Alternative polyadenylation quantitative trait methylation mapping in human cancers provides clues into the molecular mechanisms of APA

Summary

Genetic variants are involved in the orchestration of alternative polyadenylation (APA) events, while the role of DNA methylation in regulating APA remains unclear. We generated a comprehensive atlas of APA quantitative trait methylation sites (apaQTMs) across 21 different types of cancer (1,612 to 60,219 acting in cis and 4,448 to 142,349 in trans). Potential causal apaQTMs in non-cancer samples were also identified. Mechanistically, we observed a strong enrichment of cis-apaQTMs near polyadenylation sites (PASs) and both cis- and trans-apaQTMs in proximity to transcription factor (TF) binding regions. Through the integration of ChIP-signals and RNA-seq data from cell lines, we have identified several regulators of APA events, acting either directly or indirectly, implicating novel functions of some important genes, such as TCF7L2, which is known for its involvement in type 2 diabetes and cancers. Furthermore, we have identified a vast number of QTMs that share the same putative causal CpG sites with five different cancer types, underscoring the roles of QTMs, including apaQTMs, in the process of tumorigenesis. DNA methylation is extensively involved in the regulation of APA events in human cancers. In an attempt to elucidate the potential underlying molecular mechanisms of APA by DNA methylation, our study paves the way for subsequent experimental validations into the intricate biological functions of DNA methylation in APA regulation and the pathogenesis of human cancers. To present a comprehensive catalog of apaQTM patterns, we introduce the Pancan-apaQTM database, available at https://pancan-apaqtm-zju.shinyapps.io/pancanaQTM/.

Graphical abstract

We generated a comprehensive atlas of APA quantitative trait methylation sites (apaQTMs) across 21 different cancers (1,612 to 60,219 acting in cis and 4,448 to 142,349 in trans), which could be accessed at a web-based platform, Pancan-apaQTM. Our findings provided deeper mechanistic insights into the role of QTMs in diseases.

Introduction

Most mature eukaryotic mRNAs possess a string of adenine residues referred to as a poly(-A) tail, which is added during a step of post-transcriptional modification known as polyadenylation. More than 70% of human protein-coding genes contain two or more polyadenylation sites (PASs).¹ Alternative polyadenylation (APA) is a post-transcriptional process that generates distinct mRNA isoforms with alternative termination sites.² APA alters the length of 3′- untranslated region (UTR), which consequently modulates mRNA stability, translation efficiency, or subcellular localization by disrupting the binding sites of miRNA or RNA-binding proteins (RBPs).^3,4 Despite the significant strides made in identifying APA events, especially of the mechanisms governing its regulation and the consequences of its dysregulation,^{5,6,7,8,9,10,11} APA still represents an area of limited study and understanding in relation to its impact on health and disease.^12,13

Previous studies have shown that genetic variants are involved in the orchestration of APA events. For instance, the C allele of rs78378222, which is located in a PAS within the 3′UTR of TP53 (MIM: 191170), was found to impair proper termination and polyadenylation of the TP53 transcript.¹⁴ Additionally, Shulman and Elkon, using RNA sequencing (RNA-seq) data from the GTEx Project, identified 330 single-nucleotide polymorphisms (SNPs) that exhibit significant associations with PAS usage.¹⁵ As anticipated, the majority of these PAS-interrupting alleles result in reduced polymerization cleavage at their PAS, leading to the production of longer isoforms.¹⁵ Recently, in a study by Chen et al., 68 pathogenic variants in PASs were identified through a new pipeline.¹⁶ The authors also demonstrated their ability to consistently replicate known PAS variants in HBA2 (MIM: 141850) and HBB (MIM: 141900), which resulted in abnormal RNA production.¹⁶ Furthermore, Yang et al. characterized APA as a quantitative trait using genotype and APA data from The Cancer Genome Atlas (TCGA) and The Cancer 3′UTR Atlas (TC3A), revealing hundreds of thousands of SNPs associated with APA events.¹⁷ Similarly, Li and colleagues conducted a genome-wide association study (GWAS) of APA, utilizing data from the GTEx Project.¹⁸ They identified numerous variants associated with APA in normal tissues, which are highly enriched within 3′UTRs, and demonstrated their contributions to disease heritability.¹⁸

Besides genetic control of APA, epigenetic modifications have gained increasing recognition as crucial regulators of APA. In mice, it has been observed that the paternal allele of the imprinted signal peptide peptidase gene H13, when hypomethylated, tends to favor the utilization of its proximal PAS, while the distal PAS shows a higher likelihood of being coupled to the hypermethylated maternal allele.¹⁹ A similar phenomenon has been documented in the imprinted hect domain and RLD 3 (Herc3)/nucleosome assembly protein 1-like 5 (Nap1l5) locus. DNA methylation within the promoter of Nap1l5 on the maternal allele, situated within an intron of Herc3, appears to inhibit the utilization of the intronic Herc3 PAS.^20,21 However, the opposite may be observed in other genomic regions. For instance, as demonstrated by Ma et al., DNA methylation in the vicinity of proximal PASs promotes the utilization of the same proximal PAS, leading to the production of the transcript possessing the corresponding proximal alternative last exon, FANCD2-V2.²² Mechanistically, the absence of DNA methylation permits the binding of CCCTC-binding factor (CTCF) and facilitates the recruitment of the cohesin complex, thereby promoting the use of the proximal PAS in methylation-deficient double knockout (DKO) cells.²³ Additionally, DNA methylation within the intron of specific genes, such as the increase in Bonsai methylation 1 (IBM1) gene in Arabidopsis thaliana, has been found to enhance the expression of the full-length IBM1 transcript.^24,25,26 Results from a recent study also showed that the depletion of DNA-modifying factors, such as DNA methyltransferase 1 (DNMT1 [MIM: 126375]), DNA methyltransferase 3 alpha (DNMT3A [MIM: 602769]), and DNA methyltransferase 3 beta (DNMT3B [MIM: 602900]), affected the dynamic changes of APA.¹¹ Collectively, these findings likely represent only the “tip of the DNA-methylation iceberg”––our understanding of how DNA methylation influences APA is far from complete. Nevertheless, the aforementioned findings underscore the significance of conducting comprehensive investigations into the role of DNA methylation in the regulation of APA.

The role of aberrant DNA methylation in the initiation and progression of cancer is widely acknowledged. Non-mutational epigenetic reprogramming is recognized as a hallmark in cancer development.²⁷ The utilization of data from TCGA and its molecular insights offer the opportunity to interrogate the influence of DNA methylation on APA regulation. We retrieved the percentage of distal PolyA site usage index (PDUI), a measure quantifying the proportion of distal APA usage, for 21 different TCGA cancer types from TC3A, along with DNA methylation data sourced from TCGA.²⁸ Analogous to genetic variants, DNA methylation can function as either a cis-element or a trans-acting factor affecting APA, and both aspects remain relatively underexplored and poorly understood.

In this study, we conducted an analysis of apaQTMs to establish the connection between DNA methylation and APA. In addition, we assessed the impact of DNA methylation on cis-APA events using two-sample Mendelian randomization (MR) analysis. We also integrated ChIP-exo and RNA-seq data to explore the role of transcription factors (TFs) in the regulation of APA. Additionally, we performed an analysis of quantitative trait methylation sites (QTMs) in gene expression (eQTMs; methylation-expression) and splicing (spQTMs; methylation-splicing) and carried out colocalization analysis. We incorporated the QTM results into epigenome-wide association studies (EWASs) to colocalize potential causal genes associated with DNA methylation in 17 different cancer types. Our enhanced understanding of APA regulation may pave the way for new avenues of research, shedding light on the implications of DNA methylation and providing deeper mechanistic insights into the development of human diseases.

Material and methods

Identification of apaQTM

Clinical samples

We obtained clinical data, including patient age, gender, and tumor stage, for various cancer types from the TCGA data portal (https://portal.gdc.cancer.gov). Only selected samples with complete clinical information, methylation measurements, and APA data were included for further analysis. After conducting quality control on the methylation and APA data, 21 cancer types with a sample size >150 were selected for the identification of apaQTMs.

Collection and processing of DNA methylation data

Beta values of DNA methylation (TCGA, level 3) gathered from the Illumina Infinium HumanMethylation450 (HM450) platform were obtained from the UCSC Xena browser (https://xenabrowser.net/datapages/).²⁹ To ensure quality of probes, the following exclusion criteria were applied, as outlined by Gong et al.³⁰: (1) probes with beta value missing rate >0.05; (2) mapping to multiple genomic locations³¹; (3) containing known SNPs (from the 1000 Genomes phase 3,³² with MAF ≥0.01) at CpG sites; (4) on sex chromosomes; and (5) with a standard deviation <0.05. Samples with a missing rate >0.05 were also excluded. Missing beta values were imputed using the impute.knn function from the R package impute.³³ Finally, we retained an average of 230,337 CpG sites across the different cancer types for the apaQTM analysis.

