Systematic identification of genomic markers of drug sensitivity in cancer cells.

Clinical responses to anticancer therapies are often restricted to a subset of patients. In some cases, mutated cancer genes are potent biomarkers for responses to targeted agents. Here, to uncover new biomarkers of sensitivity and resistance to cancer therapeutics, we screened a panel of several hundred cancer cell lines--which represent much of the tissue-type and genetic diversity of human cancers--with 130 drugs under clinical and preclinical investigation. In aggregate, we found that mutated cancer genes were associated with cellular response to most currently available cancer drugs. Classic oncogene addiction paradigms were modified by additional tissue-specific or expression biomarkers, and some frequently mutated genes were associated with sensitivity to a broad range of therapeutic agents. Unexpected relationships were revealed, including the marked sensitivity of Ewing's sarcoma cells harbouring the EWS (also known as EWSR1)-FLI1 gene translocation to poly(ADP-ribose) polymerase (PARP) inhibitors. By linking drug activity to the functional complexity of cancer genomes, systematic pharmacogenomic profiling in cancer cell lines provides a powerful biomarker discovery platform to guide rational cancer therapeutic strategies.

INTRODUCTION

There is compelling evidence that the likelihood of a patient’s cancer responding to treatment can be strongly influenced by alterations in the cancer genome. For example, the use of drugs to selectively target the protein product of the BCR-ABL translocation in chronic myeloid leukemia (CML) has revolutionized the treatment of this disease, with 5-year survival rates of 90% in treated patients[1]. While targeting of specific genetic changes in defined patient subsets has been successful, a poorly explained range of responses to appropriately selected therapies is often still observed in patients[2,3]. Moreover, a large number of cancer drugs have not been linked to specific genomic alterations that could be used as biomarkers to specify their selective therapeutic effectiveness[4]. As drug pipelines generate new classes of compounds, systematic methods to identify predictive biomarkers during their early development could have a profound effect on the design, cost and ultimate success of new cancer drug development.

The NCI60 cell line panel and associated drug screens pioneered the approach of using cancer cell lines to link drug sensitivity with genotype data[5,6]. Cancer cell lines have subsequently been used to identify rare drug-sensitizing genotypes, including mutant EGFR, BRAF and EML4-ALK, that are highly predictive of clinical responses[2,3,7]. Here, we report the results of a large-scale screen of human cancer cell lines, incorporating detailed genomic and gene expression analysis, to systematically identify drug sensitivity biomarkers to a broad range of cancer drugs.

RESULTS

Systematic screening for therapeutic biomarkers

In order to capture the high degree of genomic diversity in cancer and to identify rare mutant subsets with altered drug sensitivity, we assembled 639 human tumour cell lines, representing the spectrum of common and rare types of adult and childhood cancers of epithelial, mesenchymal and haematopoietic origin (Fig. 1a and Supplementary Data 1). Cell lines were subjected to sequencing of the full coding exons of 64 commonly mutated cancer genes, genome-wide analysis of copy number gain and loss using Affymetrix SNP6.0 microarrays, and expression profiling of 14,500 genes using Affymetrix HT-U133A microarrays. The presence of seven commonly rearranged cancer genes and of microsatellite instability (MSI) were also investigated. The 130 drugs selected for screening covered a wide range of targets and processes implicated in cancer biology (Fig. 1b and Supplementary Data 2). They encompassed both targeted agents (n =114) and cytotoxic chemotherapeutics (n=13), including approved drugs used in clinical practice (n=31), drugs in development undergoing studies in clinical trials (n=47), and experimental tool compounds (n=52). To gain insight into drug-to-drug variation, we included multiple drugs designed against well-credentialed targets (Fig. 1b). The effect of 72 hours of drug treatment on cell viability was used to derive a multi-parameter description of drug sensitivity, including the half maximal inhibitory concentration (IC50), and the slope of the dose response curve (Supplementary Fig. 1). In total, we assayed 48,178 drug-cell line combinations with a range of 275-507 cell lines screened per drug (mean = 368 cell lines per drug; Supplementary Data 2). Clustering of compounds across cell lines based on IC50 values indicated that drugs with overlapping specificity were highly correlated, supporting the selectivity of the biological effects observed in the dataset (Supplementary Fig. 3 and Supplementary Tables 1 and 2).

