti-human IgG (Biorad), then blocked with TBST for 2 h and incubated with 1:500 dilutions of the sera in each well. Captured therapeutic antibodies were detected by HRP-conjugated anti-human Fc (Jackson; 1:10,000), with development by O-phenylenediamine in citrate-phosphate buffer, pH 5.5, and absorbance measured at 495 nm (n = 5 for injections of both control sera groups; n = 10 for injections of antiNeu5Gc serum groups). Quantification of Neu5Gc-specific IgG antibodies in Neu5Gc-immunized mice. A Neu5Gcα2-6Galβ1-4Glc-conjugate4 (1 μg per well) and serial dilutions of mouse IgG as standards (0.625–20 ng per well) were used for coating overnight, then blocked with PBST for 2 h and incubated with pooled serum from Neu5Gc-immunized mice (1:250 dilution) for 2 h at 25 °C. Binding of mouse IgG was detected using HRP-conjugated goat anti-mouse IgG-Fc (Jackson; 1:10,000 in PBST) and developed with O-phenylenediamine in citrate-phosphate buffer, pH 5.5, with absorbance measured at 490 nm. ELISA samples were studied in triplicate. Levels of anti-Neu5Gc IgG after injections of the antibodies. Cmah-null mice were injected intravenously with 4 μg antibody per gram of mouse body weight in PBS weekly for 3 weeks. Mice were bled initially, and again 1 week after the third intravenous injection. Wells of ELISA plates were coated with 1:1,000 dilutions of human (Neu5Gc-deficient) or chimpanzee (Neu5Gc-positive) serum glycoproteins (note that the only major difference between human and chimp serum glycosylation is the absence or presence of Neu5Gc; ref. 37). Alternatively, wells were coated with human or bovine fibrinogen, which carry Neu5Ac or Neu5Gc on otherwise identical N-glycans38. Wells were then blocked with TBST for 2 h followed by incubation with 1:100 dilutions of the mouse sera. Binding of the mouse antibodies was detected using HRP-conjugated goat anti-mouse IgG Fc fragment (1:10,000 in TBST). Neu5Gc-specific binding (change in absorbance at 495 nm) was determined by subtracting the background signal of the wells coated with human serum or human fibrinogen (no Neu5Gc) from the signal of chimpanzee serum–coated or bovine fibrinogen–coated wells (containing Neu5Gc). Data were obtained in triplicate (n = 5 for injection of mouse IgG; n = 4 for injection of pani­ tumumab; n = 6 for injection of cetuximab ). An approach to reduce Neu5Gc contamination in biotherapeutic products. Human 293T kidney cells were grown in DME supplemented with 10% (vol/vol) fetal calf serum. Cells were lifted from the culture plate using 20 mM EDTA in PBS and allowed to grow to 50% confluence. At this point, buffered 100 mM Neu5Gc was added to the culture in duplicate for a final 5 mM concentration, and the cells were grown in this supplemented media for 3 d. At the end of this Neu5Gc pulse, the cells were once again lifted using 20 mM EDTA in PBS, pelleted, washed once with PBS to remove any excess Neu5Gc and then suspended in 30 ml of growth medium. We added 5 ml of this cell suspension to each of five P-100 dishes. We immediately harvested the last aliquot of cell suspension, at time 0, by pelleting the cells, washing once with PBS, suspending them in 1 ml of PBS and transferring them to a 1.5-ml microcentrifuge tube. The cells were repelleted and frozen until all time points were collected. Buffered 100 mM Neu5Ac was added to each of the other five plates for the ‘Neu5Ac chase’ and an equivalent amount of media added to the ‘minus chase’ samples. We harvested cells at days 1, 2, 3, 4 and 5 by scraping them into the

doi:10.1038/nbt.1651

acid (final) and incubation at 80 °C for 3 h. Samples were passed through a Microcon-10 filter and the filtrate derivatized with DMB (1,2-diamino-4, 5-methylenedioxybenzene) reagent for analysis of sialic acids by HPLC. A similar approach was taken with CHO cells stably expressing a Siglec-Fc protein in the medium, except that the Neu5Gc pulse was omitted and the secreted glycoproteins were captured on protein A–Sepharose beads. The cells were also processed similarly, except that total cell membranes were pelleted by centrifugation. The sialic acid content of the secreted proteins and cell membranes was determined by acid hydrolysis, DMB derivatization and HPLC. The cell membranes were also studied by western blotting with the chicken anti-Neu5Gc IgY, as described above.

© 2010 Nature America, Inc. All rights reserved.

culture media, collecting by pelleting, washing once with PBS, transferring them to a 1.5-ml microcentrifuge tube, pelleting and freezing the cell pellet. At the end of the 5 d of chase, all collected cell pellets were homogenized in 300 μl of ice-cold 20 mM potassium phosphate pH 7 using a 3- to 20-s burst with a Fisher Sonicator. We precipitated glycoconjugate-bound sialic acids by adding 700 μl of 100% ice-cold ethanol (final 70% (vol/vol) correct ethanol) and incubating at −20 °C overnight. The samples were spun at 20,000g for 15 min and the supernatants transferred to clean tubes and dried on a speed vac. The precipitated glycoconjugates and dried ethanol supernatants were each suspended in 100 μl of 20 mM potassium phosphate pH 7 by sonication. Sialic acids were released from both fractions by acid hydrolysis with 2 M acetic

doi:10.1038/nbt.1651

nature biotechnology

letters

Global analysis of lysine ubiquitination by ubiquitin remnant immunoaffinity profiling

© 2010 Nature America, Inc. All rights reserved.

Guoqiang Xu, Jeremy S Paige & Samie R Jaffrey Protein ubiquitination is a post-translational modification (PTM) that regulates various aspects of protein function by different mechanisms. Characterization of ubiquitination has lagged behind that of smaller PTMs, such as phosphorylation, largely because of the difficulty of isolating and identifying peptides derived from the ubiquitinated portion of proteins. To address this issue, we generated a monoclonal antibody that enriches for peptides containing lysine residues modified by diglycine, an adduct left at sites of ubiquitination after trypsin digestion. We use mass spectrometry to identify 374 diglycine-modified lysines on 236 ubiquitinated proteins from HEK293 cells, including 80 proteins containing multiple sites of ubiquitination. Seventy-two percent of these proteins and 92% of the ubiquitination sites do not appear to have been reported previously. Ubiquitin remnant profiling of the multi-ubiquitinated proteins proliferating cell nuclear antigen (PCNA) and tubulin -1A reveals differential regulation of ubiquitination at specific sites by microtubule inhibitors, demonstrating the effectiveness of our method to characterize the dynamics of lysine ubiquitination. Protein ubiquitination occurs on a wide variety of eukaryotic proteins and affects processes ranging from protein degradation and sub­cellular localization to gene expression and DNA repair1. Ubiquitination involves the transfer of ubiquitin to a target protein using E1 ubiquitin– activating enzymes, E2 ubiquitin–conjugating enzymes and E3 ubi­ quitin ligases1. This process typically leads to the formation of an amide linkage comprising the ε-amine of lysine of the target protein and the C terminus of ubiquitin, and can involve ubiquitination at distinct sites within the same protein, although the roles of ubiquitination at distinct sites are incompletely understood. The human genome is predicted to encode 16 E1, 53 E2 and 527 E3 proteins2, which underscores the likely importance of ubiquitination in molecular signaling. In most cases, proteins suspected to be ubiquitinated have been identified based on their susceptibility to proteasome-mediated deg­ radation, as evidenced by their increased levels following application of proteasome inhibitors. These proteins are immunopurified and ubiquitin adducts are confirmed by anti-ubiquitin immuno­blotting3. Mutagenesis experiments can identify ubiquitination sites4. Global identification of ubiquitinated proteins has been performed by purifying ubiquitinated proteins, using ubiquitin-binding proteins

such as anti-ubiquitin antibodies5, or by purifying hexahistidine (His6)-tagged ubiquitin-protein conjugates6. The enriched set of pro­ teins are then proteolyzed and subjected to tandem mass spectrometry (MS/MS). However, as only one or a few lysines are typically modified in any ubiquitinated protein, most peptides do not exhibit any ubiquitin-derived modifications7. This introduces uncertainty whether they are derived from the nonubiquitinated portion of a protein or from coprecipitated proteins. Alternatively, proteolytic digests can be screened for peptides that contain remnants of ubiquitin modification. Digestion of ubiquitinconjugated proteins results in peptides that contain a ubiquitin rem­ nant derived from the ubiquitin C terminus. The three C-terminal residues of ubiquitin are Arg-Gly-Gly, with the C-terminal glycine conjugated to a lysine residue in the target. After digestion with trypsin, ubiquitin is cleaved after arginine, leaving a Gly-Gly dipep­ tide remnant on the conjugated lysine. Therefore, tryptic digests will include peptides that contain a diglycine-modified lysine, indicating the prior conjugation of ubiquitin to that region of the target protein. The diglycine-modified lysine serves as a signature of ubiquitination and also identifies the specific site of modification. Sequencing of ubiquitin remnant–containing peptides in ­ tryptic digests has been used to identify 110 ubiquitination sites from yeast expressing His6-ubiquitin7. Despite the availability of these approaches for several years, analysis of the Swiss-Prot database indicates that only 255 mammalian proteins have been reported to be ubiquitinated based on experimental evidence. In most cases, the ubiquitination sites have not been identified. Direct enrichment of ubiquitin remnant–containing peptides would facilitate the highthroughput identification of ubiquitination sites. To identify ubiquitinated proteins and simultaneously report their sites of ubiquitination, we generated an antibody that recognizes ­peptides containing the ubiquitin remnant left after trypsin ­digestion of ubiquitinated proteins. To prepare a protein antigen containing ­diglycine-modified lysines, we first reacted purified lysine-rich ­histone III-S protein with t-butyloxycarbonyl-Gly-Gly-N-hydroxysuccinimide (Boc-Gly-Gly-NHS) to introduce amide-linked Boc-Gly-Gly adducts on all amines (Fig. 1a). Nearly complete modification of the amines was confirmed by the reduction in labeling of the Boc-Gly-Gly–modified protein by the lysine-modifying reagent biotin-NHS, as assessed by anti-biotin immunoblotting (Fig. 1b). The modified protein was treated with trifluoroacetic acid (TFA) to remove the Boc moiety. Quantitative