APA data processing

Xia et al. introduced a robust bioinformatics algorithm, dynamic analyses of APA from RNA-seq (DaPars), for the identification of APA events indicated by percentage of PDUI values.³⁴ Subsequently, their team refined and updated this algorithm, i.e., DaPars2, offering a comprehensive resource on APA usage in human cancers.²⁸ We fetched APA data from TC3A (http://tc3a.org/, https://www.synapse.org/#!Synapse:syn24982198/files/). To exclude potentially problematic APA events, we applied the following criteria: (1) PDUI values with a missing rate >0.1; (2) PDUI values with a standard deviation <0.05. Samples with a missing rate <0.5 were selected for further analysis.¹⁷ In cases of duplicated samples, we calculated the average PDUI values using the avereps function from the limma package in R.³⁵ Any missing PDUI values were imputed using the impute.knn function from the R package impute, and these imputed values were used as input data for probabilistic estimation of expression residuals (PEER).

Covariate adjustment

We employed PEER to estimate hidden covariates, which encompassed potential batch effects, in relation to the PDUI values.³⁶ Selection of the number of PEER factors adhered to the protocol established by the GTEx Consortium³⁷: for sample sizes ranging from 150 to 250, 30 factors were adjusted, and for sample sizes >250, 35 factors were adjusted. The PDUI values were further quantile normalized. As part of the covariate adjustment, factors such as age, gender, and tumor stage were included in the apaQTM analysis.^17,18

Analysis of methylation and APA association

To assess the impact of methylation levels on PDUI values for each cancer type, we performed linear regressions using MatrixeQTL.³⁸ Analogous to genetic variants, DNA methylation can function as either a cis-element or a trans-acting factor that modulates APA. For cis-analysis, we defined CpG sites within a region spanning 1 megabase (Mb) distance from both ends of each PAS as candidates. All other CpG sites, including those from different chromosomes, were examined for trans-acting associations in relation to each APA event. We considered statistical significance to be achieved when p < 0.05 after applying the Benjamini-Hochberg false discovery rate (FDR) and Bonferroni correction for cis- and trans-acting apaQTM, respectively.

Annotation of apaQTMs

Genomic and cytosine-guanine (CG) island characteristics of apaQTMs

We retrieved the HM450 BeadChip annotations for each probe from this source: https://zwdzwd.github.io/InfiniumAnnotation. We tallied the CpG sites falling within seven distinct gene regions, i.e., the 1st exon (the initial exon), 3′UTR, 5′UTR, gene body, intergenic region, transcription start site (TSS)-1500, and TSS-2000. Probes that mapped to various regions, such as islands, shores (which extend 2 kilobases [kb] upstream and downstream of islands), shelves (which span 2 kb up and downstream of shores), and open sea (located more than 4 kb away from any island), were annotated. We quantified the number of both cis- and trans-apaQTMs within each region and determined statistical significance using hypergeometric test. The p values were subsequently adjusted using the Benjamini-Hochberg procedure. For reference, all CpG sites in the HM450 BeadChip were set as the background.

Identification of PAS motif-altering apaQTM

Putative PAS of genes undergoing APA were retrieved from polyA_DB.³⁹ Subsequently, we conducted a search for PAS motifs within the −20/+20 nt flanking regions of the identified cis-apaQTMs, which were included when located within 100 bp of the corresponding PAS.

Enrichment analysis of TF binding sites

To identify additional contributing factors (e.g., TFs) associated with APA regulation, we assessed the enrichment of 171 TF binding sites. We performed this analysis using the Locus Overlap Analysis (LOLA, v1.1.3).⁴⁰ CpG sites from the HM450 array were set as a reference background. Results with FDR <0.05 were considered to be statistically significant.

MR

To explore potential causal relationships among molecular events, including DNA methylation, APA, and mRNA expression, we conducted MR. MR employs germline variants, typically SNPs, as instrumental variables (IVs) for the exposure of interest. This approach is based on three fundamental assumptions: first, that the IVs are associated with exposure (e.g., a molecular event); second, that the germline variants affect the outcome of interest solely through the exposure factor; and third, that the germline variants are not associated with any confounding factors. Adhering to MR guidelines,⁴¹ we selected independent (R² < 0.1, window size = 250 kb) exposure-associated SNPs as IVs. Given the genetic architecture of these molecular events, we predominantly focused on the cis region (up to 1 Mb on both sides of the gene body or the molecular event loci) to select IVs. Palindromic variants were excluded. To identify potential causal relationships, we applied the weighted-median estimator after variant harmonization. The weighted-median estimator provides consistent estimates even when up to 50% of the IVs are invalid, i.e., violate one of the major assumptions of MR. MR was conducted only if the exposures had at least three independent IVs.

For estimating the effect of DNA methylation on cis-APA events, we employed a two-sample MR approach. We utilized methylation quantitative trait loci (mQTL) summary results obtained from 6,994 blood samples⁴² to select DNA methylation-associated SNPs that reached genome-wide significance (p < 1 × 10⁻¹⁵) as IVs. We obtained the PDUI of APA events from APAatlas,⁸ and the SNP-APA associations were estimated from GTEx whole-blood samples.⁴³

ChIP-exo/seq data collection, processing, and visualization

To investigate the binding signals of genes associated with apaQTMs in trans, we obtained a dataset comprising 887 ChIP-exo data generated by William and colleagues.⁴⁴ We applied a filter to exclude ChIP-exo/seq data that captured binding sites numbering 100 or fewer. We retained peaks with scores ranking above the 25th percentile. Subsequently, we collected the binding signals of genes associated with apaQTMs in trans. Visualization plots were generated using the R package circlize.⁴⁵ Genomic location information was based on the hg19 reference genome.

APA quantification

Raw sequencing data (in FASTQ format) was obtained from the Sequence Read Archive (SRA). Reads with high quality were included and aligned to the human genome (hg19) by utilizing HISAT2.⁴⁶ Subsequently, the aligned BAM files were sorted and converted into WIG files. These WIG files were then processed through the dynamic analysis of alternative polyadenylation (v2.0; DaPars2) algorithm.³⁴ DaPars2 directly estimated the APA usage by determining the PDUI score, which represents the fraction of distal PAS usage for the corresponding gene.

EWASs

We conducted DNA methylation-based epigenome-wide association studies (EWASs) for 17 different cancer types, using a design comparing tumor tissue to adjacent normal tissue, except for acute myeloid leukemia where a case-control design was employed. The primary data source was obtained from the EWAS data hub (https://ngdc.cncb.ac.cn/ewas/datahub). Methylation levels were measured using either the Illumina HumanMethylation450 BeadChip (450K) or the Illumina HumanMethylationEPIC BeadChip (EPIC/850K). The detailed information, including sample sizes, platforms, and basic characteristics can be accessed at this link: https://ngdc.cncb.ac.cn/ewas/datahub/download. To remove potential batch effects, including technical variations, we used Gaussian mixture quantile normalization (GMQN); wherein this method assumes that the signal intensity of each channel displays a Gaussian mixture distribution.⁴⁷ In essence, a two-state Gaussian mixture model was fitted based on a reference dataset from a single study, and the parameters from these distributions were used for rescaling signals of Infinium type I probes. Then, the Gaussian mixture model was fitted to the input type I probe signal intensities. The probabilities of type I probes were then converted to quantiles using the parameters estimated from the reference dataset. Finally, beta values were computed to represent DNA methylation levels.

Traits with fewer than 10 cases (tumor tissues) or 10 controls (normal tissues) were excluded from the EWAS analysis. Logistic regression was employed to evaluate the influence of DNA methylation on the trait, and the parameters beta and standard error were included in the subsequent analyses.

Mediation analysis of CpG sites, gene expression, and APA events

We conducted mediation analysis to assess whether the expression of APA regulatory genes can act as a mediator between CpG sites and APA events by utilizing the mediation package.⁴⁸ Specifically, we selected CpG sites that demonstrate both trans-effects in APA and cis-effects in the expression of APA regulatory genes for analysis. A P_{mediation effect} < 0.05 is considered indicative of a significant mediating effect.

Colocalization analysis

Methylation-based colocalization analysis was performed to assess the association between APA and diseases. To achieve this, we employed a commonly used approach, coloc, which relies on an approximate Bayes factor to identify gene-trait pairs by colocalizing the signals from eQTL (SNP-expression) and GWASs (SNP-trait).⁴⁹ In our study, we adapted coloc to identify APA-trait pairs by integrating signals from apaQTM (methylation-APA) and EWASs (methylation-trait). We utilized the R package coloc (v5.1.0.1) for this analysis. For each of the sentinel apaQTM, CpG sites within the surrounding region of 100 kb were used to perform the colocalization analysis. Regional-level apaQTM and EWAS signals with PP4 ≥ 0.75 and PP4/(PP4+PP3) ≥ 0.9 were reported. In addition, we estimated the colocalization of eQTMs and spQTMs with EWAS signals.

