Epigenome-wide DNA methylation analysis of small cell lung cancer cell lines suggests potential chemotherapy targets

Background Small cell lung cancer (SCLC) is an aggressive neuroendocrine lung cancer. SCLC progression and treatment resistance involve epigenetic processes. However, links between SCLC DNA methylation and drug response remain unclear. We performed an epigenome-wide study of 66 human SCLC cell lines using the Illumina Infinium MethylationEPIC BeadChip array. Correlations of SCLC DNA methylation and gene expression with in vitro response to 526 antitumor agents were examined. Results We found multiple significant correlations between DNA methylation and chemosensitivity. A potentially important association was observed for TREX1, which encodes the 3′ exonuclease I that serves as a STING antagonist in the regulation of a cytosolic DNA-sensing pathway. Increased methylation and low expression of TREX1 were associated with the sensitivity to Aurora kinase inhibitors AZD-1152, SCH-1473759, SNS-314, and TAK-901; the CDK inhibitor R-547; the Vertex ATR inhibitor Cpd 45; and the mitotic spindle disruptor vinorelbine. Compared with cell lines of other cancer types, TREX1 had low mRNA expression and increased upstream region methylation in SCLC, suggesting a possible relationship with SCLC sensitivity to Aurora kinase inhibitors. We also identified multiple additional correlations indicative of potential mechanisms of chemosensitivity. Methylation of the 3′UTR of CEP350 and MLPH, involved in centrosome machinery and microtubule tracking, respectively, was associated with response to Aurora kinase inhibitors and other agents. EPAS1 methylation was associated with response to Aurora kinase inhibitors, a PLK-1 inhibitor and a Bcl-2 inhibitor. KDM1A methylation was associated with PLK-1 inhibitors and a KSP inhibitor. Increased promoter methylation of SLFN11 was correlated with resistance to DNA damaging agents, as a result of low or no SLFN11 expression. The 5′ UTR of the epigenetic modifier EZH2 was associated with response to Aurora kinase inhibitors and a FGFR inhibitor. Methylation and expression of YAP1 were correlated with response to an mTOR inhibitor. Among non-neuroendocrine markers, EPHA2 was associated with response to Aurora kinase inhibitors and a PLK-1 inhibitor and CD151 with Bcl-2 inhibitors. Conclusions Multiple associations indicate potential epigenetic mechanisms affecting SCLC response to chemotherapy and suggest targets for combination therapies. While many correlations were not specific to SCLC lineages, several lineage markers were associated with specific agents.


Introduction
Small cell lung cancer (SCLC) is a highly aggressive neuroendocrine tumor prone to early metastasis, short survival, and limited options for effective treatment [1][2][3]. Despite an unmet need to identify new therapies, progress in SCLC treatment has been hindered by rapidly acquired resistance to therapy resulting in limited and transient response to second and third line chemotherapeutic and immunotherapeutic agents [2]. Recently, the US FDA approved the immunotherapy drugs atezolizumab, pembrolizumab, and nivolumab for the treatment of recurrent SCLC [4].
Recent molecular studies have established that SCLC lineages fall into a number of distinct subtypes, currently referred to as SCLC-A, SCLC-N, SCLC-Y, and SCLC-P, based on their differences in gene and protein expression of transcriptional molecular neuroendocrine, nonneuroendocrine, or tuft cell-like lineage regulators ASCL1, NEUROD1, INSM1, YAP1, and POU2F3, and of their downstream molecular targets [2]. Whether these SCLC subtypes respond differently to specific treatments and whether patient tumors may represent a heterogeneous mix of SCLC lineages remains a subject of active investigation [2].
There is a growing understanding that many changes associated with SCLC carcinogenesis may be driven by epigenetic processes. Genes encoding several epigenetic factors including histone acetyltransferases EP300 and CREBBP, and histone methyltransferases KMT2A and KMT2D are frequently mutated in SCLC tumors [2,5,6,11]. High expression of another histone methyltransferase gene, EZH2, is a distinct feature of SCLC when compared to normal lung tissue or other cancer categories [11]. The transcriptional master regulator, POU2F3, which defines the tuft cell-like SCLC lineage [2,12,13], is epigenetically silenced in cervical cancer via hypermethylation of the POU2F3 promoter [14], suggesting the possibility that DNA methylation-mediated regulatory mechanisms could play a role in SCLC development and progression. A survey of global methylation patterns in primary SCLC tumors and SCLC cell lines identified 73 potential gene targets enriched for binding sites of cell fate-specifying transcriptional factors [15].
Despite growing evidence for the role of epigenetic factors in SCLC cancer, understanding their influence on tumor response to treatment remains limited. Earlier studies of other cancer histologies have identified therapeutically relevant DNA methylation biomarkers in the promoter regions of MGMT, increased methylation of which is beneficial for the response to alkylating therapeutic agents and ionizing radiation in glioblastoma and colorectal cancer, and of SLFN11, methylation of which has been associated with resistance to DNA damaging agents in a variety of cancer categories [16][17][18][19][20][21][22]. Insight into epigenetic modulation of SCLC response to DNA damaging agents was provided by the discovery of epigenetic silencing of SLFN11 by EZH2 in the course of cisplatin-etoposide therapy, which may lead to treatment resistance or chemo-sensitive relapse [23]. Unfortunately, the role of epigenetic mechanisms in response to other agents and the effect of various epigenetic alterations on drug treatment response in SCLC remain largely unknown.
To provide insight into epigenetic factors which may influence the response of SCLC to treatment, an epigenome-wide DNA methylation analysis was performed. Methylation data were analyzed to determine how epigenomic states of gene regions, individual probes, and genes were associated with SCLC response to FDA-approved oncology drugs and about 400 investigational agents. These epigenome analyses utilized drug response and transcriptional profiling data obtained in our earlier study [24], in which we identified a number of gene transcripts and miRNAs associated with SCLC response to treatment. In the current report, we describe the use of high-density DNA methylation measures to identify epigenomic regions and gene targets that were strongly associated with response to a variety of therapeutic agents.

Drug response measures
We analyzed drug response measures obtained earlier in the in vitro screen of 526 FDA-approved and investigational agents using SCLC cell lines [24]. Sixty-six SCLC cell lines had both drug response information and DNA methylation measures and were included in the analysis of Spearman correlation between probe or gene region methylation levels and median values of the log(IC50) measures of drug response among cell line replicates (Supplementary Table 1). Among 526 agents, 412 drugs showed variability of response among the 66 SCLC cell lines and were included in association analysis.
Measurement of SCLC cell growth and drug response measures and the steps for quality control (QC) were described in detail previously [24]. Briefly, each agent was tested at nine concentrations (10 μM to 1.5 nM, with DMSO concentration of 0.25%), after a 96-h incubation with the cells. The statistical validity of the drug response dataset was evaluated by calculating the Z' factor for each plate in the assay, with Z' > 0.5 considered to be a high-quality assay. Concentration response data were fit with a 4-parameter curve, and median IC50 values for each agent were computed among cell line replicates.

Methylation data processing
Methylation measurements for all cell lines were generated in a single batch using Illumina Infinium Methyla-tionEPIC BeadChip (Illumina, Inc). The absence of batch effects was confirmed by comparing clustering of SCLC based on methylation data with that based on gene expression data.
Methylation data were normalized, and beta and detection p values were calculated using the minfi package [25] using default parameters, resulting in 866,091 methylation probe measurements. Methylation probe beta-values for individual cell lines with detection p values ≥ 10 −3 and the entire 1427 probes with median detection p ≥ 10 −6 were excluded from analysis. Probes overlapping with single nucleotide polymorphisms (SNPs) were filtered out according to the list of probe masking recommendations of Zhou et al. [26,27]. The final methylation dataset used in correlation analysis with drug response and with gene expression had methylation beta-values for 760,637 probes that passed all filtering.
