Enhanced genetic fine mapping accuracy with Bayesian Linear Regression models in diverse genetic architectures.

Abstract

We evaluated Bayesian Linear Regression (BLR) models with BayesC and BayesR priors as statistical genetic fine-mapping tools, comparing their performance to established methods such as FINEMAP and SuSiE. Through extensive simulations and analyses of UK Biobank (UKB) phenotypes, we assessed F1 classification scores and predictive accuracy across models. Simulations encompassed diverse genetic architectures varying in polygenicity, heritability, causal SNP proportions, and disease prevalence. In the empirical analyses, we used over 6.6 million imputed SNPs and phenotypic data from more than 335,000 UKB participants. Our results show that BLR models, particularly those using the BayesR prior, consistently achieved higher F1 scores than the external methods, but having comparable predictive accuracy. Applying the BLR model at the region-wide level generally yielded better F1 scores than the genome-wide approach, except for traits with high polygenicity. These findings highlight BLR models as accurate and robust tools for statistical fine mapping in both simulated and empirical genetic datasets.