A predictive approach to genome design for multi-trait improvement and environmental adaptability via machine learning.

Machine learning on marker-environment interactions boosts multi-trait genomic prediction across four crops and enables in silico design of cotton lines combining high yield, quality, and climate adaptability. Genomic selection (GS) accelerates breeding through early prediction of superior genotypes, but genotype-environment interactions (G×E) compromise yield stability and climate adaptation. Here, we employed machine learning to identify genome-wide loci associated with environmental responsiveness, phenotypic plasticity, and BLUP-adjusted traits, integrating them into core marker sets. Coupled with environmental covariates, these markers enabled GS models with improved cross-environment predictive accuracy. Evaluations across 13 phenotypic traits in four crops revealed LightGBM and random forest regression (RFR) as top performers, with LightGBM showing the most consistent superiority across scenarios, particularly in untested environments. In an in-silico genome-design analysis of cotton, SHAP-based interpretation within the fitted predictive model was used to propose candidate marker configurations: modification of 23-36 BLUP-associated markers in six elite fiber lines was predicted to improve boll weight (2.2-6.3%) and lint percentage (0.2-3.9%) in target environments, while additionally optimizing plasticity/environment-responsiveness markers yielded further predicted gains (3.5-7.2% and 2.3-3.5%, respectively). In 11 high-yield lines, modification of BLUP-associated markers was predicted to improve fiber length (0-2.4%) and strength (0.4-3.9%), with environment-interaction markers further enhancing traits (0.8-3.7%). Notably, 78-85% of environment-specific markers exhibited high SHAP values despite partial overlap with BLUP loci, suggesting their potential relevance to adaptive breeding. These improvements are derived from in silico model simulations and should be interpreted as candidate breeding targets rather than validated genetic gains. These results highlight machine learning as a pivotal tool for G × E-driven breeding strategies in multi-objective breeding to enable simultaneous optimization of yield, quality, and climate adaptability in crop improvement efforts.