PQG Working Group

 Quantifying and partitioning SNP effect correlation across UK Biobank traits

While traditional heritability estimation methods such as LD-score regression (Bulik-Sullivan et al. 2015 Nat Genet) assume independent SNP effects on the trait, recent works (Schoech et al. 2020 bioRxiv) determined that nearby SNPs (e.g., <100bp) have negatively correlated effects. Yet, it is unclear if the SNP effect correlation exists for other types of SNP pairs. We developed a new method, gene-level directional effect regression (GDREG), that analyzes summary association statistics and in-sample LD information to jointly estimate average SNP effect correlation and per-SNP heritability, stratified by MAF and functional category; we define SNP effect correlation based on standardized minor allele causal effect sizes of SNPs. We performed extensive simulations to confirm that GDREG obtains unbiased estimates of average SNP effect correlation under a range of different genetic architectures.

We applied GDREG to the UK Biobank imputed data across 32 independent diseases and complex traits (N=334K unrelated British-ancestry samples). We determined that SNP pairs within 100bp are significantly negatively correlated (non-significant for common pairs, -0.27±0.04 for low-freq SNP pairs), recapitulating results in previous works. Furthermore, we detected a significant negative correlation for gene-level SNP pairs, including SNP pairs on the same exon (non-significant for common SNP pairs, -0.04±0.01 for low-freq SNP pairs) and SNP pairs on the promoter of the same gene (-0.04±0.01 for common pairs, -0.21±0.03 for low-freq pairs). The negative correlation of gene-level SNP pairs cannot be fully explained by their proximity, providing additional information to the proximity annotation. We hypothesize that the negative correlation is due to linkage masking, in which negatively correlated effects of positively correlated minor alleles enable the alleles to cancel each other’s effects and thus survive the action of negative selection. In addition, we determined that low-freq SNP pairs are more negatively correlated than common SNP pairs across the annotations.