# GViT-GP: injecting the genomic relationship matrix as an inductive bias into a vision transformer via cross-attention for genomic prediction

## TL;DR

GViT-GP is a new model that uses genomic data more efficiently by incorporating biological knowledge into a vision transformer, improving accuracy in genomic prediction.

## Contribution

The novel integration of the genomic relationship matrix as an inductive bias in a vision transformer for genomic prediction.

## Key findings

- GViT-GP outperformed existing models in 16 out of 20 genomic prediction tasks across three species.
- Selective Patch Embedding and cross-attention fusion were confirmed effective through ablation studies.
- The model adapts attention to informative genomic regions, improving robustness in high-dimensional settings.

## Abstract

Genomic Prediction (GP) faces significant challenges in balancing model complexity with computational efficiency, particularly for high-dimensional genomic data under limited sample sizes.

We propose GViT-GP, a Vision Transformer architecture that injects the Genomic Relationship Matrix (GRM) as a biological prior via a dual-pathway cross-attention fusion mechanism, coupled with a Selective Patch Embedding strategy to reduce redundancy and improve data efficiency.

We evaluated GViT-GP on 20 traits across four datasets from three species (soybean, cattle, and chicken). GViT-GP outperformed established linear and non-linear baselines (including GBLUP, LightGBM, and DNNGP), achieving the best accuracy in 16/20 tasks. Ablation studies supported the effectiveness of Selective Patch Embedding and cross-attention fusion, and visualization analyses suggest adaptive attention to informative genomic regions.

These results indicate that injecting GRM-informed inductive bias improves robustness and generalization in “p ≫ n” settings. GViT-GP provides a practical, high-performance framework for capturing complex genotype–phenotype relationships in modern digital breeding.