Linking microbes to cardiovascular-associated metabolites using knowledge-based machine learning - PROJECT SUMMARY The gut microbiome produces numerous metabolites that significantly influence the development of cardiovascular disease (CVD). Examples include trimethylamine N-oxide (TMAO), short-chain fatty acids, secondary bile acids, and imidazole propionate (ImP), which directly contribute to the development of atherosclerosis, dyslipidemia, and insulin resistance. Yet our ability to predict and modulate these metabolic outputs remains severely limited. Preliminary data reveal that most gut bacterial genes lack functional annotations, creating a critical bottleneck for understanding microbiome metabolism. Moreover, different strains of the same species may differ by ~25% in gene content, and strain-level metabolic variation drives patient- specific disease risk. However, current models operate at the species level and miss this crucial variation. These limitations prevent the development of targeted microbiome interventions for CVD. My goal is to overcome these limitations using machine learning and a combination of in vitro and in silico assays. State-of-the-art protein language models (pLMs) and interpretable machine learning now offer unprecedented opportunities to decode microbial metabolism at scale and connect it mechanistically to human health. In Aim 1, I will generate high- quality metabolic annotations for >200,000 gut bacterial genomes using structure-informed pLM predictions, expanding reaction coverage by a factor of two with a focus on CVD-relevant pathways, including TMAO, ImP, and bile acid metabolism. These predictions will be validated through heterologous expression and targeted metabolomics. In Aim 2, I will develop an interpretable graph neural network incorporating ecological constraints and metabolic interactions to predict community-level metabolite production from microbiome composition, with each prediction traceable to specific microbial contributors and enzymatic pathways. In Aim 3, I will apply this framework to a deeply phenotyped cohort of 10,000 CVD patients to identify key microbial drivers of disease, perform in silico perturbation experiments, and design personalized interventions including dietary modifications, targeted probiotics, and precision antibiotics. These experiments aim to establish mechanistic links between specific microbes, metabolites, and CVD outcomes, and may result in microbiome-targeted therapies that reduce CVD morbidity and mortality within the next 3-5 years. The proposed work will leverage my expertise in computational modeling and machine learning, supplemented by new training in experimental bacterial culturing and metabolomics, as well as CVD medicine and clinical translation. An interdisciplinary mentoring team bridging computational and clinical domains, along with UCSF's exceptional microbiome and CVD research environment, will provide ideal support for my proposed scientific and professional development, enabling me to establish a successful independent research program at the intersection of microbiome science and precision CVD medicine.
