Edited by Peter K. Vogt, The Scripps Research Institute, La Jolla, CA, and approved September 11, 2012 (received for review April 30, 2012)
Small cell lung cancer (SCLC) accounts for about 15% of all lung cancers. The prognosis of SCLC patients is devastating and no biologically targeted therapeutics are active in this tumor type. To develop a framework for development of specific SCLC-targeted drugs we conducted a combined genomic and pharmacological vulnerability screen in SCLC cell lines. We show that SCLC cell lines capture the genomic landscape of primary SCLC tumors and provide genetic predictors for activity of clinically relevant inhibitors by screening 267 compounds across 44 of these cell lines. We show Aurora kinase inhibitors are effective in SCLC cell lines bearing MYC amplification, which occur in 3–7% of SCLC patients. In MYC-amplified SCLC cells Aurora kinase inhibition associates with G2/M-arrest, inactivation of PI3-kinase (PI3K) signaling, and induction of apoptosis. Aurora dependency in SCLC primarily involved Aurora B, required its kinase activity, and was independent of depletion of cytoplasmic levels of MYC. Our study suggests that a fraction of SCLC patients may benefit from therapeutic inhibition of Aurora B. Thus, thorough chemical and genomic exploration of SCLC cell lines may provide starting points for further development of rational targeted therapeutic intervention in this deadly tumor type.
Over the past years the development of targeted therapies has dramatically affected clinical treatment of lung cancer (1⇓–3). This development was sparked by the identification of mutations in EGFR (4⇓–6) that confer exquisite sensitivity to EGFR inhibitors (2, 7) and EML4-ALK fusions (8) that make tumors susceptible to ALK inhibition (3). The recent identification of FGFR1 amplification and DDR2 mutations in squamous cell lung cancer (SQLC) patients has fueled hopes that not only lung tumors of never-smokers bear therapeutically amenable genetic alterations (9, 10). However, in small cell lung cancer (SCLC) the lack of specimens suitable for deep genomic characterization has so far hampered similar efforts to identify novel therapeutically relevant genome alterations.
Among the genes recurrently affected by genomic alterations in SCLC are TP53, RB1, as well as the MYC family genes such as MYC, MYCL1, and MYCN, which are frequently amplified in a mutually exclusive manner (11, 12). The PI3-kinase (PI3K) pathway has been proposed to be a therapeutically actionable signaling cascade that is activated in SCLC (11) but the frequency of genetic alterations driving PI3-kinase activation is currently unclear (13). Furthermore, the Hedgehog (HH) pathway has been identified as a potentially druggable target in SCLC mouse models (14) but it is presently unclear whether HH signaling dependency segregates with particular genetic alterations.
Given the inherent difficulties in the rational design of potent inhibitors of MYC and other transcription factors, alternative therapeutic strategies such as inhibition of MYC-MAX dimerization and “synthetic lethality” have emerged (15⇓⇓⇓–19). As a complementary approach, screening of libraries of small molecules across genomically characterized cell line panels has revealed direct oncogene dependencies as well as synthetic lethal dependencies (20⇓⇓–23). The use of small molecules offers the advantage of immediately addressing the question of whether a given vulnerability can be chemically attacked.
To identify therapeutically relevant genome alterations in SCLC, we performed a combined genomic and chemical vulnerability analysis in a panel of 60 SCLC cell lines. This study involved the screening of a library of 267 compounds across 44 SCLC cell lines coupled to genomic characterization of these and additional cell lines.
