Multi-ancestry modeling improves fine-mapping resolution, protein prediction, and discovery for proteome-wide...

Proteomic predictive models are predominantly trained on cis-acting variants in European-ancestry cohorts, limiting power and predictive accuracy in ancestrally diverse populations. We performed cis- and trans-protein quantitative trait locus (pQTL) mapping and developed protein-prediction models using whole-genome sequencing (WGS) and plasma protein levels (Olink) across four ancestry groups from the Trans-omics for Precision Medicine (TOPMed) Multi-Ethnic Study of Atherosclerosis (MESA): European (EUR, n=1270), African (AFR, n=675), Hispanic (HIS, n=642), and Chinese (CHN, n=366), and a combined population (ALL, n=2953). African-ancestry samples demonstrated improved fine-mapping resolution relative to cohort size, yielding significantly smaller cis-credible sets than European-ancestry samples, consistent with shorter linkage disequilibrium (LD) blocks and greater allele frequency diversity in African-ancestry populations. For the first time, we benchmarked fine-mapping models SuSiE, SuShiE, MultiSuSiE, and SuSiEx with multi-ancestral cohorts, revealing a precision-recall tradeoff driven by model assumptions. Comparing protein-prediction models, multivariate adaptive shrinkage (MASHR) and ultimate deconvolution in R (UDR) outperformed elastic net (EN) regression, with trans-pQTL inclusion and fine-mapping improving prediction performance and proteome-wide association study (PWAS) discovery. Applying our models in PWAS of 10 phenotypes, we discovered 68 protein-phenotype associations in All of Us (AoU) that also replicated in Pan-UK Biobank. MASHR and UDR models identified 60% more protein-phenotype associations than EN. Notably, 32 of these associations were not previously reported in the GWAS Catalog. Overall, our study demonstrates the importance of including multiple ancestries in genomic studies to capture the full spectrum of regulatory variation and improve cross-ancestry generalizability.