Figure 1

A systematic screen in cancer cell lines identifies therapeutic biomarkers

a, The number of tumour-derived cell lines used for screening classified according to tissue type (n = 639 in total). b, The panel of 130 screening drugs classified according to their therapeutic targets, primary effector pathways, and cellular functions. A single drug may be included in multiple categories. The inset indicates the number of drugs screened against a selection of prototype cancer targets. c, A volcano plot representation of MANOVA results showing the magnitude (effect; x-axis) and significance (p-value; inverted y-axis) of all drug-gene associations. Each circle represents a single drug-gene interaction and the size is proportional to the number of mutant cell lines screened (range 1 – 334). The horizontal dashed line indicates the threshold of statistical significance (0.2 FDR, P < 0.0099). Insets I and II are magnified views of selected highly significant associations and the drug name, therapeutically relevant target(s) (in superscript), and cancer gene (in brackets) are given for each. The p-values for nilotinibABL(BCR-ABL), P = 2.54 × 10−65, and nutlin-3aMDM2(TP53), P= 2.78 × 10−37, have been capped at 1 × 10−28 in this representation.

Tumours from a particular tissue frequently have a shared set of somatic mutations. To gain insight into how this might relate to drug sensitivity, we performed an analysis to identify associations between cancer tissue type and drug sensitivity based on IC50 values. As expected, in some instances tumour-type specific sensitivity may be explained by the prevalence of cancer gene mutations (e.g. breast cancer sensitivity to inhibitors of the PI3Kinase pathway which is commonly altered in this tumour type; Supplementary Data 3 and 4). In other cases, however, our current understanding of cancer genomes could not explain the observed associations. For example, renal cell carcinoma (RCC) cells were sensitive to five SRC inhibitors (e.g. AZD0530, P < 1 × 10−4, n = 9 RCC and 294 non-RCC)[8], glioma cells to a ROCK inhibitor (GSK269962A, P < 1 × 10−6, n = 23 glioma and non-glioma 266)[9]. This analysis also identified therapeutic associations already used in the clinic with incompletely understood molecular basis such as sensitivity of myeloma cells to lenalidomide (P < 1 × 10−5, n = 3 myeloma and non-myeloma 455)[10]. For most drugs, however, sensitive cell lines were scattered across multiple cancer types.

Cancer gene mutations are biomarkers of drug response

Single gene mutations are increasingly being adopted as clinical biomarkers for the optimal application of cancer therapeutics. To identify associations between individual mutated cancer genes and drug sensitivity across the cell line panel we used a multivariate analyses of variance (MANOVA) incorporating the IC50 and slope of the dose response curve. This analysis revealed a large number of individual gene-drug associations, a subset of which (448/9039, 5%) were highly significant and are discussed here (Fig. 1c and Supplementary Data 5). Interestingly, most of the cancer genes analyzed, including those that are not known direct targets of the drugs tested, were associated with either sensitivity or resistance to at least one drug in our panel (65/69, 94%) (Supplementary Fig. 4). Similarly, sensitivity to most drugs tested was associated with a mutation in at least one cancer gene (118/130, 91%). Thus, diverse cancer gene mutations are implicated as markers of sensitivity or resistance to a broad range of anti-cancer drugs, indicating that genomic biomarkers could inform the therapeutic selectivity of many cancer drugs.

The mutated cancer genes most clearly associated with drug sensitivity are oncogenes that are direct targets of the relevant drug. For example, the BCR-ABL rearrangement conferred sensitivity to multiple ABL inhibitors (e.g. P = 2.54 × 10−65 for nilotinib, Fig. 1c and Supplementary Fig. 5)[1], several of which are approved for CML treatment. Similarly, BRAF mutation was associated with sensitivity to BRAF and MEK1/2 inhibitors (e.g. P = 1.25 × 10−24 for PLX4720, Fig. 1c and Fig. 2a, b and c)[3], including a structural analogue of Vemurafenib, which in clinical trials has extended the survival of BRAF mutation-positive melanoma patients. Additionally, ERBB2 (HER2) amplification was associated with sensitivity to EGFR-family inhibitors including Lapatinib (P < 1 × 10−7, Fig. 2d)[11], which is licensed for the treatment of HER2 positive breast cancer. We were also able to detect known associations between EGFR, FLT3, and PIK3CA mutations and drugs that target the products of these genes (Supplementary Data 5)[12,13]. A number of associations were driven by dramatic responses in small subsets of outlier cell lines. For example, two FGFR2-mutated cell lines were exquisitely sensitive to the FGFR inhibitor PD-173074 (Fig. 2e; P < 1 × 10−5)[14,15], confirming the need for large panels of cell lines to capture low frequency drug sensitizing genotypes.

