GViT-GP: Injecting the Genomic Relationship Matrix as an Inductive Bias into a Vision Transformer via Cross-Attention for Genomic Prediction

Jingxuan Li¹Wei Luo¹Honghao Yu¹Xishi Huang¹Jisi Ma¹Shijun Li²Yong Li^1*Lantao Gu^1*

• ¹Guilin Medical University, Guilin, China
• ²Huazhong Agriculture University College of Animal Sciences and Technology, Wuhan, China

Genomic Prediction (GP) faces significant challenges in balancing model complexity with computational efficiency, particularly when handling high-dimensional genomic data with limited sample sizes. To address these issues, we introduce GViT-GP, a novel Vision Transformer architecture that employs a cross-attention mechanism to inject the Genomic Relationship Matrix (GRM) as a biological prior, thereby enhancing the model's ability to learn robust representations from sparse data without the overfitting often associated with deep learning. We evaluated GViT-GP on 20 distinct traits across four datasets from three species—soybean, cattle (Cows4020 and Bulls1508), and chicken—where it demonstrated superior predictive performance against established linear and non-linear benchmarks, including GBLUP, LightGBM, and DNNGP, achieving the highest accuracy in 16 of the 20 tasks. Ablation studies confirmed the efficacy of our Selective Patch Embedding strategy in reducing computational redundancy and validated the dual-pathway cross-attention mechanism for effectively fusing SNP data with population structure information, while visualization analyses suggest that the model adaptively attends to informative genomic regions. Consequently, GViT-GP presents a robust and data-efficient framework that effectively captures complex genotype-phenotype landscapes, highlighting its potential as a practical, high-performance tool for modern digital breeding programs.