Biologically-informed Interpretable Deep Learning Framework for Phenotype Prediction and Gene Interaction Detection

The detection of epistatic effects has significant potential to enhance understanding of the genetic basis of complex traits, but statistical epistatic analysis methods are complex and labour intensive. In recent years, Deep Neural Networks (DNNs) have emerged as a powerful tool for modelling arbitrarily complex genetic interactions in relation to a phenotype; however, their utility is often limited by the challenge of interpreting their predictive reasoning. Although DNN interpretation methods exist, they are typically not designed for genomic applications, leading to hard-to-understand outputs with limited relevance to the field. To address this gap, we introduce GENEPAIR – a novel DNN interpretation framework designed specifically for genomic data, aimed at detecting putative associated gene-gene interactions for a phenotype of interest. Our approach offers several key advantages including model agnositicity, robustness to sample- and variant-level data variance, and flexibility to integrate varied domain knowledge into interpretable features. We demonstrate the efficacy of our method by applying it to a DNN trained on genetic variant data to predict Body Mass Index (BMI). The results of the analysis not only reveal single gene influences in close alignment with literature but also uncover previously unreported gene-gene interactions, demonstrating its significant potential for genomic discovery.