Kernelized approach enables explainable gene prioritizations for complex traits

Abstract

Genome-wide association studies (GWAS) have identified numerous variant-trait associations; yet, assigning effector genes to GWAS loci remains challenging. Similarity-based machine-learning methods, such as PoPS, prioritize effector genes from shared functional profiles among trait-relevant genes. These models assign a prioritization score for each gene and nominate a single effector gene within a GWAS locus. However, the scores provide limited insight into why a gene was prioritized or whether the nomination is biologically plausible. To address this gap, we introduce Kernelized Polygenic Priority Score, K-PoPS, a kernelized reformulation of PoPS that enables gene-centric explanations by decomposing each prediction into contributions from training genes. For each prioritized gene, K-PoPS reports top contributor genes and an anchor score that quantifies support from a user-defined set of trait-relevant genes. Across 38 Pan-UK Biobank traits, the full-feature OLS implementation underlying K-PoPS improved closest-gene enrichment relative to default PoPS for 26 of 37 evaluable traits. Across 25 traits with curated anchor sets, predictions supported by anchor scores were more enriched for closest-gene proxies than unsupported predictions. When applying to blood level apolipoprotein B, K-PoPS nominated SCARB1 over UBC gene, and further provided convincing explanations that support this prediction. Using explanation evidence, K-PoPS identified multiple plausible effector genes within a dilated cardiomyopathy locus, contrary to the parsimonious assumption. In summary, K-PoPS provides a post hoc framework for examining and interpreting GWAS effector-gene nominations.