We analyzed chromosomal gene copy number alterations in 60 patient-derived SCLC cell lines (Dataset S1) using Affymetrix 6.0 SNP arrays and determined significant copy number alterations using the previously described GISTIC algorithm (Dataset S2) (24, 25). Next, we compared the significant alterations present in the cell line collection to the genetic alterations of a previously described collection of 63 primary SCLC specimens (Fig. 1A) (13). Confirming an overall high similarity of SCLC cell lines and primary tumors, this analysis revealed a significant (r = 0.83) correlation of copy number alterations in both datasets (Fig. 1A and SI Appendix, Fig. S1A), similar to findings in other tumor types (20⇓–22). Our cell line collection captures hallmark events of SCLC such as recurrent deletions of RB1 and PTEN (13, 26) but also amplification of genes such as FGFR1 (11, 13). Furthermore, in both datasets we identified recurrent and focal amplification of MYCL1, MYCN, and MYC (13). High-level MYCN amplification (inferred copy number > 4) occurred in about 4–6% of cases in both datasets, whereas MYCL1 (primary samples, 8% and cell lines, 22%) (Dataset S3) and MYC amplification (primary samples, 3% and cell lines, 15%) was detected with a higher prevalence in SCLC cell lines (Fig. 1 A–C and Dataset S3) (27). Although major events such as MYC amplification are found in both datasets, overall the significant copy number changes of SCLC differ from those found in non-small cell lung cancer (NSCLC) (r = 0.57) (SI Appendix, Fig. S1 B and C). Given the high prevalence of limited stage disease (∼68%) in the cohort of patient samples and the high prevalence of advanced stage disease in the case of the cell lines (∼95%) the frequency of MYCL1 (Dataset S3) and MYC amplification is likely associated with the stage of the tumor as seen previously for MYC (Fig. 1 A–C and Dataset S3) (28, 29). However, cell line artifacts and a treatment bias might contribute to this association and cannot be formally excluded.
SCLC cell line collection reflects major genetic lesions of SCLC patients. (A) Significant copy number changes (amplifications are in the lower panel and deletions are in the upper panel) as defined by GISTIC (q values) in SCLC primary samples (green and brown) and in SCLC cells (red and blue). Selected genes are annotated. (B) Significant amplifications (extensive, red; limited, black) as defined by the GISTIC algorithm (g score) identified in SCLC tumors. (C) Copy number changes at the MYC locus are displayed for the top 10 MYC amplified of extensive- (Left) and limited (Right)-stage samples. (D) FISH analysis of a sample showing amplification of MYC (Left) and a sample with no MYC amplification (Right) (MYC, green; control, red). (E) Frequency of MYC amplification in primary samples as determined by SNP arrays in a published dataset (25) and by FISH in the independent SCLC cohort.
To confirm our findings of significant copy number changes in SCLC, we analyzed an independent cohort of 55 primary SCLC tissues for the presence of MYC amplification using FISH (Fig. 1D). In accordance with published data (25), we identified high-level amplification of the MYC gene in about 5.5% of primary SCLC samples (Fig. 1 D and E). Thus, our data suggest that our cell line collection captures major copy number alterations of SCLC.
We performed a systematic cell-based screen (44 SCLC cell lines) against a library of 267 compounds with diverse scaffolds (Fig. 2A), targeting a wide range of cellular proteins (Dataset S4 and SI Appendix, Fig. S2A) (30⇓⇓⇓–34). Compound activity was assessed across cell lines as the remaining cellular viability at two different concentrations (Dataset S5). The resulting activity profiles ranged from compounds with no activity (n = 97) at high concentrations (5–10 μM) across all cell lines to compounds with high activity at low concentrations (0.5–1 μM) across the majority of cells (e.g., IPI-504) to highly selective compounds (e.g., PD173074 and PD0325904) (SI Appendix, Fig. S2B and Dataset S5) showing activity in only a few cell lines.
Identification of therapeutically tractable alterations in SCLC. (A) All screened chemical scaffold groups are depicted. (B) Hierarchical clustering of the raw activity data across all SCLC cell lines and compounds (0.5–1 μM) showing activity (red, high; white, low) in at least one cell line (viability <50%; 0.5–1 μM). Selected genetic lesions (Lower). (black, present; gray, not present). (C) GI50 values for PD173074 after 96-h treatment in SCLC cells are displayed (red, FGFR1 amplified; black, FGFR1 nonamplified). (D) Induction of apoptosis after 72-h treatment with 1 μM of PD173074 as assessed by FACS (annexin V/PI) is displayed for two FGFR1-amplified cell lines. (E) SBC7 cells treated for 24 h with PD173074 were analyzed for protein expression of phospho-FGFR, FGFR, phospho-FRS2, PARP, and actin by immunoblotting.
Using hierarchical clustering of the raw inhibitor activity data, we identified compound groups of different scaffolds indicating common targets (Fig. 2B). For example, the mTOR inhibitor everolimus shared a cluster with the AKT inhibitor MK-2206, the PI3K inhibitor PI-103, and the spirooxindole derivative AA123, previously described to induce mitotic arrest in cellular assays (Fig. 2B and Dataset S4) (30). Our data therefore suggest that AA123 might be a scaffold that inhibits the PI3K-signaling pathway. This analysis supports the robustness of our screening approach and affords identification of unexpected cellular targets for unique compounds.
