A transformer-based genomic prediction method fused with knowledge-guided module.

Area of Science:

• Genomics
• Machine Learning
• Plant Breeding

Background:

• Genomic prediction (GP) utilizes single nucleotide polymorphisms (SNPs) to associate genetic markers with traits.
• Deep learning (DL) methods are increasingly applied to enhance GP accuracy and efficiency.
• Accelerating breeding programs relies on accurate genomic estimated breeding values for early selection.

Purpose of the Study:

• To introduce GPformer, a novel Transformer-based deep learning model for genomic prediction.
• To evaluate GPformer's performance against existing GP methods across diverse crop datasets.
• To develop and integrate a knowledge-guided module (KGM) for incorporating prior biological information into GP models.

Main Methods:

• Developed GPformer, a deep learning architecture leveraging Transformer structures for genomic prediction.
• Implemented a knowledge-guided module (KGM) to integrate genome-wide association studies (GWAS) information as prior knowledge.
• Conducted comprehensive experiments on five crop datasets, comparing GPformer with RR-BLUP, SVR, LightGBM, and DNNGP.

Main Results:

• GPformer significantly outperformed established methods (RR-BLUP, SVR, LightGBM, DNNGP) in predictive accuracy across all tested datasets.
• The knowledge-guided module (KGM) demonstrated effectiveness in enhancing GPformer's performance, as shown by ablation studies.
• GPformer exhibited robustness to hyperparameter choices and strong generalization capabilities across different phenotypes and datasets.

Conclusions:

• GPformer represents a significant advancement in deep learning for genomic prediction, offering superior accuracy and efficiency.
• The integration of KGM provides a flexible approach to incorporate biological insights, further improving predictive models.
• GPformer's stability and generalizability make it a promising tool for practical applications in crop improvement and breeding programs.