Statistical analyses

Statistical significance was determined using the Wilcoxon rank-sum test for comparisons of gene expression levels between tumor tissues and adjacent non-tumorous tissues, APA/expression/splicing levels between high and low methylation levels, and APA PDUI levels between TCF7L2-knockdown and wild-type cells. When necessary, multiple test correction was applied to reduce false positives. p < 0.05 indicates statistical significance. Effect size, reflecting enrichment values, was computed using the chi-square test. In survival analysis, we utilized the Kaplan-Meier method to plot survival curves, and adjusted p-values were obtained through multivariate Cox regression. All statistical analyses were performed using R (v4.1.0) along with the aforementioned packages. Statistical significance is supplemented in each of the respective figure legends.

Pancan-apaQTM, a web resource of pan-cancer apaQTMs

Utilizing the Shiny package,⁵⁰ we developed a user-friendly and interactive web application named Pancan-apaQTM, which can be accessed at https://pancan-apaqtm-zju.shinyapps.io/pancanaQTM/. Data are accessible and downloadable without restrictions. Tabular information is presented interactively using tables generated with the DT package.⁵¹ This data portal provides a comprehensive resource of apaQTMs across 21 human cancer types.

Results

apaQTMs across human cancers

We initiated a pilot study, with replication testing, involving TCGA colon adenocarcinoma (COAD) cases by mapping apaQTMs. The cases analyzed using the HM450 platform constituted the discovery set, while those using the HumanMethylation27 (HM27) platform were designated as validation set. Following rigorous data processing and quality control (Supplementary methods), we obtained a total of 9,912 CpG sites × 2,696 APA events to conduct further apaQTM analysis. The results revealed that 68 out of 257 (26.5%) cis CpG-APA pairs (where the CpG-PAS distance was within 1 Mb) from the discovery set were replicated in the validation set, while 548 (14%) trans pairs (with a CpG-PAS distance >1 Mb) were also recaptured (total n_trans-pairs = 3,918) with highly consistent direction of effects (99.19%, pooling cis and trans pairs together) (Figure S1; Table S1). Notably, we maintained a stringent significance threshold for trans-associations during replication (FDR <0.05). This pilot study demonstrated the reproducibility of apaQTMs in independent samples and helped refine the criteria for identifying apaQTMs.

We proceeded with a comprehensive pan-cancer analysis of apaQTMs (Material and methods). From 21 distinct types of cancer, an average of 364 samples with pooled mean of 230,084 CpG sites and 4,246 APA events were subjected to analysis to identify all potential CpG sites responsible for regulating APA. Our results unveiled ubiquitous influence of DNA methylation on APA across various cancer types. Using FDR <0.05, we observed a median of 20,416 CpG sites associated with 2,147 APA events in cis across the 21 cancers, whereas a median of 49,774 CpG sites had the potential to influence 725 APA events in trans, attaining genome-wide significance even after Bonferroni correction (Figure 1A; Table S2). Notably, low-grade glioma (LGG) displayed the most substantial impact, with as many as 60,219 cis-apaQTMs potentially perturbing 3,477 APA events, and 142,349 trans-apaQTMs had the potential to disrupt PDUI of 1,158 genes, conferring LGG with the largest number of such regulations among the 21 cancers. Moreover, our analysis identified that apaQTMs found in the cis-analysis were prominently enriched near PASs (Figure 1B). The number of significant CpG-APA pairs exhibited considerable variation across different cancer types, ranging from thousands to millions of pairs (1,730 to 89,618 for cis and 9,768 to 4,962,326 for trans) (Figure S2; Table S2). This variability is partly attributed to the varied number of APA events and sample sizes observed in these distinct cancers (Figure S3), suggesting that we may have yet to create a complete atlas of methylation-APA associations.

Figure 1.

Atlas of apaQTM across 21 cancers

(A) The number of apaQTMs and mAPAs in each cancer type. cis-apaQTMs are defined as CpG sites that affect APA in cis, cis-mAPAs represent APA events influenced by CpG sites in cis; trans-apaQTMs are CpG sites that affect APA in trans, and trans-mAPAs are APA events influenced by CpG sites in trans.

(B) Distance of cis-apaQTMs from polyadenylation sites (PAS). The locations of the cis-apaQTMs in the apaQTM analysis are shown relative to the PAS for 535,336 pairs.

(C) The average proportion of variation of APA events explained by CpG sites. The histogram shows the number of APA events for a given correlation coefficient. The bottom boxplot illustrates the overall distribution of the APA variation explained by CpG sites, which was at a median of 4.6% for cis and 13.6% for trans.

(D) DNA methylation level was found to be associated with the usage of NFYA 3′-UTR in a previous study; here, we show that four CpG sites are associated with NFYA polyadenylation in skin cutaneous melanoma (SKCM) and stomach adenocarcinoma (STAD). Color represents correlation (r), with red indicating positive correlation and blue representing negative correlation. In these examples, the boxplots demonstrate that high levels of DNA methylation at cg11290720 are associated with increased usage of distal PAS (and therefore have higher PDUI) of NYFA in both SKCM and STAD. p values were calculated by the Wilcoxon rank-sum test. ^∗∗p < 1e-04.

In addition, our analysis estimated that DNA methylation explained 4.6% of the variance in PAS usages for cis-associations and 13.6% for trans (Figure 1C). Among the CpG sites included in the analysis, the percentage of apaQTMs in cis-regulation ranged from 0.86% to 29.14%, with the lowest proportion found in acute myeloid leukemia (LAML) and the highest in LGG. The average absolute value of effect size (|r|) ranged from 0.156 to 0.346, the lowest effect size in breast cancer (BRCA) and the highest in LAML. Furthermore, we found that in cis-elements, the proportion of CpG sites positively correlated with APA ranged from 45.44% to 84.37% across the 21 cancer types. Intriguingly, in LAML, the majority (84.37%) of apaQTMs in cis-elements showed a positive correlation with APA. The average magnitude was comparable between positive and negative correlations, ranging from 0.155 to 0.346 and 0.156 to 0.345, respectively. In contrast, the proportion of apaQTMs in trans-regulation varied significantly among cancers, with the lowest proportion at 2.37% in LAML and the highest at 68.89% in LGG. The average effect size ranged from 0.272 to 0.534. Notably, among trans-acting apaQTMs, 43.61%–94.87% of CpG sites displayed positive correlations with APA, while negative correlations were observed in the range of 11.40%–91.26%. Similarly, the majority of apaQTMs in trans-elements exhibited positive correlations with APA in LAML. The average effect size was again comparable between positive and negative correlations, ranging from 0.272 to 0.531 and from 0.272 to 0.538, respectively (Table S3).

Our analysis generated a comprehensive atlas of DNA methylation and APA events. 13.37% of the APA events regulated by DNA methylation in cis were shared across more than 11 cancer types, while 62.55% were shared in 2∼11 cancers. Over 20% were specific to individual cancer types (i.e., only implicated in one cancer type), with the proportion of cancer-specific trans-APA genes (41.08%) exceeding that of cis-APA genes (Figure S4). Importantly, alteration of PAS usage of several previously reported genes, including CALM1 (MIM: 114180),⁵² CD47 (MIM: 601028),^53,54,55 and MECP2 (MIM: 300005),^56,57 was shown to influence cancer or disease prognoses, providing evidence for further research into its mechanisms. Moreover, we observed concordance with a prior study that identified APA events regulated by DNA methylation via poly(A)-sequencing, such as NFYA (MIM: 189903) (with experimental data), ATF6 (MIM: 605537), and RER1 (MIM: 620048)²³, providing additional support for our results (Figures 1D and S5; Table S4).

apaQTM identification using MR

MR was performed to identify a CpG site whose methylation level was causally associated with the PDUI level of local APA. Due to the specific genetic architecture of DNA methylation, the availability of IVs is limited. In this context, conventional inverse variance weighted (IVW)-based approaches, as advised by MR guidelines, are not frequently employed, given the potential influence of horizontal pleiotropy.⁴¹ To mitigate this concern, we adopted the weighted median estimator, a robust MR approach designed to handle pleiotropic effects. CpG sites with at least three independent mQTLs were considered eligible for this pleiotropy-robust analysis. Linkage disequilibrium (LD) clumping was performed based on an ancestry-matched reference dataset. Leveraging the summary statistics for mQTL and apaQTLs estimated from peripheral blood samples (Material and methods), we conducted a total of 1,651 MR tests involving CpG and local APA pairs. Following multiple comparison corrections, 32 pairs retained significance (FDR <0.05, Table S5). For instance, a higher methylation level of cg10805676 was found to be causally associated with a lower PDUI level of TMEM8A (MIM: 619342) in the same tissue (p = 1.38e-33). Similar patterns were observed for nearby CpG sites cg26944245 and cg02609775 with p = 9.46e-32 and p = 9.18e-20, respectively (Figure S6). Additionally, we found a nominal p value of 0.048 supporting the causal association of cg16579431 with APA in TMEM8A. A cluster of CpG-APA pairs exhibiting significant causal relationships was identified within the human leukocyte antigen (HLA) region, involving genes such as HLA-DRB1 (MIM: 142857) (cg20022036, p = 1.62e-42), HLA-DRA (MIM: 142860) (cg25645491, p = 9.02e-22), HLA-B (MIM: 142830) (cg23923934, p = 1.10e-14), HLA-C (MIM: 142840) (cg17606183, p = 1.55e-9), and HLA-DRA (cg17606183, p = 1.91e-7) (Figure S6).

