Fig. 2: Simulation study for the performance of our BayesRR-RC model implemented in the GMRM software against existing approaches for localisation of associations and genomic prediction.

a Probability of detecting genomic regions containing simulated causal variants by a Bayesian regional fine-mapping approach (GMRM: blue) versus standard mixed linear model association (MLMA) testing (BOLT: green). The column facets give the simulated heritability and rows give the number of causal variants and whether the effect sizes differed across genomic annotation groups (enrich) or were randomly assigned (random). b Correlation of a genomic predictor and a phenotype in an independent sample when the genomic predictor is created from GMRM effects sizes (blue), MLMA effect sizes using BOLT (green), and the optimal effect sizes obtained from individual-level and summary statistic models implemented in the Mega-PRS LDAK approach (purple). The column facets give the simulated heritability and the number of causal variants. The row facets give whether the effect sizes differed across genomic annotation groups (enrich) or were randomly assigned (random). The red lines give the expected prediction accuracy based on ridge regression theory. Error bars show the SD in both panels.