The Cancer Genome Atlas Program (TCGA) has previously been described [21]. Level 3 DNA methylation (beta values), ATAC-seq (log2((count+5)PM)-qn), and gene expression (log2(fpkm-uq+1)) data from the TCGA-PANCAN dataset were downloaded from the UCSC Xena browser [22]. Methylation probes with more than 50% missing values were removed and the remaining missing values were imputed using the R package impute v1.70.0 (function impute.knn) for each cancer type independently. For each cancer type, the number of neighbors used in the imputation (k) was set to 10 except for ovarian cancer in which k was set to 3 due to few samples with DNA methylation data from the Illumina HumanMethylation 450k array available.

DNA methylation and gene expression data obtained from primary pancreatic cancer tumors and normal tissue used for validation of the findings in the cell cycle- and hormone-like community were obtained from Nones et al. (GEO accession number GSE49149; Illumine Human Methylation 450k) [23] and from Sandhu et al. (SurePrint G3 Human GE 8x60k microarrays) [24] respectively.

Raw intensity (IDAT) files with methylation data from the HepG2 and PANC-1 cell lines were retrieved from encode with accession codes GSE128685 and GSE40699 respectively. Beta-values were imputed using the minfi R package v1.42.0 (function getBeta). DNA methylation data with pre-calculated beta values were obtained from: T-cells (GSE79144), monocytes (GSE68456), leukocytes (GSE69270), B-cells (GSE68456), MDAMB231 (GSE94943) and MCF7 (GSE69188). DNA methylation data from the NCI-60 cell lines were downloaded from the CellMiner database (https://discover.nci.nih.gov/cellminer/home.do) [25].

All data analyses were performed using the R software version 4.0.2 [26] unless otherwise specified. A p-value<0.05 was considered statistically significant in this study unless otherwise specified. The interaction plot between the emQTL communities was generated using Cytoscape v3.9.1 in which all 192 links from community 4 and 4808 randomly selected emQTL from communities 1–5 were included. Random links were selected by setting the seed parameter to 42 using the R function set.seed. Each community was labeled using the AutoAnnotate v1.4.0 app in Cytoscape. Pan-cancer emQTL analysis R code is available on GitHub (https://github.com/JorgenAnkill/Pan-cancer-expression-methylation-Quantitative-Trait-Loci-analysis). UpSet plots were made using the UpSet R package v1.4.0 [27].

The TCGA pan-cancer tumor samples with matching DNA methylation and gene expression data available were randomly split into a discovery and validation dataset (R function sample, seed = 42). All CpGs with an interquartile range (IQR)>0.1 (191 636 CpGs) and genes with a non-zero IQR (17 265 genes) in the discovery dataset (n = 4104) were tested for significant correlations using Pearson’s correlation. Due to the large number of associations, we considered significant only correlations with a Pearson correlation coefficient smaller than -0.5 (corresponding to p-value<1.40e-258), which is much stricter than Bonferroni-corrected p-value less than 0.05 (corresponding to p-value<1.51e-11). Significant CpG-gene associations from the discovery dataset were reanalyzed in the validation dataset (n = 4089). Associations were validated if the correlation coefficient was smaller than -0.5.

A clustering algorithm for community detection was applied to the pan-cancer emQTL with negative correlations using the condor R package v1.1.1 (function condor.cluster) [28]. The multilevel community was set as the clustering method and the absolute Pearson correlation coefficient of the pan-cancer emQTL was used as weight. For reproducibility, the initial state of the generator (seed) was set to 42.

Gene sets used for GSEA were obtained from the Molecular Signatures Database (MSigDB) v.7.4 [29]. Enrichment of genes in each gene set of interest was performed by hypergeometric testing using the C5 gene ontology- and hallmark (H) gene set collections. Enrichment with a BH-corrected p-value<0.05 was considered statistically significant.

Pathway enrichment analysis was performed on the gene sets obtained from each of the emQTL communities using the Reactome webtool (https://reactome.org) version 89 [30].