Department of Pharmacology, Weill Medical College, Cornell University, New York, New York, USA. Correspondence should be addressed to S.R.J. ([email protected]). Received 25 March; accepted 11 June; published online 18 July 2010; doi:10.1038/nbt.1654

868

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

letters a

b

Boc-G ly-Gly

-

Boc-Gly -Gly-

Histone A B C Figure 1  Generation of monoclonal antibodies that selectively recognize NH 2 Boc-Gly-Gly-NHS – + + diglycine-modified lysines. (a) The antigen used to raise antibodies was NH 2 TFA – – + N NH H + + + Biotin-NHS synthesized by modifying the ε-amines of all lysines in a histone with 2 2 t-butyloxycarbonyl-Gly-Gly-N-hydroxysuccinimide (Boc-Gly-Gly-NHS) 250 150 and then removing the Boc group by treatment with TFA. The lysines Boc-Gly-Gly-NHS 100 in the final protein contain Gly-Gly adducts on the ε-amine of all lysine 75 residues. (b) To validate the synthesis of Gly-Gly–modified histone, 50 Bo WB: anti-biotin c-G 37 we monitored the reaction of the histone with Boc-Gly-Gly-NHS by lyGl y25 detecting amines, such as those in unmodified lysine, through the NH reaction of the protein with the amine-modifying agent biotin-NHS, 15 N NH H NH and subsequent western blot analysis with an anti-biotin antibody. lyG Lysozyme β-Lact Amines in the histone were completely lost after treatment with ly Brain lysate G A B C A B C cBoc-Gly-Gly-NHS, indicating complete modification of all the Bo A B C 180 lysines in the histone. Removal of the Boc protecting group with 115 TFA TFA resulted in the formation of an amine at the N terminus of the 82 185 64 Gl Gly-Gly adduct. This step was essentially complete, as the TFA-treated 98 y49 Gl 52 37 y protein exhibited nearly complete recovery of amine reactivity. 31 NH 26 The position of the bands in the different samples is slightly 19 15 17 NH NH shifted due to the different molecular weights and number of 6 ly- NH G 14 y l positive charges in the modified and unmodified samples. The G WB: anti–diglycyl-lysine bands above 50 kDa represent impurities in the histone sample. (c) We evaluated the specificity of the GX41 monoclonal antibody by western blot analysis of β-lactoglobulin, lysozyme or rat brain lysate, in which the lysines were either unmodified (A), or modified with Boc-Gly-Gly (B) or Gly-Gly- (C) adducts, respectively.

c

© 2010 Nature America, Inc. All rights reserved.

Gl y-G

ly-

Gly-Gly-

conversion of the Boc-Gly-Gly adduct, which does not contain an amine, to Gly-Gly, which contains an amine, was confirmed by the reactivity of the TFA-treated protein with biotin-NHS (Fig. 1b). We injected the diglycine-modified histone into mice, and screened hybridoma lines for antibodies that specifically recognize proteins containing diglycine-modified lysines. Hybridoma line GX41 gener­ ated monoclonal antibodies that exhibited pronounced specificity for proteins containing the diglycine-modified lysines. The antibodies failed to interact with unmodified lysozyme or lactoglobulin (Fig. 1c), or either of these proteins after they have been modified with Boc-GlyGly. However, the antibody recognized Gly-Gly–modified lysozyme and lactoglobulin obtained after removal of the Boc group. These results indicate that the antibody recognizes Gly-Gly–modified lysines, and suggest that the antibody only recognizes Gly-Gly adducts that contain an unmodified primary amine. Similarly, the antibody exhibits negligible reactivity with rat brain lysate (Fig. 1c), or brain lysate modified with Boc-Gly-Gly, but exhibits substantial reactivity with Gly-Gly–modified proteins from brain lysate. Notably, the brain lysate includes highly abundant proteins containing internal GlyGly peptide sequences, such as β-actin, glyceraldehyde-3-phosphate dehydro­genase and α-tubulin, as well as histone H2A, which contains an internal Gly-Gly-Lys sequence. This indicates that internal Gly-Gly sequences are not recognized by the antibody. Additionally, peptides that contain Gly-Gly as the first two amino acids are not recognized (Supplementary Fig. 1). Together, these data indicate that the anti­ body recognizes Gly-Gly sequences that are present as an adduct on the ε-amine of lysine. We next investigated whether the anti–diglycyl-lysine antibody was able to immunoprecipitate peptides containing Gly-Gly–modified lysine. A flow chart for sample preparation, immunoprecipitation, and MS/MS analysis is shown in Figure 2a. We prepared a peptide containing an N-terminal Gly-Gly sequence (GGDRVYIHPFHL), and a peptide con­ taining a diglycyl adduct on lysine (Ac-SYSMEHFRWGK*PV-NH2; K* and Ac represent Gly-Gly–modified lysine and an acetyl group, respec­ tively). An equimolar mixture of the peptides was immunoprecipitated with the anti–diglycyl-lysine antibody, resulting in selective enrichment (≥50×) of the peptide containing the Gly-Gly–modified lysine (Fig. 2b). Additionally, this peptide was quantitatively immunoprecipitated with a nearly 100% yield (Supplementary Fig. 2). These experiments

demonstrated that the GX41 antibody is capable of enriching peptides containing diglycine-modified lysines and does not immunoprecipitate peptides containing a Gly-Gly sequence at their N termini. We next sought to assess the diversity of lysine ubiquitination in cultured cells. To distinguish diglycine remnants derived from ubiqui­ tin from those originating from less common ubiquitin-like proteins (such as ISG15 and NEDD8, which also leave a diglycine remnant on lysines after trypsinization8), we used HEK293 cells express­ ing His6-tagged ubiquitin. Ubiquitinated proteins were purified by immobilized metal-affinity chromatography, before proteolysis and anti–diglycyl-lysine immunopurification. Ubiquitin remnant– containing peptides were subjected to liquid chromatography (LC)MS/MS followed by database searching and spectral validation. To minimize alterations in ubiquitination levels after cell lysis, 5 mM chloroacetamide was included in lysis buffer to inhibit deubiquitinase and ubiquitin ligase activity9. To measure post-lysis ubiquitination, we spiked a lysate with excess glutathione S-transferase. This protein showed no detectable level of ubiquitination (Supplementary Fig. 3), suggesting that negligible ubiquitination occurred after cell lysis. MS/MS spectra of ubiquitin remnant–containing peptides exhibited normal y- and b-ion series, typically with a pair of ions separated by a mass of 242.14 Da, consistent with the masses of a lysine residue (128.09) and a Gly-Gly adduct (114.04 Da) on the ε-amine of lysine. Whereas most peptides contained a single diglycine-modified lysine (Fig. 2c), 17 peptides contained two diglycine-modified lysines. The majority (>92%) of ubiquitin remnant–containing peptides have a +3 or +4 charge (Supplementary Fig. 4), which reflects the additional charge from the N-terminal amine on the Gly-Gly adduct. Gly-Gly–modified lysines as the C-terminal residue of peptides were also detected (~2% of total) (Supplementary Fig. 5), and reflect use of the Gly-Gly–modified lysine as a substrate for trypsin, as described previously10. In total, we identified 374 diglycine-modified lysines on 236 ubi­ quitinated mammalian proteins. Analysis of the Swiss-Prot database suggests that 72% of these proteins were not previously known to be ubiquitinated. Similarly, 92% of the ubiquitination sites that we identified were not previously known. Among the identified proteins, 156 proteins have one ubiquitination site and 80 have two or more ubiquitination sites (Supplementary Table 1 and Supplementary Fig. 6). To validate the ubiquitination detected using the ubiquitin

