Identification of Important SNPs using Bayesian Deep Learning on Whole-Genome Arabidopsis thaliana Data

With the advent of Genome-Wide Association Studies (GWAS) and sequencing techniques, researchers can now examine the genetic basis of various traits and diseases. GWAS present their challenges, primarily originating from the high dimensionality of the Single Nucleotide Polymorphism (SNP) data. Deep Learning (DL) has been manifested as a promising method for such high-dimensional data. This article proposes a DL-based workflow, BayesDL, to identify potentially important SNPs from the Arabidopsis thaliana data. BayesDL incorporates Bayesian methodologies to DL by using filtered SNPs obtained from Chi-square and ANOVA tests. We use coefficient of variation in BayesDL to guide the identification of important SNPs. The SNPs obtained from BayesDL are subsequently validated, genes are pinpointed, and a comparison is drawn with the results from GWAS software GAPIT and TASSEL. Our results show that BayesDL not only mitigates over-fitting but also performs better than the traditional Convolutional Neural Network (CNN). Our findings suggest that integrating BayesDL and GWAS offers the potential to uncover more genes associated with the phenotypes of interest.