Clustering of DNA methylation and expression data was performed using the R package pheatmap v1.0.12 with Euclidean distance and ward.D2 clustering method unless otherwise specified. Expression values were transformed into z-scores before clustering for visualization purposes.

Genome annotation data of cis-Regulatory Elements (cCREs) from The Encyclopedia of DNA Elements (ENCODE) [31] was downloaded from the SCREEN webtool (https://screen.encodeproject.org). The registry encompasses 1,063,878 human regulatory regions from 4834 experimental datasets representing 1,518 different cell types and tissues. Nine different cCREs were included in the dataset; “pELS, CTCF-bound”, “dELS, CTCF-bound”, “PLS, CTCF-bound”, “CTCF-only, CTCF-bound”,”dELS”,”DNase-H3K4me3, CTCF-bound”, “pELS”, “PLS” and “DNase-H3K4me3”. In downstream analyses, some of these annotations were collapsed/converted into one as follows: Proximal CTCF-bound enhancer =“pELS, CTCF bound”, Distal CTCF-bound enhancer =“ dELS, CTCF-bound”, CTCF-bound promoter =“ PLS, CTCF-bound” and “DNase-H3K4me3, CTCF-bound”, Promoter =“ PLS” and “DNase-H3K4me3”, Distal enhancer =“ dELS”, pELS =“ Proximal enhancer” and CTCF =“ CTCF-only, CTCF-bound”.

Cell-type specific genome-wide segmentation data from the pancreatic cancer cell line PANC1 and the liver cancer cell line HepG2 was retrieved from Segway (https://segway.hoffmanlab.org). Segway is a method used to annotate the genome using a dynamic Bayesian network model on ChIP-seq data obtained from cell lines and/or tissues [32]. Enrichment of emQTL-CpGs in Segway-defined regulatory regions was assessed by hypergeometric testing (R function phyper) using the IlluminaMethylation450k CpGs as background. Adjusted p-values were obtained by Benjamini-Hochberg correction.

Maps of direct TF-DNA interactions were obtained from the UniBind database (https://unibind2018.uio.no) [33]. The UniBind TFBSs correspond to high-confidence direct TF-DNA interactions determined experimentally through ChIP-seq and computationally through position weight matrices (PWMs) from JASPAR [34]. The predicted TFBSs were derived from 1983 ChIP-seq experiments for 231 TFs across cell types and tissues and were predicted to have high PWM scores and be near ChIP-seq peaks.

Since the genomic coordinates of the UniBind database are based on the hg38 reference genome, all CpG positions from the Illumina 450k array were lifted over from hg19 to hg38 using the LiftOver web tool from the UCSC genome browser (https://genome.ucsc.edu) and were extended ±150bp from TFBSs. The CpG positions were then intersected with the TBFR using BEDTools v2.27.1 [35]. Since UniBind maps of TF binding sites are often derived from several ChIP-seq experiments for each TF, we merged the TF binding sites for all ChIP-seq experiments for each TF. Enrichment of CpGs in TFBRs was imputed using hypergeometric testing (R function phyper) using the IlluminaMethylation450k CpGs as background. Multiple testing was accounted for using Benjamini-Hochberg correction (R function p.adjust).

Deconvolution of the cellular composition of the tumor samples in the TCGA-PANCAN dataset was performed using the xCell [36] r-package version 1.1.0 (function xCellAnalysis). xCell is a machine learning algorithm trained on 64 stromal and immune datasets to identify specific cell types from bulk tissue using 10 808 genes as signatures. To assess the link between the xCell-derived cell type scores, the tumor samples were divided into quartile groups based on the xCell score; i.e., the level of infiltration. Differences in DNA methylation at CpGs or expression of genes between the groups were then assessed using the Kruskal-Wallis test.