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

869

a

c Relative intensity

Ub(RGG-)HN

Trypsin digestion

K Relative intensity

GG

Immunopurification

GG

60S ribosomal protein L7a GA|L|A|217K*|L|V|E/A/I/R

100

R

I

A

E

80 40

a2

y1 b

y3

y2-NH3 y

b3 2 2 20 b4 0 100 200 300

100 80

400

I

A

G

40 20 0

y1

b2

200

nLC-MS/MS analysis

Ac-SYSMEHFRWGK*PV-NH2

+Na

1,000

1,200

1,400

1,600

1,800

2,000

m/z After purification

100 80 60 40 20 0 1,000

500 600 m/z

y7

b7

700

K*

Y

800

900 1,000

Y

y2

F

y5

y

2+ 4

y3 y6

400

2+

y11

600 m/z

Ac-SYSMEHFRWGK*PV-NH2

y7

y6

2+

b4 y7

800

y8

1,000

1,200

Western blotting: ubiquitin β-14-3-3 IgG – + – +

250 150 100 75

Met-Ox

60 40 20 0

b6

60

K

Before purification GGDRVYIHPFHL

y6

2+

y9 b5

Splicing factor, arginine/serine-rich 1 38 DI|ED\V/F/Y/ K*/Y/G/A/I/R

a2 R

y5 2+

y8

2+

y7

2+

a7

LLnL

100 80

K*

L

60

IP:

b

V y4

d

Relative intensity

Figure 2  Profiling immunopurified ubiquitin remnant–containing peptides to identify ubiquitinated proteins. (a) Strategy to identify ubiquitinated proteins by immunoprecipitation of peptides containing diglycyl-lysine, followed by MS analysis. (b) Confirmation of antibody specificity using two peptides, GGDRVYIHPFHL and Ac-SYSMEHFRWGK*PVNH2. Equimolar amounts (0.3 nmol) of the two peptides were mixed and immunoprecipitated with immobilized anti–diglycyl-lysine monoclonal antibody. Matrix-assisted laser desorption ionization/time-of-flight (MALDITOF)-MS analysis for the starting material and the antibody-purified material suggests an enrichment factor of at least 50, based on the comparison of the MS signals of the two peptides before and after immunoprecipitation. (c) Representative annotated MS/MS spectra of two ubiquitin remnant–containing peptides obtained by immunoprecipitation from a HEK293 cell lysate. The sequence of the ubiquitinated peptide, including the diglycine-modified lysine (K *), is indicated and the fragment ions are labeled. The symbols \, / and | represent b-ions, y-ions, and both b-ions and y-ions, respectively. (d) Biochemical verification of the ubiquitination of six proteins. Proteins were immunoprecipitated using target-specific antibodies and the immunoprecipitate was detected by western blotting using an anti-ubiquitin antibody. IgG was used as a control for nonspecific immunoprecipitation. The proteasome inhibitor N-acetyl-LeuLeu-norleucinal (LLnL) was added to allow accumulation of the ubiquitinated protein.

Relative intensity

© 2010 Nature America, Inc. All rights reserved.

letters

IP: LLnL

Vimentin –

+

IgG –

+

250 150 100 75

50 IP: NAP1L1 IgG – + – + LLnL

IP: LLnL

250 150 100 75

250 150

PARP1 – +

IgG – +

100

50 IP:

HSP70

IgG

LLnL

– +

– +

IP: LLnL

β-Catenin IgG –

+

–

+

Met-Ox

1,200

1,400

m/z

1,600

1,800

2,000

250 150 100 75

250 150 100 75

remnant–profiling approach, we selected a subset of six proteins iden­ tified by MS and assessed whether they were ubiquitinated in cells. Lysates from HEK293 cells were immunoprecipitated with antibodies specific for the protein under investigation and immunoblotted using an anti-ubiquitin antibody (Fig. 2d). In these experiments, the HEK293 cells were not transfected with plasmids expressing His 6tagged ubiquitin. In each case, the immunopurified protein exhibits anti-ubiquitin immunoreactivity consistent with the endogenous ubiquitination of these proteins. The ubiquitination targets include disease-related proteins, such as 14-3-3ε, ataxin, β-catenin, BRCA1-associated protein and TTRAP (TRAF and TNF receptor-associated protein). The proteins identi­ fied by ubiquitin remnant profiling have roles in numerous biological processes, of which the largest number involve metabolism, cell cycle/ apoptosis and signal transduction (Fig. 3a). Additionally, we identified proteins that influence the trafficking, localization and structure of proteins, as well as regulate the immune system, consistent with previ­ ously reported roles for ubiquitination11–14. Ubiquitination of many ubiquitin-conjugating enzymes, ubiquitin ligases and 26S proteasome regulatory subunits also supports previous studies that reported the prevalence of ubiquitination of proteins involved in proteasome degradation pathways15,16. Some of the proteins found to be ubiqui­ tinated extend earlier findings regarding the role of ubiquitination in certain cellular processes. For example, although histone H2 ubiquitina­ tion has been described3, we found that histone H1, H3 and H4 isoforms are also ubiquitinated, as are subunits of histone acetyltransferases and histone deacetylases. These findings support the idea that ubiquitin

contributes to epigenetic gene regulation through multiple path­ ways. Many heat shock proteins, such as HSP70, HSP105, and HSC71, are ubiquitinated, linking ubiquitination to stress responses. Ubiquitination of several heterogeneous nuclear ribonucleoproteins reveals a role for ubiquitination in mRNA processing, metabolism, transport and splicing. Our studies also identify numerous transcrip­ tion factors, splicing factors, DNA repair proteins and kinases. This supports the well-characterized role for ubiquitination in regulating cellular signal transduction. The subcellular distribution of the detected proteins is likely to reflect, in part, the subcellular fractions that were used for MS/MS analysis. Subcellular localization analysis of the identified proteins indicates that essentially all the ubiquitinated proteins are cytosolic (Fig. 3a, right panel), which is consistent with the general observation that ubiquitination occurs primary in the cytosolic compartment of the cell12. Many of the identified proteins are localized to the nucleus, and several proteins are localized to the mitochondria, suggesting a role for ubiquitination in regulating aspects of mitochondrial function. We next wanted to gain insight into how lysine ubiquitination might be regulated at the level of primary and secondary structure. Interestingly, ubiquitin remnant–modified lysines have a slight tendency to be localized in regions enriched in small hydrophobic residues, such as alanine, leucine, isoleucine, glycine, proline and valine (Supplementary Fig. 7a). Examination of a six-amino-acid window adjacent to ubiquitinated lysines in the human proteome revealed that cysteine, histidine and lysine are found at a ~40% lower frequency than when they are adjacent to lysines in general

870

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

letters Figure 3  Bioinformatic analysis of ubiquitinated Cell cycle/apoptosis (13.0%) Cytoplasm (48.2%) proteins and ubiquitin-modified lysines. (a) Pie Small-molecule transport (2.3%) charts of biological processes and subcellular Plasma Immunity/defence (4.2%) localization of ubiquitinated proteins analyzed membrane (1.2%) Protein trafficking/localization (8.3%) Metabolism using the PANTHER and PENCE Proteome Other (15.8%) Mitochondria (49.6%) Other/unclassified (9.1%) Analyst databases, respectively. Proteins were (3.6%) Golgi (2.4%) designated ‘other’ if their localizations or Signal transduction (9.1%) Endoplamic functions were not annotated in the database. reticulum (2.4%) Nucleus (28.7%) Structure (4.4%) (b) Backbone amino acid sequence analysis of ubiquitinated peptides. A density map of All Lys All Lys 2.5 the ratios of the frequencies of each of the A 40 Ub Lys Ub Lys 0.6 C D 20 amino acids adjacent to the ubiquitinated E 2.0 0.5 F G lysines and adjacent to lysines in general was 30 H 0.4 I plotted using MATLAB. Several amino acids 1.5 K L 0.3 20 M are slightly enriched at certain positions, N P 1.0 0.2 Q such as leucine at +2, valine at −2, alanine 10 R S 0.1 at −5, glycine at +6, and tyrosine at −1 and T 0.5 V 0.0 0 W +1, determined by Rosner’s test with a 95% Y Helix Strand Coil Disordered K 0 confidence. (c) Ubiquitinated lysines (Ub –10 –5 0 5 10 Lys) possess an increased solvent accessible Relative solvent accessible area (%) area (SAA) relative to lysines in general. The distribution of SAA of both populations of lysines indicates an increase in SAA among ubiquitinated lysines. The two distributions are significantly different (Student’s t test, P < 0.001). The results were obtained from an analysis of 89 PDB structures (140 ubiquitinated lysines, 3,970 total lysines). (d) Distribution of secondary structures of all lysines and ubiquitinated lysines obtained from an analysis of 89 PDB structures. The disordered region was predicted by DisEMBL for all ubiquitinated proteins identified by our MS experiments. χ2 test: P < 0.001.

a

c

d

Relative intensity

(Supplementary Fig. 7a). Analysis involving Motif-x17 identified K*XL as a potential consensus ubiquitination site. This motif appears to be ~1.8 times more common among ubiquitinated lysines than lysines in general (Supplementary Fig. 7b). To compare all 20 amino acids for their propensity to be found at specific residues adjacent to ubiquitinated lysines, we prepared a density map that indicates the frequency of each amino acid at any of the ten proximal positions on either side of the ubiquitinated lysines, compared to the frequency of that amino acid next to lysines in general, as assessed by surveying the human proteome (Fig. 3b). This analysis shows that there is only a subtle enrichment for specific residues at some positions, such as leucine at the +2 position, valine at the −2 position, alanine at the −5 position, glycine at the +6 position, and tyrosine at the −1 and +1 positions. In contrast, an analysis of ubiquitinated proteins in yeast7 indicates an significant enrichment of aspartic acid, glutamic acid, histidine and proline at some positions (Supplementary Fig. 7c). To determine whether the sequence of the immunogen affected the specificity of the immunoprecipitated peptides, we generated a similar density map to present the frequency of each amino acid adjacent to the Gly-Gly–modified lysines in the immunogen. Although there are

1.0

Lys0

Lys8

DLSHIGDAVVISCA164K*DGVK Lys0:Lys8 = 1.08 ± 0.05

0.8 0.6 0.4 0.2 0 524

526

528

530

532

534

m/z

Relative intensity

© 2010 Nature America, Inc. All rights reserved.

