A survey on deep learning for polygenic risk scores.

Area of Science:

• Genetics
• Computational Biology
• Artificial Intelligence

Background:

• Polygenic risk scores (PRS) predict disease predisposition using multiple genetic variants.
• Current PRS often use linear models, limiting predictive accuracy and failing to capture full genetic variability.
• Deep learning neural networks offer potential for modeling non-linear genetic relationships.

Purpose of the Study:

• To survey and categorize deep learning approaches for modeling polygenic risk scores.
• To highlight the assumptions, strengths, and weaknesses of various neural network architectures in PRS.
• To identify challenges and suggest future directions for deep learning-based PRS development.

Main Methods:

• Literature survey of deep learning applications in polygenic risk scores.
• Categorization of neural network architectures (e.g., sequence-based, graph neural networks, autoencoders).
• Analysis of modeling assumptions, predictive power, and interpretability.

Main Results:

• Sequence-based models, graph neural networks, and biologically informed networks show promise for enhancing PRS predictive power.
• Autoencoders with latent representations improve performance across diverse ancestries.
• Lack of standardized benchmarks and interpretability challenges hinder progress.

Conclusions:

• Deep learning architectures offer significant potential to improve polygenic risk score accuracy.
• Establishment of reporting standards and benchmarks is crucial for advancing deep learning-based PRS.
• Careful consideration of interpretability is necessary when inferring causation from deep learning models.