AI-guided discovery and expansion of the human bile acid landscape

Bile acids represent one of nature's most chemically diverse and therapeutically important metabolite families. However, the human bile acid metabolome remains largely uncharted, with analytical limitations obscuring potentially thousands of biologically active species present in cells and biofluids. Here, we present MetLLM, a transformer-based language model that learns bile acid modification patterns to prioritize candidate bile acid analogs, and MetSpace, an in silico bile acid chemical landscape containing 3,900 prioritized candidates. Liquid chromatography-high resolution mass spectrometry analysis of human feces confirmed multiple newly detected and structurally novel bile acids, including MS2141 and MS1281. In vitro functional validation showed that MS2141 modulates M2 receptor-associated Gi/o signaling and alters electrophysiological parameters in human induced pluripotent stem cell-derived cardiomyocytes. This work expands the searchable bile acid metabolome and provides a computational framework for systematic metabolite discovery.