1.0

Lys0

YYLAP254K*IEDEEGS Lys0:Lys8 = 1.47 ± 0.09

0.8

Lys8

0.6 0.4 0.2 0 814

816

818 m/z

820

822

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

00

80

–1

0

0

–8

60

0

–6

40

–4

20

0–

20

Fraction

Percentage

b

marked amino acid preferences adjacent to lysine in the immunogen (Supplementary Fig. 7d), these preferences are not seen in peptides pulled down by the anti–diglycyl-lysine antibody (Supplementary Fig. 7d). This suggests that the sequence of the immunogen used to generate our immunoaffinity reagent does not substantially bias the sequences of the immunoprecipitated peptides the antibody recovers. We found that ubiquitinated lysines have a slight tendency to appear on protein surfaces in preferred structural contexts. Structural infor­ mation is available in Protein Data Bank (PDB) for 89 of the proteins identified in this study. Measurements of the solvent-accessible area of lysines in these proteins indicate that ubiquitinated lysines tend to be exposed slightly more to solvent than other lysines (Fig. 3c, Student’s t test, P < 0.001). If lysines with >50% surface exposure are considered solvent exposed18, 60% of the ubiquitinated lysines are exposed, which is more than for lysines in general (45%). Overall, ubiquitinated lysines are ~6.5% more exposed than all the lysines. This is in agreement with a ubiquitination site survey for yeast 19. Interestingly, in some cases, the ubiquitinated lysine is fully buried (e.g., Supplementary Fig. 8). In these proteins, ubiquitination may be regulated by stimuli that induce the exposure of the lysine to the Figure 4  Colchicine differentially regulates the ubiquitination of two lysines in PCNA. HEK293 cells were grown in SILAC medium containing either light (Lys0) or heavy (Lys8) lysine, and transfected with a plasmid expressing His6-ubiquitin. Whereas Lys0-labeled cells were treated with 10 μM colchicine, Lys8-labeled cells were treated with vehicle for 16 h. Identical amounts of cells from each treatment were mixed and processed for MS analysis of ubiquitin remnant–containing peptides. The relative ratio of MS signals between Lys0- and Lys8-labeled peptides was used for relative quantification of the change in ubiquitination at K164 and K254. The observed ratio was normalized to the change in PCNA protein abundance in the two samples by measuring two unmodified PCNA peptides in the initial mixed cell lysate (Supplementary Fig. 11). The observation that the ion intensity of the novel ubiquitination site (K254) is about 20% of that of K164 suggests that its ubiquitination may be less common or more transient than K164. This may explain why it was not detected previously in mutagenesis studies33. All data are the averages of experiments repeated three times. Note that the peptide ubiquitinated at K254 is the C-terminal tryptic peptide of the protein so that the last amino acid is neither K nor R, and the charge state of this peptide is +2.

871

© 2010 Nature America, Inc. All rights reserved.

letters surface. Analysis of the local secondary structure surrounding all lysines and ubiquitinated lysines indicates that ubiquitinated lysines prefer helical structures compared to all lysines, although ubiquiti­ nation sites can also be found in other structural contexts (Fig. 3d). Additional crystal structures of proteins that are susceptible to ubiqui­ tination are needed to fully assess the solvent exposure and structural contexts of ubiquitinated lysines. Recently, a large number of lysine acetylation sites have been dis­ covered by proteomic approaches20–23. Although only 0.6% of lysines are predicted to be acetylated based on yeast studies24, >20% of the lysines that we found to be ubiquitinated are also sites of acetylation. For example, all the ubiquitinated lysines in H2B, H3.1 and H4 were reported to be acetylated. In the case of tubulin α-1A, four of the six ubiquitinated lysines were reported to be acetylated. The surprisingly high degree of concordance of lysine ubiquitination and acetylation sites suggests that acetylation of a specific lysine residue could serve as a means to prevent lysine ubiquitination25, or vice versa. A BLAST analysis of ubiquitination sites in human proteins against mouse, rat and yeast revealed that modified lysines are statistically more con­ served between these species than lysines in general (Supplementary Fig. 9). This suggests that the pathways leading to the ubiquitination of these sites may be evolutionarily conserved. In cases where a protein is ubiquitinated at more than one site, it is particularly challenging to monitor how the ubiquitination at the indi­ vidual sites is independently regulated. We therefore examined two proteins exhibiting multi-ubiquitination: tubulin α-1A and PCNA, a protein that regulates cell cycle progression26 and has been linked to tumorigenesis27. We labeled His6-ubiquitin-expressing HEK293T cells with either light (Lys0) or heavy (Lys8) lysine to quantify ubi­ quitination using the SILAC (stable isotope labeling by amino acids in cell culture) approach28 (Supplementary Fig. 10). We treated cells for 16 h with either vehicle (Lys8) or 10 μM colchicine (Lys0), an inhibitor of microtubule polymerization that affects progression through the cell cycle29, before mixing, lysing and processing cells as described in the Online Methods. We then analyzed the samples by nanoLC-MS to quantify ubiquitination at the PCNA ubiquitination sites that we had previously identified using MS/MS based on their retention time, mass-to-charge ratio (m/z) and charge states. We quantified relative ubiquitination at each modification site by normalization using pro­ tein abundance, as measured by the averaged light-to-heavy ratio of unmodified peptides detected from initial mixed cell lysate before any affinity purification30 (Supplementary Fig. 11). Interestingly, whereas the ubiquitination of K164 was unaffected by colchicine treatment, the ubiquitination of K254 was increased by 47% (Fig. 4). We also examined the multi-ubiquitination of tubulin α-1A. Treatment with colchicine resulted in a similar ~80% decrease in the ubiquitination of K326, K336 and K370. Surprisingly, treatment with vinblastine, which also disrupts microtubules, albeit through a distinct mechanism31,32, resulted in an opposite effect on ubiquiti­ nation, with a ~40% increase in ubiquitination at each of these sites (Supplementary Figs. 12 and 13). These results highlight how some ubiquitination sites may be ubi­quitinated in a dynamic manner, for example, in response to specific signals, whereas other ubiquitination sites may be ‘constitutive’. In the case of both PCNA and tubulin α-1A, ubiquitin remnant profiling provided insights into how distinct ubiquitination sites respond to different experimental treatments in a manner not readily available using currently available approaches. The ubiquitin remnant–profiling approach described here ­provides a simple and robust strategy to identify and quantify sites of ubi­ quitination in cells. It could be used to identify ubiquitination pat­ terns in cells and tissues with altered expression of ubiquitin ligases,

deubiquitinating enzymes, as well as to profile changes in ubiquitina­ tion elicited by various signaling molecules, drugs and disease states. Although the present data used cells expressing His6-tagged ubiquitin to reduce the likelihood of obtaining diglycine-modified peptides from ISG15- and NEDD8-modified proteins, ubiquitin-modified proteins could readily be enriched using immobilized ubiquitin-binding proteins, such as S5a, or ubiquitin antibodies5 in cells and tissues not amenable to transfection.

