The NCI-60 Methylome and Its Integration into CellMiner

DNA methylation is a heritable epigenetic event occurring at cytosines 5′ of guanosines (CpG) and catalyzed by DNA methyltransferases (DNMT), which transfer a methyl group from S-adenosyl methionine to the 5 position of the cytosine ring (1). DNA methylation is involved in multiple epigenetic processes ranging from transcriptional downregulation, X chromosome inactivation, embryonic development, and genomic imprinting (2, 3). “CpG islands” in the 5′ regulatory regions of many genes (∼56%), are involved in transcription downregulation and histone deacetylase recruitment (4, 5). The pattern of methylation has both allele-specific and evolutionary conservation, as well as tissue-specific and cell-specific variations (6–8).

DNA methylation defects have been associated with multiple diseases, including (i) global methylation defects for diabetes, obesity, fetal alcohol syndrome, and aging, (ii) imprinting disorders for Angelman syndrome, Prader–Wili syndrome, Beckwith–Wiedemann syndrome and autism, (iii) genetically driven methylation defects for Fragile X syndrome, dyslexia, Rett syndrome, centromere instability, Sotos, Weaver, and Kleefstra syndomes, and (iv) candidate gene methylation defects for obesity and type I and II diabetes (9, 10).

DNA methylation defects are also linked to cancers (9). Global loss of DNA methylation has been associated with genomic instability, loss of imprinting, and reactivation of transposable elements (11). Loss of genomic imprinting of IGF2 has been associated with increased risk of liver, lung, intestinal, and colon cancers (3). Concurrent with global hypomethylation, specific hypermethylation of tumor suppressor genes, including CDKN2A, MLH1, VHL, and CDH1, occurs, leading to their transcriptional repression (3, 9).

The NCI-60 was the first cancer cell line database established and it remains the largest drug and most complete source of molecular data (12, 13). In the current study, we provide whole genome methylation levels from 485,577 probes across the NCI-60 cancer cell line panel. We detail probes that are associated with genes, and provide easy access to them using our CellMiner\Cell line signature web-based application (14). This allows direct comparison and integration with other molecular and activity data, examples of which are included. We also provide a novel form of visualization of the level of influence on gene transcript levels of DNA methylation and copy number, and examples of the relevance of DNA methylation for predicting the activity of DNA-targeted agents.

We combined data for the NCI-60 cell lines from both the National Cancer Institute (NCI) and the Cancer Epigenetics and Biology Program (CEBP; ref.15). Both the NCI and CEBP obtained the cell lines from the Developmental Therapeutics Program (DTP; refs.12, 16, 17). For the NCI dataset, cells were grown and DNA isolated as described previously (18, 19). For the CEBP dataset, cells were grown and DNA isolated as described previously (15).

NCI bisulfite conversion and DNA sample handling were done as described previously (20). The Infinium HumanMethylation450k BeadChip Kit (Illumina, Inc, catalog #WG-314-1001) was used with standard protocol for both the CEBP and NCI studies (20, 21). CEBP bisulfite conversion and DNA sample handling was done as described previously (15).

NCI samples used P values of detection as filters for each probe. Probes with values of P > 0.01 were treated as missing. For dataset comparison, we did average linkage hierarchical clustering of all cell line probe intensity profiles using 1−Pearson correlation distance. This and all subsequent statistical analysis was done in the R statistical environment (22). We used the probe-wise average of the cell line replicates from the two datasets for all subsequent analysis.

Probe beta values by cell line are calculated as:

(Intensity of the methylated probe)/[(Intensity of the methylated probe) + (Intensity of the unmethylated probe)]

Probes were assessed for location with respect to genes and proximal CpG islands defined by Illumina. Probes were designated as category 1 or 2 with category 1 considered to be most informative. For subsequent comparisons, category 1 probes were used if available, and if not then category 2. In either case, gene methylation values resulted from averages of those probes. For genes with multiple transcriptional start sites, the transcript with the most negative correlation to the methylation probes was identified and used.