To identify genetic predictors for the activity of the screened compounds, we used signal-to-noise–based feature selection combined with the K-nearest-neighbor (KNN) algorithm (22) (Dataset S6). This analysis revealed that PTEN loss predicts cytotoxic activity of the HSP90 inhibitor IPI-504 and its close homolog 17-AAG (P = 0.02 and P = 0.01; Fisher’s exact) (Dataset S6). Surprisingly, PTEN loss did not predict efficacy of PI3K inhibitors (Dataset S6). Overall, these results suggest that in the clinical setting PTEN loss may be a genetic marker for the efficacy of HSP90 inhibitors but not PI3K inhibitors in SCLC.
Next, we identified FGFR1 amplification as a predictor for the activity (P = 0.05) of the FGFR inhibitor PD173074 (Dataset S6). FGFR1 amplification is a recurrent genome alteration in SQLC, associated with FGFR dependency in some lung cancer cell lines (9, 11, 35). To test whether FGFR1 amplification is also linked with cytotoxic activity of FGFR inhibition in SCLC, we determined the GI50 values for a subset of seven SCLC cells (Fig. 2C). One of the two FGFR1-amplified cell lines, DMS114, was previously shown to be sensitive to PD173074 (GI50 = 0.46 µM) (9) (Fig. 2C). By contrast, the other FGFR1-amplified cell line, SBC7, was resistant to PD173074 and no apoptosis was induced upon treatment with the FGFR inhibitor (Fig. 2 C and D and SI Appendix, Fig. S3A). These data suggest that not all FGFR1-amplified SCLC tumors may be dependent on FGFR1 activity. Because these cells lack expression of PTEN (SI Appendix, Fig. S3B), we speculate that PTEN loss may contribute to modulation of FGFR inhibitor sensitivity. Consequently, despite inducing dephosphorylation of the adaptor protein FRS2 in SBC7 cells (Fig. 2E), PD173074 induced apoptosis in DMS114 cells but not in SBC7 cells (Fig. 2D).
Using the KNN algorithm, we next sought to identify compounds with specific activity in MYC-amplified cell lines. We found the PLK1 inhibitor BI2536, the ROCK1 inhibitor GSK269962A, as well as the pan-Aurora inhibitor VX680 to be specifically active in these cells (Datasets S5, S6, and S7) (36). To test whether Aurora kinase inhibitor efficacy is linked to MYC amplification in SCLC, we determined the GI50 values of structurally diverse Aurora kinase inhibitors VX680, MLN8237, PHA680632, and ZM447439 (Fig. 3A and Dataset S7). We observed a significant (P = 0.004 MLN8237; P = 0.003 PHA680632; P = 0.01 VX680; P = 0.01 ZM447439) enrichment of MYC-amplified cells in the subgroup of sensitive cells (<1 μM) for all four Aurora kinase inhibitors (Fig. 3A). We next assessed the ability of VX680 (37) to induce apoptosis in a subset of 11 cell lines. Using flow cytometry, we observed robust induction of apoptosis after 24–48 h of treatment with VX680 in MYC-amplified cell lines but not in those lacking MYC amplification (Fig. 3B). Furthermore, Aurora inhibition induced a collapse of the mitochondrial membrane potential, specifically in MYC-amplified SCLC cells (SI Appendix, Fig. S4A). Inhibition of Aurora kinases also led to significantly faster G2/M-arrest in MYC-amplified SCLC cell lines, compatible with a generally increased proliferation rate in these cells (Fig. 3C). To test whether MYC-amplified SCLC cells are generally vulnerable to induction of a G2/M-arrest, we measured apoptosis after treatment with the prometaphase-arresting agent nocodazole. No difference between the induction of apoptosis with nocodazole or VX680 was observed in MYC-nonamplified SCLC cells, whereas in MYC-amplified cells, we measured a significant difference (Wilcoxon test; P = 0.025) between nocodazole and VX680 (SI Appendix, Fig. S4B), suggesting that VX680-induced cell cycle arrest is not the main driver of cytotoxicity in MYC-amplified cells.
