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

Area of Science:

• Genetics
• Artificial Intelligence
• Bioinformatics

Background:

• Polygenic risk scores (PRS) are crucial for summarizing genetic information.
• Deep learning (DL) has shown limited impact on PRS generation previously.
• This study explores DL's potential to enhance PRS creation.

Purpose of the Study:

• To investigate how deep learning (DL) can improve the generation of polygenic risk scores (PRS).
• To assess DL model performance in recreating human-programmed PRS and generating multiple PRS from a single model.
• To evaluate DL's ability to handle missing genetic data and performance constraints.

Main Methods:

• Training DL models on existing PRS using UK Biobank data.
• Evaluating DL models for PRS recreation and multi-PRS generation.
• Assessing DL model performance with reduced training data and missing single nucleotide polymorphisms (SNPs).

Main Results:

• DL models achieved near-perfect generation of multiple PRS with minimal performance loss, even with reduced training data.
• For missing SNPs, DL models improved case-population separation (AUC 0.847) compared to traditional PRS (AUC 0.798).
• DL models demonstrated transferability and longevity.

Conclusions:

• Deep learning (DL) can accurately generate polygenic risk scores (PRS), including multiple scores from a single model.
• DL models show promise in improving PRS generation, especially when dealing with missing genetic data.
• Further advancements may require additional input data.