A survey on deep learning for polygenic risk scores

Abstract Polygenic risk scores (PRS) combine the effects of multiple genetic variants to predict an individual’s genetic predisposition to a disease. PRS typically rely on linear models, which assume that all genetic variants act independently. They often fall short in predictive accuracy and are not able to explain the genetic variability of a trait to the full extent. There is growing interest in applying deep learning neural networks to model PRS given their ability to model non-linear relationships and strong performance in other domains. We conducted a survey of the literature to investigate how neural networks model PRS. We categorize deep learning-based approaches by their underlying architecture, highlighting their modeling assumptions, likely strengths and potential weaknesses of the architectures. Several categories of neural network architectures exhibited promising signs for the improvement of PRS’ predictive power, namely sequence-based architectures, graph neural networks and those that incorporated biological knowledge. Additionally, the use of latent representations in autoencoders has improved predictive performance across diverse ancestries. However, a lack of existing model benchmarks on consistent datasets and phenotypes makes it challenging to understand the extent to which different architectures improve performance. Interpretability of deep learning-based PRS is also challenging with great care required when inferring causation. To address these challenges, we suggest the establishment and adherence to reporting standards and benchmarks to aid the development of deep learning-based PRS to find quantifiable trends in neural network architectures.