Predicted cis-regulatory elements-gene associations were obtained from the GeneHancer database v5.0 [37]. Enhancer to gene associations was obtained from GeneCards Suite Version 5.0 [37]. The genomic coordinates were lifted over from hg38 to hg19 using the LiftOver web tool from the UCSC genome browser (https://genome.ucsc.edu). Predicted Integrated Methods for Predicting Enhancer Targets (IM-PET) interactions from the PANC-1, HepG2, A549, HCC1954, MCF7, HCT116, HELA, and K562 cell lines were retrieved from the 4D Genome data portal (https://4dgenome.research.chop.edu) [38]. IM-PET is a computational algorithm used to predict chromatin interaction loops between enhancers and promoters by using genomic features including distance restrictions, enhancer-target promoter activity profile correlation, TF and target promoter correlation, and co-evolution of enhancer and target promoter [39]. Experimentally defined Chromatin Interaction Analysis by Paired-End sequencing (ChIA-PET) data defining long-range chromatin interactions in the MCF7 breast cancer cell line was obtained from ENCODE (Accession number ENCR000CAA [40]). BEDTools v2.27.1 was used to intersect gene and CpG positions with the genomic intervals defining the feet of the chromatin loops for the GeneHancer, IM-PET and ChIA-PET data. An emQTL was considered in a chromatin loop if the CpG and gene were found in different feet of the same chromatin loop. To avoid bias from the array design, enrichment of emQTL in chromatin loops was calculated using a hypergeometric test (R function phyper) using all possible loops between CpGs and genes (i.e., CpG-gene pairs) on the same chromosome on the 450K array as background. Visualization of the chromatin interactions was made using the Gviz v1.40.1 [41] and GenomicRanges v1.48.0 [42] R packages and genomic interactions were visualized in Gviz using the GenomicInteractions v1.30.0 [43] R package.

Allele-Specific Copy Number Analysis of Tumors (ASCAT) [44] is a computational tool that estimates tumor purity by inferring the fraction of tumor cells in a sample. This is done by analyzing the proportion of tumor-specific and normal-specific allele frequencies in DNA copy number data. ASCAT compares the observed DNA copy number data to a reference profile, taking into account DNA contamination, ploidy of cancer cells and the presence of subclonal populations of cells with different copy number alterations. Tumor purity estimates for the TCGA samples were obtained from the NCI Genomic Data Commons (GDC) data portal upon request (https://portal.gdc.cancer.gov) [45].

The ability of the pan-cancer emQTL to predict prognosis was assessed by building Cox regression with ridge penalty to contract models for survival predictions by estimating the hazard function (Ridge Cox model) [46]. Principal component analysis (PCA) was performed on the z-scores of the expression and methylation data for each cancer type in TCGA using the R package stats v3.6.2. In each community, the first two principal components (PCs) of emQTL-CpGs and the first two PCs of the emQTL-genes were extracted, thus 24 PCs were extracted from the six communities. We regressed patients’ survival in each cancer type on the 24 PCs using the Ridge Cox model (R package glmnet v4.1.6). The survival prediction metric Uno’s C-index (R package survAUC v1.1.1) was obtained using 50 5-fold (80–20 split) cross-validation to overcome overfitting. The seed parameter was initialized as 123 using the R function set.seed. Survival plots were made using the survminer R package v0.4.8.

To generate Kaplan-Meier survival curves, we repeated the analysis using a 60–40 split between training and test data, and the trained model was applied to the test data. Kaplan-Meier survival curves and Log-rank tests were performed using the R functions Surv and survfit from the R package survival v3.5.3.

To identify robust associations between DNA methylation and gene expression in cis and trans, we correlated genome-wide levels of DNA methylation and gene expression from the TCGA-PANCAN dataset (see Material and Methods; S1 Fig for workflow outline). The TCGA pan-cancer dataset was split into a discovery (n = 4104) and a validation cohort (n = 4089; see Material and methods). 1 192 567 (87%) out of the 1 363 565 significant CpG-gene associations discovered were significant in the validation cohort. Most correlations had negative Pearson coefficient values (64.8% negative vs 35.2% positive; S2 Fig), indicating that methylation at a CpG increases when the level of expression for the correlated gene decreases, which can reveal a potentially putative inhibitory effect of CpG methylation on gene expression. The validated emQTL included 5472 genes and 45 406 CpGs. We found cis-associations (same chromosome; 83 035 associations) to be significantly enriched compared to trans associations (1 037 323 associations; hypergeometric test p-value<2.2e-16; fold enrichment (FE = 1.42). The validated associations are hereafter called emQTL. Although causality on gene expression regulation cannot be determined from statistical correlations, the emQTL provides insight into the overall degree of coordination between CpG methylation and gene expression globally, considering both potential direct and indirect regulation.

