MACHINE: a robust and scalable multi-ancestry fine-mapping method using a continuous global-local shrinkage prior

Fine mapping aims to identify causal genetic variants with nonzero phenotypic effects. Leveraging genome-wide association study (GWAS) data from diverse ancestries enhances fine-mapping accuracy and resolution by exploiting differences in linkage disequilibrium (LD) and increasing sample sizes. However, existing multi-ancestry fine-mapping methods rely on discrete priors and assume that all causal variants are shared across ancestries - an assumption that may not hold in practice. Although MESuSiE accounts for both shared and ancestry-specific causal effects, it requires a priori specification of prior probabilities for causal variant sharing. Moreover, methods based on discrete priors are prone to sub-optimal convergence. To address these limitations, we introduce Multi-AnCestry Heritability INducEd Dirichlet decomposition (MACHINE), a flexible Bayesian framework that employs a continuous prior to model both shared and ancestry-specific effects without restrictive assumptions. Importantly, we propose an approach to control false discovery rate (FDR) for fine mapping with GWAS summary statistics and out-of-sample LD matrices, a challenge not addressed by existing multi-ancestry fine-mapping methods. We further improve fine-mapping performance by incorporating functional annotations of variants using generalized LD score regression (g-LDSC). Simulation studies across diverse genetic architectures demonstrate robustness and superior FDR control of MACHINE + g-LDSC compared to existing methods. In the real data analyses, we applied MACHINE + g-LDSC to four lipid traits and schizophrenia, identifying previously unknown causal variants and depicting their genetic architectures across ancestries.