Epigenome-wide analysis of association of DNA methylation with chemosensitivity In order to compute gene region-averaged methylation beta-values for the Infinium MethylationEPIC BeadChip dataset, we developed an R program which followed the algorithm which had previously been developed by other authors for the IMA software package [28] for the analysis of Illumina 450K Infinium methylation array data. Briefly, we used the Infinium MethylationEPIC Bead-Chip annotation of each probe [29] according to the UCSC genome browser data to compute gene regionaveraged methylation values for each of the following gene regions: TSS1500 (200-1500 bases upstream of the transcriptional start site, or TSS), TSS200 (0-200 bases upstream of the TSS), 5′UTR (within the 5′ untranslated region, between the TSS and the ATG start site), first exon, gene body (between the ATG start site and the stop codon), and 3′UTR (within the 3′ untranslated region, between the stop codon and poly A signal). Methylation of different gene regions was considered separately in association analyses and was not combined. The probes annotated as belonging to more than one region and/or more than one gene were included in calculation of average methylation values of each of their respective annotated gene regions. Statistical analysis was performed using the R environment v. 3.5.3. The resulting methylation values were computed for 108,795 regions belonging to 26,239 genes, transcripts, and miRNA listed in the Infinium MethylationEPIC Bead-Chip manifest annotation [29]. Chromosomal regions (cytoband) of each probe were identified according to the UCSC genome annotation database for the hg19 (GRCh37) assembly of the human genome based on probe coordinates in the Infinium MethylationEPIC BeadChip annotation [30].
Spearman correlation analysis of methylation measures with log(IC50) was performed for methylation betavalues of each individual methylation probe, and also for methylation values averaged among the probes within each of the six gene regions (TSS1500, TSS200, 5′ UTR, first exon, gene body, and 3′ UTR). Sixty-six SCLC cell lines, which had both drug response data and methylation measures, were included in correlation analysis. All analyses described in this report included only SCLC cell lines and did not include NSCLC or mesothelioma cell lines. Significance of correlation of methylation of gene regions with drug response was evaluated using the Benjamini-Hochberg adjustment procedure for false discovery rate (FDR) [17] using all p values from correlation tests of all 412 drug agents with variable drug response and 108,795 gene regions. For individual methylation probes, we used a fixed threshold of p < 9.42 × 10 −8 for a single Illumina Infinium MethylationE-PIC BeadChip array methylation probe, according to recently published recommendations [31]. In addition, we also compiled a broader list of top genes associated with drug response by combining the probes satisfying a more liberal threshold of p < 5 × 10 −7 , in an analogy with previously reported criteria of 10 −6 for Illumina 450K array which had fewer probes than the Illumina Infinium MethylationEPIC BeadChip array [32]. We refer to the FDR-adjusted p values as p FDR and the original p values prior to FDR adjustment as p O . Special attention was paid to the significant associations involving the upstream gene regions, which are most likely to contain the promoter regions and regulatory regions affecting gene expression and to individual probes located in the upstream gene regions.
The overlap between top results for different agents and genes was visualized using Venn diagrams which were constructed with the help of the public online version of DisplayR [33].

Association of methylation of candidate genes with selected antitumor agents
In addition to epigenome-wide analysis of association of SCLC DNA methylation with drug response, we also focused more closely on possible epigenetic mechanisms of response to 44 anticancer agents (Supplementary Table 2). This list included agents that exhibited higher efficacy in subgroups of SCLC cell lines in the in vitro single agent screen, as well as agents with potential promise for activity against SCLC based on in vitro or preclinical results from other studies or based on their inclusion in SCLC clinical trials [24]. We examined association of log(IC50) of these agents with methylation of individual probes and gene regions of 78 genes representing drug-specific targets and 48 additional genes involved in drug target pathways (Supplementary Table 2). In addition, we analyzed association of methylation of log(IC50) of each of the 44 agents with methylation of individual probes and gene regions of 159 proteincoding genes that included genes with relevance to SCLC lineage determination; SCLC lineage markers; genes that carry frequent mutations or genome alterations in SCLC; genes which are commonly inactivated, overexpressed, or epigenetically modified in SCLC tumors or in specific SCLC subtypes; as well as genes previously reported as being involved in pathways leading to SCLC pathogenesis; or those suggested as being relevant to SCLC response to chemotherapy [3,5,11,13,15,24,[34][35][36][37][38][39]. The list of these genes is provided in the legend to Supplementary Table 2. Each candidate gene was represented by multiple probes and up to six gene regions (TSS1500, TSS200, 5′UTR, first exon, gene body, and 3′UTR), which were analyzed independently from each other. The resulting correlation p values were FDR adjusted by combining the results for all 44 agents, separately for 10,515 methylation probes in or near the candidate genes and for 1376 gene regions in candidate genes. The adjustment for multiple testing and interpretation of the significance of the results derived from the analysis of candidate genes and regions were done separately and independently from the adjustment for multiple testing and interpretation of significance in the analysis at the epigenome-wide level, described above. The results obtained using these two approaches were presented separately.
Association of DNA methylation with gene and miRNA expression and correlation of transcripts with drug response We examined how association of SCLC DNA methylation of individual probes and gene regions with drug response may be related to expression of genes and miRNAs located in the same genome regions. For this purpose, we used gene expression and miRNA measurements generated and processed by an earlier study of our group [24], which generated transcript expression data using Affymetrix GeneChip®Human Exon 1.0 ST Arrays (NCBI GEO accession number GSE73160) and NanoString miRNA expression measurements (NCBI GEO accession GSE73161). Experimental and computational procedures for mRNA and miRNA data collection, processing, QC, data normalization, and adjustment for batch effects were reported previously [24,38]. We used mRNA expression measures normalized using Robust Multi-Array Average (RMA) and summarized at the whole transcript level using AROMA [40]; we also utilized miRNA data which were normalized and log 2 + 1 transformed [24]. Expression data for the total of 18,690 transcripts and 800 miR-NAs were adjusted separately for batch effects using the ComBat function of the sva package [41]. The validity of adjustment was confirmed by hierarchical sample clustering using the hclust function of R v. 3.3.0. Pearson correlation was used to evaluate association of log 2 -transformed normalized expression values of each transcript and each miRNA with log(IC50) of each drug agent. In addition, we used Spearman correlation (Supplementary Table 1) to examine how the methylation beta-values of each of the top methylation probes and average methylation betavalues of gene regions associated with drug response were correlated with log 2 -transformed normalized expression measures of genes and miRNAs located in the same genome regions, based on Illumina Infinium MethylationE-PIC BeadChip microarray annotation according to the UCSC genome browser data. Here and below, ρ stands for Spearman correlation coefficient and r stands for Pearson correlation coefficient.

Copy number data
In cases when DNA methylation directly affects gene expression without copy number changes, a negative association between DNA methylation measures, most commonly in the upstream gene region, and gene expression may be expected [42]. Gene copy number gain commonly results in its overexpression, whereas copy number loss could lead to lower expression levels. Copy number events have been reported to result in positive or negative correlations between DNA methylation and gene expression measures depending on the probe location, with positive correlations more common in the gene body [42]. To examine possible causes of positive associations between DNA methylation and gene expression, we verified copy number information from the Cancer Cell Line Encyclopedia (CCLE) resource at the Broad Institute [43,44] for selected genes in which one or more probes and/or gene regions were significantly associated with drug response in our data, and the same probes and/or gene regions were strongly positively correlated with gene expression values. Thirty-three SCLC cell lines with available methylation, transcript expression, miRNA expression, and drug response measurements in our dataset also had copy number data available from CCLE. Gene level copy number data had been generated by the CCLE Consortium using Affymetrix 6.0 SNP arrays, with segmentation of normalized log 2 ratios of the copy number estimates performed using the circular binary segmentation algorithm [43,44].
Analysis of association of TREX1 expression and methylation with drug response using data from other resources Due to the absence of TREX1 gene expression measurements among the transcript clusters derived from the Affymetrix GeneChip®Human Exon 1.0 ST Array, for this gene we used Affymetrix Human Genome U133 Plus 2.0 microarray measurements (probe 34689_at) available from the CCLE legacy portal [22,45] for the 36 cell lines that were included both in our dataset and in the CCLE data. These microarray measures had an excellent correlation with TREX1 RNA-seq expression measurements available from CCLE [43,44] (Spearman correlation coefficient ρ ≥ 0.9135, Pearson r ≥ 0.9041, p ≤ 3.80 × 10 −20 for all tests in SCLC cell lines and across cancer categories).
