Allelic imbalance of chromatin accessibility in cancer identifies candidate causal risk variants and their mechanisms

Abstract

Understanding the functional impact of non-coding variants remains a major challenge in cancer genetics. In the case of non-coding germline risk associations from GWAS, many associations are in high LD and cannot be resolved by statistical fine-mapping. In the case of non-coding somatic drivers from tumor sequencing, existing datasets are generally underpowered for discovery with frequency-based methods. For both domains, methods that can infer the functional impact of individual variants within regulatory elements are thus urgently needed.

Here I studied allelic imbalance of chromatin accessibility in 406 ATAC-Seq samples across 23 cancer types. I employed a statistical model of allelic imbalance that aggregates signals across individuals while modeling individual-level somatic copy number variation. I discovered 7,262 germline allele-specific accessibility QTLs (as-aQTLs) and found that they are highly enriched for cancer risk heritability across seven common cancer GWAS (e.g., prostate cancer as-aQTLs with a 145±35.7 (p=6.3x10-5) fold enrichment for prostate cancer risk) and are largely explained by genetic variants that directly alter transcription factor binding and gene expression. To connect as-aQTLs to putative risk mechanisms, I introduced the Regulome-Wide Associations Study (RWAS). RWAS identified accessible peaks genetically associated with cancer risk at >70% of known breast and prostate loci (compared to % for a conventional Transcriptome-Wide Association Study) and discovered novel risk loci in all examined cancer types.

To estimate allele-specificity for variants that could not be tested, I developed siamAS, a predictor of allelic imbalance that uses a novel “Siamese” neural network approach that trains two mirrored networks on allele-specific features. I trained siamAS on my allele specificity data incorporating >7,000 features from multiple variant effect predictors. In a hold-out data set, siamAS achieved a classification AUC of 0.87, which substantially outperformed any individual predictive feature. Finally, I applied siamAS to variants within regulatory elements and non-coding somatic mutations in TCGA. siamAS identified germline variants that are enriched for cancer risk heritability and non-coding somatic mutations that result in allelic imbalance.

In summary, my results establish cancer as-aQTLs, RWAS, and siamAS as powerful tools to study the genetic architecture of cancer risk.