Characterizing APA events regulated by DNA methylation

Interestingly, we observed that APA genes subject to potential regulation by CpG sites (i.e., mAPA) in cis were significantly enriched in various biological processes, such as translation, RNA localization, and transport. Conversely, genes regulated by trans-elements were notably enriched in cellular components (Figure S7). In addition, we curated lists of oncogenes and tumor suppressor genes (TSGs) from oncoKB and The Cancer Gene Census and identified shared genes from these resources for further enrichment analysis. Our results revealed that mAPA genes exhibited enrichments in both oncogenes and TSGs across several cancer types, including BRCA, cervical squamous cell carcinoma and endocervical adenocarcinoma (CESC), head and neck squamous cell carcinomas (HNSC), lung squamous cell carcinoma (LUSC), skin cutaneous melanoma (SKCM), stomach adenocarcinoma (STAD), urothelial bladder carcinoma (BLCA), LGG, liver hepatocellular carcinoma (LIHC), and lung adenocarcinoma (LUAD) (Figure S8).

We compared the APA events potentially regulated by CpG sites with those affected by SNPs, previously identified in cancers.¹⁷ On average, about 1.0% of the total CpG sites considered in our analysis were classified as cis-apaQTMs and 23.1% as trans-apaQTMs (with |r| ≥ 0.3); only 0.33% were cis-apaQTLs, and 0.25% were trans-apaQTLs.¹⁷ Hence, in comparison, the number of apaQTMs identified in our study significantly outnumbered the apaQTLs identified in TCGA, especially for those in trans-regulation. Importantly, the majority of APA events, accounting for an average of 90.27% in cis- and 95.47% in trans-regulation, controlled by DNA methylation did not overlap with those controlled by genetic variants (Figure S9A; Table S6). When assessing the effect size, apaQTLs exhibited an estimated effect size of 0.128 for cis-associations and 0.167 for trans-associations, with |r| ≥ 0.3. After excluding those with |r| < 0.3 (Figure S9B), the effect size for cis- and trans-apaQTMs were 0.115 and 0.143 (mean values), respectively. These results imply that both cis- and trans-apaQTMs may play crucial roles. In addition, it is noteworthy that many apaQTMs are involved in the regulation of APA events across various cancer types, underscoring their likely relevance to some common pathways in cancer pathogenesis.

Annotation of the apaQTMs identified

As expected, apaQTMs were enriched in the 3′UTR regions, and the significance was more pronounced for cis-acting CpGs than those in trans. Moreover, cis-acting CpGs exhibited enrichment in gene body regions, CpG shelves, and shores, while trans-acting CpGs were relatively more enriched in intergenic regions and open seas (Figure 2A).

Figure 2.

Features of apaQTM

(A) Enrichment of cis-apaQTMs and trans-apaQTMs at positions relative to genomic regions and CpG islands. The y axis represents the fold change (FC) of the enrichment. Point size indicates significance of enrichment. Colors in points indicate the cancer types. Red boxplot represents the enrichment outcomes for apaQTMs in cis-acting regions, whereas the green boxplot represents apaQTMs in trans-regions. As shown, cis-acting CpGs exhibited enrichment predominantly within gene body regions, CpG shelves, and shores, while trans-acting CpGs were enriched within the intergenic regions (IGRs) and open seas. CpG shores: the region within 2 kb upstream and downstream of CpG island; CpG shelf: 2–4 Kb from CpG islands; open sea: other regions except for islands, shores, and shelves. TSS, transcription start site. TSS200: region from TSS to −200 nt upstream of TSS; TSS1500: region between −200 and −1,500 nt upstream of TSS.

(B) Comparisons of apaQTMs with eQTMs and spQTMs, in cis and trans separately. Colors in different bars represent cancer types; the darker colors in each bar indicate the number/proportion of apaQTMs that overlap with eQTMs or spQTMs, while the lighter colors show the total number of identified apaQTMs. Although a considerable proportion of eQTMs and spQTMs overlap with apaQTMs, there remains a significant portion of apaQTMs that do not exert their influence through either gene expression or splicing regulation.

(C) Overview of the apaQTMs that altered the usage of PAS motif across cancers. The x axis displays the cancer type and the y axis lists the number of apaQTMs potentially influencing the PAS signal. Of note, the canonical PAS signal (AATAAA) was found to be frequently implicated.

(D) Enrichment analysis for transcription factor (TF) binding regions among cis- and trans-apaQTMs. Of the 171 binding regions of TFs from ENCODE, 48 of them were enriched in relation to cis-apaQTMs, while 71 of them to trans-apaQTMs. The x axis shows log₂OddsRatio (calculated by LOLA), and the y axis lists the TFs from ENCODE. Point size indicates enrichment significance.

Given that DNA methylation is known for its regulatory functions in gene expression and splicing, we conducted analyses for gene expression quantitative trait methylation (eQTM) and splicing quantitative trait methylation (spQTM) analyses (Supplementary methods). Our results indicated that apaQTMs and spQTMs favored acting in trans, whereas eQTMs were predominantly inclined to cis (Figure S10). This is noted to be akin to how DNA methylation is more frequently associated with non-coding RNAs in trans rather than cis.⁵⁸ We compared apaQTMs and eQTMs and found that, on average, approximately 43.45% cis-apaQTMs were associated with gene expression in 21 cancers. This percentage ranged from 12.98% in pancreatic adenocarcinoma (PAAD) to 83.64% in BRCA. In comparison, the proportion of trans-apaQTMs that overlapped with trans-eQTMs was relatively lower, averaging only 8.08% and ranging from 2.15% to 19.25%. When comparing apaQTMs with spQTMs, our results showed that, on average, about 16.35% of cis-apaQTMs were also cis-spQTMs, with the range spanning from 4.37% to 42.50%. A similar pattern emerged in the trans-analysis, with approximately 18.91% of trans-apaQTMs concurrently associated with splicing (Table S7). In summary, a significant proportion of cis-apaQTMs appeared to be linked to gene expression, while a relatively smaller fraction of trans-apaQTMs exhibited overlap with eQTMs. Concerning alternative splicing, slightly more trans-apaQTMs were shared with spQTMs, which could be attributed to the fact that type III APA events often coincide with alternative splicing events.⁵⁹ These findings collectively provide support for the hypothesis that DNA methylation is intricately involved in APA, alternative splicing, and gene expression (Figure 2B).

Considering genetic control of DNA methylation, we conducted an overlap analysis of apaQTMs identified in all cancers with those associated with mQTLs (n = 11,165,559) in human blood samples, as determined by Hawe et al.⁴² Our analysis revealed that the fraction of cis-apaQTMs regulated by mQTLs in blood ranged from 19.78% to 29.90%, while for trans-apaQTMs, it ranged from 16.56% to 33.16% (Figure S11A). Similarly, the proportion of cis-eQTMs regulated by mQTLs in blood ranged from 20.22% to 30.46%, and for trans-eQTMs, it varied from 9.64% to 20.11%. As for cis-spQTMs, the range was from 21.85% to 30.85%, and for trans-spQTMs, it was 14.15%–24.84%. However, due to limited sample sizes from TCGA,¹⁷ a lower proportion of mQTL-related CpG sites were observed to be cis-apaQTMs (ranging from 1.44% to 6.16%) and trans-apaQTMs (ranging from 0.21% to 14.20%) (Figure S11B). Similar trends were observed for spQTMs and eQTMs.

Clues underlying apaQTMs

Recognition of the PAS motif is essential to produce 3′UTRs of varying lengths. To gain deeper insights into the mechanisms underlying methylation-APA associations, we retrieved 214 cis-apaQTMs within 100 bp upstream/downstream of each putative poly(A) site from the database (PolyA_DB)³⁹ and searched 15 common PAS motifs within the −20/+20 nt flanking regions of the selected cis-apaQTMs (Material and methods).⁶⁰ This investigation revealed a total of 32 cis-apaQTMs that may dictate the usage of PAS hexamers (n = 139) across 21 different cancers. These disruptions frequently altered the recognition of the canonical PAS sequence (AATAAA) (Figure 2C). These findings indicate that a small subset of identified cis-apaQTMs is associated with APA by affecting well-established PAS recognition motifs, highlighting that the mechanisms governing methylation-APA associations remain largely understudied.