Figure 2

Biomarkers of drug sensitivity and resistance

a, Gene-specific volcano plots of drug sensitivity associated with BRAF mutations in cancer cell lines (n = 54). b-k, Scatter plots of cell line IC50 (uM) values from selected drug-gene associations. IC50 values are on a log scale comparing mutated or non-mutated (WT) cell lines. Each circle represents the IC50 of one cell line and the red bar is the geometric mean. The drug name is indicated above each plot and therapeutic drug target(s) are bracketed.

We also found associations between the presence of inactivating mutations in tumor suppressor genes and several drugs, which in some instances provide insight into the interplay between tumour suppressors and the cellular machinery in mediating drug sensitivity. For example, mutation of TP53, an important regulator of apoptosis and cell cycle arrest in response to cellular stress, confers resistance to Nutlin-3a (P < 1 × 10−36), an inhibitor of the MDM2 E3-ligase which negatively regulates p53 protein levels (Fig. 2f)[16]. Similarly, mutational inactivation of RB1, a key repressor of cell cycle progression in normal cells, confers resistance to PD-0332991 (P < 1 × 10−10), an inhibitor of the upstream cyclin-dependent kinases (CDKs) 4 and 6, which drive cell cycle progression by inhibiting pRb through phosphorylation (Fig. 2g)[17]. Conversely, mutational inactivation of CDKN2A, encoding the CDK inhibitory protein p16, was associated with sensitivity to PD-0332991 (P < 1 × 10−11; Supplementary Data 5)[17], presumably because CDKN2A mutated cells have an enhanced requirement for signaling through the CDK4/6 - pRb signaling pathway.

In other instances genomic associations appear related to enrichment of mutations in a particular tissue type. The association of BRAF and NRAS mutation with sensitivity to obatoclax mesylate, a pro-apoptotic drug that targets BCL2 family anti-apoptotic proteins (BCL2, BCL-XL and MCL1), likely results from the enrichment of these mutations in melanoma, since drug sensitivity among melanoma cell lines was not correlated with presence or absence of these mutations (Supplementary Fig. 6). The tissue-specific effect of obatoclax may be related to inhibition of the melanoma survival mediator MCL1[18], since sensitivity of melanoma lines to ABT-263, another BCL-2 inhibitor which does not target MCL1, was not correlated with BRAF or NRAS mutation. Moreover, an ABT-263 insensitive melanoma cell line can be sensitized to this drug by siRNA-mediated depletion of MCL1 (Supplementary Fig. 7).

The genomic associations identified for 13 clinically approved cytotoxic chemotherapeutics in our panel were generally less significant than for targeted drugs, indicating that single gene biomarkers may be less informative for this class of drugs with broad action across many cancers (Supplementary Fig. 8 and 9). Intriguingly, we did not find general associations between targeted or cytotoxic drug sensitivity patterns and mutations in TP53. It may be that functional inactivation of p53, through mutations or abrogation of signaling pathways that regulate its activity, is an almost universal feature of cancer cell lines and thus differential drug sensitivity between mutant and non-mutant cell lines is not observed[19].

Several other novel gene-drug associations were identified that cannot be readily explained on the basis of our current knowledge of signaling pathways and may reflect unappreciated biological relationships. Mutation of NOTCH1 was associated with sensitivity to ABT-263 (P <1 × 10−9; Fig. 2h, Supplementary Fig 10), perhaps due to decreased expression of BCL2 family members in NOTCH1 mutant cell lines (Supplementary Fig. 11). Amplification of CCND1 (CyclinD1) or loss of SMAD4 were associated with sensitivity to multiple EGFR-family inhibitors including lapatinib and BIBW2992; and for SMAD4 this correlated with elevated EGFR gene expression (Fig. 2i and Supplementary Fig. 12). Inactivation of STK11 (LKB1; P < 0.01), thought to relieve repression of mTOR, was associated with sensitivity to the HSP90 inhibitor 17-AAG. Additionally, loss of FBXW7 was associated with sensitivity to the histone deacetylase (HDAC) inhibitor MS-275 (P < 1 × 10−5; Fig. 2j), and TET2 loss with sensitivity to the WEE1/CHK1 inhibitor 681640 (P < 1 × 10−4; Fig. 2k). These associations, and others presented here (Supplementary Data 5), represent candidate biomarkers of drug sensitivity and may ultimately be useful for the deployment of targeted therapies in cancer.

