Demonstrating the Application of Explainable Deep Learning in Human Genetic Studies

Genome-wide association studies (GWAS) have been instrumental in discovering these loci, and they have successfully identified numerous disease-associated loci for other complex traits and diseases. Despite the stunning success of GWAS at identifying this many loci, only a relatively small number of disease/trait-associated single nucleotide polymorphisms (SNPs) have been functionally characterised due to various scientific, technical, methodological and funding challenges. The fact that the vast majority of GWAS risk variants with low penetrance are situated in non-coding regions that affect gene regulation6 poses many methodological challenges in fine-mapping the causal variant, identifying the genes it regulates and inferring molecular mechanisms of diseases. By unravelling the disease mechanisms and providing fundamental biological insights, improved evidence-based and personalised treatments can be offered. Existing in vivo models lack the sensitivity in observing phenotypes for low penetrance risk variants as biological robustness, intrinsic to these systems, can confound the results. New approaches are therefore required for rapid translation of GWAS findings to medical interventions. We propose a novel solution using explainable deep learning (EDL), to accelerate the process of GWAS gene discovery and mechanistic understanding of disease. This approach can reveal the complex mechanisms leading to genotype-phenotype associations, a fundamental requirement to model ubiquitous systemic properties such as polygeny, pleiotropy and robustness intrinsic to complex biological systems. We will build upon existing proof-of-concept EDL models and, for the first time, demonstrate its application in human genetic studies using disease-specific GWAS data.