Interpretable machine learning meets systems biology to decode genotype-phenotype maps

ABSTRACT Resolving causal genes from quantitative trait loci (QTL) remains fundamentally limited by linkage disequilibrium. We developed an interpretable machine learning framework that captures higher-order nonlinear genotype-phenotype relationships and allows conditional evaluation of genetic variants, enabling statistical decorrelation of linked loci. Applied to Saccharomyces cerevisiae segregants across chemical stress conditions, our method achieved >75% prediction accuracy and identified known causal genes, including MKT1 (genotoxic stress) and IRA2 (osmotic stress). SHAP-based analysis recovered 56% of the validated pleiotropic genes, compared with 36% by conventional contingency testing. Integration with genome-scale metabolic models revealed pathway enrichments distinguishing high-growing strains, including carbon transport, glycolysis, and oxidative phosphorylation. Notably, gene regulatory network analysis identified a novel function for PDR8 in protein mannosylation and cell wall integrity—functions extending beyond its role in drug resistance. This framework demonstrates that interpretable machine learning, coupled with systems biology, transforms QTL associations into mechanistic biological insight.