Refining polygenic risk score models through fine mapping and functional gene modules

Abstract

Polygenic risk score (PRS) models are a valuable tool for assessing risk of developing a disease, given an individual's genotype. The unique combination of germline variants in each genome can be informative in risk calculations, especially in complex polygenic traits that are affected by a broad set of variants, many of which may contribute a small amount to the disease individually but combine to have significant effects. Recent studies have refined the basic linear PRS model by using functional annotations and linkage disequilibrium modeling to shift towards variant effect sizes that reflect genetic mechanisms. Here, I refine the standard PRS method to incorporate the complex, non-linear interactions between variants. I perform fine mapping on the input genome-wide association study (GWAS) variant effects using a convolutional neural network trained on regulatory feature annotations. I further extend the PRS method by calculating scores for a set of genes that form functional gene expression modules and using this partitioned PRS to learn different disease subtypes. Module-based PRS could improve clinical utility of PRS results by providing a fine-grained method for stratifying different aspects of disease risk, especially those that are based on biological mechanisms occurring in specific cell-types.