For comparisons between methylation levels and transcript expression z scores, we used the methylation and transcript “Cell line signatures” (16). Expression versus methylation correlations, heatmaps, and histograms were generated using The R Project for Statistical Computing (22). For the transcript versus methylation analyses by gene category, genes were divided on the basis of correlation value of r ≤ −0.5. Enrichment of Gene Ontology Consortium classifications were accessed using GoMiner with an FDR cut-off value of <0.05 and a minimum 10 identified genes per category (23, 24).

DNA copy number patterns were determined as described previously (25). Their gene “Cell line signatures” can be accessed at CellMiner by gene using “CellMiner\NCI-60 Analysis Tools\Cell line signature” (16).

For each of the 15,798 genes with all three forms of data available (transcript, methylation, and copy number levels) a linear regression model was fit, with both copy number and methylation as independent variables and transcript expression as the dependent variable. The model provided coefficients for the copy number and methylation that gave the lowest squared error between fitted values and true expression. We separated individual contributions of these two factors for gene expression prediction using the method of relative importance (26), using the lmg method (27) from the R package relaimpo to compute individual R² values. Total (or combined) R² is the summation of these two. Square roots of the R² values were multiplied by the sign of the coefficients of the factors in the combined model to get the value of R.

Genetic variants were assessed using the “CellMiner\NCI-60 Analysis Tools\Cell line signature\Genetic variant summation” tool (12, 13, 16). The form of the data used for SMARCB1 in the linear regression analysis is the “Amino acid changing.” The form of the data used for MLH1, MSH2, and MSH6 is the “Protein function affecting.”

The box and whisker plots and linear regression analysis were done using The R Project for Statistical Computing (22). For the linear regression analysis, the form of KDM2B, CREBBP, and SMARCB1 “Genetic variant summation” data used is the “Amino acid changing.” Protein expression for PARP1 and EP300 were accessed using “CellMiner\NCI-60 Analysis Tools\Cell line signature\Protein mean values” (14, 16).

Correlations between median methylation values and cytogenetic measurements of instability were determined using “Pattern comparison,” accessed at “CellMiner\NCI-60 Analysis Tools\Pattern comparison,” within “Miscellaneous phenotypic parameters.” Drug “Cell line signatures” can be accessed at CellMiner by National Service Center (NSC) number using “CellMiner\NCI-60 Analysis Tools\Cell line signature\Drug activity z scores” (14, 28).

A comparison of the two NCI-60 whole genome DNA methylation datasets independently prepared at both the NCI and the CEBP (Barcelona, Spain) was done (15). Cross correlations of probes-specific methylation levels for matched cell lines in the two datasets yielded very high correlations for all cell lines, with a range between 0.919 and 0.995, and an average of 0.978. Three representative cell lines probe set comparisons are shown in Fig. 1A. Also, clustering analyses of all cell lines across the two datasets (NCI vs. CEBP) showed that in all cases, the nearest neighbor cell line was the duplicate cell line from the independent dataset (Fig. 1B). Only two of the cell lines ME:MDA-N and RE:CAKI-1 did not have duplicates in the CEBP analysis. Together, these analyses demonstrated the consistency of the NCI-60 methylome data. Because of their high concordance, the NCI and CEBP NCI-60 whole genome DNA methylation datasets were merged by taking the matched cell line average methylation. This form of the data was used for the remaining analyses in this article, as well as for uploading the data to CellMiner.

In the NCI dataset analysis, independent from the CEBP team (15), methylation probes that were gene-specific were selected. To this end, the 485,577 methylation probes on the Infinium HumanMethylation450k BeadChip array were assessed on the basis of location with respect to CpG islands and genes (Supplementary Fig. S1A–S1C) and their relationship with gene expression (Supplementary Fig. S1D). This resulted in a 2-tier probe categorization (Supplementary Fig. S1C) yielding DNA methylation signatures for 17,552 genes. The complete gene-probe information is included in Supplementary Table S1. For the purpose of comparing gene transcript levels with methylation, category-1 probes were the most informative. Category-2 probes were found to be less informative (less well correlated), but still relevant for some genes. As a result, the NCI-60 methylation data for 17,552 genes may be directly retrieved from the CellMiner website (updated to version 2.0) in several forms as described previously (14, 16, 28). A link providing online instructions for generating DNA methylation Cell line signatures is available (29). The graphical output “Cell line signatures” for three examples, VHL, SLFN11, and IRF6, are presented in Fig. 2. 