Complex genomic correlates of drug sensitivity

In most instances sensitivity of cancer cells to drugs is likely to depend on a multiplicity of genomic and epigenomic variables. Indeed, single gene-drug associations were only rarely able to explain the range of drug sensitivities observed across cell lines for any given drug (Fig. 2). We thus applied elastic net (EN) regression[20], a penalized linear modeling technique, to identify cooperative interactions among multiple genes and transcripts across the genome and defined response signatures for each drug. EN identified 26,938 feature-drug associations (Supplementary Data 6) from which 534 associations corresponding to 69 different drugs were highly significant (defined as −2.95> effect(e)>2.79 and frequency (f) > 0.76; Fig. 3a and Supplementary Fig. 13 and Supplementary Data 7).

Figure 3

Multi-feature genomic signatures of drug response

a, The top drug-feature associations identified by the EN are plotted for their frequency and effect size. Associations are colored black for expression features, red for mutations, blue for copy number, and green for tissue. b-c. Heatmaps of highly significant EN features associated with response to b, dasatinib (inhibitor of SRC,ABL) and c, 17-AAG (HSP90 inhibitor) for the 14 most sensitive (purple) and resistant (yellow) cell lines. For each cell line mutation features are at the top of the heatmap shown in black (present) or gray (absent), followed by expression features (blue corresponds to lower expression, red to higher expression). To the left of each feature is a bar indicating the absolute value of the effect size. Bars in purple are negative effects, indicating features associated with sensitivity, and bars in yellow are positive effects, indicating features associated with resistance. The natural log IC50 values are represented at the bottom. For clarity, only the top 4 features associated with sensitivity and resistance to 17-AAG are shown.

In many instances transcriptional features showed correlations with drug sensitivity that were equal to or stronger than those observed with gene mutation (Fig. 3a and Supplementary Table 3). For example, while sensitivity to the EGFR/ERBB2 inhibitor lapatinib correlated with ERBB2 expression and mutation, the strongest correlate for this drug was actually expression of the matrix metalloproteinase MMP28 (e = −29.28, f=1) (Fig. 3a). Notably, for most drugs, including those with clear linkage to cancer gene mutations, EN modeling identified multi-feature signatures of drug sensitivity. For example, together with BRAF mutation, sensitivity to RAF or MEK1/2 inhibitors was recurrently associated with 67 features. These features included expression of SPRY2 and DUSP4/6, which are known regulators of MAP-Kinase signaling (Supplementary Fig. 14)[21,22]. Interestingly, expression of 8 genes identified as markers of sensitivity to the MEK inhibitor AZD6244 significantly overlapped with a 18-gene signature of sensitivity to this drug (hypergeometric test of the overlap significance: P = 3 × 10−9)[23]. In some cases, EN modeling identified complex patterns of biomarkers corresponding to distinct subsets of sensitive cancer cell lines. Thus, sensitivity to Dasatinib, an inhibitor of multiple kinases including ABL and SRC, correlated with both BCR-ABL translocation and with multi-gene transcriptional signatures that were expressed in sensitive cell lines lacking that gene fusion (Fig. 3b).

EN modeling also identified transcriptional correlates of sensitivity for drugs without a known sensitizing mutational event. This included expression of LAG3, which correlated with sensitivity to the SGK inhibitor GSK-650394 (e= −29.9, f =1) and the correlation between expression of the NADPH dehydrogenase family member NQO1 with sensitivity to the HSP90 inhibitor 17-AAG (e = −22.21, f =1, Fig. 3c). Consistent with these findings NQO1 was previously shown to metabolize 17-AAG into a more potent inhibitor of HSP90[24].

A small number of features were recurrently associated with altered sensitivity to drugs from different classes indicating that they may be broadly involved in mediating drug sensitivity by diverse mechanisms such as drug efflux (e.g. ABCB1; Supplementary Data 6). To give further insight into this dataset, we have mapped the EN drug signatures onto the target of the drugs (Supplementary Data 7) and onto 457 known cancer genes (http://www.sanger.ac.uk/genetics/CGP/Census/; Supplementary Data 8). Collectively, these observations illustrate that in many instances multi-feature genomic signatures incorporating markers related to mutations, tissue lineage, cellular differentiation states and cellular pathways have the potential to expand and refine our current understanding of drug sensitivity.

