RESUMEN
Comprehensively studying metabolism requires the measurement of metabolite levels. However, in contrast to the broad availability of gene expression data, metabolites are rarely measured in large molecularly-defined cohorts of tissue samples. To address this basic barrier to metabolic discovery, we propose a Bayesian framework ("UnitedMet") which leverages the empirical strength of RNA-metabolite covariation to impute otherwise unmeasured metabolite levels from widely available transcriptomic data. We demonstrate that UnitedMet is equally capable of imputing whole pool sizes as well as the outcomes of isotope tracing experiments. We apply UnitedMet to investigate the metabolic impact of driver mutations in kidney cancer, identifying a novel association between BAP1 and a highly oxidative tumor phenotype. We similarly apply UnitedMet to determine that advanced kidney cancers upregulate oxidative phosphorylation relative to early-stage disease, that oxidative metabolism in kidney cancer is associated with inferior outcomes to combination therapy, and that kidney cancer metastases themselves demonstrate elevated oxidative phosphorylation relative to primary tumors. UnitedMet therefore enables the assessment of metabolic phenotypes in contexts where metabolite measurements were not taken or are otherwise infeasible, opening new avenues for the generation and evaluation of metabolite-centered hypotheses. UnitedMet is open source and publicly available (https://github.com/reznik-lab/UnitedMet).