Deep integrative models for large-scale human genomics.

Area of Science:

• Genomics
• Computational Biology
• Precision Medicine

Background:

• Polygenic risk scores (PRSs) are crucial for precision medicine, but current linear models primarily capture additive genetic effects.
• Existing PRS methods are limited in handling complex genetic interactions and integrating diverse data modalities.

Purpose of the Study:

• To develop a novel deep learning framework (EIR) for enhanced PRS prediction.
• To incorporate a specialized model (genome-local-net, GLN) for large-scale genomics data.
• To enable multi-task learning, data integration, and model explainability.

Main Methods:

• Developed the EIR deep learning framework with the GLN model for large-scale genomics.
• Applied the GLN model to individual-level UK Biobank data.
• Integrated genotype, blood, urine, and anthropometric data for PRS construction.

Main Results:

• The GLN model achieved competitive performance against established neural networks, particularly for complex genetic traits.
• GLN outperformed linear PRS methods for Type 1 Diabetes by modeling non-additive genetic effects and epistasis.
• Integrating multi-modal data improved PRS prediction performance for 93% of 290 diseases and disorders.

Conclusions:

• The EIR framework, particularly the GLN model, offers a powerful approach for PRS prediction, capturing complex genetic architectures.
• Multi-modal data integration significantly enhances PRS accuracy across a wide spectrum of diseases.
• EIR provides a flexible and explainable tool for advancing precision medicine through advanced genetic risk prediction.