Considering that numerous TFs have been reported and validated to be involved in APA processes,^11,61 we proceeded to assess the enrichment of apaQTMs within the binding regions of 171 TFs from the ENCODE project (Material and methods). This analysis revealed significant enrichment of cis-acting apaQTMs in the binding regions of 48 TFs, and for trans-apaQTMs, there were 71 TFs exhibiting enrichment (FDR <0.05). These included TFs such as CTCF, TCF7L2, EZH2, and SUZ12 (Figure 2D). Leveraging the data from GEO: GSE135328,⁶² we compared the PAS usage in colon cancer cell lines with and without TCF7L2 (MIM: 602228 knockdown (Material and methods). Our findings showed that TCF7L2 knockdown induced broad alterations in PAS usage for 429 APA events (Table S8; Figure S12). Of note, a prior study had experimentally shown that the absence of DNA methylation promotes proximal PAS usage through CTCF,²³ suggesting a positive relationship between methylation levels and PDUI. To examine the role of CTCF in methylation-APA regulation, we investigated the enrichment of CpGs that positively correlated with APA in cis, as well as cis-apaQTMs with negative correlations. Surprisingly, we found that cis-apaQTMs were enriched in CTCF binding regions regardless of the direction of correlation (i.e., be it positive or negative), depending on the cancer types. In particular, CTCF was enriched in cis-apaQTMs having positive correlation with APA events in pheochromocytoma and paraganglioma (PCPG), LGG, STAD, and kidney renal papillary cell carcinoma (KIRP). Conversely, in LGG and STAD, CTCF was enriched in cis-apaQTMs with a negative relationship (Figure S13). In total, 13,918 positive and 7,638 negative cis-apaQTMs with CTCF binding sites were identified. Of these, 3,270 CpG sites had both positive and negative relationships with APA, depending on the cancer type. Taken together, we provide evidence suggesting that DNA methylation disrupts APA events by influencing the binding affinity of TFs in a site- and/or cancer-type-specific manner.

Prioritization of cis-apaQTM genes by integrative analysis

To pinpoint notable apaQTMs governing cis-regulation of 3′UTR usage, we incorporated evidence from four analyses as criteria: apaQTM, MR for methylation APA (MR apaQTM), differential methylation between tumor and normal samples (dmCpG; material and methods), and survival analysis-associated apaQTMs (Figure 3A; Table S9, material and methods), whereby fulfilling each criterion confers 1 point to the priority score (hence ranging 1 to 4). In this integrated analysis, we identified 12 mGenes (i.e., genes corresponding to apaQTMs) that exhibited concordant evidence across all four aspects (Figure 3B). One of the genes that piqued our interest was TMEM8A, which encodes a phospholipase A2 with specificity for glycosylphosphatidylinositol (GPI). TMEM8A stood out due to the number of apaQTMs that overlapped with MR apaQTMs (Table S10). We detected four CpG sites with the potential to cis-regulate the 3′UTR usage of TMEM8A, as indicated by evidence from both apaQTM and MR analyses. It is worth mentioning that these CpG sites exhibited relatively weak pairwise correlations (Figure S14), suggesting their independent associations. Further examination using TF footprint data revealed that one of these four CpG sites, namely cg16579431, shared strong binding peaks with nuclear factor kappa B (NF-κB) and CTCF. In addition, the downstream sequence of cg26944245 contained several PAS motifs such as AAAAAA (Figure 3C). Among these four CpG sites, three (cg26944245, cg10805676, and cg16579431) exhibit a negativecis effect on PAS usage, and cg02609775 demonstrates a positive cis effect.. In cancers like BRCA, prostate adenocarcinoma (PRAD), and PCPG, samples with elevated methylation levels of cg16579431 tended to preferentially use the proximal PAS of TMEM8A (p = 4.2e-9; Figure S15). Additionally, cg16579431 was specifically associated with APA and not with splicing or gene expression (Figure 3D). Moreover, the lengthening or shortening of PAS usage of TMEM8A was significantly associated with clinical outcomes in several cancer types (Figure S16). These findings collectively suggest that DNA methylation-mediated regulation of APA events of TMEM8A could play a pivotal role in tumor development and/or disease prognoses. Utilizing the dataset from GSE171073, we identified 86 pairs of cis-regulating CpG sites and APA events that exhibited significant alterations in BRCA patient-derived xenograft (PDX) model samples upon treatment with decitabine, a DNMT inhibitor. Notably, among these markedly changed CpG sites, 30 (34.9%) sites attained 3 points, which indicate that these CpG sites are associated with the occurrence of APA and subsequent consequences of its dysregulation. Taken together, these findings further substantiate the regulatory relationships between these CpG sites and APA events.

Figure 3.

Prioritization of cis-mGenes

(A) A schematic of the prioritization. To prioritize genes possessing apaQTM(s) that regulate 3′UTR usage in target genes (i.e., cis-mGenes), evidence from 4 criteria, i.e., apaQTM, Mendelian randomization (MR), dmCpG in tumor and normal tissues, and associations with prognosis, each conferring a score of 1, were integrated to produce a priority score (ranging from 1 to 4).

(B) The number of mGenes for increasing priority scores from 1 to 4. Symbols of genes with a score of 4 are shown; 12 mGenes with such attributes were found.

(C) Top: schematic illustration of TMEM8A (hg19). The top five tracks show the RefSeq gene structure, the polyA sites predicted by DaPars v2, the poly(A) sites collected by polyA_db, apaQTM locations, and the PAS motifs around the CpG site cg26944245. Green alphabets highlight the PAS motifs, and the CG in orange represents the CpG site, cg26944245. Middle: the x axis shows the corresponding chromosome location of the CpG sites, while the y axis indicates -log₁₀ p values. Blue dots indicate -log₁₀ p values of apaQTM analysis, red for methylation-APA MR analysis, and orange for eQTM analysis. Bottom: ChIP-seq peak of NF-κB and CTCF in chr16:420,000–437,500 (data from ENCODE), corresponding to both panels above, which clearly demonstrate the interaction of cis-apaQTMs with TFs such as NF-κB and CTCF.

(D) Relationship between methylation level of the cg16579431 site and APA, expression, and splicing level of TMEM8A. The boxplots show cg16579431 being specifically associated with APA but with neither splicing nor gene expression. p values were calculated by Wilcoxon rank-sum test. ^∗∗p < 0.01; NS, no significance.

DNA methylation regulated APA in trans

We tallied the frequency of APA events regulated by genes with trans-apaQTMs across 21 different cancers. The top 100 genes are shown in Figure 4A. By referring to TREND-DB, a comprehensive database curated by Marini and colleagues showing the genome-wide impact on APA of 174 proteins involved in various aspects of transcriptional, co- and post-transcriptional gene regulation, and epigenetic modifications, as well as the ensuing processes,¹¹ 154 of them were identified as putative trans-regulators in our study. For instance, DNMT3A, one of the top 100 genes, was implicated in the regulation of APA events for 726 genes. Knock down of DNMT3A in BE(2)-C neuroblastoma cells altered the APA events of 193 genes, as showcased by Marini et al. in the same study (Table S11).¹¹

Figure 4.

The genes broadly involved in the regulation of APA events in trans

(A) Top 100 trans-mGenes. The circle barplot displays the top 100 mGenes that regulate the frequency of APA events, decreasing in clockwise order.

(B) The relationship between PDUI of MMAB or SDC2 with methylation levels of their respective CpG sites. Left panel shows DNA methylation level at cg19256292 was associated with the alternative usage of MMAB 3′UTR in low grade glioma (LGG). The right panel shows DNA methylation level at cg15150970 was associated with the alternative usage of SDC2 3′UTR, also in LGG.

(C) Circos plot for DMNT3A summarizing (1) the DNA-binding sites of DNMT3A identified by ChIP-exo/seq studies (black bars); and (2) the genomic distribution of CpG sites associated in trans (inner connections) with APA events at the DNMT3A DNA-binding sites. Orange lines represent dAPA events in noDox vs. Dox, and blue lines represent dAPA events in ZF-D3A-wt vs. ZF-D3A-mut.

(D) The relationship between PDUI of AK2 with cg15624624. Boxplot shows high DNA methylation level at cg15624624 was associated with the usage of AK2 3′UTR as implicated in thyroid cancer (THCA).

(E) The TCF7L2 binding peak in AK2. The first track shows the ChIP-seq data of TCF7L2 around AK2 in HCT116 cells, and the second track represents the p values, which were obtained from ENCODE with accession no. ENCSR000EUV.

(F) Network of proteins interacting with TCF7L2. Green circles represent mAPA regulated by TCF7L2 in trans and orange circles represent the genes that bind to TCF7L2. Orange-green circles represent the genes both associated with TCF7L2 in trans and binding to TCF7L2. p values were calculated by Wilcoxon rank-sum test. ^∗∗p < 1e-11; ^∗∗∗p < 2.2e-16.

