Machine learning reveals complex genetics of fungal resistance in sorghum grain mold.

Plant disease resistance is often a complex, polygenic trait, making its genetic dissection with traditional genome-wide association studies (GWAS) challenging. Grain mold in sorghum, a devastating disease caused by a fungal complex, exemplifies this complexity. We hypothesized that a machine learning (ML)-driven GWAS, employing diverse phenotypic representations from a panel of 306 sorghum accessions, could more effectively unravel the genetic basis of resistance. Phenotypic data, including raw disease scores, a 'difference phenotype' (inoculated vs. control), and principal components, were analyzed using Boosted Tree and Bootstrap Forest models, demonstrating strong explanatory power for phenotypic variance when trained on the entire dataset. This ML-GWAS approach confirmed a highly polygenic architecture for grain mold resistance, identifying numerous SNPs across the sorghum genome. Notably, several SNPs were consistently associated with resistance across multiple analytical models and phenotypic representations. These robustly identified SNPs were frequently located near genes with predicted functions integral to plant defense. Gene ontology (GO) analyses of the candidate gene set confirmed enrichment in categories supporting roles in pathogen recognition, DNA repair, and stress response modulation, indicating a multifaceted defense mechanism. This study provides valuable candidate genes for breeding sorghum with enhanced grain mold resistance and offers a refined methodological framework for dissecting complex traits in this crop. The successful application of this ML-based strategy in sorghum suggests its potential utility for studying similar complex traits in other plant species.