872

VOLUME 28  NUMBER 8  AUGUST 2010  nature biotechnology

Methods Methods and any associated references are available in the online version of the paper at http://www.nature.com/naturebiotechnology/. Accession code. MS/MS data and the identifications are deposited in the open access public repository PRIDE (http://www.ebi.ac.uk/ pride/) with the accession code of 12018. Note: Supplementary information is available on the Nature Biotechnology website. Acknowledgments We thank T. Neubert and G. Zhang (New York University) for useful suggestions, P. Zhou (Weill Cornell Medical College, WCMC) for the His6-ubiquitin plasmid, U. Hengst, A. Deglincerti, R. Almeida and B. Derakhshan for the assistance during initial cell culturing, S. Gross and Y. Ma (WCMC Mass Spectrometry Core Facility) for helpful discussion in MS/MS analysis, F. Campagne, L. Skrabanek, J. Sun (WCMC Institute for Computational Biomedicine) for instructions and assistance in bioinformatic analysis. The mass spectrometry work was performed at the WCMC Mass Spectrometry Core Facility using instrumentation supported by US National Institutes of Health (NIH) RR19355 and RR22615. This work was supported by grants from Weill Cornell, NIH (MH086128) (S.R.J.), and a pharmacology cancer training grant from the National Cancer Institute (T32CA062948) (G.X. and J.S.P.). AUTHOR CONTRIBUTIONS S.R.J. and G.X. conceived and designed the study. G.X. and J.S.P. conducted the experiments, and G.X. and S.R.J. analyzed the data. S.R.J. and G.X. wrote the manuscript. COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. Published online at http://www.nature.com/naturebiotechnology/. Reprints and permissions information is available online at http://npg.nature.com/ reprintsandpermissions/. 1. Hershko, A. & Ciechanover, A. The ubiquitin system. Annu. Rev. Biochem. 67, 425–479 (1998). 2. Xu, P. & Peng, J. Dissecting the ubiquitin pathway by mass spectrometry. Biochim. Biophys. Acta 1764, 1940–1947 (2006). 3. Ericsson, C., Goldknopf, I.L. & Daneholt, B. Inhibition of transcription does not affect the total amount of ubiquitinated histone 2A in chromatin. Exp. Cell Res. 167, 127–134 (1986). 4. Galluzzi, L., Paiardini, M., Lecomte, M.C. & Magnani, M. Identification of the main ubiquitination site in human erythroid alpha-spectrin. FEBS Lett. 489, 254–258 (2001). 5. Tomlinson, E., Palaniyappan, N., Tooth, D. & Layfield, R. Methods for the purification of ubiquitinated proteins. Proteomics 7, 1016–1022 (2007). 6. Beers, E.P. & Callis, J. Utility of polyhistidine-tagged ubiquitin in the purification of ubiquitin-protein conjugates and as an affinity ligand for the purification of ubiquitin-specific hydrolases. J. Biol. Chem. 268, 21645–21649 (1993). 7. Peng, J. et al. A proteomics approach to understanding protein ubiquitination. Nat. Biotechnol. 21, 921–926 (2003). 8. Srikumar, T., Jeram, S.M., Lam, H. & Raught, B. A ubiquitin and ubiquitin-like protein spectral library. Proteomics 10, 337–342 (2010). 9. Hershko, A., Heller, H., Elias, S. & Ciechanover, A. Components of ubiquitin-protein ligase system. Resolution, affinity purification, and role in protein breakdown. J. Biol. Chem. 258, 8206–8214 (1983). 10. Denis, N.J., Vasilescu, J., Lambert, J.P., Smith, J.C. & Figeys, D. Tryptic digestion of ubiquitin standards reveals an improved strategy for identifying ubiquitinated proteins by mass spectrometry. Proteomics 7, 868–874 (2007). 11. Rechsteiner, M. Ubiquitin-mediated pathways for intracellular proteolysis. Annu. Rev. Cell Biol. 3, 1–30 (1987). 12. Bonifacino, J.S. & Weissman, A.M. Ubiquitin and the control of protein fate in the secretory and endocytic pathways. Annu. Rev. Cell Dev. Biol. 14, 19–57 (1998).

letters 24. Basu, A. et al. Proteome-wide prediction of acetylation substrates. Proc. Natl. Acad. Sci. USA 106, 13785–13790 (2009). 25. Yang, X.J. & Seto, E. Lysine acetylation: codified crosstalk with other posttranslational modifications. Mol. Cell 31, 449–461 (2008). 26. Prosperi, E. Multiple roles of the proliferating cell nuclear antigen: DNA replication, repair and cell cycle control. Prog. Cell Cycle Res. 3, 193–210 (1997). 27. Mayer, A. et al. The prognostic significance of proliferating cell nuclear antigen, epidermal growth factor receptor, and mdr gene expression in colorectal cancer. Cancer 71, 2454–2460 (1993). 28. Ong, S.E. et al. Stable isotope labeling by amino acids in cell culture, SILAC, as a simple and accurate approach to expression proteomics. Mol. Cell. Proteomics 1, 376–386 (2002). 29. Jordan, M.A. Mechanism of action of antitumor drugs that interact with microtubules and tubulin. Curr. Med. Chem. Anticancer Agents 2, 1–17 (2002). 30. Wisniewski, J.R. et al. Constitutive and dynamic phosphorylation and acetylation sites on NUCKS, a hypermodified nuclear protein, studied by quantitative proteomics. Proteins 73, 710–718 (2008). 31. Gigant, B. et al. Structural basis for the regulation of tubulin by vinblastine. Nature 435, 519–522 (2005). 32. Ravelli, R.B. et al. Insight into tubulin regulation from a complex with colchicine and a stathmin-like domain. Nature 428, 198–202 (2004). 33. Unk, I. et al. Human SHPRH is a ubiquitin ligase for Mms2-Ubc13-dependent polyubiquitylation of proliferating cell nuclear antigen. Proc. Natl. Acad. Sci. USA 103, 18107–18112 (2006).

nature biotechnology  VOLUME 28  NUMBER 8  AUGUST 2010

873

© 2010 Nature America, Inc. All rights reserved.

13. Kirkpatrick, D.S., Denison, C. & Gygi, S.P. Weighing in on ubiquitin: the expanding role of mass-spectrometry-based proteomics. Nat. Cell Biol. 7, 750–757 (2005). 14. Sun, L. & Chen, Z.J. The novel functions of ubiquitination in signaling. Curr. Opin. Cell Biol. 16, 119–126 (2004). 15. Etlinger, J.D., Li, S.X., Guo, G.G. & Li, N. Phosphorylation and ubiquitination of the 26S proteasome complex. Enzyme Protein 47, 325–329 (1993). 16. Peters, J.M. Subunits and substrates of the anaphase-promoting complex. Exp. Cell Res. 248, 339–349 (1999). 17. Schwartz, D. & Gygi, S.P. An iterative statistical approach to the identification of protein phosphorylation motifs from large-scale data sets. Nat. Biotechnol. 23, 1391–1398 (2005). 18. Ahmad, S. & Gromiha, M.M. NETASA: neural network based prediction of solvent accessibility. Bioinformatics 18, 819–824 (2002). 19. Catic, A., Collins, C., Church, G.M. & Ploegh, H.L. Preferred in vivo ubiquitination sites. Bioinformatics 20, 3302–3307 (2004). 20. Choudhary, C. et al. Lysine acetylation targets protein complexes and co-regulates major cellular functions. Science 325, 834–840 (2009). 21. Gnad, F. et al. PHOSIDA (phosphorylation site database): management, structural and evolutionary investigation, and prediction of phosphosites. Genome Biol. 8, R250 (2007). 22. Kim, S.C. et al. Substrate and functional diversity of lysine acetylation revealed by a proteomics survey. Mol. Cell 23, 607–618 (2006). 23. Zhao, S. et al. Regulation of cellular metabolism by protein lysine acetylation. Science 327, 1000–1004 (2010).

© 2010 Nature America, Inc. All rights reserved.

