Deep learning accelerated correlation of genotypic and phenotypic data

Abstract

The prediction of phenotypic values based on genetic data is referred to as genomic prediction (GP). Genome-wide association studies (GWAS), on the other hand, look for correlations between genotypic markers (single nucleotide polymorphisms, SNPs) and phenotypic traits like grain yield and plant height in order to discover the key SNPs responsible for those traits. This study aims to address the distinct challenges of both GP and SNP identification. The rrBLUP and BLINK models are widely used for GP and GWAS, respectively. However, rrBLUP can only model simple linear relationships between genotype and phenotype, and BLINK often results in false positives when identifying SNPs. To address these challenges, we use machine learning approaches capable of capturing complicated, non-linear patterns, hence improving genomic prediction performance and SNP identification. In this study, we evaluate popular ML model support vector regression (SVR) and its variants as well as the transformer-based GPformer, for their ability to improve predictive performance. Motivated by the di!culty of identifying significant SNPs in high dimensionalty low sample size SNP data, we initially create a hybrid model that combines the regression power of SVR with the feature interaction strength of self-attention. Building on this breakthrough, we then reimagine the SNP sequence as a two-dimensional, image like representation, a strategy that reveals spatial patterns in genomic variation by taming the curse of dimensionality and enabling potent image-based learning models.