Functional gene embeddings improve rare variant polygenic risk scores

Abstract

Rare variant association testing is a promising approach to identify effector genes for common traits. However, ensuring sensitive and robust rare variant association testing is challenging due to the scarcity of high-impact rare-allele carriers. Here we introduce FuncRVP, a Bayesian rare variant association framework that addresses this issue by leveraging functional gene embeddings, i.e. multidimensional representations of gene function. FuncRVP models the accumulated effects of rare variants on traits as a weighted sum of rare-variant gene impairment scores. A prior, learnt from data, regularizes the weight of each gene depending on the location of the gene in a functional gene embedding. We investigated this approach on 41 quantitative traits from unrelated UK Biobank exome samples. For the gene impairment score, we found that DeepRVAT scores strongly outperformed loss-of-function variant counts. Moreover, effective weight regularizations were obtained when using a prior on gene weight variance, but not on gene weight expectation. Compared to linear regressions on significantly associated genes, FuncRVP typically improved trait prediction, and did so more effectively for traits with higher burden heritability. Moreover, FuncRVP led to more robust gene effect estimates and to increased gene discoveries notably among genes that are more genetically constrained. Altogether, these results demonstrate that the integration of functional information across genes improves rare-variant phenotype prediction and gene discovery.