Evaluation of Bayesian Linear Regression Models as a Fine Mapping tool

Abstract

Bayesian linear regression (BLR) models consider the underlying genetic architecture of complex phenotypes by specifying different prior distributions for SNP effects allowing heterogenous distribution of the true genetic signals. Our goal is to evaluate BLR models with BayesC and BayesR prior distributions for fine mapping on simulated and real binary and quantitative phenotypes, and compare them to the state-of-the-art external models: FINEMAP, SuSIE-RSS, SuSIE-Inf and FINEMAP-Inf. Evaluation of models was based on the F1 classification score for simulations, and predictive accuracy for the UK Biobank (UKB) phenotypes. We used over 533K genotyped SNPs (simulations) and 6.6 million imputed SNPs (UKB phenotypes), from over 335K White British Unrelated samples in the UK biobank. We simulated phenotypes from low (GA1) to moderate (GA2) polygenicity, heritability(h^2) of 30% and 10%, causal SNPs (π) of 0.1% and 1% sampled genome-wide, and prevalence (PV) of 5% and 15%. Summary statistics and in-sample linkage disequilibrium were used to fit models in regions defined by lead SNPs. BayesR improved the F1 score, averaged across all simulations, by by 27.26%, 26.96%, 18.40%, 15.42%, and 13.32% relative to FINEMAP-Inf, BayesC, FINEMAP, SUSIE-RSS and SUSIE-Inf. Prediction R2, averaged across all the UKB quantitative phenotypes, with BayesR decreased by 5.32% and 3.71% compared to SuSIE-Inf and FINEMAP-Inf, whereas increased by 7.93% and 8.3% compared to SuSIE-RSS and BayesC. Prediction AUC, averaged across all the UKB binary phenotypes, with BayesR increased by 0.40%, 0.16%, 0.08%., and 0.05% relative to SuSIE-RSS, BayesC, FINEMAP-Inf, and SuSIE-Inf. These findings suggest that the performance of the BLR models was comparable to the state-of-the-art external models. The performance of BayesR prior was closely aligned with SuSIE-Inf and FINEMAP-Inf models. Results from both simulations and application of the models in the UKB phenotypes suggest that the BLR models are efficient fine mapping tools.

Genome-wide fine-mapping improves identification of causal variants

“…These limitations of current fine-mapping methods can be addressed through conducting a finemapping analysis using a genome-wide Bayesian mixture model (GBMM). GBMMs, which have been widely used for predicting breeding values in agricultural species [21][22][23] and complex trait phenotypes in humans [24][25][26][27] , have recently emerged as a method of GWFM 28,29 . Compared to conventional GWAS and region-specific fine-mapping approaches, GBMMs consider genomewide SNPs simultaneously, which are all utilised to estimate the genetic architecture and functional prior 27,28 .…”

Genome-wide fine-mapping improves identification of causal variants

“…These limitations of current fine-mapping methods can be addressed through conducting a finemapping analysis using a genome-wide Bayesian mixture model (GBMM). GBMMs, which have been widely used for predicting breeding values in agricultural species [21][22][23] and complex trait phenotypes in humans [24][25][26][27] , have recently emerged as a method of GWFM 28,29 . Compared to conventional GWAS and region-specific fine-mapping approaches, GBMMs consider genomewide SNPs simultaneously, which are all utilised to estimate the genetic architecture and functional prior 27,28 .…”

Expanded utility of the R package, qgg, with applications within genomic medicine

“…Tools such as PLINK ( Chang et al 2015 ), GCTA ( Yang et al 2011 ), LDpred ( Vilhjálmsson et al 2015 , Privé et al 2021 ), LDAK ( Speed and Balding 2014 , Speed et al 2017 , Zhang et al , 2021 ), and PRSice ( Euesden et al 2015 ) have changed the way researchers around the globe conduct genetic analyses of human complex traits and diseases. Here we present a major update of the R package qgg ( Rohde et al 2020 ), which has now been expanded to include a range of commonly used methods such as LD Score Regression (LDSC), adjustment of marker effect using correlated trait information ( Rohde et al 2022 ), a range of different Bayesian shrinkage models for gene mapping ( Shrestha et al 2023 ), and construction of single- and multiple trait polygenic scores (PGS). With a user-friendly interface, qgg offers a unified tool for quantitative genetic analysis of complex traits and diseases.…”