Multi-task genomic prediction using gated residual variable selection neural networks.

Area of Science:

• Genomics
• Machine Learning
• Quantitative Genetics

Background:

• High-throughput sequencing enables genome-wide prediction (GWP), but integrating pedigree data traditionally requires computationally intensive methods.
• Existing methods struggle to extend pedigree-genomic integration to flexible machine learning approaches.

Purpose of the Study:

• To enhance genomic prediction accuracy and interpretability by implementing gated residual variable selection neural networks (GRVSNN) for multi-task learning.
• To integrate low-rank pedigree information with genomic markers using GRVSNN, comparing its performance against conventional regression and deep learning (DL) models.

Main Methods:

• Developed and applied a gated residual variable selection neural network (GRVSNN) model for multi-task genomic prediction.
• Integrated pedigree-based relationship matrices (low-rank information) with genomic markers.
• Evaluated GRVSNN on real-world datasets from loblolly pine, mouse, and pig.

Main Results:

• GRVSNN significantly outperformed traditional models like Bayesian regression and LassoNet in predictive accuracy.
• Achieved lower mean squared error (MSE) and higher Pearson (r) and distance correlation (dCor) between predicted and true phenotypes.
• GRVSNN demonstrated improved interpretability by selecting fewer genetic markers and pedigree loadings.

Conclusions:

• The GRVSNN framework offers a computationally effective method for integrating pedigree and genomic data to improve genomic prediction.
• Its multi-task prediction capability holds potential for advancing selection in agriculture and disease prediction in precision medicine.