Gsformer: a dual-architecture deep learning framework with CNN-self-attention and sparse-attention for genomic...

Area of Science:

• Genomics
• Bioinformatics
• Animal and Plant Breeding

Background:

• Genomic selection (GS) utilizes genome-wide SNPs for modern breeding.
• Deep learning models show superior predictive performance over traditional GS methods.
• Existing deep learning approaches for GS require further optimization for complex genetic architectures.

Purpose of the Study:

• Introduce Gsformer, a novel deep learning framework for phenotype prediction.
• Model complex genetic architectures using two distinct Gsformer architectures: CSA and NSA.
• Evaluate Gsformer's predictive performance across diverse animal and plant species.

Main Methods:

• Developed Gsformer with two architectures: CSA (CNNs + self-attention) and NSA (native sparse attention).
• Evaluated Gsformer on six datasets (pig, cattle, chicken, mouse, wheat, maize).
• Compared Gsformer against five established GS methods (DNNGP, MLP, LightGBM, SVR, GBLUP).

Main Results:

• Gsformer consistently ranked among the top two models across all datasets.
• Gsformer-CSA improved cattle fat percentage prediction; Gsformer-NSA excelled in predicting chicken egg weight, pig weight, and mouse anxiety.
• Gsformer-NSA demonstrated comparable or superior performance to Gsformer-CSA with hyperparameter tuning, with lower computational cost.

Conclusions:

• Gsformer provides a flexible and powerful framework for genomic prediction in animals and plants.
• Gsformer-NSA is recommended for breeding applications prioritizing computational efficiency.
• Identified influential SNPs and biologically relevant pathways for growth and body size traits in pigs using SHAP analysis.