Enhanced power and transferability for genetics-driven metabolomic biomarker discovery in admixed American cohorts

Despite metabolomics transforming our understanding of risk factors and aetiology of metabolic diseases, profiling is rarely performed for people of non-European ancestries, on whom much of metabolic disease burden falls. Metabolome-wide association studies (MWAS) can be performed using genetic scores to predict metabolomic traits, helping address these inequities; however, their performance in populations of admixed American (AMR) ancestries is unexplored. We evaluated 141 genetic scores, developed in an INTERVAL Study sample of European (EUR) genetic ancestries, in the Mexico City Prospective Study (MCPS; n=132,336), obtaining a median predictive R2 of 0.027. Training Bayesian ridge models within MCPS substantially improved performance, with a median R2 of 0.083 on a withheld 20% subset. MCPS-trained models also outperformed INTERVAL-trained models among UK Biobank participants of AMR ancestries (n=600; median R2: 0.070 vs. 0.046). Finally, among AMR participants of the All of Us cohort, using MCPS-trained (vs. INTERVAL-trained) models to predict metabolomic traits yielded five times as many significant associations (FDR-corrected P<0.05) across three cardiometabolic diseases: ischaemic heart disease, type 2 diabetes, and chronic kidney disease. The genetic scores are openly available at the OmicsPred portal (www.OmicsPred.org), enabling better-powered analyses in diverse AMR cohorts and helping reduce global inequities in omics research.