For validation of drug sensitivity associations with TREX1 methylation and expression, we analyzed correlations of molecular measurements with drug response in 40 SCLC cell lines that had drug sensitivity data available from the Genomics of Drug Sensitivity in Cancer (GDSC) dataset [46,47] and TREX1 DNA methylation and gene expression measures available from CCLE. The independent TREX1 methylation dataset in CCLE was generated using reduced representation bisulfite sequencing (RRBS). These data included the TREX1 promoter region within 1 kb upstream of the TSS, promoter CpG clusters, and promoter CpG islands, as provided by the CCLE project [43] and described in detail in a recent report [44]. For TREX1 gene expression measures, we used CCLE Affymetrix Human Genome U133 Plus 2.0 microarray data (probe 34689_at) [22,45]. Drug sensitivity measurements (GDSC1 and GDSC2 datasets) were obtained from the Genomics of Drug Sensitivity in Cancer (GDSC) resource [46,47].
Association of methylation and gene expression with drug response in relation to SCLC lineage classification To examine whether patterns of DNA methylation and transcript expression that were significantly correlated with drug response were also associated with SCLC lineage subgroups, we analyzed Spearman and Pearson correlation of DNA methylation of individual probes and gene regions with expression of six lineage SCLC markers, ASCL1, ASCL2, NEUROD1, INSM1, YAP1, and POU2F3 [2,48]. We also used hierarchical clustering of SCLC cell lines based on these six lineage markers to examine whether patterns of DNA methylation and gene expression in the genes of interest were different among SCLC clusters. Clustering of SCLC cell lines according to their lineage marker expression was performed using the "average" (UPGMA) option of the hclust command in the R environment based on Euclidian distances, with subsequent annotation of SCLC cell line cluster assignments according to a previous report [2] when such annotation was available.

Results
Below, we first present the results of the epigenomewide association analysis of individual probes and gene regions with all agents. We discuss the strongest associations of methylation of individual probes with drug response. We further discuss whether those associations were also in agreement with the correlations of methylation of regions of the same genes with drug response and whether such associations could be explained by the effect of DNA methylation on gene expression. We also highlight some of the top correlations of gene regions with drug response. In a separate section, we report the associations of methylation of the probes and regions in the candidate genes with response to candidate drug agents. We highlight their strongest correlations and also discuss specific genes of particular biological interest. Detailed information about all significant associations is provided in Supplementary Tables 3, 4, 5, 6, 7 and 8. Additional details are provided in Supplementary Data 1, 2, 3 and 4.
Association of DNA methylation of probes and gene regions with drug response at the epigenome-wide level of significance Spearman correlation analysis between the beta-values of methylation probes that passed QC and SNP filtering and log(IC50) of drug agents identified 294 strong correlations with p < 9.42 × 10 −8 ; all of them had Spearman correlation coefficient |ρ| > 0. 6 (Supplementary Tables 3  and 4). The summary of genes containing the top probe correlations with drug agents (p < 10 −8 ) is presented in Table 1. Table 2 provides the list of significant (p FDR < 0.05) and nearly significant (p FDR < 0.1) gene regions associated with drug response.
The strongest probe correlation satisfied the Bonferroni-adjusted threshold for multiple testing of EPIC array probes with 412 agents (p O < 2.29 × 10 −10 ). It involved the probe cg13178916 in the body of SNED1 and resistance to the histone deacetylase (HDAC) inhibitor 4SC-202 (ρ = 0.6927, p O = 1.16 × 10 −10 ; Table 1). Due to similar mechanisms of action of various agents, their associations with DNA methylation are likely not independent, and the Bonferroni threshold is likely to be excessively stringent. When using p O < 9.42 × 10 −8 , the probe cg13178916, which had the range of beta-values from 0.096 to 0.843, was also associated with microtubuledisruptive agents BAL-101553 and vinblastine (Supplementary Tables 3 and 4; Supplementary Data 1). Even though other SNED1 probes did not satisfy the significance threshold for multiple testing, 52 probes were associated with resistance to 4SC-202 with p O < 0.05, including 44 probes with p O < 0.01 (ρ ≥ 0.3197; Supplementary Data 2A). The probes cg10717312 and cg07644939, located immediately adjacent to cg13178916, were among the 7 probes most strongly associated with 4SC-202 (p O < 5 × 10 −5 , ρ ≥ 0.4935; Supplementary Data 1 and 2A). SNED1, a Sushi, Nidogen, and EGF-like Domain 1 extracellular matrix protein, is associated with progression and metastasis of mammary carcinomas and with poor outcomes in ER − /PR − breast cancer [49]. Deletion of the chromosomal region 2q37.3 containing SNED1 is a recurring event in cancer, and in ovarian cancer cell lines, it was associated with resistance to the HDAC inhibitor vorinostat [47]. Similar to 4SC-202, multiple SNED1 probes were weakly associated with resistance to vorinostat (Supplementary Data 2B). Methylation of cg13178916 was weakly positively associated with transcript expression (Spearman ρ = 0.2178; Table 1; Supplementary Table 4). Expression of the SNED1 transcript NM_001080437 (Affymetrix cluster ID 2536071) was weakly but significantly associated with multiple HDAC inhibitors, although its association with resistance 4SC-202 and vorinostat was weak and did not reach statistical significance (Supplementary Data 2C). The number of cell lines with SNED1 deletion in our data was insufficient to derive any conclusions about its association with drug response or DNA methylation (Supplementary Data 2D).
Associations of TREX1 promoter methylation and expression with log(IC50) of vinorelbine were validated in SCLC cell lines from the CCLE and GDSC datasets (− 0.6506 ≤ Spearman ρ ≤ − 0.4275, 0.0006 ≤ p O ≤ 0.0207 for significant correlations of various measures of TREX1 promoter methylation and log(IC50) of vinorelbine from    Among multiple transcripts matching the same gene, the best matching transcript was identified by finding all Affymetrix transcripts with matching gene names and by selecting the most significant transcript from Spearman correlation between log 2 of transcript expression and methylation of each probe within the gene/gene region that had been associated with log(IC50) of that agent with p O < 5 × 10 −7 . We also provide the NCBI or Ensembl ID of the transcript with a matching gene name that had the most significant Pearson correlation between log2 of transcript expression with log(IC50) of the agent. For those gene regions where probes with p O < 5 × 10 −7 were assigned to multiple genes in the Infinium MethylationEPIC BeadChip manifest annotation [29] according to the UCSC genome annotation, the output summary for multiple genes was combined, as shown in the Gene Regions column for the regions combining ETV6 and RNU6-19P output and SLC16A6 and ARSG output. An expanded list of genes and combined genome regions containing multiple genes, each with at least one methylation probe satisfying p O < 9.42 × 10 −8 , is provided in Supplementary Table 4. Correlation results of log 2 of expression of the most strongly correlated transcript with probe methylation and with log(IC50) are provided for those genes that had matching transcript data in the Affymetrix GeneChip®Human Exon 1.0 ST Array.
