LCLs from unrelated CEU samples were phenotyped for cellular sensitivity to the four chemotherapeutic drugs included in our study: carboplatin 20 , cisplatin 21 , daunorubicin 22 , and etoposide 16 . We conducted genome-wide association scans using drug IC₅₀ as a quantitative trait.

A total of 5,238 CNVs from an array-based study 4 were evaluated in genome-wide association studies (GWAS) against cellular sensitivity drug phenotypes. Of these CNVs, 77% are deletions (0, 1, or 2 copy number), 16% are amplifications (2, 3, or 4 copy number), and the remainder are multi-allelic (greater than 3 diploid copy number genotypes) 4 . At the nominally significant threshold of P < 0.05, we identified 67 CNVs associated with carboplatin IC₅₀, 70 CNVs with cisplatin IC₅₀, 73 CNVs with daunorubicin IC₅₀, and 113 CNVs with etoposide IC₅₀.

We further evaluated the genomic characteristics of these drug susceptibility-associated CNVs for their size and type (deletion versus amplification). In general, there is little (Pearson) correlation between the size of a CNV and its association with cellular sensitivity to carboplatin (r = 0.020), cisplatin (r = 0.008), daunorubicin (r = 0.054) and etoposide (r = 0.024). We did, however, observe that the top CNVs associated (P < 0.05) with IC₅₀ for daunorubicin are significantly smaller (average size of 10.6 kb) than expected (average size of 14 kb) from the full set of CNVs included in our study; etoposide-associated CNVs are, in contrast, close to expectation (average size of 13.4 kb). The CNVs associated with carboplatin and cisplatin IC₅₀ (average size of 11.2 kb and 11.4 kb, respectively) are significantly smaller than expected.

Sixty-two of the 67 carboplatin-associated CNVs (P < 0.05) are biallelic (that is, deletions or amplifications); the remaining five CNVs are multi-allelic CNVs (that is, defined as having more than three CNV genotypes). Deletions are significantly more frequent (85%) than duplications among the carboplatin-associated CNVs. Similarly, all but 4 of the 70 cisplatin-associated CNVs (P < 0.05) are biallelic. The top cisplatin-associated CNVs are significantly more likely to be deletions (87%) than duplications.

All but 9 etoposide-associated CNVs (P < 0.05) are biallelic; 69 out of the 73 daunorubicin-associated CNVs (P < 0.05) are biallelic. Nearly 82% of the daunorubicin-associated CNVs and 82% of the etoposide-associated CNVs are deletions.

We observed that no exons overlap the genomic regions defined by the top associated CNVs for each anticancer drug included in our study, suggesting that these CNVs do not act to disrupt coding sequence. We thus hypothesized that these CNVs act to influence gene regulation. We evaluated the functional import of the drug susceptibility-associated CNVs by considering their effect on global gene expression. At an expression association threshold of P < 0.0001, 60% (N = 40) of the top CNVs associated with carboplatin (P < 0.05) were found to be expression quantitative trait loci (eQTLs). Interestingly, two of the top carboplatin-associated CNVs (CNVR3882.1 on chromosome 8 and CNVR666.1 on chromosome 2) predict the expression of SELL. We found that SELL expression level is also significantly correlated with carboplatin IC₅₀ (P = 0.02) in the CEU samples. We identified several target genes of carboplatin-associated CNVs (as eQTLs) whose expression levels were significantly correlated (after multiple testing correction 23 , false discovery rate (FDR) <0.05) with carboplatin IC_50, including PHGDH, MYO1B, TGFBR2, and PRF1. Similarly, nearly 56% (N = 39) of the cisplatin-associated CNVs (P < 0.05) were associated with the transcript level of genes as eQTLs. We found a target gene, MAST4, for two cisplatin-associated CNVs (CNVR2968.1 on chromosome 6 and CNVR7881.1 on chromosome 20). MAST4 trends toward significance (P = 0.06) with cisplatin IC₅₀ in the CEU samples. A target gene (C4A at P = 8.2 × 10^-6) for a cisplatin-associated CNV eQTL (CNVR4748.1 on chromosome 10) was found to be significantly correlated (after multiple testing correction 23 , FDR <0.05) with cisplatin IC₅₀.

