Deciphering colorectal cancer genetics through multi-omic analysis of 100,204 cases and 154,587 controls of European and East Asian ancestries

Colorectal cancer (CRC) is a leading cause of mortality worldwide. We conducted a genome-wide association study meta-analysis of 100,204 CRC cases and 154,587 controls of European and Asian ancestry, identifying 205 independent risk associations, of which 50 were unreported. We performed integrative genomic, transcriptomic and methylomic analyses across large bowel mucosa and other tissues. Transcriptome- and methylome-wide association studies revealed an additional 53 risk associations. We identified 155 high confidence effector genes functionally linked to CRC risk, many of which had no previously established role in CRC. These have multiple different functions, and specifically indicate that variation in normal colorectal homeostasis, proliferation, cell adhesion, migration, immunity and microbial interactions determines CRC risk. Cross-tissue analyses indicated that over a third of effector genes most likely act outside the colonic mucosa. Our findings provide insights into colorectal oncogenesis, and highlight potential targets across tissues for new CRC treatment and chemoprevention strategies.

The data submitted here are expression and methylation models with LD reference data for the transcriptome-wide (TWAS), methylome-wide (MWAS) and transcript isoform-wide association study (TIsWAS) as described in the manuscript "Deciphering colorectal cancer genetics through multi-omic analysis of 100,204 cases and 154,587 controls of European and East Asian ancestries". Details of the methods are presented in the method section and supplementary information file.

TWAS analysis

Gene expression models for the six in-house expression datasets were generated using the PredictDB v7 pipeline for a total of 1,077 participants. Elastic net model building with 10-fold cross-validation was performed independently for each dataset. The elastic net models for GTEx v8 Colon Transverse were obtained from the PredictDB data repository (http://predictdb.org/) and had been generated using the same pipeline. Models were computed using HapMap2 SNPs ±1Mb from each gene, together with covariate factors estimated using PEER32, clinical covariates when appropriate (age, sex and, where appropriate, case-control status, type of polyp and anatomic location in the colorectum), and three PCs from the individual dataset’s SNP genotype data.

Transcript-based TWAS analyses (TIsWAS) were likewise performed by using transcript-level data from the SOCCS, BarcUVa-Seq and GTEx Colon Transverse datasets.

MWAS analysis Methylation beta values were calculated based on the manufacturer’s standard, ranging from 0 to 1. Quality control and data normalization were performed in R using the ChAMP software pipeline for the EPIC and 450K arrays. Briefly, we filtered out failed probes with detection P > 0.02 in >5% of samples, probes with <3 reads in >5% of samples per probe and all non-CpG probes. Samples with failed probes >0.1 were also excluded from downstream analyses. We discarded all probes with SNPs within 10bp of the interrogated CpG (from 1,000 Genomes Project, CEU population)34, and probes that ambiguously mapped to multiple locations in the human genome with up to two mismatches33. We only considered probes mapping to autosomes and those overlapping between the EPIC and the 450K arrays. Normalization was achieved using the Beta MIxture Quantile (BMIQ) method. Per probe methylation models were created using the PredictDB pipeline on the normalized methylation matrix and the genotypes as per TWAS eQTL analysis. To optimize power, we restricted our analysis to 263,341-238,443 (for the 450K array) and 377,678 (for the EPIC array) probes annotated to Islands, Shores and Shelves, and discarded “Open Sea” regions.