Cluster ID transcript cluster ID for the Affymetrix GeneChip®Human Exon 1.0 ST Array, ρ Spearman correlation coefficient, r Pearson correlation coefficient, CDK cyclin-dependent kinase, HDAC histone deacetylase, LPAAT-β lysophosphatidic acid acyltransferase-β, PLK-1 polo-like kinase 1, SERM selective estrogen receptor modulator GDSC included data for the Aurora kinase inhibitors ZM447439, tozasertib, and alisertib (MLN-8237). Alisertib/MLN-8237 was also included in our dataset, and its activity was associated with increased methylation of all TREX1 regions other than the 3′UTR (p O ≤ 0.0231), although they did not reach statistical significance after adjustment for multiple testing at the epigenome-wide level and accounting for all agents. The strongest associations with alisertib sensitivity in our data were observed for TSS200, 5′UTR, first exon, and gene body (− 0.5676 ≤ ρ ≤ − 0.4995, 6.70 × 10 −7 ≤ p O ≤ 1.96 × 10 − 5 ; data not shown). Methylation of 11 TREX1 probes (out of 19 total; Supplementary Data 3) in our data was also associated with sensitivity to MLN-8237 but did not reach significance after adjustment for multiple testing (− 0.5837 ≤ ρ ≤ − 0.4235, 2.70 × 10 −7 ≤ p O ≤ 0.0004; data not shown). Consistent with our data, several TREX1 promoter methylation measures and TREX1 expression were associated with alisertib sensitivity in GDSC2 (e.g., − 0.4996 ≤ ρ ≤ − 0.4292, 0.0178 ≤ p O ≤ 0.0590 for promoter CpG island methylation; r = 0.4365, p O = 0.0258 for expression). In GDSC1, TREX1 promoter methylation was associated with sensitivity to tozasertib and ZM447439 (− 0.5739 ≤ ρ ≤ − 0.4468, 0.0007 ≤ p O ≤ 0.0326 for the methylation measures most strongly associated with both agents), and expression was associated with sensitivity to tozasertib (r = 0.5805, p O = 0.0023). These correlations in multiple datasets suggest that TREX1 methylation and expression are associated with SCLC response to Aurora kinase inhibitors, antimitotic agents, and a number of additional drug categories.
Supplementary Table 4 provides a gene level summary for 182 gene-drug correlations for those genes in our dataset that had one or more probes with p O < 9.42 × 10 −8 . It also summarizes matching transcript correlation with methylation probes in or near these genes that satisfied a less stringent criterion of p O < 5 × 10 −7 .   Table 2 associated with gene regions of multiple genes and/or genes from Table 2, regions of which were associated with multiple agents with p FDR < 0.1. Drug names are provided in blue font, whereas the names of associated genes are provided in italic black font. Circle sizes are proportionate to the number of associated genes. Overlapping regions show genes associated with multiple agents or agents associated with multiple genes. The full list of gene region-agent associations with p FDR < 0.1 including additional single gene-single agent associations is provided in Table 2 Spearman correlations of 1180 methylation probes satisfied p O < 5 × 10 −7 , with |ρ| > 0.57, and 519 gene-drug correlations included genes with probes with p O between 9.42 × 10 −8 and 5 × 10 −7 (data not shown). Some of these less significant results may be of biological interest. For example, methylation of the probe cg17159843 in the TSS200 of ERBB2 was associated with resistance to the BET bromodomain inhibitor, (+)-JQ1 (ρ = 0.5757, p O = 4.25 × 10 −7 ; data not shown). Methylation of cg17159843 was negatively correlated with expression of the ERBB2 transcript NM_001005862 (ρ = − 0.4963, p = 2.25 × 10 −5 ), and increased ERBB2 expression was associated with sensitivity to JQ1 (Pearson r = −0.3112, p = 0.0104). TSS200 and the first exon of ERBB2 were also strongly associated with log(IC50) of JQ1 (ρ = 0.5195 and 0.5345, p O = 3.79 × 10 −6 and 7.84 × 10 −6 , respectively), however, not statistically significant after adjustment for multiple testing of all 108,795 gene regions and 412 agents (p FDR ≥ 0.23). Upregulated expression of ERBB2 was previously associated with acquired multidrug resistance in SCLC [53,54]. HER2-positive (HER2 + ) breast cancer cells were sensitive to JQ1, and BET bromodomain inhibitors may alleviate acquired resistance of HER2 + breast cancer cells to lapatinib [55].
Combinations of BET inhibitors with other agents in SCLC with increased ERBB2 expression may be a potential way to overcome drug resistance. Indeed, addition of the BET inhibitor MK-8628 increased the killing of SCLC lines in combination with etoposide or topotecan in a study of triple drug combinations [38]. Table 2 provides the list of gene regions associated with log(IC50) with p FDR < 0.1 at the epigenome-wide level (with Spearman |ρ| > 0.607). Figure 2 provides a Venn diagram of the subset of these associations with p FDR < 0.1 that involved multiple agents per gene or multiple genes per agent. An expanded list of regions that satisfied a less stringent criterion of p FDR < 0.15 at the epigenome-wide level (with |ρ| ≥ 0.579) is provided in Supplementary Table 5. As discussed above, TREX1 regions were associated with multiple agents. The 3′ UTR of MLPH and CEP350 was also associated with multiple agents with p FDR < 0.1 ( Fig. 3; Supplementary  Figure 1). The 3′UTR of MLPH was associated with the Aurora kinase inhibitor AZD-1152 and the polo-like kinase (PLK) inhibitor BI-2536. The 3′UTR of CEP350 was associated with ABT-348, which inhibits multiple kinases including Aurora kinases [56], and the Aurora kinase inhibitor SNS-314 (ρ ≥ 0.610). At p FDR < 0.15, the 3′ UTR of both MLPH and CEP350 was associated with the PLK-1 inhibitor TAK-960, MLPH was associated  Table 3 based on individual probes, with p FDR < 0.1. Drug names are provided in blue font, whereas the names of associated genes are provided in italic black font. The size of the circles is proportionate to the number of associated genes. Overlapping regions show genes associated with multiple agents or agents associated with multiple genes with the Aurora kinase inhibitor SCH-1473759, and CEP350 with a TAK-960 analog (ρ ≥ 0.587; Supplementary Table 5; Supplementary Figure 1). Association of TREX1, MLPH, and CEP350 methylation with multiple Aurora kinase inhibitors is notable, as Aurora kinase inhibitors were highly effective against SCLC cell lines and induced a partial response in SCLC patients in a clinical study [24,57]. Associations with MLPH and CEP350 may be based on functional interactions, as regulatory roles of Aurora kinases include mitotic regulation of microtubule interactions, mitotic spindle assembly, and centrosome maturation [58,59].
MLPH, melanophilin, has a role as a microtubule plusend tracking protein and interacts with microtubule plus end-binding protein EB1, which is involved in spindle symmetry in mitosis [60][61][62]. CEP350, centrosomal protein 350, is a putative substrate of Aurora kinases as a part of centrosomal maturation machinery [59]. While the methylation of the 3′UTR regions of MLPH and CEP350 was strongly associated with drug response, its biological consequences require further investigation, as each region was represented by a single probe. Methylation of the 3′ UTR of MLPH was strongly positively associated with MLPH expression (ρ = 0.5396, p O = 2.94 × 10 −6 ). The 3′ UTR of CEP350 was very weakly negatively associated with its expression (ρ = − 0.1887, p O = 0.1227; Supplementary Table 5).

Association of methylation of probes and regions in candidate genes with drug response
Analysis of the association of methylation probes and regions in candidate genes with 44 antitumor agents listed in Supplementary Table 2 revealed 29 probe-drug associations involving 18 genes and 16 agents with p FDR < 0.1 (Spearman |ρ| > 0.516|). The summary of these associations at the gene level, representing 24 gene-drug associations, is provided in Table 3  Associations of the probe cg08937075 in EPAS1, which encodes the hypoxia-inducible factor 2α (HIF2α), with the Aurora kinase inhibitors AMG-900 and SNS-314 and of cg09637363 in CCND1 with the Aurora A kinase/tyrosine kinase inhibitor ENMD-2076 had p FDR < 0.05 (Table 3; Supplementary Table 7). Multiple other probes in both genes had strong correlations with these agents (Supplementary Table 7). The probe cg08937075 in the EPAS1 body was significantly associated with response to AMG-900 not only when testing among candidate genes, but also at the epigenome-wide level among all Illumina EPIC array probes (p O = 1.68 × 10 −8 , ρ = 0.6278; Supplementary Table 3; Supplementary Figure 2). In total, 13 probes in the body of EPAS1, upstream of that gene, or in the 3′UTR had modest or strong correlations with resistance to AMG-900 (Supplementary Data 2D). Among candidate drug-gene associations, EPAS1 probes had FDR < 0.1 in associations with six agents (Table 3, Fig. 3), including resistance to Aurora kinase inhibitors AMG-900, SNS-314, and MLN-8237; multiple kinase inhibitor including Aurora kinase ABT-348, and the PLK-1 inhibitor GSK-461364 (0.5165 ≤ ρ ≤ 0.6278), and sensitivity to the Bcl-2 inhibitor ABT-737 (ρ = − 0.5204). EPAS1 gene regions were not significantly associated with drug response (the strongest correlation, with etoposide, had p FDR = 0.509, ρ = 0.4122; data not shown). Methylation of EPAS1 probes which were associated with drug response was significantly positively correlated with EPAS1 expression (ρ = 0.6800, p O = 3.39 × 10 −10 for cg08937075), and both methylation of EPAS1 probes and EPAS1 expression had the same direction of associations with drug response (Table  3; Supplementary Figure 2), which suggests that EPAS1 expression may be regulated by mechanisms other than DNA methylation. While EPAS1 expression is associated with response to the KSP/Eg5 inhibitor ARQ-621 [24], however, the correlation of EPAS1 methylation with response to that agent was weak (ρ ≤ 0.2147 for different regions), was in same direction as that for gene expression, and did not reach statistical significance. We did not find any association of EPAS1 copy number with probe methylation, transcript expression, or response to the Aurora kinase inhibitors AMG-900 and SNS-314 (Spearman ρ and Pearson r < 0.2; data not shown), and the molecular basis for the association between methylation of EPAS1 probes and drug response remains unclear.
