KANG2P enables interpretable genotype-to-phenotype prediction through genetically anchored molecular-state learning

Genotype-to-phenotype (G2P) mapping represents a central scientific challenge in genetics, necessitating advancements in both prediction performance and mechanistic interpretability. The complexity of biological systems creates a substantial gap between genotypic variations and phenotype. While multi-omics integration strategies strive to bridge this gap, they face limitations including prohibitive costs and challenges in acquiring large-scale cohorts. To address these constraints, we present KANG2P, an interpretable genotype-to-phenotype framework that bridges these settings by transferring genetic regulatory information learned from external reference panels to target cohorts. Modality-specific autoencoders learn compact representations of genotype and genetically predicted molecular profiles, which are subsequently integrated by a Kolmogorov--Arnold Network to model the nonlinear path from genetic variation through molecular regulation to phenotype. Across human neurodegenerative diseases and crop agronomic traits, KANG2P outperformed conventional and deep-learning G2P models. Notably, the benefit of genetically predicted molecular profiles extended beyond KANG2P itself, as their integration also improved established G2P models. Beyond predictive accuracy, KANG2P recovered established amyotrophic lateral sclerosis genes and newly prioritized SLC1A2 as the leading genome-wide candidate. Functional analysis of its conserved C. elegans orthologue, glt-6, revealed locomotor impairment and heightened paraquat sensitivity, providing in vivo support for the disease relevance of the SLC1A2-glt-6 axis. Together, these findings establish KANG2P as a scalable, accurate and experimentally grounded framework for interpretable G2P mapping across species.