Another study by Mendoza and colleagues employed an artificial zinc finger-DNMT3A fusion protein (ZF-D3A-wt) to induce methylation at thousands of promoters upon treatment with doxycycline (Dox).⁶³ As for negative control, they generated an identical construct containing four amino acid mutations on the catalytic domain of DNMT3A, rendering it incapable of cytosine methylation (ZF-D3A-mut).⁶³ To further the findings by Mendoza et al.,⁶³ we obtained their RNA-seq data (GEO: GSE165891) and estimated the PDUI scores by DaPars2.³⁴

We identified 533 APA events with significantly different PDUI values between noDox and Dox samples (p < 0.05, Table S12). Considering the potential effects of ZF-D3A binding, we compared the differences in PDUI values between ZF-D3A-wt (WD, withdraw) and ZF-D3A-mut samples, finding 318 APA events to be significantly altered (p < 0.05, Table S13). Remarkably, 16 of these 32 genes displayed differential methylation regions induced by Dox treatment, including SDC2 (encoding syndecan-2) (MIM: 142460) and MMAB (metabolism of cobalamin associated B) (MIM: 607568). The APA events of SDC2 and MMAB were significantly associated with DNMT3A in trans, as revealed by our analyses (Figure 4B). Additionally, we corroborated these findings using ChIP-exo/-seq signals (GEO: GSE151287), confirming DNMT3A binding to the genes whose APA events were associated with DNMT3A in trans.⁴⁴ Seventy-seven genes exhibited binding signals for DNMT3A (Figure 4C), among which MMAB is included. Our results provide support for the hypothesis that DNMT3A binding to MMAB leads to changes in DNA methylation status, consequently altering PAS usage. Additional validation is warranted to verify these observations.

Besides, TCF7L2 was found to be involved in the regulation of APA events among 753 genes. The association of TCF7L2 with APA of AK2 (MIM: 103020) (encoding adenylate kinase 2) was detected in ten different cancer types (Figure 4D). Moreover, changes in PAS usage of AK2 were significantly associated with clinical outcomes in several types of cancer (Figure S17). ChIP-seq data further showed TCF7L2 binding to AK2 (Figure 4E). Furthermore, when TCF7L2 was knocked down in HT29 cells, the PDUI values differed significantly (Figure S12). Adenylate kinases are known for modulating the adenine nucleotide composition by catalyzing the reversible transfer of phosphate groups among adenine nucleotides.⁶⁴ To better illustrate the interactions of TCF7L2 and other proteins, we obtained 68 proteins interacting with TCF7L2 using GPS-Prot (http://gpsprot.org/)⁶⁵ and integrated with APA events associated with TCF7L2 and ChIP-seq signals (Figure 4F).

We also explored a mechanistic model in which DNA methylation influences gene expression levels, subsequently regulating the usage of distal or proximal PAS in a trans manner. We combined evidence from trans-apaQTMs, cis-eQTMs, and APA events influenced by gene expression (Figure 5A). We identified 2,404 genes whose DNA methylation levels were associated with APA events in trans, along with their own gene expressions in cis, wherein gene expression levels influenced PDUI values. These selected genes were enriched in processes such as RNA splicing, catabolic processes, and translational initiation (Figure S18). Among these genes, 18 are known APA regulators, including PABPN1 (MIM: 602279),^66,67,68 MBNL1 (MIM: 606516)^69,70, CPSF7,⁷¹ PCF11 (MIM: 608876),^72,73 and RBBP6 (MIM: 600938).⁷⁴

Figure 5.

Integrative analysis of gene expression on APA events and association in trans

(A) Depiction of our hypothesis that DNA methylation sites regulate APA events in trans through control of target gene expression. By combining evidence from trans-apaQTMs, cis-eQTMs, and regression analysis of gene expression on APA events, we found some genes regulated by methylation may play an important role in APA.

(B) Bar plot of intersection size of genes corresponding to trans-apaQTMs with known factors. The bar plot shows intersection size of these genes with known APA regulators, transcription factors, and RBPs.

(C) Circos plot for MBNL1 summarizing (1) the DNA-binding sites of MBNL1 identified by ChIP-exo/-seq studies (black bars) and (2) the genomic distribution of CpG sites associated in trans (inner connections) with APA events at the MBNL1 DNA-binding sites.

(D) The interrelationship of DNA methylation sites, gene expression, and the APA event. The Sankey plot shows the DNA methylation sites (DNAm, left), genes (middle), and the APA events (right). In this study, we chose MBNL1, PABPN1 as examples of known APA regulators, and OAS2, OAS3 as examples of RBPs. Here, we listed the CpG sites that regulate gene expression and thus affect APA events. Meanwhile, the expression of the gene is also associated with APA events.

We then asked whether known APA regulatory genes mediate the regulation of APA by CpG sites in BRCA (Figure S19A). We conducted a mediation analysis and identified a total of 2,146 CpG sites showing 28,324 associations with APA events and the expression of APA regulatory genes. Of these associations, 13.19% (3,736 out of 28,324) of these apaQTMs were found to affect APA mediated by gene expression of APA regulatory factors (Table S15). In this regulatory process, the top three APA regulatory factors acting as intermediaries were CDK12, CPSF7, and PTBP1. CDK12, also as known as cyclin-dependent kinase 12, phosphorylates the C-terminal domain of RNA polymerase II, thereby participating in the regulation of transcription processes.⁷⁵ Furthermore, CDK12 is intricately linked to DNA damage repair with a specific involvement in homologous recombination (HR).⁷⁶ CDK12 deficiency can lead to alternative usage of APA, which disrupts the expression of HR repair genes, resulting in increased DNA damage and genomic instability.⁷⁷ This defect induces a cellular condition resembling deficiencies in BRCA1 (MIM: 113705) and BRCA2 (MIM: 600185) genes, known as HR deficiency or BRCAness.⁷⁸ On the bright side, several studies and clinical trials suggest that CDK12 (MIM 615514) mutations may exhibit higher sensitivity to certain treatment strategies, such as poly(ADP-ribose) polymerase (PARP) inhibitors.^79,80,81 Our analysis reveals that a total of 544 pairs of CpG sites regulate APA through the intermediary role of CDK12, which is comprised of 108 CpG sites and 93 APA events. Furthermore, our results reveal that certain CpG sites can regulate downstream APA events by modulating the expression of APA regulatory genes on which they are located. This mode of regulation is observed in APA regulatory genes such as MBNL1, MBNL2 (MIM: 607327) and PAPOLG (MIM: 616865) (Figure S19B). Overall, our findings indicate that CpG sites have complex regulatory roles in APA regulatory networks, influencing the occurrence and regulation of downstream APA events through the influence of APA regulators. This provides a more comprehensive perspective for a deeper understanding of APA regulatory mechanisms, highlighting the multilayered regulatory role of methylation in this process.

Besides known APA regulators, growing evidence has demonstrated that TFs and RBPs can regulate APA.^2,11,61,73 Of the 2,404 genes, we discovered candidate RBPs (381 genes encoding RBPs, such as C1D [MIM: 606997] and H3F3C [MIM: 616134]) and TFs (132 genes intersecting with TFs, including the ZNF and HOX families) that may play a role in regulating polyadenylation (Figure 5B; Table S14). For example, PABPN1, which encodes poly(A) binding protein nuclear 1, is a well-known regulator of APA. We observed that PABPN1 was associated with 18 APA events, including AP2B1 (MIM: 601025) and MRPL42 (MIM: 611847), both of which were found to exhibit significant shifts in PAS usage in PABPN1 knockdown samples.⁸² Notably, five CpG sites (cg21526750, cg18454133, cg10250660, cg08188555, and cg07197059) were found to potentially influence the PDUI of PATL1 through their effects on the expression of PABPN1. MBNL1, a member of the MBNL protein family (consisting of MBNL1, MBNL2, and MBNL3) with RNA-binding sites, was linked to alternative PAS selection secondary to depletion of Mbnl proteins.⁶⁹ We found that MBNL1 might lead to alternative PAS selection of SUPT6H (MIM: 601333) (Figure 5C), a phenomenon previously observed in cells with deficient Mbnl activity.⁶⁹ Two CpG sites in MBNL1 (cg00661523 and cg14423778) showed evidence of potentially influencing the APA of SUPT6H through their effect on the expression of MBNL1.

In addition to these reported APA regulators, we identified other candidate RBPs and TFs that may play a role in regulating polyadenylation. For example, OAS2/OAS3, members of the 2-5A synthetase family, were found to potentially affect the PDUI of NFYA, with five DNA methylation sites associated with this process. Among TFs, we noticed that a cluster of methylation sites had effects on the expression of the HOX family genes, thereby influencing PAS selection in CMPK1 (MIM: 191710) (Figures 5D and S20). Dysregulated APA events of SUPT6H, NFYA, and CMPK1 have been previously identified in methylation-deficient cells by poly(A)-seq.²³

Collectively, our results provide valuable insights into the novel roles of TFs and RBPs in APA regulation (Figure S21) and clarify the intricate relationships between DNA methylation, gene expression, and APA.

Theorized regulatory mechanisms underlying cis- and trans-apaQTMs

In cis-apaQTMs, DNA methylation sites may regulate APA positively or negatively (i.e., positive correlation indicates higher methylation levels in a gene with a preference for distal PAS, while negative correlation indicates higher methylation levels in a gene though with a preference for proximal PAS) (Figure 6). For trans-apaQTMs, we propose three possible models (Figure 6). Firstly, CpG sites on methyltransferases (e.g., DNMT3A), which influence their own expression, would theoretically result in broad impact on the methylation levels of downstream genes, subsequently altering PAS usages of target genes, whether through cis or trans. Secondly, methylation sites on RBP genes or cleavage and polyadenylation-specific factors, which are known regulators of APA events, should similarly exhibit broad effects. These CpG sites regulate gene expression of their own, and thus would also broadly impact PAS usage. Lastly, CpG sites on TFs impact the expression of these factors themselves, thereby influencing APA. Concerning trans-regulation of non-coding RNAs, we have not yet delved into this aspect and plan to incorporate it into our future studies.