CCND1, encoding cyclin D1, is overexpressed in some SCLC tumors [5]. The probes in its upstream region and gene body had the opposite directions of association with drug response and with CCND1 expression (Supplementary Table 7). The probes cg09637363 and cg19964454 in its body were associated with sensitivity to the Aurora kinase inhibitors ENMD-2076, AS-703569, SCH-1473759, CYC-116, and MLN-8237 (− 0.5777 ≤ ρ ≤ − 0.4855, 0.0391 ≤ p FDR ≤ 0.1366; Supplementary Table 7). Methylation of both probes cg09637363 and cg19964454 was strongly positively associated (ρ = 0.5972 and 0.5617, p O = 1.20 × 10 −7 and 9.24 × 10 −7 , respectively; data not shown) with expression of one of the two CCND1 transcripts with available data, NM_053056 (Affymetrix cluster ID 3338192). The significance of the positive correlation of the probes in the gene body with expression is unclear, although it is consistent with a report of multiple positive correlations of some gene body probes with expression in the Cancer Genome Atlas (TCGA) data [42,63]. Expression of the CCND1 transcript 3338192 was modestly associated  In contrast, the second CCND1 transcript (Affymetrix cluster ID 3380065) was not associated with any of the above agents or with methylation of the probes correlated with these agents (Pearson |r| and Spearman |ρ| ≤ 0.2, p ≥ 0.1). Therefore, possible associations between CCND1 methylation and drug response could be mediated by the transcript 3338192. Analysis of the possible effect of CCND1 copy number on these associations did not provide conclusive results as only one cell line with available copy number data, DMS 114, had a high-level amplification of CCND1.
Analysis of association of regions of candidate genes with 44 agents identified five gene-drug pairs involving four genes which had p FDR < 0.1, with Spearman ρ ≥ 0.51 (Table 4). All associations indicated increased drug resistance for higher methylation of the respective gene regions. Two associations, of the TSS1500 of PTGFRN with the dual PI3K/mTOR inhibitor BEZ-235 (dactolisib) and of the KDM1A body with the PLK-1 inhibitor TAK-960, had p FDR < 0.05. Methylation of the regions of three out of four genes (PTGFRN, KDM1A, and MDM2) listed in Table 4 was significantly negatively associated with expression of their transcripts (− 0.4786 ≤ Spearman ρ ≤ − 0.2436, 4.81 × 10 −5 ≤ p O ≤ 0.0487). This may indicate a negative regulatory effect of DNA methylation on transcript expression. Accordingly, increased expression of these three genes was associated with sensitivity to the agents listed in Table 4 (− 0.4017 ≤ Pearson r ≤ − 0.2325, 0.0008 ≤ p O ≤ 0.0752). Supplementary Table 8 provides an expanded list of associations between gene regions and drug response satisfying p FDR < 0.15.
Methylation of the body of KDM1A, which encodes lysine demethylase 1A (LSD1), an epigenetic histone modifier, was significantly associated with resistance to the PLK-1 inhibitor TAK-960 (Spearman ρ = 0.5486, p FDR = 0.0352; Table 4). Using a less stringent threshold (p FDR < 0.15), it was also associated with resistance to another PLK-1 inhibitor GSK-461364, and to the KSP inhibitor SB-743921 (ρ = 0.4619 and 0.4722, respectively; p FDR = 0.1347 for both agents, Supplementary Table 8). Methylation of the KDM1A body was significantly negatively correlated with expression of its transcript with Affymetrix cluster ID 2325002 (NCBI locus ID NM_001009999; Spearman ρ = − 0.2436, p = 0.0487; Table 4), and increased KDM1A transcript expression was weakly correlated with sensitivity to TAK-960 (Pearson r = − 0.2325, p = 0.0583; Table 4). LSD1 is overexpressed in SCLC and is the target of selective LSD1 inhibitors which are currently being pursued in the clinical setting [2,3]. LSD1 directly regulates the transcription of PLK-1 [64], which could suggest a potential molecular mechanism of association between increased methylation and reduced expression of KDM1A and resistance to PLK-1 inhibitors.
All other gene region associations with p FDR < 0.1 involved upstream regions. Methylation of the TSS1500 of PTGFRN was significantly associated with resistance to the dual PI3K/mTOR inhibitor BEZ-235 (ρ = 0.5565, p FDR = 0.0352; Table 4). PTGFRN, a frequently mutated gene in SCLC, encodes the prostaglandin receptor F2 inhibitor that inhibits the binding of prostaglandin F2α to its receptor [5]. In agreement with the correlation of the TSS1500 region, the probe cg08361238 in that region was also associated with BEZ-235 (ρ = 0.5040, p FDR = 0.1184; Supplementary Table 7). Methylation of the TSS1500 of PTGFRN was significantly negatively correlated with expression of the PTGFRN transcript NM_020440 (cluster ID 2353717; ρ = − 0.4786, p = 4.81 × 10 −5 ). Accordingly, increased expression of that transcript was associated with sensitivity to BEZ-235 (Table 4).
Methylation of the TSS200 of MDM2 was associated with resistance to Bcl-2 inhibitors ABT-263 (Navitoclax) and ABT-737, with ρ = 0.52 and p FDR = 0.0617 for both agents (Supplementary Table 8). The probe cg04667586 in the TSS200 and the gene body was also associated with log(IC 50 ) of ABT-263 (ρ = 0.5070, p FDR = 0.1184; Supplementary Table 7). TSS200 was significantly negatively associated with expression of the MDM2 transcript NM_002392 (cluster ID 3421300, r = − 0.4269, p = 0.0004), which was correlated with sensitivity to both Bcl-2 inhibitors (ρ = − 0.4017, p = 0.0008 for ABT-263 and r = − 0.3867, p = 0.0012 for ABT-737). While MDM2 inhibitors alone were not effective in the in vitro screens of SCLC cell lines [24,38] and TP53 is nearly universally inactivated in SCLC [3,5], surprisingly, the combination of the MDM2 inhibitor JNJ-27291199 with etoposide and carboplatin resulted in enhanced cytotoxicity against SCLC cell lines [38].   Methylation of the TSS1500 of IGFBP5, which encodes an endogenous IGF1R inhibitor and is expressed at low levels in the POU2F3-regulated tuft cell-like SCLC subtype and at high levels in the subtype with high ASCL1 expression [12,13], was associated with resistance to the mTOR inhibitor INK-128 (sapanisertib; ρ = 0.5107, p FDR = 0.0896; Table 4). This association may not be mediated by IGFBP5 transcription, as methylation of the TSS1500 region of IGFBP5 was weakly positively associated with expression of the IGFBP5 transcript NM_ 000599 (cluster ID 2598828), which was not significant (ρ = 0.2035, p = 0.1013; Table 4). Using less stringent criterion of p FDR < 0.15, methylation of IGFBP5 was also associated with resistance to another mTOR inhibitor, BEZ-235, and with Aurora kinase inhibitors ENMD-2076, SNS-314, and MLN-8237 (0.4549 ≤ ρ ≤ 0.4881, Supplementary Table 8). Neither individual probes nor gene regions of either IGF1R or POU2F3, the products of which are involved in regulatory cascades downstream or upstream of IGFBP5 [12,13], were strongly associated with any agents.