Inhibition of Aurora kinases leads to cell death and apoptosis in MYC-amplified cells. (A) GI50 values (Dataset S7) for the Aurora kinase inhibitors MLN8237, PHA680632, VX680, and ZM447439 (96-h treatment) in a subset of 34 SCLC cell lines (*MYC amplified). P values (Fisher’s exact test) are displayed. (B) Induction of apoptosis after treatment with 1 μM of VX680 (12–48 h) as assessed by FACS (annexin V/PI) is displayed for six MYC-amplified and five MYC-nonamplified cell lines. Inset shows representative pictures of GLC1 and SBC6 cells treated with either control or VX680. (C) Depicted is the fraction of cells in the G2/M-phase as measured by flow cytometry (PI signal) in six MYC-amplified (black) and five MYC-nonamplified cell lines (white). Error bars represent SD. (D) GLC1, N417, and SBC6 cells treated with VX680 (24 h) were analyzed for protein expression of Aurora A, phospho-HH3, PARP, phospho-AKT, AKT, MYC, and actin by immunoblotting. Lysates for detection of phospho-AURKA/B/C were sonificated.
VX680 treatment led to dephosphorylation of Aurora A/B/C as well as of histone H3 (HH3), a surrogate marker of Aurora B signaling, in all cell lines at concentrations in the range of the determined GI50 values (Fig. 3 A and D). Aurora inhibition was paralleled by poly(ADP-ribose)-polymerase (PARP) cleavage and a reduction of phosphorylation of AKT in MYC-amplified cells (GLC1 and N417) but not in MYC-nonamplified SBC6 cells (Fig. 3D). The observed induction of apoptosis translated into inhibition of tumor growth of VX680-treated (60 mg/kg) mice engrafted with MYC-amplified cells (GLC1) (SI Appendix, Fig. S5). By contrast, no growth inhibition was observed in nude mice engrafted with MYC-nonamplified SW1271 cells (SI Appendix, Fig. S5). Of note, increasing concentrations of VX680 led to enhanced levels of p-AKT in the control cell line, indicating that in MYC-amplified cells a lack of feedback loops that activate the PI3K pathway may contribute to the Aurora dependency (Fig. 3D). Depletion of Aurora A was recently shown to destabilize MYCN protein in MYCN-amplified neuroblastoma; the catalytical activity of Aurora A was, however, not required (15). By contrast, in MYC-amplified SCLC, inhibition of Aurora kinase activity did not affect MYC protein levels (Fig. 3D). Overall, these data indicate that Aurora kinase activity is specifically required for the survival of MYC-amplified SCLC cells and that the mechanism differs from Aurora dependency in MYCN-amplified neuroblastoma.
MYC is known to be involved in the pathogenesis of diverse cancer types (38). To test the relevance of MYC amplification in SCLC, we silenced expression of MYC (SI Appendix, Fig. S6A) in a subset of SCLC cell lines and assessed their viability (Fig. 4A). We observed a strong association between MYC amplification and MYC dependency resulting in induction of cell death in MYC-amplified but not in MYC-nonamplified cells (Fig. 4A). The MYC dependency of MYC-amplified cells (SI Appendix, Fig. S6B) was also linked to gene expression of MYC but not AURKA and AURKB (SI Appendix, Fig. S6C) and independent of MYC, Aurora A, and Aurora B protein expression (SI Appendix, Fig. S6D). Knockdown of MYC led to induction of apoptosis in MYC-amplified cells as assessed by PARP cleavage in immunoblotting assays (Fig. 4B).
MYC-amplified SCLC cells depend on MYC and Aurora B protein expression. (A) Viability of SCLC cells after transduction with MYC-shRNA selected from a set of four shRNA constructs (SI Appendix, Fig. S6A) compared with controls is displayed. Error bars represent SD. (B) GLC1, H211, N417, and SBC6 cells after transduction with either control or MYC-shRNA were analyzed for MYC and PARP protein by immunoblotting. (C) N417 (Left) and SW1271 (Right) cells after transduction with either control or AURKA/B-shRNA constructs were analyzed for protein expression of Aurora A/B, PARP, and MYC by immunoblotting. (D) Viability of SCLC cells after transduction with AURKB-shRNA or AURKA-shRNA (E) compared with controls was assessed. (F) N417, GLC1, SW1271, and SBC4 cells treated with AZD1152 (24 h) were analyzed for protein expression of phospho-Aurora A/B/C, Aurora A/B, and actin by immunoblotting. Lysates for detection of phospho-AURKA/B/C were sonificated. (G) Viability of N417, GLC1, (MYC-amp, red) SW1271, and SBC4 (MYC-nonamp, black) cells after 96-h treatment with barasertib (AZD1152) was determined with celltiter-glo (CTG) assays. Error bars represent SD.