The distribution of global methylation levels for 485,577 probes showed marked differences among individual cell lines (Fig. 3A). Median values ranged from 17% for melanoma MALME-3M to 84% for colon carcinoma HCT-116. To determine genes that are major contributors in such global differences, the global methylation pattern was compared with the epigenetic, chromatin, and histone functional groups (defined in Supplementary Table S2) for: (i) amino acid changing genetic variants, (ii) protein function affecting genetic variants, (iii) gene transcript levels, and (iv) protein levels, using the “Pattern comparison” web application (14, 28).

Significant correlations (P < 0.01) were found for 24 genes, including 5 with literature connection to DNA methylation: UHRF1, MTA1, HIST1H1A, PARP1, and EP300. However, none of the 24 significant correlations found could individually predict more than 21% of the median methylation pattern, implying multivariate causation. Using linear regression, it was found that all of these except HIST1H1A made significant contribution to modeling a relationship to the median methylation pattern. After removing HIST1H1A, four other genes (KDM5C, KDM2B, CREBBP, and SMARCB1) were added back, one at a time, based on their appearing in both the epigenetic and chromatin functional categories (Supplementary Table S2). Of these, KDM5C and SMARCB1 were found to contribute to the model significantly (P = 0.00046) by comparison to the four-gene model. The resulting scatter plot of true versus fitted values in Fig. 3B showed an r² of 0.62. The six-gene model for prediction of global methylation status included PARP1 (P = 3.7 × 10⁻⁶, r = −0.062), MTA1 (P = 5.9 × 10⁻⁶, r = 0.066), EP300 (P = 1.1 × 10⁻³, r = −0.030), KDM5C (P = 1.2 × 10⁻³, r = −0.051), SMARCB1 (P = 0.017, r = −0.0049), and UHRF1 (P = 0.041, r = 0.021).

Pattern comparison using the median probe methylation values (Fig. 3A) identified significant negative correlation to six parameters of genomic instability from cytogenetics (30). These were modal chromosomal number, numerical complexity, structural heterogeneity, fraction of abnormal chromosomes that experience numerical heterogeneity, fraction of normal chromosomes that experience numerical heterogeneity, and numerical heterogeneity, with correlations of −0.341, −0.362, −0.362, −0.377, −0.389, and −0.425 respectively. These correlations indicate that as the levels of DNA methylation decrease, the levels of genomic instability increase.

Comparison of gene methylation to transcript expression levels across the NCI-60 identified 44 GO categories enriched for genes with significant methylation versus expression correlations (see rows in Fig. 4A; refs. 23, 24). As these 44 categories had overlapping genes, they were organized into the seven groups shown in Fig. 4A (right), including: metabolic processes, blood coagulation, cell migration and mobility, cell adhesion and assembly, white blood cell proliferation, activation of immune response, cell death, and signaling.

Curated gene lists generated from literature for different pathways related to cancer [including epithelial–mesenchymal transition (EMT), tumor suppressors, oncogenes, apoptosis, DNA repair chromatin and mitochondria; see Supplementary Table S2] were also tested. For all the genes in each gene category, the correlation between transcript and methylation levels was computed. The data presented in Fig. 4B show the median Pearson correlation for each curated gene category (vertical lines), as well as for the GO categories shown in Fig. 4A (red font). The histogram of correlation distribution for all the 17,144 transcripts with expression and DNA methylation data is shown as the shaded area (see Supplementary Table S2 for individual values).

