Diabetes is characterized by metabolic dysregulation. Metabolomics captures interactions of cellular processes and environmental exposures to promote disease and can improve mechanistic understanding to identify clinically relevant targets. However, use is limited by cost and sample availability. To facilitate investigation, machine learning approaches were used in the Insulin Resistance Atherosclerosis Family Study (n=943 Mexican Americans) with empiric metabolites (Metabolon HD4) and GWAS data to generate genetically regulated metabolite estimation models for broad application. 950 metabolites were heritable. Ridge and LASSO regression had cross validated correlation >0.1 for 448 metabolites. In an independent set, LASSO (enforcing sparsity) outperformed Ridge (full polygenic architecture) regression. Estimation was extended to the GUARDIAN Consortium (n∼4377) to assess the association of genetically predicted metabolites with glucose homeostasis and adiposity. Multiple associations (FDR P In conclusion, we have developed algorithms for estimating 448 metabolites using genetic data that had good performance in an independent dataset. Estimation models were used to assess association with glucose homeostasis and adiposity highlighting known and novel associations and demonstrating utility. Future work will expand genetic coverage to include rare variants and phenotypic associations to further characterize metabolic dysregulation. Disclosure N.Allred: None. L.J.Raffel: None. T.A.Buchanan: None. R.M.Watanabe: None. J.I.Rotter: None. L.E.Wagenknecht: None. C.D.Langefeld: None. H.K.Im: None. H.C.Ainsworth: None. X.Guo: None. A.W.Mackay: None. F.Nyasimi: None. O.Melia: None. Y.Liang: None. D.W.Bowden: None. K.Taylor: None. Funding R01DK118062