To validate our findings, we silenced expression of AURKA and AURKB in a panel of SCLC cell lines (Fig. 4C). We had observed preferential but subtle dephosphorylation of Aurora B over Aurora A in two of the cell lines treated with VX680 (Fig. 3D, Center and Left), compatible with requirement of Aurora B in the context of MYC amplification in SCLC. Supporting this notion, knockdown of AURKB but not of AURKA induced PARP cleavage in MYC-amplified SCLC cell lines (Fig. 4C). Note that the AURKA-targeted hairpin 3 also has off-target effects against AURKB (Fig. 4C, Left). The observed dependency on AURKB expression was correlated with MYC dependency (r = 0.96) (SI Appendix, Fig. S6E). As a consequence, knockdown of AURKB (Fig. 4D) but not AURKA (Fig. 4E) resulted in a reduction of cell viability in MYC-amplified SCLC cells. As a further confirmation of preferential dependency of MYC-amplified SCLC cells on Aurora B, we next tested barasertib-hQPA (AZD1152), a compound with ∼1,000-fold selectivity against Aurora B compared with Aurora A (39). AZD1152 induced marked dephosphorylation of Aurora B and partially Aurora C but not Aurora A at concentrations of 0.1 μM in all cell lines (Fig. 4F). However, AZD1152 treatment led to a reduction of cell viability at low nanomolar concentrations only in the two MYC-amplified but not in MYC-nonamplified cells (Fig. 4G). Thus, further extending results obtained in genetically engineered cells (36), the observed activity of Aurora kinase inhibition is predominantly mediated by inhibition of Aurora B in MYC-dependent cells.
The findings of this combined genomic and chemical vulnerability screen support the use of this approach to develop strategies for genetically tailored therapies against SCLC. We demonstrate that our cell line panel captures the major copy number alterations of primary SCLC tumors, thereby allowing extrapolating to actual patient populations. Furthermore, the diversity of the screened inhibitors provides a first broad assessment of pathway dependencies across representative SCLC genotypes. Building on previous studies (11, 40), the scaling of both the number of cell lines and the number of compounds afforded identification of vulnerabilities associated with infrequent genome alterations, such as amplifications of FGFR1 and MYC. In the case of FGFR1 amplification, further studies will be required to clarify the frequency of FGFR1 dependency in SCLC, as other genetic lesions may play a role in the responsiveness to FGFR inhibition.
Furthermore, we describe and functionally characterize the dependency of MYC-amplified SCLC tumors on Aurora B (41). We find that MYC-amplified tumors depend on the kinase activity of Aurora B for their survival (36). Currently a series of Aurora kinase inhibitors (e.g., MLN8237 and PHA739358) including the Aurora B-selective inhibitor barasertib (AZD1152) are undergoing clinical evaluation in phase I/II studies (42, 43). Our data provide a rationale for the testing of these compounds in genetically defined SCLC patient groups.
Previous studies did not support a role for Aurora dependency in NSCLC cell lines of different genotpyes (22) or implied different MYC family genes in a mixed lung cancer panel (40), thus indicating that the role of amplified MYC and its dependency on AURKB (36) may differ between different cancer subtypes. Thus, our data provide a lineage-specific extension of previous data generated in lymphoma mouse models driven by an active transgene of MYC (36). Compatible with this notion is the recent finding that BRAF mutations associate with BRAF inhibitor sensitivity in melanoma but not colorectal cancer (44, 45). Our data not only point out the differences in the biology of distinct subtypes of lung cancer but also underscore the differences in oncogenic signaling of MYC gene family members: although amplifications of MYCL1 and MYCN amplifications occur in a mutually exclusive fashion with MYC amplification (suggesting genetic epistasis), they do not segregate with vulnerability to Aurora B inhibition in SCLC.
Although MYC-amplified SCLC represents only a small subset of lung cancer patients, recent experience with the successful introduction of ALK inhibitors for the treatment of about 2–3% ALK fusion positive adenocarcinomas suggest that genetic stratification is feasible and beneficial, even in small subgroups (3). Our study provides a framework for preclinical testing of genetically encrypted vulnerabilities in SCLC. In support of this notion, our initial screen has already yielded actionable targets for further preclinical validation and possible clinical testing. We therefore hope that—in the longer term—our approach might help to improve the disappointing survival rates of these patients.
