Atlas of epistasis

Abstract

We performed a genome-wide epistasis search across 502 phenotypes in case control matched cohorts from the UK Biobank. We identified 152,519 genome wide significant interactions in 68 distinct phenotypes, and 3,398 interactions in 19 phenotypes were successfully replicated in independent cohorts from the Finngen consortium. Most interactions (79%) involved variants that did not present significant marginal association and might explain part of the missing heritability for these diseases. In 10 phenotypes we show the presence of epistasis between common variants with intermediate to large effect size (OR>2) supporting the hypothesis that common diseases are modulated by common variants. Most of the variants in interactions (82%) were more than 1Mb apart and cis-epistasis was hardly found outside the HLA region. Functional annotation of the variants suggests that most mechanisms behind epistasis occurs at the supra pathway level and that intra-gene or intra-pathway epistasis is rare. Surprisingly we find a significant biais toward antagonistic epistasis, representing 60% to 95% of interactions. In type 1 diabetes, hypothyroidism, disorders of mineral absorption, rheumatoid arthritis, asthma, and multiple sclerosis more than 50% of interactions were completely compensating the effect of the marginally associated variant. In psoriasis we identified an interaction between a stop gain variant in CCHCR1 with two missense variants in MUC22 and HSPA1L leading to a 3 fold increase of the effect of CCHCR1 variant on disease risk. Our study shows that there is still much to discover in epistasis and we provide the full summary statistics results to researchers interested in studying epistasis.

Deep integrative models for large-scale human genomics

“…When filtering disease traits for where GLN and LASSO had better performance compared with covariates and difference of at least 0.01 ROC-AUC, we found Figures 17 and 18). Interestingly, the GLN model performed markedly better on T1D, rheumatoid arthritis, multiple sclerosis, psoriasis and ulcerative colitis, all autoimmune traits in which studies have shown indication of interaction effects 49,50,[57][58][59][60][61] . For instance, for rheumatoid arthritis, the GLN model had a ROC-AUC of 0.665 while the LASSO had a ROC-AUC of 0.624 on the test set and the covariate only models achieved a ROC-AUC of 0.622 and 0.634 for the LASSO and NN based models, respectively (Supplementary Figure 19).…”

Novel Alzheimer’s disease genes and epistasis identified using machine learning GWAS platform

“…Part of the missing heritability in LOAD might be explained by non-additive interactions [10], which are ignored by GWAS studies. Indeed, a genome-wide replicated scan has found epistasis to be a ubiquitous phenomenon across multiple phenotypes [11]. Epistatic interactions have long been implicated in complex genetic disease, including neurological diseases [12] and LOAD itself [13].…”

“…University Of California, San Francisco, USA 11. University Of Southern California, Los Angeles, USA 12.…”

Genotype Pattern Mining for Pairs of Interacting Variants Underlying Digenic Traits

“…In the Methods section, we outline some approaches to the detection and analysis of epistatic interactions in genetics, with particular emphasis on data mining and machine learning methods [11]. A recent investigation of quantitative trait loci in mice significantly demonstrated multiple epistatic interactions, while single-locus analyses were far from significant [12], and largescale investigations into epistatic effects on disease phenotypes are currently underway [13].…”