The epithelial and mesenchymal gene categories, consisting of 25 and 27 genes, respectively, were assembled previously (31). These categories showed the most significant correlation between DNA methylation and transcript levels, with medians of −0.639 and −0.525 for the epithelial and mesenchymal genes, respectively. A wide gap was found prior to the next groups of genes with significant correlation. Those included the genes found by GO analysis (see Fig. 4A), and additional categories, including tumor suppressor genes, which were tightly grouped as a cluster of 23 functional groups, within a range of −0.259 to −0.148. In contrast, the median value of miRNAs showed a lack of significant correlation to expression (r = 0.010). Together, these analyses demonstrate that DNA methylation drives gene expression for selected pathways, such as the EMT, and to a lesser extend for tumor suppressors in the NCI-60.

Supplementary Table S2 contains the Pearson correlations between DNA methylation and transcript expression for 17,144 transcripts, including 16,155 genes, 494 open reading frames, 167 loci, 256 miRNAs, 67 long intergenic non-protein–coding RNA, and five long noncoding RNAs that contain miRNAs in their introns. The ability to retrieve the methylation and transcript expression patterns for individual genes (through CellMiner) enables their comparison. Figure 5 shows representative transcript versus methylation scatter plots constructed from the CellMiner data, including their correlations, for genes from the functional categories listed in Fig. 4B.

The expression of the mesenchymal gene vimentin (VIM) was significantly driven by promoter methylation, as were 4 epithelial genes (claudins 7 and 4, the epithelial splicing regulator protein 2 gene ESRP2, and RAB25, a member of the RAS oncogene family). Among tumor suppressors, examples of highly significant correlations include APC (the adenomatous polyposis coli gene), VHL (the Von Hippel-Lindau gene, inactivated in RE:RXF-393), BRCA1 (inactivated in 3 of the ovarian cell lines: OVCAR-4, OVCAR-8 and NCI/ADR-RES), IRF6 (the interferon regulatory factor 6), and CDKN2A (cyclin-dependent kinase inhibitor 2A). Both IRF6 and CDKN2A are inactivated by promoter methylation in multiple NCI-60 cell lines. For the apoptotic pathway, promoter methylation was found significant for some key genes such as BIM (BCL2L11), BID, the cell surface death receptor (FAS), caspase 8, and heat shock 70 kDa protein 1B (HSPA1B).

Notably, for the genes encoding epigenetic factors, only the expression of TET1, the gene encoding Ten-Eleven Translocation Methylcytosine Dioxygenase 1, a key demethylating enzyme, was found driven by promoter methylation (Fig. 5, 4th row, on left). Examples of DNA damage response genes driven by promoter methylation include MLH3 (MutL homolog 3), HLTF (helicase-like transcription factor/SMARCA3), TDP1 (tyrosyl-DNA phosphodiesterase 1), MGMT (O-6-methylguanine-DNA methyltransferase), and SLFN11 (Schlafen 11). MGMT and SLFN11 (Fig. 5, 5th row, on right) will be discussed below in the context of their pharmacologic relevance (32–34).

The majority of the genes (15,798 genes) can be queried for concurrent analyses of DNA methylation, transcript expression, and DNA copy number in the NCI-60 (Supplementary Table S3; ref.14). In Fig. 6A, “Cell line signatures” obtained directly from CellMiner for DNA methylation, copy number (14, 25), and transcript expression (28), are presented for three exemplary genes for which transcript levels are driven by DNA methylation, copy number of both.

Expression of RAB25, a member of the RAS family involved in membrane trafficking and cell polarity and epithelial phenotype, showed high correlation with methylation (as do other EMT genes, see Figs. 4B and 5, top row), but not DNA copy number (r = −0.978 and 0.069, respectively). In contrast, expression of POLG, a housekeeping gene encoding the mitochondrial replicative DNA polymerase, showed no correlation with methylation but instead a high correlation with copy number (r = 0.042 and 0.758, respectively). The high impact of copy is probably related to the frequent gain or loss of the locus 15q25/26, which also encodes important repair genes including FANCI and BLM as well as IDH2 and CHD2 (readily detected by the “Pattern comparison” tool of CellMiner). Expression of the tumor suppressor gene CDKN2A encoding p16^INK4A and p14^ARF showed high correlation to both DNA copy number and methylation (r = 0.622 and −0.558, respectively). Moreover, the methylation and copy number profiles across the NCI-60 showed remarkable mirror images (Fig. 6A, compare 2nd and 3rd bar graphs from the right), with the same cell lines lacking a gene copy also showing hypermethylation on the other allele and r = 0.849.