Restricting our analysis to biallelic CNVs, we found, through simulations, that the top CNVs, for each platinating agent, are significantly enriched for eQTLs relative to frequency-matched SNPs (enrichment P < 0.05). The eQTL enrichment holds at a lower P-value threshold (P = 10^-6) used to define an eQTL, showing the robustness of our observation to the definition of eQTL. See Materials and methods for details on the simulation procedure.

Of the top CNVs associated with etoposide IC₅₀ (P < 0.05), 76% (N = 86) were found to be eQTLs. Of these CNV eQTLs, eight share UBA1 as a target gene (P < 9.7 × 10^-5). Two target genes (PLEKHG6 at P = 1.3 × 10^-6 and WSB2 at P = 8.0 × 10^-5) for etoposide-associated CNV eQTLs (CNVR4784.1 on chromosome 10 and CNVR1874.1 on chromosome 4, respectively) were found to be significantly correlated with etoposide IC₅₀ (FDR < 0.01) 23 . Nearly 52% (N = 38 of the top CNVs associated with daunorubicin IC₅₀ (P < 0.05) were eQTLs. We identified two daunorubicin-associated CNVs (CNVR479.2 and CNVR332.1 on chromosome 1) predicting the expression of HIST1H4A (P < 6.7 × 10^-5); we also found the expression level of this gene to be correlated (P = 3.7 × 10^-8) with daunorubicin IC₅₀ in the CEU samples. We identified several target genes (including PAPLN at P = 3.3 × 10^-5 and KLF12 at P = 6.1 × 10^-5) for daunorubicin-associated CNV eQTLs (CNVR2616.1 on chromosome 5 and CNVR948.1 on chromosome 2, respectively) whose expression levels were significantly correlated (after multiple testing correction 23 , FDR < 0.05) with daunorubicin IC₅₀.

As in the case of the platinating agents, we found, through simulations, that the top CNVs for each topoisomerase II inhibitor are more likely to be eQTLs than frequency-matched SNPs (enrichment P < 0.05).

Given the observed high proportion of deletions among CNVs associated with cellular sensitivity to chemotherapeutic agents, we sought additional functional support for the role of CNVs as transcriptional regulators from whole genome sequencing data coming out of the 1000 Genomes project, which characterized the CNV deletions with Gencode/ENCODE transcripts 14 . The resulting enlarged catalog of CNVs (with an initial focus on deletions) included CNVs of size 50 bp or larger mapped at single nucleotide resolution. We identified 376 transcripts to which CNV deletions were annotated 14 (by Gencode/ENCODE) as influencing (cis-regulating) transcription and/or translation. We proceeded to test the 376 transcripts for their role in predicting cellular sensitivity to chemotherapeutics. At P < 0.05, we found 21 transcript correlations with carboplatin, 15 with cisplatin, 23 with daunorubicin, and 21 with etoposide (see Table 1). Three transcripts (MOXD1, RCC1, SULF2) were significant after multiple testing adjustment (p _adj < 0.05, Bonferroni). Remarkably, the three transcripts were the only CNV deletions associated with all four agents at the nominal P < 0.05 threshold (Figure 1).

We investigated to what extent the CNVs associated with cellular sensitivity to chemotherapeutic agents may already be interrogated by SNP-based GWAS through linkage disequilibrium 6 . We found that the top CNV (CNVR1616.1) associated with carboplatin IC₅₀ (P = 5 × 10^-4) is not well-tagged by SNPs. Indeed, the best proxy SNP for this CNV on chromosome 3 is rs967422 (at r ² = 0.075). We found that the same CNV is also associated with cisplatin IC₅₀ (P = 6.5 × 10^-3). Another cisplatin-associated CNV (P = 5.5 × 10^-3), CNVR7870.1, is also not well-tagged; the best proxy SNP, rs915049, tags the CNV at a low r ² = 0.11. In each case, the best proxy SNP, in contrast to the 'tagged' drug susceptibility-associated CNV, shows no evidence of being associated with cellular sensitivity to the drug even at the nominal threshold of P = 0.05.

