Exploring the application of deep learning methods for polygenic risk score estimation.

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Polygenic risk scores (PRS) consolidate genetic data for clinical applications.
• Machine learning (ML) has had limited impact on genomics and PRS development.
• This study explores ML's potential to enhance PRS generation.

Approach:

• ML models were trained on existing PRS using UK Biobank data.
• The study investigated ML's ability to replicate human-programmed PRS.
• ML's capacity to handle missing data and performance constraints was assessed.

Key Points:

• ML models achieved near-perfect PRS generation, including multi-PRS prediction.
• Performance remained high even with reduced training data.
• ML models improved case-population separation for missing SNPs (AUC 0.847 vs. 0.798).

Conclusions:

• ML accurately generates PRS, including multiple scores from a single model.
• ML models demonstrate transferability and longevity.
• ML significantly improves PRS generation with missing SNPs, with input data as a performance bottleneck.