Linear regression assessment of influence of both DNA methylation and copy number on transcript expression at the whole-genome scale is shown in Fig. 6B. In the scatter plots of R (the coefficients of determination times the sign of the coefficient) for both DNA copy number (the x-axis), and methylation (the y-axis), values of 1 or −1 indicate perfect prediction in the positive or negative sense, respectively, and 0 no predictive value. The “All genes” plot (Fig. 6B, left) illustrates the cumulative importance of both DNA methylation and copy number, as indicated by the presence of 59% of the points in the bottom right quadrant. These points are both negative for DNA methylation (indicating negative predictive power for transcript expression), and positive for DNA copy number (indicating positive predictive power for transcript expression).

Figure 6B (2nd panel from left) shows that restricting the plot to the epithelial genes (including RAB25, CLDN4, ESRP2, CLDN7; see Figs. 4B and 5) demonstrates that methylation is their predominant regulator. The mitochondrial and tumor suppressor plots indicate more balanced influences from both DNA methylation and copy number for these categories. The R values are precalculated for any gene of interest (with data) in Supplementary Table S3. These analyses demonstrate which genes' transcript levels in what cell lines are driven by copy number and/or methylation levels in particular cell lines.

Figure 6C extends the integrative approach combining DNA copy number, methylation, transcript expression, and mutational data obtained from whole sequencing of the NCI-60 (12, 14) for MLH1, MSH2, and MSH6, to provide the genomic underpinnings for the reported presence of microsatellite instability in 8 of 9 cell lines (12, 35). For MLH1, there is DNA copy number loss in SKOV3, DNA methylation and reduced transcript expression in KM12, reduced transcript expression in IGROV1 and SKOV3, and predicted function affecting amino acid changes in HCT116, HCT15, CCRF-CEM, IGROV1, and DU145. For MSH2, there are predicted function affecting amino acid changes in SK-MEL2, NCI-H522, and DU145. For MSH6, there is reduced transcript expression in IGROV1, and predicted function affecting amino acid changes in IGROV1 and DU145. Only the instability of LE:MOLT-4 remains unexplained. The example of the mismatch repair pathway demonstrates the converging contribution of the four genomic parameters now available in the NCI-60 (methylation, copy number, expression, and deleterious mutations), and the importance of data integration.

MGMT encodes methylguanine methyltransferase, an enzyme that removes O-6-methylguanine, the most cytotoxic DNA methylation adduct produced by temozolomide, a commonly used oral drug in glioblastomas (36). Cancer cells deficient for MGMT are exquisitely sensitive to temozolomide (34) and MGMT promoter methylation is a positive prognostic indicator for temozolomide treatment in glioblastoma (36). The scatter plot of MGMT methylation versus expression levels (see Fig. 5, bottom right) showed significant correlation (r = −0.48). DNA promoter methylation levels above 40% was found associated with MGMT expression levels at background levels (←0.6) for 81% of those cell lines (Fig. 5 and Supplementary Fig. S2B). DNA promoter methylation levels less than 40% are associated with (76%) of expressed cell lines (≥0.6). Promoter hypermethylation leading to transcriptional inactivation extended beyond the 6 glioma (CNS) cell lines, with colon (KM12) and the leukemia (SR) both having high methylation and background expression. However, only 5 of the 21 NCI-60 cell lines with low MGMT expression (≤0.06) have significant methylation (above 50%). Thus, the incomplete association between MGMT expression and methylation parameters imply the use of epigenetic silencing by DNA methylation level as a prognostic indicator for temozolomide treatment is useful, but incomplete.