EWS-FLI1 is a biomarker of PARP inhibitor sensitivity

We identified a highly significant association between the EWS-FLI1 rearrangement that is characteristic of Ewing’s sarcoma tumours and sensitivity to olaparib (AZD2281), an inhibitor of the poly-ADP ribose polymerase (PARP)(P = 1.03 × 10−26, Fig. 1c and 4a). Screening of a structurally distinct PARP inhibitor, AG-014699, across a large panel of cell lines confirmed the sensitivity of Ewing’s sarcoma cell lines (geometric mean IC50 for EWS-FLI1 = 4.7 uM versus 64 uM for wild-type, P < 0.0001 Mann-Whitney test, n = 291 cell lines; Fig. 4a, Supplementary Fig. 15 and Supplementary Data 9). Cells from Ewing’s sarcoma were more sensitive to olaparib (P = 2.84 × 10−11) than cells from other tumour types, including sarcomas of bone and soft tissue (Supplementary Fig. 15 and Supplementary Data 3). PARP inhibitors have activity in BRCA1/2 mutant cancers due to the defects in homologous recombination present in these tumors and their consequent reliance on alternative DNA damage repair pathways that are targeted by these inhibitors[25]. A comparison of olaparib and AG-014699 sensitivity in a panel of cell lines using a 6-day viability assay and colony formation experiments confirmed the marked sensitivity of Ewing’s sarcoma to PARP inhibitors, an effect that was comparable to that observed in BRCA-deficient cells (Fig. 4b and c, and Supplementary Fig. 16 and 17)[26]. Furthermore, treatment with olaparib selectively induced apoptosis in Ewing’s sarcoma compared to control cells (Fig 4d). Unlike in Ewing’s Sarcoma, we did not observe an association between BRCA1/2 mutations and sensitivity to PARP inhibitors in the 3-day screening format, which is likely due to a requirement for several rounds of division in these cells to accumulate toxic levels of DNA damage.

Figure 4

Ewing’s sarcoma cell lines are sensitive to PARP inhibition

a, The IC50 values of WT and EWS-FLI1 fusion positive cell lines to olaparib and AG-014699. b, Dose response curves to olaparib following 6-days constant drug exposure. Cell lines are classified according to tissue sub-type. c, Colony formation assays were performed for 7-21 days over a range of olaparib concentrations (0.1, 0.32, 1, 3.2 or 10 uM) and the concentration at which the number of colonies is reduced >90% for each cell line is indicated. d, Olaparib induced apoptosis in Ewing’s sarcoma cell lines following 72 hours treatment. e, Sensitivity to olaparib of EWS-FLI1 and FUS-CHOP transformed mouse mesenchymal cells compared to the SK-N-MC cell line (which harbors the EWS-FLI1 fusion). f, Sensitivity to olaparib of A673 cells transiently transfected with (siEF1) and without (siCT) EWS-FLI1 specific siRNA. All error bars are s.d from triplicate measurements except for b where error bars have been removed for clarity.

To assess whether the sensitivity to PARP inhibitors is due to the presence of the EWS-FLI1 rearrangement or intrinsic to the mesenchymal precursor cell type from which Ewing’s sarcoma arise, we compared the sensitivity to olaparib in mouse mesenchymal cells transformed with either EWS-FLI1 or the related liposarcoma-associated translocation FUS-CHOP[27,28]. EWS-FLI1 transformed cells showed sensitivity comparable to human Ewing’s sarcoma cells, while the FUS-CHOP-transformed cells were relatively resistant (Fig. 4e and Supplementary Fig. 18). Moreover, expression of EWS-FLI1 in NIH3T3 cells conferred increased sensitivity to olaparib (Supplementary Fig. 19), whereas olaparib sensitivity was partially reverted by transiently depleting EWS-FLI1 from Ewing’s sarcoma cells (Fig 4f and Supplementary Fig. 20). Higher FLI1 expression was also correlated with sensitivity to olaparib even when considering only non-Ewing’s sarcoma cell lines (r= −0.32 between IC50 and FLI1 expression, n =413, P = 1.68 × 10−11), suggesting that the sensitivity to olaparib of Ewing’s sarcoma lines might be related to EWS-FLI1 transcriptional activity.