Figure 6.

Theorized regulatory mechanisms underlying cis- and trans-apaQTMs

For cis-apaQTMs (distance between CpG site and PAS ≤1 Mb), DNA methylation sites may regulate APA positively or negatively (i.e., positive correlation being distal PAS preference in a gene with higher methylation levels, and in contrast, negative correlation being proximal PAS preference in a gene with higher methylation levels). For trans-apaQTMs (distance between CpG site and PAS >1 Mb or lie on different chromosomes), we propose three hypotheses based on analysis of apaQTMs, eQTMs, and regression analysis of gene expression with APA: (1) methylation sites on DNA methylation regulators, such as DNMT3A, modulate their own expression and alter PAS usage of target genes by widely perturbing methylation levels of genes; (2) CpG sites on RBPs or cleavage and polyadenylation-specific factors, which are known regulators of APA events, modulate gene expression of these factors, which in turn lead to the change of PAS usage; (3) similarly, CpG sites on TFs can affect gene expression of TFs themselves, thus altering PAS usage. Lastly, similar to cis-apaQTMs, trans-apaQTMs also can have positive or negative correlation with APA events (e.g., genes with higher DNA methylation levels have lower gene expression, which induce proximal PAS usage). PAS, poly(A) site; pA1, proximal PAS; pA2, distal PAS; RBP, RNA binding protein; TF, transcription factor.

Colocalizing eQTMs, spQTMs, and apaQTMs with EWAS for cancers

We repurposed the SNP-based colocalization approach to identify disease-associated APA, gene expression, and splicing events, using CpG sites as marks.⁴⁹ Consistent with the assumptions of the original colocalization framework, we operated under the premise that each tested region has a single causal CpG site. A high regional level PP H4 would indicate that the apaQTM/eQTM/spQTM is colocalized with the EWAS signal. Moreover, we could pinpoint the potential causal CpG site with the site-level PP H4.

Our analysis identified a large number of QTMs that shared the same putative causal CpG sites across five different cancer types (Figure 7A). Expectedly, we observed the highest number of genes among eQTMs that were colocalized with specific cancers. For example, over 900 genes were colocalized with BRCA, implicating a causal relationship between these eQTMs with BRCA. Similarly, substantial number of genes related to apaQTMs and spQTMs (57 and 48, respectively) were found to be colocalized with EWAS signals associated with BRCA. Notably, 46 out of 57 APA events exhibited specific colocalization with BRCA (Figure 7B). As an example, our results reveal that a transcript of the FOXP1 (MIM: 605515) gene (GenBank: NM_032682) specifically colocalizes with BRCA. This finding highlights the potential significance of FOXP1 in the development of BRCA, providing valuable clues for further exploration of its molecular mechanisms. Collectively, these results underscore the significance of QTMs, including apaQTMs, for carcinogenesis and provide clues on causal genes. Given these results, further research into the role of APA in cancer development is strongly warranted.

Figure 7.

Colocalization of QTMs with cancers

(A) Barplot for colocalization analysis of EWAS signals and 5 cancers. Total number of colocalized EWAS signals (y axis) for 5 cancers (x axis).

(B) Number of QTM-colocalizing genes, colored based on the gene colocalized with each type of QTM. Blue represents genes of eQTMs colocalizing with spQTMs and traits. Yellow indicates the genes are shared by three types of QTMs colocalizing with traits. Pink represents genes of apaQTMs colocalizing with spQTMs and traits. Green represents genes of apaQTMs colocalizing with eQTMs and traits. Red represents genes of apaQTMs colocalizing with traits only.

Pancan-apaQTM: A web resource of associations between CpG sites and APA

Pan-cancer apaQTM (Pancan-apaQTM) is a database we developed to provide a comprehensive resource of apaQTMs across 21 human cancer types (https://pancan-apaqtm-zju.shinyapps.io/pancanaQTM/). As of now, Pancan-apaQTM offers users the following: (1) to search for cis-apaQTMs and trans-apaQTMs across diverse cancer types, (2) to search for cis-mAPA genes and trans-mAPA genes across different cancer types, and (3) to download all the results across different cancer types. For instance, on the “Search for apaQTMs” and “Search for APA genes” tabs, we provide a variety of flexible search options. Users can select cancer types (e.g., COAD) and cis-/trans-modes of regulation and input CpG sites (e.g., cg08269402) or gene symbols to conduct searches for apaQTMs. Additionally, users have the capability to explore APA events and associated apaQTMs by just entering gene symbols (e.g., TMEM8A) within specific cancer types and selecting cis-/trans-modes of regulation. Furthermore, to meet the diverse research and analysis requirements of users, we offer a convenient download option. Users can download cis-apaQTMs and trans-apaQTMs for each cancer type from the “Download” section. Detailed information regarding data collection and processing can be found under the “Help” tab. Pancan-apaQTM encourages users to share their feedback via email, and the email addresses of correspondents can be found under the “Contact” tab.

Discussion

We present ample evidence demonstrating the extensive involvement of DNA methylation in the regulation of APA events in various human cancers. Like genetic variants, methylated DNA may be regarded as a cis-element or a trans-acting factor that modulates APA. The variation of APA explained by DNA methylation was found to be comparable with the variation explained by SNPs. We generated a comprehensive atlas of apaQTMs in 21 different cancer types, ranging from 1,612 to 60,219 in cis and 4,448 to 142,349 in trans. The apaQTMs could be searched and retrieved from the Pancan-apaQTM database (https://pancan-apaqtm-zju.shinyapps.io/pancanaQTM/). Notably, among the cancers, we observed the highest number of apaQTMs as well as mAPA events in LGG. This observation aligns with previous research emphasizing the integral role of APA in LGG.^68,83,84 Because these hyper-methylated loci are enriched in LGG,⁸⁵ DNA methylation-based profiling of gliomas has led to its inclusion in the clinical classification of gliomas in the 2021 5th edition of the WHO Classification of Tumors of the Central Nervous System, underscoring the ever expanding importance of understanding DNA methylation and its implications for human health. In line with previous reports, our findings strongly suggest that DNA methylation plays a far greater role than previously thought and provide strong impetus to further the work of elucidating the association between DNA methylation and APA.

Recently, Huang and colleagues conducted apaQTM analysis in human cancers and constructed a web-based platform, CAFuncAPA.⁸⁶ This platform offers association results between APA and DNA methylation. However, neither a deep functional characterization of the apaQTMs nor the molecular mechanisms governing APA regulation was explored. To the best of our knowledge, our study represents the pioneering comprehensive investigation into the cis- and trans-regulation of apaQTMs. Previous studies have identified genetic variants enriched in 3′UTRs.¹⁸ Consistently, the apaQTMs acting in cis were enriched in 3′UTRs, which are known to frequently contain cis-regulatory elements.⁸⁷ Consistent with the differences in the number of eQTMs and eQTLs,⁴² we observed in our study that the number of apaQTMs exceeded that of apaQTLs identified in the TCGA dataset and is especially so for those in trans-regulation. It is important to highlight that the majority of APA events potentially regulated by DNA methylation differed from those associated with genetic variants. This implies that a large proportion of interindividual differences in APA within a given cancer type can be attributed to variations in methylation levels rather than genetic variations. In our findings, approximately 25% of the cis-apaQTMs were CpG sites whose DNA methylation levels were under genetic control, as identified in previous mQTL studies conducted in blood samples. The CpG sites less influenced by genetic variants were primarily responsible for the regulation of APA events. Given that both genetic variants and epigenetic modifications participate in transcriptional regulation, it is plausible that epigenome dynamics contribute to precise control over gene expression, operating at both the subcellular and supracellular levels.

The present findings contribute to our comprehension of APA regulation and broaden our understanding of the alternative roles of DNA methylation. Although the impact of DNA methylation on complex traits is well documented, deciphering the functional consequences of CpG sites lacking canonical methylation-expression associations (i.e., non-eQTMs) presents challenges. DNA methylation is known to partake in modulation of splicing and post-transcriptional regulation, the latter of which includes APA. Our observations suggest that DNA methylation predominantly exerts its effects through trans-regulation rather than cis-regulation in the context of splicing and APA, since a higher number of trans-spQTMs and trans-apaQTMs were identified compared to cis-eQTMs mapped. However, Kim et al.⁸⁸ observed that proportionally more eQTMs were trans-acting in airway epitheliums. The discrepancy between their results and ours could be attributed to the use of different multiple testing correction methods. They adopted FDR correction, which is less stringent than the Bonferroni approach we used for adjustment. In scenarios where APA occurs in the upstream region of a 3′UTR and is subjected to regulation similar to alternative splicing,⁸⁹ our analysis revealed the extensive occurrence of spQTMs, apaQTMs, and eQTMs. These findings have the potential to uncover the unexplored biology of EWAS hits, thereby igniting interest in the functional interrogation of CpG sites beyond eQTMs. Results from a very recent publication⁸⁸ and our study jointly emphasize the significance of considering CpG sites as either cis-acting or trans-regulation elements. Our study exemplifies how integrating DNA methylation data with other omics datasets in normal tissues or disease contexts can provide a more comprehensive view of molecular phenotypes and their possible associations.