A second gene, which has recently been causally linked with response to a broad spectrum of DNA-damaging drugs (including topoisomerase inhibitors, PARP inhibitors as single agents or in combination with temozolomide, cisplatin, alkylating agents, and DNA synthesis inhibitors) is SLFN11 (Schlafen 11; refs.15, 32, 33, 37,38). SLFN11 encodes a nuclear protein with putative helicase activity that blocks cell-cycle progression and dampens DNA repair (32, 39). SLFN11 was among the genes with the highest correlation (at the top 94th percentile) between methylation and expression (see Fig. 5, bottom right). Notably, approximately 38% of the NCI-60 cell lines do not express SLFN11 above background level (32). Of these, lack of SLFN11 expression is linked to methylation in approximately half of the cell lines (Fig. 5). Neither SLFN11 or MGMT expression had significant association to DNA copy number (P < 0.01, Supplementary Table S3).

To test whether SLFN11 promoter methylation was linked with resistance to DNA-damaging agents, the whole NCI-60 drug database (≈21,000 compounds including 108 FDA-approved and 70 clinical trial drugs) was tested. Drug activity correlations revealed that SLFN11 methylation was significantly correlated with resistance to multiple clinically relevant drugs that cause DNA damage, including alkylating agents (cisplatin, carboplatin, melphalan), topoisomerase I (topotecan, LMP400) and II (etoposide) inhibitors, DNA synthesis inhibitors (gemcitabine, fludarabine, cytarabine, and hydroxyurea), PARP inhibitors (talazoparib and olaparib), and bleomycin (Table 1). As expected (32, 37), no correlation was observed for tubulin inhibitors (paclitaxel and docetaxel) and protein kinase inhibitors (erlotinib, crizotinib, and vemurafenib), consistent with the selective implication of SLFN11 for cytotoxic response to DNA-damaging drugs.

Here, we provide readily usable and accessible genome-wide data for the NCI-60 cancer cell line methylome based on two independent determinations (Fig. 1). Assignment of salient CpG sites for each gene (Supplementary Table S1) enables the extraction of those data using our CellMiner tools (14, 29). We provide examples of how the DNA methylation data may be integrated with the extensive genomic and pharmacologic databases for the NCI-60 (Figs. 4, 5 and 6). We anticipate that these data and tools will enable exploration of (i) the relevance of DNA methylation as a key regulator of gene expression, (ii) its relationships with other forms of molecular data, and (iii) genomic biomarkers for precision therapeutics. Determination of methylation status is a preferred approach for clinical samples, because it provides robust information with which to evaluate cancer genomes (as it uses DNA rather than RNA, which tends to be unstable).

The examples we provide demonstrate that genome-wide access to DNA methylation gives new insights in tumor biology and regulatory mechanisms for gene expression. Figures 4–6 demonstrate that promoter methylation is a prominent regulator for EMT, a major determinant for tumor invasion and resistance to therapy, which is being targeted by DNA methyl transferase inhibitors (40). On the other hand, for tumor suppressors, both methylation and gene deletion drive gene expression. A salient example is the cyclin-dependent kinase inhibitor 2A gene CDKN2A, which encodes the two major tumor suppressors, p16^INK4A and p14^ARF. As shown in Figs. 5 and 6, approximately 40% of the NCI-60 fail to express CDKN2A. That almost half of cancer cells suppress CDKN2A expression is consistent with the larger MIT-Broad-CCLE dataset where approximately 40% of the 1,000 cell lines only express background (no) CDKN2A (33). In addition, examination of CDKN2A gene copy number shows that 20 of the NCI-60 cell lines have 9p21 deletions (Fig. 6A, right). Furthermore, integrating the NCI-60 data for CDKN2A methylation and gene copy number shows a high correlation between the two genomic parameters (r = −0.85; P = 9.7 × 10⁻¹⁸), This demonstrates that, in the NCI-60, cancer cells commonly inactivate CDKN2A by biallelic loss of CDKN2A (one allele by promoter methylation and the other by 9p21 chromosome deletion).