To identify coordinated emQTL, we performed bipartite network analysis using Complex Network Descriptions Of Regulators (CONDOR) [28], identifying communities of highly connected emQTL-CpGs and emQTL-genes. Since our study focused on the events where DNA methylation had a putative inhibitory effect on gene expression, we considered negative correlations in the community detection analysis (772 730 emQTL). We detected six pan-cancer emQTL communities (Fig 1A and 1B and S1A and S1B Table). To elucidate the biological functions of the emQTL communities, we performed gene set enrichment analyses (GSEA) using gene sets obtained from the Molecular Signatures Database (MSigDB) [29]. We found the communities to be enriched for genes involved in metabolic processes (Community 1), cell cycle (Community 2), neuroendocrine-like (NE-like) functions (Community 3), hormone-signaling (Community 4), immune response (Community 5) and epithelial-mesenchymal transition (EMT; Community 6; Fig 1C and S1C Table). Pathway enrichment analysis was performed and validated that the genes in each emQTL community were enriched in their respective functions as identified in the GSEA. However, significant enrichment was not observed for the hormone-signaling community, which is likely due to the lack of specific pathways for hormone-signaling in Reactome (S1D Table and S3 Fig). The number of CpGs covering the genes in the metabolism, cell cycle and hormone like communities showed similar distribution as all genes in the genome, showing that our analysis does not introduce notable bias potentially caused by CpGs being unevenly distributed across genes (S4 Fig).

We re-analyzed all emQTL within each cancer type and showed that there is a correlation within most cancer types for most communities (i.e., high methylation is associated with low expression also within individual cancer types), which points to that the identified biological functions are linked to epigenetic dysregulation both within and across cancer types (Fig 1D).

To assess the functional context of the pan-cancer emQTL-CpGs, we determined their enrichment in specific categories of regulatory regions by considering the pan-tissue candidate cis-Regulatory Elements (cCREs) predicted by ENCODE [31]. We found the emQTL-CpGs from all communities significantly enriched in enhancer regions (Fig 2A and S1E Table). Moreover, the CpGs in communities 1 to 5 were enriched in promoter regions. Since TFs can bind to promoter and enhancer regions to modulate gene expression, we aimed to identify which TFs may bind to these regions using TF-DNA interaction data obtained from the UniBind database [33]. We found the emQTL-CpGs in each community to be enriched in TF binding regions (TFBR) of TFs associated with functions coherent with the GSEA of the communities (Fig 2B and S1F Table). In the metabolism community, CpGs were enriched in binding regions of hepatocyte nuclear factors (HNFs) and NR2F2/6 known to be associated with metabolism. In the cell cycle community, CpGs were enriched in binding regions of Fos and Jun family TFs known to be involved in proliferation. In the hormone-like community, CpGs were enriched in binding regions of ER and AR, as well as FOX and GATA family TFs, known to be important in hormone signaling in cancer. Taken together, our analysis highlight the cancer-relevant role of proliferation-, metabolism- and hormone signaling-associated TFs. We also observed that the enriched TFs in each community shared binding regions, especially in the cell cycle bicluster (S5 Fig).