The probes cg25627144 and cg18124721 in the TSS200 of the DLL3 (delta-like ligand 3) gene and the entire TSS200 region were associated with the Bcl-2 inhibitor ABT-737 (ρ = 0.5506 and 0.4850, p FDR = 0.0515 and 0.1366 for probes; ρ = 0.4762, p FDR = 0.1260 for TSS200; Table 3, Supplementary Tables 6, 7 and 8), and cg25627144 was also associated with another Bcl-2 inhibitor, ABT-263 (ρ = 0.4926, p FDR = 0.1282; Supplementary Table 7). DLL3, encoding a Notch pathway regulator and a promising clinical target in SCLC treatment, is upregulated in ASCL1-high SCLC tumors; in contrast, DLL3 expression is diminished in RB wild-type tumors and in SCLC tumors not expressing neuroendocrine markers [2,3,39]. Methylation of all DLL3 probes was significantly negatively associated with expression of the DLL3 transcript, NM_016941 (cluster ID 3833122; ρ = − 0.4426, p = 0.0002 for the strongest association). Expression of that transcript was associated with sensitivity to ABT-737 (r = 0.4139, p = 0.0005). Association of lower DLL3 methylation and increased expression with sensitivity to Bcl-2 inhibitors suggest that the use of the Bcl-2 inhibitors in the ASCL1-high SCLC lineage may be considered. While methylation of individual probes and of the upstream region (TSS200 and TSS1500) of ASCL1 was modestly positively correlated with ABT-737, this association was not significant after adjustment for multiple testing (ρ ≤ 0.3687, p O ≥ 0.0023, p FDR ≥ 0.4024 for probes; ρ = 0.2862, p O = 0.0198, p FDR ≥ 0.4024 for the TSS200). Response to ABT-737 was associated with multiple genes (Tables 3 and 4; Supplementary Tables 6, 7 and 8), indicating that it may involve a complex interplay of factors and may not be limited to specific SCLC lineages.
Despite this and other modest non-significant correlations, methylation of master SCLC lineage regulators ASCL1, NEUROD1, and POU2F3 [2] was not significantly associated with any agents after adjustment for multiple testing. Further examination of SCLC lineage drivers [2,3] showed that methylation of the probe cg20782778 in the body of YAP1 (encoding yesassociated protein 1 which regulates the Hippo pathway [2,65]) was associated with resistance to the mTOR inhibitor rapamycin; ρ = 0.4904, p O = 2.91 × 10 −5 , p FDR = 0.1282; Supplementary Tables 6 and 7. Methylation of that probe was negatively correlated with expression of the YAP1 transcript NM_001130145 (cluster ID 3346453; ρ = − 0.4549, p = 0.0001; Supplementary Table  6), which was weakly negatively associated with log(IC50) of rapamycin (r = − 0.2218, p = 0.0712). We observed additional associations in the same direction of the probe cg20782778 in the gene body and of the upstream regions of YAP1 with resistance to other mTOR inhibitors; however, they did not reach statistical significance after adjustment for multiple testing (ρ = 0.4345, p O = 0.0003, p FDR = 0.2215 for cg20782778 and log(IC50) of MK-8669; ρ = 0.4128, p O = 0.0006, p FDR = 0.2253 between TSS1500 and log(IC50) of sapanisertib; data not shown). Our results suggest an association between decreased YAP1 methylation, increased YAP1 transcription, and increased sensitivity to mTOR inhibitors. They are in agreement with a report that the YAP1-high SCLC subtype may be sensitive to mTOR inhibitors [65]. That study also found an association between increased expression of another Hippo pathway regulator, TAZ, and rapamycin sensitivity in SCLC [65]. In our data, associations of TAZ methylation and expression with rapamycin sensitivity were in the same direction as those for YAP1, but weaker. Methylation of TSS200 of TAZ was weakly associated with resistance to rapamycin (ρ = 0.2890, p O = 0.0186), in the same direction as that of YAP1. TAZ expression was weakly associated with log(IC50) of rapamycin in the same direction as that of YAP1 (r = − 0.2813, p = 0.0211).
Amplification and overexpression of MYC, which encodes c-Myc, has been associated with SCLC sensitivity to Aurora kinase inhibitors and with a loss of neuroendocrine markers and the ASCL1-negative and NEUROD1-positive lineage of SCLC [2,5,8,10], including an inverse correlation between MYC and ASCL1 expression in this dataset [24]. In contrast to MYC, expression of MYCL1 is elevated in ASCL1-high SCLC tumors [2]. In our dataset, multiple probes in the bodies of MYC and MYCL1 and in the upstream region and the body of MYCN showed modest associations (0.4060 ≤ |ρ| ≤ 0.5432) with multiple agents. While some of these correlations were not significant after adjustment for multiple testing (8.68 × 10 −5 ≤ p O ≤ 0.0006, p FDR ≥ 0.1657; data not shown), MYC probes cg08526705 and cg00163372 were associated with resistance to the PLK-1 inhibitor TAK-960 (statistically significant for cg08526705 in the body of MYC and TAK-960, ρ = 0.5432, p O = 2.45 × 10 −5 , p FDR = 0.0540; Table 3, Supplementary Tables 6 and 7) and to Aurora kinase inhibitors SCH-1473759 and GSK-461364, consistent with earlier studies of SCLC and with interaction of Aurora kinase A with c-Myc [8,10,58]. In MYCL1, cg00295382 was associated with TAK-960, and in MYCN, cg06520300 and cg17360299 were associated with resistance to the Wnt inhibitor salinomycin (VS-507), and cg04609952 was associated with sensitivity to the mTOR inhibitor MK-8669. In agreement with individual probes, methylation of the 5′UTR region of MYCN was correlated with resistance to VS-507 (salinomycin; ρ = 0.4552, p O = 0.0001, p FDR = 0.1406; Supplementary Table 8). MYC, MYCL1, and MYCN are recurrently amplified in SCLC [8,66], which results in their overexpression and could potentially increase their methylation measures. Consistent with the effect of MYC amplification, all MYC probes associated with drug response were positively correlated with expression of the MYC transcript NM_002467, including some statistically significant associations (ρ = 0.2652, p = 0.0314 for cg00163372; Table  3). Intriguingly, the probe cg00295382 of MYCL1 and the 5′UTR of MYCN were significantly negatively associated with gene expression (ρ = − 0.6744, p = 5.37 × 10 −10 for cg00295382 of MYCL1 and ρ = − 0.4603, p = 0.0001 for the 5′UTR of MYCN). Accordingly, expression of the MYCN transcript NM_005378 (Affymetrix cluster ID 3349293) was associated with sensitivity to salinomycin (ρ = − 0.2376, p = 0.0529; Supplementary  Table 8), in the opposite direction from the 5′UTR methylation. The 5′UTR of MYCN contains important regulatory elements for promoter activity in neuroblastoma [67]. This suggests a possibility that while MYCN amplification may have the main effect on MYCN overexpression, in some cases methylation of its 5′UTR may affect its expression in SCLC and influence the response to salinomycin.
In recent studies, gene and protein expression of SLFN11 have emerged as significant predictors of SCLC response to DNA-damaging agents including topoisomerase I and II inhibitors, PARP inhibitors, and platinum compounds [3,23,24,38,[68][69][70][71]. The SLFN11 protein is epigenetically silenced by EZH2 via H3K27me3 histone methylation, leading to resistance to a combined therapy of etoposide and a platinum compound [23]. While association of SLFN11 expression with sensitivity to DNA-damaging agents in this dataset was highly significant [24,38], correlations of methylation of SLFN11 probes and regions with drug response had p FDR > 0.1. The strongest associations of SLFN11 probes and regions, with p FDR < 0.15, are listed in Supplementary Tables 6, 7 and 8. These include correlations of methylation of the gene body with sensitivity to BMN-673 (talazoparib), teniposide, and topotecan (− 0.4871 ≤ ρ ≤ − 0.4617, 0.1163 ≤ p FDR ≤ 0.1347; Supplementary Table 8). In addition, methylation of the upstream regions and of the body of SLFN11 was associated with a variety of topoisomerase inhibitors and PARP inhibitors with p FDR > 0.15, including correlations of topotecan, teniposide, and talazoparib with the 3′UTR, and correlations of the 3′UTR and gene body with etoposide, MK-4827, olaparib, and PF-01367338 or rucaparib (− 0.4436 ≤ ρ ≤ − 0.4110, p FDR ≥ 0.1731).