DNA methylation is also relevant to precision therapeutics, and, in this study, examples of two genes are presented: SLFN11 and MGMT. Both genes encode key factors that determine response to widely used DNA-damaging agents, which remain a major component of the cancer armamentarium but lag behind protein kinase inhibitors in terms of predictive biomarkers. A relationship between SLFN11 transcript levels and pharmacologic response has been demonstrated in both the NCI-60 and CCLE cancer cell lines (15, 32, 33, 41,42). This relationship is both causal and broad in scope, affecting topoisomerase I inhibitors, topoisomerase II inhibitors, alkylating agents, and DNA synthesis inhibitors (32, 37). SLFN11 expression has recently been shown to be driven by ETS transcription factors (43), which explain high SLFN11 expression in Ewing sarcoma (33, 43). Epigenetic inactivation of SLFN11, which accounts at least in part for frequent lack of expression of SLFN11 in many cancer cell lines including commonly used ones such as HeLa, U2OS, and HCT116 (38), and has recently been shown to have a robust and causal influence on resistance to platinum-derived drugs such as cisplatin and carboplatin (15). The current study expands this finding to eleven clinically relevant drugs (Table 1). Correlations between the SLFN11 methylation levels and DNA-damaging drug activities demonstrate the potential for using DNA methylation of SLFN11 in addition or in place of RNA- and immunofluorescence-based assays for measuring SLFN11 expression, and testing the usefulness of SLFN11 as a novel predictive biomarker for drug activity in the clinical setting.

Temozolomide is approved for the treatment of glioblastomas because of the frequent inactivation of MGMT by promoter methylation in those tumors (34, 36, 44). Methylome and gene expression analyses of the NCI-60 reveals that MGMT expression is frequently suppressed beyond CNS cancer cell lines (2 of the 6 breast cancer cell lines, 2 of the 7 colon, 2 of the 6 leukemia, 4 of the 10 melanomas, three of the 9 lungs, and one of the 8 kidney cancer cell lines; Fig. 5; Supplementary Fig. S2B). Yet, promoter methylation explained only 5 of these 20 cell lines with low or no MGMT expression. These observations suggest that MGMT promoter methylation (36, 44) is actually insufficient to predict temozolomide activity, and other assays (transcripts or protein measurements) should be used to monitor patient candidates for temozolomide not only for glioblastomas but also outside of brain tumors.

In summary, the NCI-60 methylome adds to the preexisting molecular and pharmacologic databases, which are publicly available and usable by non-informaticists at the CellMiner website (14). They provide an additional translationally relevant piece in the molecular puzzle of understanding and predicting transcriptional regulation, and for the broader interplay among cancer-associated molecular and pharmacologic parameters.

S. Varma has ownership interest (including patents) in HiThru Analytics. No potential conflicts of interest were disclosed by the other authors.

Conception and design: W.C. Reinhold, D.J. Goldstein, P.S. Meltzer, M. Esteller, Y. Pommier

Development of methodology: W.C. Reinhold, S. Varma, P.S. Meltzer, Y. Pommier

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): W.C. Reinhold, H. Stevenson, H. Heyn, V. Nogales, S. Moran, J.H. Doroshow, P.S. Meltzer, M. Esteller, Y. Pommier

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): W.C. Reinhold, S. Varma, V. Rajapakse, A. Luna, K.W. Kohn, Y. Wang, S. Moran, P.S. Meltzer, Y. Pommier

Writing, review, and/or revision of the manuscript: W.C. Reinhold, S. Varma, V. Rajapakse, A. Luna, K.W. Kohn, D.J. Goldstein, J.H. Doroshow, P.S. Meltzer, M. Esteller, Y. Pommier

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): S. Varma, M. Sunshine, J.H. Doroshow, P.S. Meltzer, M. Esteller, Y. Pommier

Study supervision: W.C. Reinhold, Y. Pommier

Other (providing funding for methylation analysis): D.J. Goldstein

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.