Enhancers can regulate gene expression through the binding of TFs and the formation of chromatin loop interactions with their target gene. As all emQTL communities were enriched for enhancer regions, we assessed whether the emQTL-CpGs and genes were linked through predicted loops from GeneHancer [37]. Our analysis revealed that the emQTL in communities 1 to 4 were significantly enriched in GeneHancer loops (Fig 2C, Hypergeometric test p-value = 5.50e-11, 2.74e-05, 3.59e-03, 7.97e-04 respectively; S6 Fig and S1G Table). No statistically significant enrichment in loops was observed for the emQTL in the immune and EMT communities. To further validate the enrichment in chromatin loops, we utilized IM-PET data from eight cancer cell lines and assessed enrichment for CpG-gene pairs in all communities (S6 Fig; S1G Table). These data show that CpG-gene pairs in the cell cycle-, metabolism- and hormone-like communities were enriched in chromatin loops in most of the available cell lines. Finally, in experimentally-defined chromatin interaction data (ChIA-PET) from the MCF7 breast cancer cell line, we observed enrichment in loops for all communities except the EMT community (S6 Fig and S1G Table). The results from the emQTL communities are summarized in Table 1.

To assess the ability of the pan-cancer emQTL to predict prognosis, we built Cox regression with a ridge penalty to construct models for survival predictions by estimating the hazard function (Ridge Cox model) [46]. Specifically, we performed principal component analysis (PCA) on the z-scores obtained from methylation and expression data for the emQTL-CpGs and genes to standardize the data on the same scale (see Material and Methods). In each community, the first two principal components (PCs) of emQTL-CpGs and -genes were extracted. We regressed patients’ survival in each cancer type by Ridge Cox model and obtained Uno’s C-index. For comparison, we extracted the first two PCs from the emQTL-CpGs and emQTL-genes respectively, i.e., without considering emQTL communities, and regressed patients’ survival in each cancer type on the four PCs from all the emQTL communities by Ridge Cox model. Uno’s C index was estimated to evaluate the models in which a value below 0.5 indicates a poor model and a value above 0.7 indicates a good model. The PCs from the pan-cancer emQTL considering the six communities show the predictive power of patients’ survival (Fig 3A), and had significantly higher C-indexes than the four PCs from the pan-cancer emQTL without considering the communities (p-value = 1.41e-11; S7 Fig). Fig 3B–3E shows examples of differences in survival outcomes between the predicted high- and low-risk group for the ocular melanoma, adrenocortical cancer, mesothelioma, and lower-grade glioma (C-index = 0.81, 0.75, 0.75 and 0.68 respectively). Altogether, our analysis highlights the ability of the pan-cancer emQTL to predict survival outcomes. We hypothesize that the emQTL has an important role in cancer-related regulation of transcriptional networks.

To shed light on the pathogenic molecular processes associated with the cell cycle community, we performed hierarchical clustering of DNA methylation levels at CpGs and expression of genes in the cell cycle community using the TCGA-PANCAN dataset. We found the methylation and expression profiles to reflect the tissue of origin of the tumors (Fig 4A and 4D; Chi-square test p-value<2.2e-16 (three patient subclusters)) which is consistent with previous reports [67]. We also categorized tumors into broader histological/tissue types classes, and observe a clear organization of tumors according to this classification. Moreover, we observed that the cell cycle community CpGs are hypomethylated in most cancer types compared to normal adjacent tissue (Fig 4B). The hypomethylation is specifically observed at the TFBRs of FOSL1/2 and JUN (S8A Fig). We further assessed whether the CpGs and genes associated with the cell cycle community were located in open chromatin regions using ATAC-seq data [68]. As expected, the cancer types harboring open chromatin around the cell cycle community CpGs were also the cancer types with the lowest level of DNA methylation at these sites (Fig 4C). Concomitantly, the genes associated with the cell cycle community show significantly higher expression in tumor cells than in normal cells for most cancer types (Fig 4E). Finally, the cancer types with the highest expression level of the cell cycle community genes correspond to the ones with the highest open chromatin signal (Fig 4F). The top five hub genes (genes with most associations to CpG methylation) were CEP55, TUBA1C, FAM83D, AURKA and CDCA8 (S1B Table), all linked to regulation of the cell cycle and microtubule cytoskeleton organization during mitosis. These lines of evidence show that the cell cycle community CpGs are hypomethylated in most cancer types and are found in accessible enhancer regions forming interactions with transcription start site (TSS) within open chromatin regions regulating cell cycle-related genes.