ONLINE METHODS

Antigen synthesis and antibody production. Lysine-rich histone from calf thymus (type III-S, Sigma) was dissolved in 100 mM NaHCO3 buffer (10 ml) at pH 10. 500 μl t-butyloxycarbonyl-Gly-Gly-N-­hydroxysuccinimide (50 mM, Boc-Gly-Gly-NHS, ref. 34) in DMSO was added to histone ­solution and the reaction was carried out at 25° C for 1 h by constant shaking on a plate rotator. This step was repeated three additional times and sample B was obtained. For deprotection of the Boc group, neat trifluoroacetic acid (6 ml, TFA, Sigma) was added and the solution was shaken for 2 h at 25° C. The reaction was stopped by neutralizing with 10 M NaOH dropwise on ice (sample C). Sample A, B and C were dialyzed four times against 20 mM ­acetic acid followed by lyophilization. The degree of the reaction was assessed by anti-biotin (Sigma) western blot analysis after samples A, B and C were reacted with 5 mM biotin-NHS (Sigma) for 10 min. The same protocol was used to prepare Boc-Gly-Gly– and Gly-Gly–modified β-lactoglobulin, hen egg white lysozyme, rat brain lysate and peptides (DRVYIHPFHL and Ac-SYSMEHFRWGKPV-NH2) for antibody evaluation. The antigen was injected into mice for antibody production, and hybri­ doma clones were made by Promab. Cells of monoclonal clones were grown in MegaCell Dulbecco’s Modified Eagle’s Medium (MegaCell DMEM, pH 7.2, Sigma) supplemented with 10% FBS (FBS), 50 μg/ml of kanamycin, 1 mM glutamine, and cells were split and cell culture supernatant was ­ collected every week. Hybridoma clone GX41 was obtained after screening a panel of hybrido­ mas to assess their utility in detecting diglycine-modified lysines. Antibodies from each hybridoma clone were first evaluated by western blot analysis using Gly-Gly–modified β-lactoglobulin, lysozyme and rat brain lysate. Clones were selected based on the absence of reactivity with unmodified protein and lysates, absence of reactivity with proteins and lysate modified with Boc-Gly-Gly, and reactivity with Gly-Gly–modified proteins and lysate. The top five clones that were further characterized were based on their ability to recognize the largest number of bands in the Gly-Gly–modified rat brain lysate. Antibodies from these clones were purified and used for immunoprecipitation of ubiquitin remnant–containing peptides from His6-ubiquitin–expressing HEK 293 cells, and tandem MS identification of tryptic ubiquitinated peptides to assess the degeneracy of antibodies. Only clone GX41 pulled down peptides that con­ tained each of the 20 amino acids N-terminal to the modified lysine and each of the 20 amino acids C-terminal to the modified lysine, suggesting that the antibody can bind peptides which contain the diglycyl-lysine in a wide range of sequence contexts, which was supported by subsequent characterization of the amino acid context of the diglycyl-lysine obtained from a larger data set of ubiquitin remnant peptides (Fig. 3b and Supplementary Fig. 7a). The GX41 anti–diglycyl-lysine monoclonal antibody was found to be IgG1κ isotype. This antibody was used for all the experiments in this study. Antibody purification and coupling. Gly-Gly–modified β-lactoglobulin was coupled to Affi-Gel 10 resin (Bio-Rad) in a concentration of 5 mg protein/ml resin in a pH 8 HEPES buffer overnight in 4 °C. The resin was quenched by 1 M Tris-HCl (pH 8), washed with three volumes of 10 mM citric acid (pH 3) and PBS. Cell culture supernatant (50 ml) from monoclonal cell lines was loaded six times into an 8-cm column with 1 ml Affi-Gel 10 resin coupled with Gly-Gly–modified β-lactoglobulin in 4 °C using a peristaltic pump. The resin was washed three times with 6 ml of 2× PBS and three times with 6 ml of PBS. The antibody was eluted four times with 0.5 ml 10 mM citric acid (pH 3) and immediately neutralized by 50 μl of 1 M HEPES (pH 8). The pH was adjusted to 8.5 and the antibody was concentrated by a 15 ml filter device (30 kDa molecular weight cutoff, Millipore). The antibody concentration was measured by Bradford protein assay (Bio-Rad). Typically, 0.1~0.2 mg of antibody was coupled to 20 μl Affi-Gel 10 resin according to the method described above. The antibody resin was stored in PBS buffer with 0.1% sodium azide at 4 °C. Cell culture and sample preparation. Human embryo kidney (HEK) 293 cells were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM, Invitrogen) supplemented with 4.5 g/l glucose, 10% FBS, 100 units/ml penicillin G and 100 μg/ml streptomycin. When the confluence reached ~50%, cells were transfected with 10 μg of a His6-tagged ubiquitin plasmid per 10-cm Petri dish using the calcium phosphate transfection method. Cells were used 1 d

nature biotechnology

after transfection and treated with vehicle or proteasome inhibitor 25 μM LLnL (Calbiochem) in DMSO and incubated for 16 h before harvest. The His6-ubiquitin is expressed at a fraction of the level of endogenous ubiquitin (Supplementary Fig. 14) suggesting that it is unlikely to perturb endogenous ubiquitin pathways. The expression of tagged ubiquitin has been widely used in proteomics studies of protein ubiquitination7,35,36. Twenty 10-cm Petri dishes were cultured and cells were washed twice with ice-cold PBS. The cells were detached, collected and centrifuged at 1,000g for 5 min at 4 °C. To increase coverage of ubiquitinated proteins, crude lysates, as well as subcellular fractions, including nuclear, membrane and cytosolic frac­ tions, were prepared for analysis. For the crude lysate, the cell pellet was lysed and His6-tagged proteins were purified by Ni-NTA resin (Qiagen) in native and denaturing conditions according to the manufacturer’s protocol. The lysis buffer contained 5 mM chloroacetamide to alkylate cysteines and to inhibit ubiquitin ligases and deubiquitinases9. The membrane fraction was obtained by centrifuging at 100,000g for 60 min after removing the nuclear pellet in the presence of 250 mM sucrose. The pellet from nuclear and membrane fraction was dissolved in 8 M urea with 1% triton X-100 and 0.1% SDS and the proteins were purified by Ni-NTA resin in the presence of 10 mM β-mercaptoethanol. After immobilized metal affinity purification, ubiquitinated proteins are sig­ nificantly enriched (Supplementary Fig. 15). All the samples after Ni-NTA purification were concentrated on an Amicon YM10 filter device (Millipore) and separated by SDS-PAGE. Gel pieces were treated with 10 mM dithiothreitol at 50 °C for 30 min, followed by 55 mM chloroacetamide at 25 °C for 45 min, using methods described previously 37, except that chloroacetamide was used in place of iodoacetamide. In-gel diges­ tion and peptide extraction were performed as described37. The lyophilized peptide mixture was dissolved in 300 μl of buffer contain­ ing 150 mM NaCl, 50 mM Tris-HCl (pH7.4) and 2 mM EDTA. The sample was boiled in a water bath for 10 min to deactivate residual trypsin activity. The peptide mixture was incubated with 20 μl antibody resin for 4 h in 4 °C, loaded on a micro-spin column (Pierce) six times, washed three times with 2× PBS and three times with PBS, and eluted six times with 20 μl 10 mM citric acid (pH 3). The eluted peptide mixture was concentrated to 20 μl for tandem MS analysis. For the MALDI-TOF-MS experiment, a sample containing ~0.3 nmol of each peptide, GGDRVYIHPFHL and Ac-SYSMEHFRWGK*PV-NH2, was pre­ pared and subjected to immunoprecipitation using the agarose-immobilized antibody described above. For SILAC quantification, five 10-cm dishes of HEK293T cells were grown in the media containing either light lysine (Lys0: 12C614N2-Lys) or heavy lysine (Lys8: 13C615N2-Lys) (Cambridge Isotope Labs) using previously described procedures for SILAC experiments38. The cells were transfected with His6ubiquitin plasmid as described above, and treated with vehicle or drugs (10 μM colchicine or 1 μM vinblastine, Sigma) in the presence of LLnL (PCNA: 25 μM for 16 h; tubulin α-1A: 50 μM for 30 min). The cells were mixed and purified under denaturing condition as described above without fractiona­ tion. To normalize the ubiquitinated peptides by unmodified peptides in the cell lysate, a small amount of initial mixed cell lysate was digested by trypsin followed by tandem MS analysis30. Mass spectrometric analysis. For MALDI-TOF-MS, samples were desalted by Millipore C18 ZipTip according to manufacturer’s protocol and eluted in a 2 μl solvent with 50% acetonitrile and 0.1% TFA in the presence of 10 mg/ml α-cyano-4-hydroxycinnamic acid (Sigma). The masses of the samples were analyzed in the reflector mode by Voyager-DE PRO MALDI-TOF-MS (Applied Biosystems). The samples purified from cell lysate were analyzed by nanoLC Q-TOF MS/MS (Agilent) to obtain peptide sequence information using settings as described previously39. Briefly, 8 μl of peptide mixtures were loaded onto an enrichment column with 97% solvent A and 3% solvent B with a flow rate of 3 μl/min. Solvent A consists of 0.1% formic acid (Fluka) and solvent B of 90% acetonitrile (Fisher) and 0.1% formic acid. Peptides were eluted with a gradient from 3% to 40% solvent B in 20 min, followed by a steep gradient to 90% solvent B in 5 min at a flow rate of 0.3 μl/min. Mass spectra were acquired in the positive-ion mode with automated data-dependent MS/MS on the five most intense ions from precursor MS scans and every selected

doi:10.1038/nbt.1654

© 2010 Nature America, Inc. All rights reserved.