Interestingly, methylation of the probes in the upstream regions was associated with drug resistance, whereas probes in the body and the 3′UTR of SLFN11 were associated with sensitivity to DNA damaging agents. For example, cg18108623 in the TSS1500 had ρ = 0.4989 and 0.4811 for teniposide and topotecan, respectively, whereas cg18124488 in the gene body had ρ = − 0.4871 with topotecan (0.1251 ≤ p FDR ≤ 0.1484; Supplementary Table 7). We also observed many modest non-significant associations of methylation of probes in the TSS1500 and TSS200 upstream regions with resistance, and of probes in the gene body and 3′UTR with sensitivity to teniposide, talazoparib, etoposide, topotecan, olaparib, MK-4827, and PF-01367338 (− 0.4638 ≤ ρ ≤ − 0.4002 for probes in the gene body and the 3′UTR, 0.4079 ≤ ρ ≤ 0.4679 for probes in the 5′UTR; p FDR ≥ 0.1657; data not shown). Despite multiple associations of probes in the body and in the 3′UTR with sensitivity to DNA-damaging agents, methylation of the gene body was very strongly positively correlated with expression of the SLFN11 transcript NM_001104587 (cluster ID 3753500, ρ = 0.7475, p = 5.61 × 10 −13 ). In contrast, methylation of the probes cg03668718, cg18108623, and cg18608369 in the TSS1500 and TSS200 was not only associated with resistance to DNA-damaging agents, but it was also strongly negatively associated with transcript expression (− 0.7134 ≤ ρ ≤ − 0.6129, 1.  Figure 3G). Our results suggest that methylation of the SLFN11 upstream regions but not of its gene body or the 3′UTR may downregulate its expression, increasing drug resistance. They are consistent with our earlier analysis of NCI-60 cell lines, which did not include SCLC, that showed the effect of methylation of these and additional SLFN11 probes in the CpG island in the upstream promoter region on resistance to platinum drugs [20].
BCL2 is one of the most highly differentially methylated genes in SCLC, with variable methylation and expression among SCLC subtypes; by contrast, it is epigenetically silenced in normal lung tissue [11]. Within its gene body, several probes showed different directions of association with response to Bcl inhibitors GX15-070 and ABT-737 and the Aurora A kinase/tyrosine kinase inhibitor ENMD-2076 in the analysis of candidate genes and preselected agents, and one probe showed a trend for association with the CDK inhibitor R-547 at the epigenome-wide level (Supplementary Data 2D and Supplementary Tables 6 and 7).

Discussion
Our study did not pre-select any DNA methylation probes. We analyzed probes with high-quality measures based on the detection p values and filtered out the probes overlapping with SNPs [26,27]. After filtering, all the remaining probes were included in computation of gene region-averaged methylation values. We reported all top probes and all gene regions that were strongly associated with drug response and satisfied the significance thresholds. Our analyses were performed at the epigenome level and also for candidate genes and drug agents relevant to SCLC pathogenesis and potential treatment. Alternative approaches previously employed in several analyses of other cancer cell line methylome datasets included prioritization and selection of probes based on their proximity to the location of genes, promoter regions, and CpG islands, and/or association with transcript expression, and the focus on gene-averaged or promoter region-averaged methylation measures [19,44]. The application of different analytical approaches to large epigenome datasets may help find novel targets through different prioritization and ranking of the significant results, and it may help provide additional ways for validation of the findings. Our analysis of TREX1 showed that while our novel initial findings of association of gene expression and drug response with DNA methylation measures were based on all probes and gene regions without prior pre-selection, these associations were validated by the correlations of the CCLE and GDSC data which were based on promoter-enriched methylation measures.
Our study focused on significant associations of DNA methylation-based measures in probes and gene regions with drug response. We further explored whether such associations could be explained by the strength and the direction of the correlations of DNA methylation with gene expression. The genes most likely to have the direct functional effect of DNA methylation on treatment response were those genes that had methylation data strongly associated with chemosensitivity and also had negative correlations between methylation and transcript expression, accompanied by significant correlations between expression and drug response in the opposite direction from methylation. For those genes that had a strong positive correlation between DNA methylation and gene expression, further investigations are needed, including the detailed analysis of their possible copy number gain or loss, which would similarly affect both DNA methylation and expression measures, and/or the transcriptional regulation of gene expression via mechanisms other than DNA methylation. Previous complementary approaches to the studies of drug response have prioritized associations based on the strength of correlation between gene expression with drug response [38] or used the correlation between methylation and gene expression as selection criteria for methylation measures [19]. The availability of both expression and methylation data, along with drug response measures for this rich dataset, provides opportunities for further integrative analyses. In a parallel effort, we have been developing a public online resource, SCLC-CellMiner [72] which  [43,44]. Red vertical arrows highlight the TREX1 DNA methylation and gene expression values for the SCLC cell lines integrates various molecular measurements for SCLC cell lines including whole exome sequencing, gene expression, miRNA expression, methylation, and copy number data, along with drug response information presented as drug activity. This resource will allow users to query and visualize the correlations for genes of interest.
Our association analysis of DNA methylation with cell line response to chemotherapy agents elucidated very strong associations of TREX1 methylation and expression with SCLC sensitivity to Aurora kinase inhibitors, antimitotic agents, a CDK inhibitor, and an ATR inhibitor. DNase III, which is encoded by TREX1, generates single-stranded DNA (ssDNA) during resolution of chromatin bridges under telomere crisis in malignant cells, and this TREX1 activity has been proposed to play a role in chromothripsis [73,74]. TREX1 is also involved in the maintenance of immune tolerance to self-DNA of the cytosolic DNA-sensing pathway, acting in an opposing manner to the STING regulator [75]. The absence of TREX1 or inactivating mutations in TREX1 result in the upregulation of the cGAS-STING-mediated type I interferon response and systemic inflammation [76]. These roles of TREX1 may provide an explanation for the association of TREX1 methylation and expression with response to Aurora kinase inhibitors, agents targeting microtubule assembly, and with a CDK inhibitor, as these agents interfere with different aspects of cell cycle, mitosis, and cell proliferation. Association of TREX1 methylation with SCLC response to the ATR inhibitor Cpd45 may be consistent with the previously reported TREX1 upregulation in cancer cells in response to DNA damaging agents and to UV light, and roles in sensitivity and response to DNA-damaging agents, DNA repair, and/or DNA degradation after drug treatment [50][51][52], including degradation of ssDNA fragments by TREX1 and TREX1 interaction with PARP1 (poly(ADP-ribose) polymerase-1) in response to DNA damage [77,78].
SCLC cell lines had among the highest TREX1 methylation and among the lowest TREX1 expression in the CCLE (Fig. 4), suggesting the possibility that low TREX1 in SCLC may lead to vulnerability, increasing SCLC susceptibility to treatment by Aurora kinase inhibitors. Therefore, TREX1 may represent a novel molecular marker or target in SCLC.
