Integrating predictive models and GWAS to identify candidate loci and genes for agronomic traits in rice
Abstract
Rice is a staple crop whose improvement relies on breeding advances; precision agriculture demands predictive loci/genes for agronomic traits to innovate rice production. To cut experimental costs and boost efficiency, this study built predictive models using seedling leaf metabolomes to forecast rice agronomic traits and decode trait correlations. We integrated 11 agronomic traits and 840 metabolites from 524 rice germplasms, plus 17 agronomic traits of 3,000 varieties. Five algorithms (RF, LightGBM, SWR, LASSO, CART) were combined to build multi-model prediction systems, offsetting defects of single models for accurate complex trait prediction. GWAS on predicted phenotypes detected nine genetic hotspots. Results showed LASSO and CART had weak generalization, while LightGBM, RF and SWR delivered trait-specific predictive performance. Seven candidate genes within hotspots were validated via variation annotation, haplotype and tissue expression analyses. Distinct from laborious, environment-prone traditional phenotyping, these models realize rapid, stable high-throughput trait prediction via early metabolic markers. The uncovered loci and genes lay groundwork for dissecting molecular regulatory networks linking rice agronomy and metabolism.