precursor peak was analyzed twice within 3 min. In some runs, a list of previ­ ous identified peptides was excluded for MS/MS fragmentation. Database search of MS/MS spectra for peptide and protein ­identification. Analysis of MS/MS spectra for peptide and protein identification was ­p erformed by protein database searching with Spectrum Mill software (Rev A.03.02, Agilent) against the Swiss-Prot database (v57.2, May 5, 2009) containing a concatenated reverse database with the same entries and the same length for each protein, as described40. The use of a decoy database to evaluate the false-positive rate for modified peptides may underestimate the false identifications as protein modifications can greatly expand the search space. Raw spectra were first extracted to MS/MS spectra that could be assigned to at least four y- or b-series ions. Scans with the same precursor within a mass window of ±0.4 m/z were merged within a time frame of ±15 s, charges up to a maximum of 7 were assigned to the precursor ion and the 12C peak was determined by the Data Extractor. Key search parameters were a minimum matched peak intensity of 50%, a precursor mass tolerance of ±20 p.p.m., and a product mass tolerance of ±40 p.p.m. A fixed modifica­ tion was carbamidomethylation (same modification as chloroacetamide) for cysteines and variable modifications were Gly-Gly modification for lysines and oxidation for methionines. It should be noted that there are potentially a large number of naturally occurring sequence variants in mammals, but very limited data in the databases on these sequences. These variants may be missed or misidentified if the sequence variation lies in the same pep­ tide that contains the diglycine modified–lysine. The maximal number of diglycine modifications was set as two. Trypsin was selected as enzyme for sample digestion and four missed cleavages were allowed during the database search. The threshold used for peptide identification was a Spectrum Mill score of ≥ 9, an SPI% (the percentage of the scored peak intensity) of ≥ 50% and the difference between forward and reverse scores of ≥2. Under these criteria, the false-positive rate is <1%. In the peptide list (Supplementary Table 1), only the highest scoring member of each peptide group is shown and only peptides with a charge state of 2, 3, 4 and 5 are reported. Finally, all MS/MS spectra were manually validated and the spectra with low quality of fragmentations were discarded. For SILAC experiments that quantified the relative ubiquitination at spe­ cific sites in tubulin α-1A and PCNA, quantification was performed using peak intensities from the MS data on peptides that were previously identified by MS/MS (included within Supplementary Fig. 6 and the manually anno­ tated MS/MS spectra are shown in Supplementary Fig. 16). Peptides were identified based on LC-retention time, m/z and charge state as described41,42. Quantification of relative ubiquitination at each modification site was obtained by normalization to protein abundance, as measured by the averaged lightto-heavy ratio of unmodified peptides detected from initial mixed cell lysate before any affinity purification as described previously30. The two modifications isobaric to the Gly-Gly seen on the Delta Mass PTM website using a 0.02 Da mass window margin of error, which is appropriate for the Q-TOF instrument used in this study, are asparagyl- and hydroxy­ prolyl-modifications. The ornithyl modification is 0.04 Da larger than the diglycine modification. These rare modifications could be mistaken for the diglycine remnant on lysine. Alternatively, if these residues are adjacent to a lysine, the MS/MS ion series could be misinterpreted to contain a GlyGly–modified lysine. However, none of the diglycine-modified lysine resi­ dues reported in this study were found in a peptide context that could have matched an alternate tryptic peptide with a lysine adjacent to an asparagine or hydroxyproline. This paper describes the identification of several hundred diglycine­modified peptides. Because of the large number of diglycine-modified peptides identified by Spectrum Mill, manual verification of every spectra proved to be impractical. Other than the spectra obtained for the tubulin -α-1A and PCNA experiments (Supplementary Fig. 16), we have not manually verified all the diglycine-modified lysine residues listed in Supplementary Table 1. To provide an alternate estimate of the accuracy of diglycyl-lysine modification identification in peptides, we used a combination of decoy database searching (described above), as well as the use of a consensussearching approach to report peptides that were identified by Spectrum Mill and at least one other search program. To identify Spectrum Mill-­identified

doi:10.1038/nbt.1654

peptides that were also matched by other search algorithms, the pkl files for ubiquitinated peptides identified by Spectrum Mill were extracted, converted and merged to a single file in Mascot Generic Format (mgf). The file was searched with other online search programs: Mascot, X!Tandem and Phenyx, against the Swiss-Prot human database. The Mascot score and expectation value, X!Tandem log (e), and Phenyx p-value are reported in Supplementary Table 1. The cutoff values for positive identifications are: 29 for Mascot score or 0.05 for Mascot expectation value; −1.0 for X!Tandem log (e), and 10−4 for Phenyx p-value. A Venn diagram depicting the number of common and unique ubiquitinated peptides obtained from the three search programs is presented in Supplementary Figure 17. To increase the con­ fidence of our identification, only the peptides which were found by at least two search programs were considered as likely ubiquitinated peptides and considered in bioinformatic studies. It should be noted that peptide identification from MS/MS spectra using computational approaches has the potential of producing false positives. Any lysine residue predicted to be modified should be verified experimentally to conclusively demonstrate that this site is indeed ubiquitinated. Biochemical validation of ubiquitination. Some identified ubiquitinated proteins were verified by the conventional immunoprecipitation and immuno­ blot approach. The target proteins were precipitated using specific antibodies: 14-3-3 (Santa Cruz), β-catenin (Millipore), HSP70 (Labvision), NAP1L1 (Abgent), PCNA (Labvision), vimentin (Developmental Studies Hybridoma Bank). Western blot analysis was carried out against anti-ubiquitin antibody (Santa Cruz and Assay Designs). Full blots for Figure 2d are provided in Supplementary Figure 18. Bioinformatic analysis. Protein biological processes were analyzed and clus­ tered by PANTHER43 after converting the Swiss-Prot accession numbers of identified ubiquitinated proteins into RefSeq protein accession numbers by the Database for Annotation, Visualization and Integrated Discovery (DAVID)44 online gene ID conversion function. In total, 385 biological processes were found in the database and the category was further grouped into eight classes. Protein subcellular localization of the ubiquitinated proteins was extracted from the database provides by PENCE Proteome Analyst 45. Note that some proteins were predicted to have multiple subcellular localizations whereas some proteins have no subcellular information available in the database. The density map for the diglycine-modified lysines was calculated as fol­ lows: a subset of protein sequence, 10 amino acids on either side of modified lysines, was extracted from the whole protein sequence. The frequency of each of the 20 individual amino acids at each position from −10 to +10 was calculated for diglycine-modified lysines and this value was normalized to the frequency of the same amino acid at the same position using all lysines in the human Swiss-Prot database to obtain a relative ratio. If the ratio in one position (say −1) for a specific amino acid (say Pro) is >1, there is a commensurately higher likelihood for Pro at the −1 position to be adjacent to a ubiquitinated lysine. The highest relative ratio detected was 2.3 and the range of the color map was set from 0 to 2.5. The density map was prepared by MATLAB. The enriched amino acids were obtained by determining the outliers with a 95% confidence using the Rosner’s test46. To access the structural features of ubiquitinated lysine residues for human proteins, we searched crystal structures for all the ubiquitinated proteins in protein database bank (PDB). In total, 89 PDB structures (Supplementary Table 2) contained lysines that we found are susceptible to ubiquitination (140 modified lysines and 3970 total lysines). In cases when multiple PDB structures for a ubiquitinated protein were reported, the structure with best quality was used. The secondary structure types for lysines were determined using the program DSSP47. H and G were considered to be helix, E and B to be strand, S, T and others for coil. The fraction of each secondary structure type of modified lysines was compared to that of all the lysine residues in 89 PDB structures. The disordered region was predicted by DisEMBL48 for all identified ubiquitinated proteins and the information for modified lysines and all lysines was extracted. The relative solvent-accessible area for the modified and all lysines in 89 crystal structures was calculated using NACCESS49 with a probe of 1.4 Å, which corresponds to the size of a water molecule.

nature biotechnology

42. Mortensen, P. et al. MSQuant, an open source platform for mass spectrometry-based quantitative proteomics. J. Proteome Res. 9, 393–403 (2010). 43. Thomas, P.D. et al. PANTHER: a library of protein families and subfamilies indexed by function. Genome Res. 13, 2129–2141 (2003). 44. Dennis, G. Jr. et al. DAVID: database for annotation, visualization, and integrated discovery. Genome Biol. 4, 3 (2003). 45. Lu, Z. et al. Predicting subcellular localization of proteins using machine-learned classifiers. Bioinformatics 20, 547–556 (2004). 46. Rosner, J. Test of auditory analysis skills (TAAS) in helping children overcome learning difficulties: a step-by-step guide for parents and teachers (Academic Therapy, New York, 1979). 47. Kabsch, W. & Sander, C. Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 22, 2577–2637 (1983). 48. Linding, R. et al. Protein disorder prediction: implications for structural proteomics. Structure 11, 1453–1459 (2003). 49. Hubbard, S.J., Campbell, S.F. & Thornton, J.M. Molecular recognition. Conformational analysis of limited proteolytic sites and serine proteinase protein inhibitors. J. Mol. Biol. 220, 507–530 (1991).

© 2010 Nature America, Inc. All rights reserved.

34. Derrien, D. et al. Muramyl dipeptide bound to poly-l-lysine substituted with mannose and gluconoyl residues as macrophage activators. Glycoconj. J. 6, 241–255 (1989). 35. Kirkpatrick, D.S., Weldon, S.F., Tsaprailis, G., Liebler, D.C. & Gandolfi, A.J. Proteomic identification of ubiquitinated proteins from human cells expressing His-tagged ubiquitin. Proteomics 5, 2104–2111 (2005). 36. Xu, P. et al. Quantitative proteomics reveals the function of unconventional ubiquitin chains in proteasomal degradation. Cell 137, 133–145 (2009). 37. Shevchenko, A., Wilm, M., Vorm, O. & Mann, M. Mass spectrometric sequencing of proteins silver-stained polyacrylamide gels. Anal. Chem. 68, 850–858 (1996). 38. de Godoy, L.M. et al. Status of complete proteome analysis by mass spectrometry: SILAC labeled yeast as a model system. Genome Biol. 7, R50 (2006). 39. Xu, G., Shin, S.B. & Jaffrey, S.R. Global profiling of protease cleavage sites by chemoselective labeling of protein N-termini. Proc. Natl. Acad. Sci. USA 106, 19310–19315 (2009). 40. Elias, J.E. & Gygi, S.P. Target-decoy search strategy for increased confidence in large-scale protein identifications by mass spectrometry. Nat. Methods 4, 207–214 (2007). 41. Silva, J.C. et al. Quantitative proteomic analysis by accurate mass retention time pairs. Anal. Chem. 77, 2187–2200 (2005).

nature biotechnology

doi:10.1038/nbt.1654

careers and recruitment

Second quarter biotech job picture Michael Francisco

I © 2010 Nature America, Inc. All rights reserved.