Our analysis of DNA methylation measures and drug response utilized SCLC cell line data. Previous studies have shown that SCLC cell lines have higher overall DNA methylation than do primary tumors [11]. The absence of primary tumor samples is a limitation of our study. Despite reported methylation differences between the cell lines and primary tumor data and the lack of patient specimens in our analysis, associations identified in our study using cell line data, including those for TREX1, suggest novel therapeutic targets for potential clinical use. A broader intriguing question is whether TREX1 would be a useful molecular target for treatment of other cancers. A combined analysis of breast cancer patients in the METABRIC (Molecular Taxonomy of Breast Cancer International Consortium) cohort suggested a lower survival probability in patients with diminished TREX1 expression [77]. However, the differences in patient survival in that combined dataset could also be attributed to other molecular differences among various breast cancer subtypes. By contrast, analysis of a BRCA1-deficient, p53-deficient genetically engineered mouse model of triple negative breast cancer (TNBC) and analysis of TNBC cell lines showed that activation of the cGAS/STING pathway within the tumor increased the efficacy of PARP inhibitors through tumor cell response and increased immune activation in the tumor microenvironment, with more profound benefits in homologous recombination repair-deficient tumor cells [79]. Similarly, the STING pathway is activated in SCLC tumor cells in response to PARP inhibition and CHK1 inhibition [80]. These observations provide an additional argument that TREX1, an antagonist of the STING-activated molecular pathway, may present an attractive target in SCLC treatment and possibly in other cancer categories.
Lineage-specific differentiation of SCLC tumors based on expression of several master regulators [2] has previously been associated with distinct methylation patterns [11,15]. While many associations of SCLC methylation patterns with drug response were not specific to particular SCLC lineages in our dataset, several lineage-specific associations, e.g., that of YAP1 methylation and expression with rapamycin confirming the results from an earlier study [65], and methylation and expression of DLL3, which is upregulated in ASCL1-high tumors, with Bcl-2 inhibitors were identified. Among non-neuroendocrine lineage markers, methylation of EPHA2 was associated with Aurora kinase inhibitors and a PLK-1 inhibitor, whereas that of CD151 was associated with Bcl-2 inhibitors.
We did not observe any strong association pattern between methylation and expression of TREX1 regions with SCLC lineage categories. Methylation of TREX1 TSS1500 and TSS200 regions was weakly associated with ASCL1 expression (Pearson r = − 0.309, p O = 0.0117 for TSS200), whereas TREX1 expression was weakly associated with NEUROD1 expression (r = − 0.368, p O = 0.02953). Interestingly, whereas c-Myc directly upregulates expression of Aurora kinase genes and promotes a neuroendocrine-low subtype of SCLC which is sensitive to Aurora kinase inhibitors [10,57], the expression of TREX1 was not associated with transcriptional levels of either MYC, MYCL, or MYCN (|r| < 0.13), even though methylation of the 5′UTR and the first exon of TREX1 was weakly associated with MYC expression (r = 0.328 and 0.326, p O = 0.0076 and 0.0072, respectively; data not shown).
DNA methylation in SCLC has been reported to be strongly associated with EZH2 expression [11]. Decreased methylation of the 5′UTR of EZH2 and lower EZH2 expression were more common in SCLC cell lines from subtypes other than SCLC-A (Supplementary Figure 4C and  D). Considering that no other EZH2 regions were associated with drug response or with SCLC lineage differentiation, and that methylation of the 5′UTR and EZH2 expression were positively correlated, interpretation of the associations of the 5′UTR requires further biological investigation. We observed correlations between EZH2 expression and expression of YAP1 (transcript cluster ID 3346453; r = − 0.397, p O = 0.0010), MYCN (2470838; r = 0.328, p O = 0.0071) and, most strongly, MYC (3115504; r = − 0.512, p O = 1.098 × 10 −5 ), which suggest a potential mechanism linking expression of SCLC master lineage regulators and epigenetic pathways of lineage differentiation; however, these correlations do not provide evidence that this mechanism could involve the methylation of the EZH2 promoter region.
Consistent with earlier analyses of the NCI-60 cell lines [19,20] and with growing evidence for the role of epigenetic mechanisms in SLFN11 inactivation [21,69,81], we found that increased methylation of the SLFN11 promoter was negatively correlated with SLFN11 expression and increased resistance of SCLC lines to DNA-damaging agents (Supplementary Figure  3). Association of SCLC promoter methylation with gene expression and with drug sensitivity are in agreement with the findings of He et al. [82], who showed that increased methylation of the SLFN11 promoter in colorectal cancer cell lines and primary tumors led to downregulation of SLFN11 expression, increased tumor resistance to cisplatin, and decreased the 5-year overall patient survival and 5-year progression-free survival.
Our study utilized a large comprehensive dataset of methylation measures derived from the Illumina Infinium MethylationEPIC BeadChip array, which substantially increased the density of available methylation measures for SCLC compared with data from earlier SCLC studies which utilized lower density measures [11,15]. We analyzed a large set of drug screening measures and genome-wide transcript and miRNA expression measures. We conducted a comprehensive analysis of association of pre-treatment SCLC methylation and expression with response to multiple approved and investigational chemotherapy agents. We focused on the gene targets identified by strong associations of methylation signals with drug response as the primary biological leads.
Our findings of methylation signals correlated with drug response confirmed multiple previously identified SCLC associations with drug response from gene expression studies, e.g., those for SLFN11 and YAP1. Several genes, e.g., EPAS1 (HIF2A), EZH2, and BCL2, that were previously suggested to be involved in SCLC drug response based on their expression, protein activity, drug action, and/or distinct methylation patterns in SCLC [11,23,24,83,84] showed complex patterns of associations between methylation and expression (EPAS1 and EZH2) or between methylation and drug response (BCL2). This suggests that the expression and activity of their products may be affected by other factors in addition to DNA methylation. We observed a number of correlations which involved candidate therapy agents that had been previously identified in earlier studies which used in vitro data, animal models, and clinical data. We report multiple associations between such agents and genetic methylation markers, which could potentially be utilized to stratify SCLC patients for therapy response. We also report multiple novel associations of gene methylation with drug responses. These results suggest new therapeutic targets and drug combinations for SCLC treatment which may be beneficial for specific epigenetic tumor landscapes. Some of the newly discovered associations, e.g., those for TREX1, may suggest potential novel direct drug targets in tumors with residual TREX1 expression, and also add to the accumulating body of evidence about the importance of upregulation of the cGAS/STING pathway in treatment of cancer patients [79,80,85]. Similarly, other associations reported in our study may suggest novel therapy targets, indicate potential broader molecular pathways of interest, and provide biomarkers for patient stratification. Due to the dynamic nature of DNA methylation and expression, and rapid acquisition of chemoresistance by SCLC cells in response to treatment, it would be important to continue future investigations of drug response by comparing pre-and post-treatment DNA methylation and expression levels and to further analyze the dynamic nature of DNA methylation, expression, alternative splicing, and post-transcriptional modifications that may drive the chemoresistance of tumor cells.

Conclusions
We completed a comprehensive analysis of a large SCLC DNA methylation dataset and examined associations of de novo DNA methylation with response to 526 chemotherapeutic agents. Our analysis of methylation data confirmed previously known association results between gene expression and drug response for several known genes, e.g., between SLFN11 and DNA damaging agents and between YAP1 and rapamycin. We discovered multiple novel associations which suggest potential targets for single agent treatment and drug combination therapies. Methylation and expression of TREX1 were associated with response to multiple drug categories, which may suggest a possible mechanism of vulnerability to Aurora kinase inhibitors for SCLC, which has low TREX1 expression compared to other cancer categories. These associations also suggest that targeting TREX1 may provide important therapeutic benefits in those SCLC tumors which have residual TREX1 expression. expression data, computation of region-specific methylation levels, and analyzed statistical correlations between SCLC probe and gene methylation, expression, and drug sensitivity. SulV performed bioinformatic analysis of distribution of DNA methylation values and carried out online database analysis of gene synonyms for integration of methylation, gene expression, and drug response data. DS participated in bioinformatic analysis of copy number variation and expression data of top gene hits. MK performed bioinformatic data integration with the NCI Small Cell Lung Cancer Project. JK and BAT selected molecular components of drug target pathways and drafted the manuscript. JK, DS, WCR, YP, and BAT participated in the biological interpretation of the study results. TS carried out the processing and QC of drug screening data and their integration into this study. DE oversaw the drug screening effort. JC performed DNA extraction and QC. AR oversaw the experimental effort of SCLC cell line growth and sample preparation for expression and methylation analyses and drug screen. SudV, PSM, WCR, and YP participated in generation, QC, and normalization of methylation data and in validation of the study findings. All authors edited the manuscript and read and approved its final version.