Hierarchical clustering of DNA methylation levels at the hormone-like and metabolism communities CpGs separated the tumors based on the tissue of origin, but also indicated variability in DNA methylation within each cancer type (S9A and S9B Fig). The CpGs in both communities were significantly enriched in enhancer regions and were less methylated in most cancer types compared to normal adjacent tissue (S9C and S9D Fig). Loss of methylation was also observed specifically at the TFBRs of FOXA1, ESR1 and AR at the CpGs in the hormone-like community and at the HNF4A/G and RXRA TFBRs in the metabolism community (S8B and S8C Fig). By assessing the chromatin accessibility around the hormone-like and metabolism community CpGs, we found lower DNA methylation levels at the CpGs to be associated with more accessible chromatin (S9E and S9F Fig). Similarly, to the methylation data, the expression of genes in the hormone-like and metabolism-related communities clustered according to the tissue of origin, and variability in expression within the cancer types was also observed (S10A and S10B Fig). We next compared the expression of the hormone-like and metabolism community genes between tumor and normal tissue and found genes in both communities to be significantly higher expressed in tumor tissue compared to normal adjacent tissue for most cancer types (S10C and S10D Fig). High expression of the hormone-related and metabolism-related genes was also associated with more accessible chromatin (S10E and S10F Fig). The top five hub genes for the hormone-like community were KRT18, JUP, KRT8, SH2D3A and PDLIM1 (S1B Table); KRT18 and KRT8 are luminal markers in breast cancer while JUP (plakoglobin), SH2D3A and PDLIM1 are relatively undescribed in the context of cancer. The top five hub genes for the metabolism community were HNF4A, SLC39A5, AMN, UGT2A3 and SLC17A4 (S1B Table). HNF4A encodes a TF involved in development of beta cells in the pancreas, and AMN (amnionless) is involved in vitamin uptake. The roles of SLC39A5, UGT2A3 and SLC17A4 are relatively undescribed in cancer. The emQTL in both communities were enriched in inferred GeneHancer enhancer-promoter loops, thereby suggesting a possible functional regulation of hormone-related and metabolism genes. Altogether, our findings suggest that enhancer methylation is linked to the upregulation of hormone- and metabolism-related genes through chromatin looping across cancer types.

Adipocytes are a prominent cell type of the tumor microenvironment in solid tumors and express high levels of metabolic genes involved in energy storage and mobilization. To ensure the signal from the metabolism community was not caused by infiltration of non-tumor cells such as adipocytes, we correlated the average DNA methylation levels at the CpGs genes in the metabolism community with the ASCAT tumor purity estimates [44]. No statistically significant correlation was observed between methylation and tumor purity in TCGA (S11 Fig, p-value = 0.90). A significant correlation was observed between the average expression of the metabolism genes and tumor purity (p-value = 8.54e-06, r-value = -0.05); however, the Pearson correlation coefficient r was close to zero, suggesting that the variations in methylation and expression profiles of the metabolism community are caused by differences between the cancer cells.

