Modeling gene interactions in polygenic prediction via geometric deep learning.

Area of Science:

• Genetics
• Computational Biology
• Bioinformatics

Background:

• Polygenic risk scores (PRS) are crucial for precision medicine but traditional linear models struggle with complex genotype-phenotype relationships.
• Existing PRS methods often fail to capture the intricate, nonlinear biological interactions underlying complex diseases.

Purpose of the Study:

• To develop an interpretable geometric deep learning framework, PRS-Net, for enhanced genetic risk prediction and biological discovery.
• To model the nonlinearity of biological systems and gene-gene interactions for improved disease prediction.

Main Methods:

• PRS-Net utilizes a graph neural network (GNN) to model gene-gene interactions after deconvoluting genome-wide PRS.
• An attentive readout module is incorporated for enhanced model interpretability.
• The framework was tested across multiple complex traits and diseases.

Main Results:

• PRS-Net demonstrated superior prediction performance compared to conventional PRS methods.
• The model successfully identified disease-relevant genes and biological pathways.
• Achieved enhanced accuracy in predicting genetic risk for complex diseases.

Conclusions:

• PRS-Net offers a powerful, interpretable tool for both genetic risk prediction and biological discovery in complex diseases.
• The framework advances precision medicine by effectively modeling complex biological systems.
• Highlights the potential of geometric deep learning in understanding disease etiology.