Deep learning-based phenotype prediction analysis of genotype-environment interactions and mining of environmentally stable germplasm and elite loci in cotton

Abstract

Abstract This study investigates the complex regulatory mechanisms of genotype-environment interactions (GEI) in cotton phenotype formation and explores the genetic basis of environmental adaptation through an integrated analytical approach. The research methodology encompasses four key components: (1) deep learning model construction, (2) phenotypic plasticity analysis, (3) environmental adaptation assessment, and (4) genome-wide association study (GWAS). Based on the multi-head self-attention mechanism and the deep feature interaction, we constructed the AttGEI-Net deep learning framework. The model demonstrates remarkable predictive performance with an average accuracy of 0.96 in fixed environments, though this decreases to 0.39-0.44 in novel environments, revealing fundamental differences between genotype-dominated and environment-dominated prediction scenarios. A total of 10,215 significant SNP loci is identified by GWAS, including 2,705 Main-SNPs, 41 phenotype plasticity loci (PP-SNPs), and 9,022 environmental adaptation loci (EvA-SNPs). The regulation of phenotypes by these loci has a distinct hierarchical character: the basic genetic architecture (Main-SNPs) maintains the basic expression of traits, the PP-SNPs mediates the immediate response of phenotypes to environmental changes, and the EvA-SNPs constitutes the highest-level adaptive regulatory network that coordinates the expression of multiple traits by integrating environmental signals. Shared loci of interpretability analyses of model and GWAS may be the key genetic basis adapting to different environments. Broadly adapted varieties in the Yellow River basin (e.g., F096, L090, etc.) can be used as the backbone parents for suitability breeding.

Code and data availability

The paper's authors publicly released the AttGEI-Net model code used for cotton phenotype prediction and interpretability analysis on GitHub. The phenotype/trait datasets themselves are not publicly deposited (available only on request), and the genomic deposits are molecular omics data, which do not qualify.