Unlike the DNA methylation levels at the CpGs and expression levels at the genes in the cell cycle, metabolism, and hormone-like communities, CpGs and genes in the immune community did not segregate the tumors equally well according to cancer type (S12A–S12D Fig). Since immune cells are prominent in the tumor microenvironment and have different DNA methylation and gene expression profiles compared to cancer cells, we hypothesize that this community of emQTL was influenced by and reflects variation in immune cell infiltration in tumors. To address this, we used the xCell deconvolution [36] algorithm to quantify the level of immune infiltration using gene expression data from TCGA. We divided the tumors into quartile groups based on the level of immune infiltration. We observed a significant difference in DNA methylation levels at the CpGs in the immune community between the quartile groups in which low methylation at CpGs in the immune-community was associated with high immune infiltration and high DNA methylation with low infiltration (S13A Fig, p-value<2.2e-16). Furthermore, DNA methylation levels at the CpGs in the immune community obtained from isolated immune cells, such as leukocytes, monocytes, T-cells, and B-cells, reflected similar methylation levels at the CpGs as the tumors with high immune infiltration. In contrast, cancer cell lines from the National Cancer Institute (NCI-60) Human Tumor Cell Lines Screen showed similar methylation levels as the quartile group with the lowest infiltration. One exception was leukemia cell lines which is reasonable as the cell of origin for this cancer type is lymphocyte and therefore shares similarities with the infiltrating immune cells. We also observed a positive association between immune infiltration and expression of genes in the immune community, i.e., high expression of the immune community genes is linked to high immune infiltration and vice versa (S13B Fig). To further confirm the link between immune infiltration and the immune community, we obtained copy-number-derived tumor purity estimates from ASCAT [44] and assessed the link between the xCell-derived immune score and tumor purity. We found a negative association between tumor purity and immune cell infiltration, i.e., low tumor purity is associated with high infiltration of immune cells, and vice versa (S13C Fig). Altogether, these results suggest that the emQTL in the immune community is caused by variations in immune cell infiltration.

The smallest emQTL community was the EMT community consisting of 31 CpGs and 43 genes representing 120 emQTL. Similar to the immune community, DNA methylation at the CpGs and expression of genes did not segregate the tumors according to cancer types compared to the cell cycle, metabolism and hormone-like communities (S14A–S14D Fig). Since fibroblasts are a prominent mesenchymal cell type of the tumor microenvironment of many solid tumors and carry out functions such as migration and remodeling of the extracellular matrix (ECM) [69], we assessed whether the EMT community could arise from heterogeneity in tumor composition in regards to fibroblast infiltration. We employed the xCell deconvolution tool [36] to estimate the relative quantity of fibroblast infiltration in each tumor sample using gene expression. By dividing the tumor samples into quartile groups based on the severity of fibroblast infiltration, we found that expression of the EMT community genes was positively associated with fibroblast infiltration (S15A Fig, p<2.2e-16), i.e., high expression of EMT community genes is linked to high fibroblast infiltration. We also observed a significant negative association between DNA methylation and fibroblast infiltration (S15B Fig, p-value<2.2e-16). Furthermore, the DNA methylation levels of the EMT community CpGs in tumors with severe fibroblast infiltration showed similar methylation profiles as purified fibroblasts. Conversely, all tumor cell lines from the NCI-60 panel showed methylation levels similar to the low infiltration group. To further confirm that the EMT community was linked to fibroblast infiltration, we assessed whether the quartile groups based on the severity of fibroblast infiltration were linked to the ASCAT tumor purity estimates. A significant and negative association was observed between fibroblast infiltration and tumor purity, i.e., high fibroblast infiltration is linked to low tumor purity, and vice versa (S15C Fig). Taken together, our findings suggest that the EMT community of emQTL is linked to heterogeneity in the severity of fibroblast infiltration in tumors.

To study the differences in terms of methylation and expression of the NE-like community CpGs and genes, we performed unsupervised hierarchical clustering of DNA methylation and gene expression levels of the CpGs and genes in the NE-like community. We found the DNA methylation and expression profiles of the CpGs and genes in the NE-like community to cluster by cancer type (S16A and S16B Fig). Among the cancer types, the neuron-related cancer types such as lower-grade glioma (LGG), pheochromocytoma & paraganglioma (PCPG), and glioblastomas (GBM) had pronouncedly different DNA methylation and gene expression profiles compared to the other cancer types. These cancer types showed the highest expression of the NE-like community genes and the lowest DNA methylation at the CpGs in the NE-like community (S16C and S16D Fig). Moreover, these cancer types had less compact chromatin around the NE-like community CpGs and genes than other cancer types (S16E and S16F Fig).