In the case of the topoisomerase II inhibitors, of the CNVs showing association with both etoposide and daunorubicin (P < 0.05), we found two - CNVR7205.1 and CNVR3293.1 - that are only modestly tagged (by rs563079 at r ² = 0.77 and rs17166803 also at r ² = 0.77, respectively). Neither rs563079 nor rs17166803 is associated with etoposide or daunorubicin IC₅₀. In contrast, CNVR2930.1, which is one of two etoposide-associated CNVs predicting the expression of CCND1 (expression P = 2.4 × 10^-7), is perfectly tagged (r ² = 1) by rs9500270. We identified a daunorubicin-associated CNV (CNVR2766.1; P = 3.7 × 10^-3) for which the best proxy SNP, rs10484327, tags the CNV at only r ² = 0.11.

PACdb 24 is a large-scale, publicly available genomic database, which to date holds the results of our SNP-based GWAS on the following chemotherapeutic agents: carboplatin, cisplatin, etoposide, daunorubicin, and cytarabine. PACdb implements a structured repository for incorporating other datasets, including information on other drugs, gene expression profiling, and cellular phenotypes. GWAS were initially conducted using SNP genotype data made available by the International HapMap project. We expanded PACdb to include the results of our CNV-based GWAS on carboplatin, cisplatin, etoposide, and daunorubicin. Furthermore, the results of eQTL mapping of HapMap CNVs to transcriptional expression are made available in the eQTL repository SCAN. Figure 2 shows a schematic diagram of our approach to the discovery of CNVs associated with sensitivity to drug and to the identification of such CNVs that act as eQTLs; it also illustrates the genomic resources we have made publicly available to the scientific community.

We evaluated to what extent the top CNV associations for a given drug would overlap with the top CNV associations for another drug belonging to the same chemotherapeutic drug class, defined in terms of mechanism of action. At the suggestive threshold of P < 0.05, of the CNVs showing association with carboplatin IC₅₀, 16% (n = 11) were also associated with cisplatin IC_50. Thus, we see a significant overlap (P = 7.7 × 10^-10) between the (two) sets of CNVs associated with cellular sensitivity to the platinating agents. Figure 3 illustrates a duplication (CNVR7826_full on chromosome 20) that is associated with sensitivity to carboplatin (Figure 3a) and to cisplatin (Figure 3b); note that the observed genotype associations with the platinums have concordant direction. Furthermore, the CNV is an eQTL predicting the expression of GSR (P = 4.67 × 10^-5) and SPARC (P = 4.70 × 10^-5). Remarkably, the expression levels of these target mRNAs, GSR (P = 0.045) and SPARC (P = 0.004), are correlated with carboplatin IC₅₀; similarly, GSR (P = 0.005) and SPARC (P = 0.005) are correlated with cisplatin IC₅₀. Glutathione reductase (GSR) has been implicated in several studies of platinum sensitivity 25 26 .

In the case of the topoisomerase II inhibitors, 12% of the etoposide-associated CNVs were found to associate with daunorubicin IC₅₀, and the observed overlap is still quite significant (P = 2.7 × 10^-10). The slightly greater percentage of overlap for the platinating agents is not due to higher phenotypic correlation (platinating agents (r = 0.52) versus topoisomerase II inhibitors (r = 0.69)).

We sought additional experimental support for the genes targeted by multiple CNVs associated with drug susceptibility. We identified two etoposide-associated CNV eQTLs that share CCND1 as a target gene (expression P = 2.4 × 10^-7). The over-expression of CCND1 has been shown to be associated with the up-regulation of the GST-π gene, increasing the sensitivity of a cancer cell line to etoposide 27 . We found CCND1 expression to be significantly correlated with etoposide IC₅₀ (P = 7.8 × 10^-6) in the CEU samples. After multiple testing correction, the gene remained significant (q-value = 0.0027). We subsequently conducted functional validation of the role of CCND1 expression in altering sensitivity to etoposide by performing real-time quantitative-PCR assays in an independent set of 52 CEPH LCLs (Figure 4; see Table S1 in Additional file 1 for the real-time PCR data on CCND1). Consistent with the direction of effect in the CEU samples, increased CCND1 mRNA levels resulted in increased IC₅₀ (P = 0.05) in the validation set. Thus, increasing CCND1 expression confers resistance to etoposide.