Due to the fact that DNA methylation status affects the binding of factors responsible for polyadenylation or PAS recognition, our analysis of TF binding site enrichment has provided fresh insights into the mechanistic understanding of APA regulation. Both cis-apaQTMs and trans-apaQTMs showed enrichment in the binding regions of numerous TFs. An illustrative case is TCF7L2, a gene intricately involved in the trans-regulation of APA events for hundreds of genes. Expectedly, perturbing TCF7L2 through knockdown led to widespread alterations in PAS usage, underscoring the crucial role of TCF7L2 in shaping APA events in malignancies. TFs may exert their influence directly or indirectly, potentially via the recruitment of methyl-CpG-binding domain proteins. Research by Kaluscha et al. has provided evidence indicating that direct inhibition of transcription factor binding is the prevailing mode of gene and repeat repression mediated by DNA methylation.⁹⁰ Nonetheless, the precise mechanisms and the extent to which DNA methylation governs APA regulation certainly necessitate further investigation through experimental studies.

Additionally, TCF7L2 is implicated in a wide array of human diseases, encompassing type 2 diabetes and various cancers.⁹¹ As a critical TF within the Wnt/β-catenin signaling pathway, TCF7L2 assumes regulatory roles over a diverse spectrum of gene expressions. Here, our findings implicate a new function of TCF7L2 in APA regulation. In multiple cancer types, we observed a trans-association between TCF7L2 and the APA event of AK2, suggesting that TCF7L2 binding to AK2 may induce changes in PAS utilization. Furthermore, this APA event within AK2 emerged as a prognostic indicator in multiple cancer types, implicating the potential involvement of AK2 in cancer pathogenesis. Adenylate kinases are known to modulate adenine nucleotide composition by catalyzing the reversible transfer of phosphate groups among adenine nucleotides.⁶⁴

Ogorodnikov and colleagues studied the depletion of 174 proteins involved in various processes of transcriptional, co- and post-transcriptional gene regulation, and epigenetic modifications, as well as their subsequent processes, evaluating their impacts on APA.^11,73 Notably, the vast majority of these genes (88.5%) were identified as trans regulators in the present study. Within our analysis, we uncovered several established methylation regulators, such as DNMT3A, TET1, and KDM2B, all of which were implicated in the trans-regulation of hundreds of APA events. Marini et al. found that the knock down of DNMT3A in BE(2)-C neuroblastoma cells perturbed the APA events of 193 genes.¹¹ A fusion of the DNMT3A catalytic domain and an artificial zinc finger was designed to manipulate DNA methylation states by a doxycycline inducible system.⁶³ To further the findings by Mendoza et al., we obtained RNA-seq data (GEO: GSE165891) from them and estimated the PDUI scores by DaPars2. We observed hundreds of APA events with significant differences after doxycycline treatment. This suggests that DNMT3A may exert its influence by binding to and controlling the methylation status of numerous target genes, consequently altering PAS usage. Additionally, DNA methylation can directly obstruct the binding of APA-related factors, including DNA methylation-sensitive proteins, potentially leading to shifts in PAS utilization.

We also identified several trans-apaQTMs associated with well-known genes participating in the regulation of a number of APA events. For instance, the coding gene for poly(A) binding protein nuclear 1 (PABPN1), which is an RNA-binding protein known to control APA events, displayed trans-apaQTMs. Previous research has demonstrated that under PABPN1 knockdown conditions, there is a decrease in intronic polyadenylation and a global shortening of 3′UTRs.^82,92 Previous studies have demonstrated that PABPN1 affects PAS selection. PABPN1 exerts its influence by directly interacting with PAS regions, competing with cleavage and polyadenylation complexes, thereby dictating the usage of PAS. Similarly, muscleblind-like (MBNL) proteins, known to bind to hundreds of pre-mature and mature mRNAs, play a role in the regulation of APA, mRNA stability, and subcellular localization. As illustrated by Batra et al., cells with depleted Mbnl1 and Mbnl2 proteins tend to preferentially use proximal PAS sites for polyadenylation.⁶⁹ In our study, we found that CpG sites within the MBNL1 and MBNL2 genes were associated with alterations in more than three hundred APA events. It is plausible that changes in DNA methylation status affect gene expression of MBNL1 and MBNL2, subsequently influencing the choice of PAS sites in target mRNA molecules.

Furthermore, we have provided multiple lines of evidence elucidating the potential mechanism of cis-regulation of TMEM8A APA events via DNA methylation in cancers. Specifically, we identified a CpG site, cg16579431, which displayed an association with altered PAS utilization. MR analysis further substantiates a causal role of cg16579431 in APA regulation. Importantly, this CpG site is located within the binding region of CTCF and NF-κB, as confirmed by ChIP-seq data. We speculate that the DNA methylation status of cg16579431 influences the binding of CTCF, thus perturbing the recruitment of other regulator factors. It is noteworthy that cg16579431 exhibited a specific association with APA, but not splicing or gene expression in breast cancer. TMEM8A, also known as PGAP6, is a phospholipase A2 with specificity to GPI.⁹³ As a member of the transmembrane protein family, it is broadly expressed on the surface of various cell types.⁹³ Our knowledge and investigations of TMEM8A have thus far been limited, and our findings served to highlight the potential significance of TMEM8A in the pathogenesis of BRCA.

In a recent publication, Sun et al. observed that a fraction (0.80%) of CpG pairs exhibit co-methylation with a Pearson’s correlation coefficient |r| > 0.8 in tumor samples.⁹⁴ As a limitation of our study, we did not examine these co-methylation pairs in our current analysis, acknowledging the fact that our investigation only covered a small fraction of the 30 million chip-based CpG sites within the human genome. In future studies, whole-genome bisulfite sequencing will provide a full spectrum of epigenetic regulation of transcriptional events. Because variation in cell-type composition can confound eQTM mapping,⁴² Liu et al. incorporated general covariates and five PEER factors into their final eQTM model, as the eQTM discovery rate tended to stabilize after correcting for more PEER factors.⁹⁵ Accounting for impact of cell types should follow correction for PEER factors. The application of multi-omics profiling at the single-cell level can help to elucidate the active role of DNA methylation in molecular processes like APA, addressing the influence of tissue heterogeneity.

In the context where all genetic instruments are positioned within the same gene region, we adhered to the MR guidelines and adopted the weighted median approach for genetic-based causal inference. While the summary-data-based Mendelian randomization and heterogeneity in dependent instruments (SMR-HEIDI) approach is commonly used to investigate gene-trait associations in the context of MR, the framework is unable to discriminate between “pleiotropy” and “causality.” The weighted median approach, however, is robust against outliers and furnishes a reliable estimate of potential causality when more than half of the IVs are valid.

Our study characterized a comprehensive landscape of how DNA methylation significantly impacts transcriptional regulation, especially on APA events. These findings offer fresh insights into the epigenetic control of APA events and assert the role of several master regulators. Further experimental investigations are required to fully elucidate the precise molecular mechanisms involved. Taken together, our findings expand our understanding of the biological function of DNA methylation underlying complex traits and may pave the way for the identification of new and viable therapeutic targets.

Data and code availability

All data used in this paper are publicly available (see web resources) or controlled by dbGaP (accession number phs000424.v6.p1). Usage of data is in compliance with the data use agreements of each respective source. The results generated in our study can be accessed in our web portal (https://pancan-apaqtm-zju.shinyapps.io/pancanaQTM/). The codes generated during this study are available at https://github.com/Zymaniswenzai/Pancan-apaQTM/.

Published: February 16, 2024

Web resources

• APAatlas, https://hanlab.uth.edu/apa/

• dbGaP, https://www.ncbi.nlm.nih.gov/gap/

• ENCODE, https://www.encodeproject.org/

• EWAS Data Hub, https://ngdc.cncb.ac.cn/ewas/datahub/

• GEO, https://www.ncbi.nlm.nih.gov/geo/

• GTEx Portal, https://gtexportal.org/home/

• OMIM, https://www.omim.org/

• PanCan-mQTL, http://bioinfo.life.hust.edu.cn/Pancan-meQTL/

• polyA_DB, https://exon.apps.wistar.org/PolyA_DB/v3/.

• R, https://www.r-project.org/

• SNP2APA, http://gong_lab.hzau.edu.cn/SNP2APA/

• Synapse, https://www.synapse.org/

• TCGA SpliceSeq, https://bioinformatics.mdanderson.org/TCGASpliceSeq/PSIdownload.jsp/

• TC3A, http://tc3a.org/

• UCSC Xena browser, https://xenabrowser.net/