Altogether, our findings suggest that the NE-like community is associated with the cell of origin. The brain-related cancer types have upregulated neural-related genes through low DNA methylation at enhancers. However, this cannot fully explain this community, as there are also associations between DNA methylation at the CpGs and expression of the genes in the NE-like community within non-brain related cancer types (Fig 1D). The methylation and expression levels also differ between tumor and normal adjacent tissue for most cancer types (S16C and S16D Fig). To assess whether this could be caused by variations in the tumor microenvironment, we obtained the tumor microenvironment score from xCell [36]. We assessed the link between DNA methylation of CpGs and the expression of genes in the NE-like community to the severity of infiltration. Due to the prominent differences in methylation and expression profiles between the neuron-like cancer types (LGG, PCPG, and GBM) and the other cancer types, we assessed the link between methylation, expression, and tumor infiltration in the neuron-like (NL) and non-neuron-like (NNL) tumors independently. We found the expression of the NE-like community genes to be positively associated with the microenvironment score from xCell in the NNL tumors (S17A Fig; p-value = 1.29e-121). Moreover, a negative association between the tumor microenvironment score and DNA methylation was observed for the NNL, and the methylation levels from the NCI-60 cancer cell lines reflected the methylation levels in the low infiltration group (S17B Fig; p-value = 3.58e-110). A negative association was observed between the tumor microenvironment score and tumor purity (S17C Fig). Altogether, our findings suggest that the variability in DNA methylation at the CpGs and expression of genes in the NE-like community within the cancer types are caused by variations in the cell composition of the tumor microenvironment in the NNL cancer types. For the NL tumors, there was a negative association between the tumor microenvironment score and expression of the genes in the NE-like community (S17D Fig, p-value = 3.31e-70) and a positive association with the level of DNA methylation (S17E Fig). A negative association between the tumor microenvironment score and tumor purity was also observed, and the main difference in purity was between the tumors with severe infiltration and those with low, moderate, and high infiltration (S17F Fig). Altogether this suggests that the NE-like community is mainly caused by differences in the tissue of origin between the tumors and that differences in tumor composition cause variability in expression and DNA methylation within the cancer types.

Since emQTL can be both in cis and trans, and some might rather be passenger emQTL rather than cancer-driving emQTL, we sought to perform emQTL analysis in a knowledge-guided manner to identify candidate emQTL facilitating cancer-promoting functions through direct regulation. By knowing the characteristics of the transcriptional networks linked to the regulation of the cancer-related emQTL communities (i.e., the hormone-like, cell cycle, and metabolism communities), we can perform a knowledge-guided analysis by applying several criteria to the selection of emQTL based on the characteristics of the communities (see methods). Using the knowledge-guided emQTL approach, we identified 1770 proliferation-promoting, 1762 metabolism-promoting, and 1062 hormone-signaling-promoting candidate emQTL in cancer (S1H Table). After identifying candidate emQTL we assessed the link between DNA methylation at the candidate emQTL-CpGs and expression of their linked emQTL-gene for each cancer type and found most candidate emQTL to display correlations within several cancer types (S1H Table). One example of a proliferation-promoting candidate emQTL involves cg23320499 and CAMP Responsive Element Binding Protein Like 2 (CREBL2) found in the GO cell cycle gene set (Fig 5A). CREBL2 associates with the emQTL-CpG cg23320499 and is located in a predicted enhancer in the binding site of JUND and CEBPB. Moreover, DNA methylation at cg23320499 was negatively correlated with the expression of CREBL2 in several cancer types such as thymoma (p-value = 8.71e-11, r-value = -0.6) and bile duct cancer (Fig 5B; p-value = 1.54e-02, r-value = -0.5). In bile duct tumors, a significantly lower methylation level at cg23320499 and higher expression of CRBEL2 were observed compared to normal adjacent tissue (Fig 5C and 5D, respectively). These results suggest that we can identify promising candidate proliferation-promoting emQTL using the knowledge-guided emQTL approach.

Next, we sought to study the role of the pan-cancer emQTL of the cancer-related communities (i.e., cell cycle, metabolism, and hormone-like communities) in further detail in a cancer-specific manner as case examples to show the link between DNA methylation at the emQTL-CpGs and target gene expression.