An Interpretable Deep Learning Framework for Genomic Prediction of Drought Tolerance in Rice

Abstract

Drought tolerance in rice is a complex quantitative trait that is difficult to predict from high-dimensional genomic data. We developed an interpretable deep-learning framework using phenotypic and genome-wide SNP genotype data from 426 rice accessions. Seven drought-response agronomic traits were integrated through principal component analysis and membership functions to derive a continuous drought-tolerance index. Four genomic representations were evaluated, and a convolutional neural network (CNN) was coupled with multi-head self-attention and Bernoulli input masking. With maximal information coefficient (MIC) screening, the proposed model achieved the highest mean Pearson correlation coefficient (PCC) of 0.6114, representing relative improvements of 5.39% and 6.01% over DeepGS and DNNGP, respectively. Attention-derived positional importance scores prioritized genomic signals centered on SNPs located within Os03g0221300, Os03g0372500, and Os12g0586300. Exploratory PC-adjusted additive allelic-dosage analyses in the same panel provided additional association evidence for all three prioritized loci. In this rice panel, the framework combined genomic prediction with model-derived candidate-locus prioritization and exploratory population-structure-adjusted marker–phenotype association analysis.