Crop-GPA 2.0: Decoding crop genotype-phenotype associations through cross-species and trait-aware hierarchical genomic representations

Cross-species prediction of trait-associated SNPs requires transferable models but remains limited by poor model transferability and uneven genomic resources, particularly for species with limited genomic data. Building upon Crop-GPA 1.0, we developed Crop-GPA 2.0 to enable transferable SNP-level genotype–phenotype association prediction across crop species. Crop-GPA 2.0 combines hierarchical genomic representations of multi-scale sequence and DNA structural features with transfer learning pre-trained on cross-species genomic diversity and trait-associated SNP information. Across multiple crop species, Crop-GPA 2.0 improved cross-species prediction performance, remained robust under limited training data, and identified biologically meaningful SNPs supported by previously reported quantitative trait locus (QTL) evidence. Crop-GPA 2.0 provides a scalable framework for trait-aware SNP-level genotype–phenotype association modeling, enabling cross-species discovery of functional variants for crop genomics and precision breeding.