n the second quarter of 2010, biotech and pharmaceutical postings on the three representative job databases tracked by Nature Biotechnology (Tables 1 and 2) largely stayed the same from the previous quarter (Nat. Biotechnol. 28, 527, 2010). Noteworthy increases in job openings were seen from instrument systems and consumables manufacturers Life  able 1 Who’s hiring? Advertised openings at the 25 largest biotech T companies Number of advertised openingsb Companya

Number of employees

Monster

Biospace

Naturejobs

Monsanto

21,700

0

0

Amgen

16,800

29

29

1

Genentech

11,186

6

26

100

Genzyme

11,000

63

0

0

9,700

71

89

0

Life Technologies

31

PerkinElmer

7,900

52

0

0

Bio-Rad Laboratories

6,600

12

17

0

Biomerieux

6,140

9

0

0

Millipore

5,900

15

11

0

IDEXX Laboratories

4,700

14

0

0

Biogen Idec

4,700

38

104

1

Technologies (Carlsbad, CA, USA), PerkinElmer (Waltham, MA, USA), Bio-Rad Laboratories (Hercules, CA, USA) and Illumina (San Diego). Table 3 shows selected downsizings within the life science industry. Nature Biotechnology will continue to follow hiring and firing trends throughout 2010. Table 2 Advertised job openings at the ten largest pharma companies Number of advertised openingsb Companya Johnson & Johnson Bayer GlaxoSmithKline Sanofi-Aventis Novartis Pfizer Roche Abbott Laboratories AstraZeneca Merck & Co.

Number of employees 119,200 106,200 103,483 99,495 98,200 86,600 78,604 68,697 67,400 59,800

Monster 522 78 5 12 144 2 35 67 71 0

Biospace 8 27 1 1 94 81 31 42 7 10

Naturejobs 19 3 2 4 20 88 20 1 4 0

aData obtained from MedAdNews. bAs searched on Monster.com, Biospace.com and Naturejobs.com, 21 July 2010. Jobs may overlap.

Table 3 Selected biotech and pharma downsizings Number of employees cut Details

Gilead Sciences

3,441

0

26

0

Company

WuXi PharmaTech

3,172

0

0

0

3,041

0

0

1

Albany Molecular Research

80

Qiagen

Restructured its US operations, including reducing head count by about 10% and suspending operations at one of its research laboratories in Rensselaer, New York.

Cephalon

2,780

0

0

0

Cell Therapeutics

36

Biocon

2,772

0

0

0

Reduced head count by 29% to 88 to conserve cash, with the cuts coming mostly from sales and marketing.

Celgene

2,441

13

10

0

GTC Biotherapeutics

50

Will restructure and reduce head count by 46% to 59 to save cash.

Biotest

2,108

0

11

0

40

Actelion

2,054

4

1

0

Helicos BioSciences

Reduced head count by 50% to 40 and plans to refocus its business on molecular diagnostics development.

InterMune

60

Amylin Pharmaceuticals

1,800

9

3

0

Elan

1,687

7

2

0

Reduced head count by 40% to 85, with the cuts coming predominantly in the commercial and discovery research areas.

Illumina

1,536

2

27

9

Albany Molecular Research

1,357

0

0

0

Myriad Pharmaceuticals

Vertex Pharmaceuticals

1,322

40

58

2

Novartis

CK Life Sciences

1,315

0

0

0

Lonza Group

193 21

Will restructure Novartis Pharmaceuticals and reduce head count at the US unit. Thirty-five percent of the cuts will be achieved by not filling vacant positions. The cuts will primarily come from “headquarter-based functions,” with minimal impact on the commercial sales organization.

6,000

Announced plans to restructure its global manufacturing plant network and reduce manufacturing head count by 18% to 27,000 over the next five years. Plans to close eight sites in Puerto Rico, Ireland and the US and reduce operations at another six sites.

defined in Nature Biotechnology’s survey of public companies (27, 710–721, 2009). bAs searched on Monster.com, Biospace.com and Naturejobs.com, 21 July 2010. Jobs may overlap.

Sanofi-Aventis Takeda Pharmaceutical

Restructured and reduced head count by 13% to about 140 to focus on its cancer pipeline.

383

aAs

Pfizer

Reducing head count by 6% to 2,899 at its R&D and production site in Visp, Switzerland, to save cash.

400 ~1,900

Cuts will primarily come from US sales force, which previously had 5,700 employees. Will reduce head count by about 10% to reduce costs in its fiscal year 2010 ending March 31, 2011.

Source: BioCentury.

Michael Francisco is Senior Editor, Nature Biotechnology

nature biotechnology volume 28 number 8 august 2010

875

© 2010 Nature America, Inc. All rights reserved.

people

Biogen Idec (Cambridge, MA, USA) has named George Scangos (left) as its new CEO as well as a member of the board of directors, replacing the recently retired Jim Mullen. Scangos joins Biogen Idec from Exelixis, where he has served as president and CEO since 1996. Previously, he spent 10 years at Bayer, leaving as president of Bayer Biotechnology. “George’s appointment is the culmination of the board’s comprehensive selection process to identify the best leader to take Biogen Idec to the next level,” says chairman William D. Young. “Science is at the heart of our business, and George has an exceptional scientific background, as well as significant operational expertise and a strong leadership track record.”

Nile Therapeutics (San Francisco) has appointed Richard B. Brewer as executive chairman. Brewer brings over 35 years of operational, financial and business development expertise to Nile. He currently serves as chairman of Arca Biopharma and was previously CEO and president of Scios, COO of Heartport and senior vice president of US marketing at Genentech. BioVex (Woburn, MA, USA) has appointed Kapil Dhingra to its board of directors. Dhingra spent nearly ten years at Hoffmann-La Roche, culminating in his appointment as vice president and head of oncology clinical development. Myriad Genetics (Salt Lake City, UT, USA) has announced the appointment of Gary A. King to the newly created position of executive vice president of international operations. King has over 25 years of life sciences experience, most recently as CEO of AverDx. Prior to AverDx, he was vice president, international operations at Biosite. Dean Mitchell (left) has been named president and CEO of Lux Biosciences (Jersey City, NJ, USA). Mitchell was formerly president and CEO of Alpharma and Guilford Ph ar m a c e ut i c a l s .

876

He is also a nonexecutive board member of ISTA Pharmaceutics, Intrexon and Talecris Biotherapeutics. Diagnostic kit developer Ingen Biosciences (Chilly-Mazarin, France) has appointed Karine Mignon-Godefroy as director of research and development. She joins Ingen from the blood virus division of Bio-Rad, where she was director of international projects. Before Bio-Rad, she held the post of R&D manager at BMD. Frank Morich, CEO of NOXXON Pharma (Berlin) has announced his intention to leave the company effective August 15 to take up the position of executive vice president, international operations of Takeda Pharmaceutical Company. Iain Buchanan, a director of NOXXON, will assume the role of interim CEO and will support the board during its search for a permanent replacement. Buchanan has over 30 years of experience in the pharma and biotech industry, most recently as CEO of Novexel. Marine biotechnology company Aquapharm Biodiscovery (Oban, UK) has named Tim Morley as CSO. Morley has over 20 years exper­ ience in the pharmaceutical industry, including previous positions as research and strategic project director at Quotient Biodiagnostics, vice president preclinical sciences at Ardana Bioscience and senior director molecular and cellular pharmacology at Vernalis.

Exelixis (S. San Francisco, CA, USA) has announced the appointment of Michael Morrissey as president and CEO, succeeding George Scangos. Morrissey will also become a member of the board of directors. He joined Exelixis in 2000 and served as executive vice president, discovery before his appointment as president of research and development in January 2007. Illumina (San Diego) has announced the appointment of Nicholas J. Naclerio to the position of senior vice president, corporate development. Naclerio formerly served as cofounder and executive chairman of Quanterix, raising $15 million in venture financing to launch the company. In addition, Illumina has named to its board of directors Gerald Möller, who currently serves as an advisor at HBM Bio Ventures, a Swiss investment firm. Previously, Möller spent 23 years at Boehringer Mannheim and Roche, where he held a number of leadership positions including CEO of the worldwide Boehringer Mannheim Group and head of global development and strategic marketing, pharmaceuticals for Roche. BrainStorm Cell Therapeutics (New York and Petach Tikva, Israel) has named Liat Sossover as CFO. Sossover has served in senior financial positions at a number of publicly traded and private companies, most recently as vice president, finance at ForeScout Technologies. James F. Young has been appointed to the board of directors of 3-V Biosciences (Menlo Park, CA, USA). He currently serves on the board of directors of Novavax. Previously, he served as head of MedImmune’s R&D organization and was directly involved in the development of approximately 20 clinical programs. Patrick J. Zenner has been elected to the board of directors of Par Pharmaceutical (Woodcliff Lake, NJ, USA). Zenner retired in January 2001 from Hoffmann-La Roche, where he served as president and CEO since 1993. He currently serves as chairman of the board of ArQule and Exact Sciences and as a director of West Pharmaceutical Services.

volume 28 number 8 august 2010 nature biotechnology