Mutations in BRCA1 or BRCA2 are not present in these Ewing’s sarcoma lines (Supplementary Data 1), and we have observed no evidence to indicate that the DNA damage response is defective in Ewing’s sarcoma cells (data not shown). However, for reasons that are currently unclear, the EWS-FLI1 translocation was associated with sensitivity to cytotoxic drugs, including DNA damaging agents such as camptothecin (P < 1 × 10−5), cisplatin (P < 1 × 10−4) and mitomycin-C (P < 0.001)(Supplementary Fig. 21 and 22). Together with the report of olaparib sensitivity in prostate cancer cell lines expressing the ETS gene fusion, TMPRSS2-ERG[29], our data support the potential utility of ETS gene fusions as biomarkers of sensitivity to PARP inhibitors. Notably however, unlike the effect reported in prostate cancer, we observe potent cell death with olaparib treatment alone in Ewing’s sarcoma cells. These observations raise the possibility that PARP inhibitors have utility in the treatment of Ewing’s Sarcoma, a tumour of children and young adults with a 15% five year survival rate in patients with metastatic disease or relapse after chemotherapy[30].

DISCUSSION

High throughput cancer cell line screening for drug sensitivity patterns provides a strategy to identify appropriate cancer subtypes and biomarkers that may guide the early phase clinical trials of multiple novel compounds under development. The validity of this approach is supported by its effective identification of drug-genotype associations that have already been established clinically, and it sets the stage for clinical testing of novel therapeutic biomarkers, such as the association between the EWS-FLI1 translocation in Ewing’s sarcoma cells and sensitivity to PARP inhibitors. The data release accompanying this report, as well as the ongoing public web resource from future screenings (Genomics of Drug Sensitivity in Cancer; www.cancerRxgene.org), will hopefully enhance the discovery and validation of additional predictive cancer biomarkers.

While the large number of cell lines screened facilitates representation of rare cancer genotypes and mitigates against the effects of individual samples, the dataset presented here is limited both by the number of available genotypes, as well as the number of targets interrogated by currently available drugs. Despite the apparent utility of using tumour-derived cell lines grown in two dimensional culture, it is likely that experimental models that better mimic the tumour environment would in some instances further improve our understanding of drug sensitivity and provide additional insights. Nonetheless, we can discern an initial landscape of drug sensitivity patterns across a broad set of different cancer types and genomic backgrounds.

The identification of “outliers” which are exquisitely sensitive to a drug as a result of a specific genetic abnormality within a targeted pathway remains the most compelling vision for targeted cancer therapies. BCR-ABL positive CML, BRAF mutant melanoma and EGFR mutant positive lung cancer and drugs that target the protein products of these genes are now well-established associations. The observation of PARP inhibitor sensitivity by EWS-FLI1 positive Ewing’s sarcoma cell lines points to the likelihood of new potent gene-drug associations as novel chemical and genomic space are explored. Even in the absence of “outlier” effects, pharmacogenomic profiling reveals a wealth of biomarkers that may prove useful for patient stratification. Although further work is needed to assess their potential clinical utility, in some instances these biomarkers may help explain heterogeneity in clinical responsiveness even among preselected patient populations.

This work, as well as an accompanying report[31], provides a systematic and extensive view of the genomics underlying the sensitivity of human cancer cell lines to the diverse array of cancer drugs currently in use or under development. The emergent picture is of a complex network of biological factors that affect sensitivity to the majority of cancer drugs. This underscores both the challenge of identifying pre-selected patient populations for targeted therapies, as well as the opportunity to improve existing therapies and identify new therapeutic avenues by identifying more predictive biomarkers.

METHODS SUMMARY

Cells were treated with 9 concentrations (2-fold dilutions) of drug for 72 hours before measuring cell number relative to controls. A MANOVA was used to examine how drug IC50 and slope values associate with tissue type, the mutation status of 64 cancer genes (including gene amplifications and homozygous deletions), rearrangements and MSI. The elastic net utilised the same genomic datasets as the MANOVA and also incorporated additional copy number data from a total of 426 cancer genes, transcriptional profiles, and tissue type to identify feature associated with drug response as measured by cell line IC50.