We obtained unrelated HapMap phase II CEU (Utah residents with ancestry from northern and western Europe) samples from Coriell Institute for Medical Research (Camden, NJ, USA). Cell lines were maintained in RPMI 1640 media supplemented with 15% fetal bovine serum (Hyclone, Logan, UT, USA) and 1% l-glutamine. The cell lines were passaged three times per week at a concentration of 350,000 cells/ml at 37°C in a 95% humidified 5% CO₂ atmosphere. Cellular sensitivity to drugs was measured in these cell lines with increasing concentrations of drug (from carboplatin, cisplatin, daunorubicin, and etoposide). Cell growth inhibition was evaluated using the alamarBlue™ assay (BioSource International Inc., Camarillo, CA, USA), as previously described 21 . IC₅₀ (the concentration required to inhibit 50% of cell growth) was determined by curve fitting of percent cell survival against concentrations of the drug.

Recent population-based surveys have mapped thousands of CNVs with increasing resolution. Such surveys have opened up approaches for modeling the relationship between structural variation and complex traits. Efforts to catalog these CNVs are necessary to clarify the functional impact of these variants. Here we utilize the recent comprehensive survey of CNVs 4 larger than 1 kb in size in the human genome, assayed in human LCLs from CEU (Utah residents with ancestry from northern and western Europe) samples. To investigate further the effect of deletions and to confirm our findings on the role of drug-associated CNVs as eQTLs, we studied the relationship between cis-regulated transcripts (from Gencode/ENCODE functional annotation) and cellular sensitivity to chemotherapeutics from a recent comprehensive study based on whole genome sequencing data of the 1000 Genomes Project 14 , which mapped CNVs of 50 bp or larger in size at nucleotide resolution.

For each CNV, genotypes were tested for association with cellular sensitivity to each of the drugs separately. Linear regression was performed between the copy number genotype (as the independent variable) and log₂-transformed IC₅₀ (as the dependent variable). Analysis was done in the statistical computing and graphics software R; the lm function was used to fit linear models.

Similarly, to examine the relationship between transcript level and drug-induced cellular sensitivity, a linear model was constructed for each drug, as previously described 19 , between log₂-transformed gene expression and log₂-transformed IC₅₀. Generally, for multiple testing adjustment, FDRs were calculated using the q-value 23 package in R. Unless otherwise stated, an FDR cutoff <0.05 was used as the statistical significance threshold.

SCAN 29 is an online database that makes publicly available the results of our eQTL studies, initially on single base polymorphisms. Global mRNA expression was assayed using the Affymetrix GeneChip Human Exon 1.0 ST Array 30 . To map CNVs as genomic loci influencing the transcriptome, we conducted linear regression on over 13,000 transcript clusters and the set of CNVs identified in the HapMap LCLs 31 .

We performed simulations to evaluate enrichment for eQTLs among the CNVs associated with cellular sensitivity to the drugs included in our study. To empirically generate the null distribution, we randomly generated sets of SNPs of matching minor allele frequency as the original list of CNVs (see Figure S1 in Additional file 2 for MAF distribution of the biallelic CNVs included in our study), as previously described 32 . To enable us to perform simulations conditional on MAF, we constructed non-overlapping MAF bins, each of width 0.05, using the MAFs of the SNPs in the HapMap CEU samples. The null sets were drawn from the combined platform SNPs (Affymetrix 6.0 and Illumina 1M) as well as from the entire set of HapMap CEU SNPs. The observed count is then compared to the empirically generated distribution to get an empirical P-value for the enrichment.

We obtained 52 unrelated non-HapMap CEPH (Centre d'Etude du Polymorphisme Humain) samples (independent of the discovery cohort consisting of HapMap CEU samples) from Coriell Institute for Medical Research. Cellular sensitivity to etoposide phenotype was quantified as described above with increasing concentrations of etoposide treatment (0.02 μM, 0.1 μM, 0.5 μM, and 2.5 μM for 72 hours). IC₅₀ was determined for each cell line. CCND1 mRNA levels were evaluated using a real-time quantitative PCR assay in the samples using TaqMan Gene Expression Assays (Applied Biosystems, Foster City, CA, USA) on the Applied Biosystems 7500 real-time PCR system. Primer/probes were obtained from Applied Biosystems. The human beta 2M (huB ₂ M, beta-2 microglobulin; NM_004048; Applied Biosystems catalog number 4326319E) was used as endogenous control. Relative quantification of gene expression utilized the 2 ^(-ΔΔCt) method 33 .