All patients gave written informed consent sample analysis. The tumor specimens have been collected under local institutional review board approval (University of Cologne, Cologne, Germany) and genomic analyses were performed as described elsewhere (13). GISTIC analyses of a group of 60 SCLC cell lines were performed as described previously (24, 25). For a subgroup of 36 cell lines, previously published copy number data (www.sanger.ac.uk/genetics/CGP/CellLines) were used. The human genome build hg18 was used. All raw copy number data have been deposited in the Gene Expression Omnibus database (accession no. GSE40142). The cell line collection of 44 unique small cell lung cancer patient-derived cells was used for cell-based screening against 267 inhibitors using CellTiterGlo as a growth inhibition assay. Detailed analysis was performed on a subset of 51 compounds that showed activity in at least one cell line at concentrations of 0.5–1 μM. Annexin V and propidium iodide (PI) staining was used to measure apoptosis in flow cytometry assays. For the subgroup of cell lines where the raw copy number data were not generated in house, genotyping of a panel of 11 SNPs was performed to control the respective annotation. Cell viability was measured at two concentrations in triplicates and compared with DMSO controls. Calculation of the P values was performed using a Wilcoxon rank sum test, a two-tailed t test, or a Fisher’s exact test implemented in “R”. Pharmacodynamic response of signaling was measured by immunoblotting of cellular lysates of treated cells using phospho-specific antibodies. Not all bands were detected at the same membrane due to overlapping protein sizes. RT-PCR assays were performed with SYBR Green and primers for the respective genes and GAPDH as control. For gene silencing, lentiviruses were produced with pLKO.1-Puro–based vectors. After transduction, cell viability was measured by measuring cell numbers of quadruplicates and normalized to viability of cell lines transduced with control constructs. All animal procedures were performed in agreement with the animal protection committee and the local authorities and treatment was performed as described previously (22).
We thank Drs. Christian Reinhardt and Hamid Kashkar for support and AstraZeneca for providing barasertib-hQPA (AZD1152). This work was supported by the European Union-Framework Programme CURELUNG (HEALTH-F2-2010-258677 to R.K.T.); by the Deutsche Forschungsgemeinschaft through TH1386/3-1 (to R.K.T. and M.L.S.) and through SFB832 (TP6 to R.K.T. and TP5 to L.C.H.); by the German Ministry of Science and Education as part of the National Network for the Study of the Genome (NGFN) plus program (Grant 01GS08100 to R.K.T.), by the Max Planck Society and by the Behrensweise Foundation (M.I.F.A.NEUR8061 to R.K.T.); and by an anonymous foundation (R.K.T.). M.L.S. is a fellow of the International Association for the Study of Lung Cancer.
↵1M.L.S. and F.D. contributed equally to this work.
Author contributions: M.L.S., F.D., M.P., and R.K.T. designed research; M.L.S., F.D., J.S., H.B.-W., C.M., M.K., S.H., F.M., J.M.H., P.M.S., and L.C.H. performed research; A.R., P.A.E., R.T.U., A.P.A., V.V.V., T.N., H.W., and D.R. contributed new reagents/analytic tools; M.L.S., F.D., M.P., J.S., H.B.-W., C.M., M.K., A.R., S.H., F.M., J.M.H., D.S., P.A.E., R.T.U., A.P.A., V.V.V., P.M.S., T.N., H.W., R.B., D.R., L.C.H., and R.K.T. analyzed data; and M.L.S., F.D., and R.K.T. wrote the paper.
Conflict of interest statement: R.K.T. received consulting and lecture fees from Sanofi-Aventis, Merck KGaA, Bayer, Lilly, Roche, Boehringer Ingelheim, Johnson & Johnson, AstraZeneca, Atlas-Biolabs, Daiichi-Sankyo, and Blackfield as well as research support from AstraZeneca, Merck, and EOS. R.K.T. is a founder and shareholder of Blackfield, a company involved in cancer genome services and cancer genomics-based drug discovery.
This article is a PNAS Direct Submission.
Data deposition: The SNP array data reported in this paper have been deposited in the Gene Expression Omnibus (GEO) database, www.ncbi.nlm.nih.gov/geo (accession no. GSE40142).
This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1207310109/-/DCSupplemental.
Freely available online through the PNAS open access option.
