Systematic benchmarking of multi-ancestry fine-mapping under variant heterogeneity and LD mismatch.

Increasingly diverse, multi-ancestry genome-wide association studies (GWASs) provide unprecedented opportunities to better understand the genetic architecture of complex traits. In particular, differences in linkage disequilibrium (LD) patterns across populations can be leveraged to improve the resolution and reliability of statistical fine-mapping. However, cross-ancestry effect heterogeneity, population-specific variant support, and LD mismatches pose substantial challenges, motivating a systematic evaluation of multi-ancestry fine-mapping methods to provide practical guidance for analyses across diverse populations. We conducted a comprehensive benchmarking of nine methods, spanning joint probabilistic models that explicitly model cross-ancestry LD and effect sharing, and univariate-based integration approaches that aggregate ancestry-specific results. We evaluated fine-mapping accuracy, robustness, and computational efficiency using realistic simulations derived from UK Biobank genotype data, with real-data validation in the All of Us Research Program. Our study revealed that joint modeling approaches achieve improved power and resolution for causal variants shared across populations. In comparison, post-hoc integration is more robust in recovering causal variants present in only a subset of populations by accommodating differences in variant availability. Moreover, methods are generally underpowered in the presence of systematic discrepancies between GWAS summary statistics and external LD reference panels. Motivated by these complementary performance profiles, we propose the multi-ancestry fine-mapping decision strategy, which avoids a priori method selection by adaptively selecting approaches based on locus-level evidence of ancestry-specific variants. Together, our benchmarking and decision strategy offer practical guidance for robust multi-ancestry fine-mapping and highlight the importance of joint modeling that effectively accounts for ancestry-specific variant availability and LD mismatch.