MSACL 2026 Abstract

Graph Neural Network-Enabled Discovery of Hidden Enzymes Driving Microbial Bile Acid Metabolism

Min Soo Kim (1), Shuchang Tian (1), Benjamin D. Anderson (1), Bethany Swencki-Underwood (1,2), Andrew D. Patterson (1,3,4), Weston Whitaker (3,4), Jordan E. Bisanz (1,3)
(1) Department of Biochemistry and Molecular Biology, Pennsylvania State University, University Park, PA, (2) Department of Biology, The Pennsylvania State University, University Park, PA, (3) One Health Microbiome Center, Huck Life Sciences Institute, University Park, PA, (4) Department of Veterinary and Biomedical Sciences, Center for Molecular Toxicology and Carcinogenesis, The Pennsylvania State University, University Park, PA

Min Soo Kim (Presenter) Pennsylvania State University

Relevant Financial Disclosures (within past 24 months, reported on Jun 10, 2026) No relevant financial relationship(s) to disclose.

INTRODUCTION:
Bile acids (BAs) are among the most clinically significant microbially modified metabolites in the human gut, with secondary BA profiles varying substantially across individuals and influencing immune function, metabolic disease, and intestinal homeostasis. Mass spectrometry-based metabolomics can capture this chemical diversity with precision, but interpreting which members of a complex microbial community drive specific BA transformations remains a fundamental challenge. Homology-based functional annotation systematically fails to detect enzymes that have diverged below sequence identity thresholds, leaving a substantial fraction of BA metabolizing capacity invisible.

METHODS:
A graph neural network (GNN) was developed using defined synthetic microbial communities (SynComs) with paired community composition and LC/MS-based BA quantitation to learn which bacterial species associate with specific BA transformations without homology-based functional annotation. The framework was adapted for a paired metagenomic-metabolomic human study encompassing 1,290 bacterial species and 23 BA features. Permutation-based importance analysis identified bacterial species associated with enrichment of immunomodulatory 3-epiDCA, and candidate species were experimentally validated for production of the metabolite. Subsequent in silico structure-based screening determined putative genes within validated species that were responsible for enzymatic activity, and the hits were experimentally validated through heterologous expression in E. coli.

RESULTS:
The GNN trained on metabolomics data from human cohort maintained robust predictive performance, identifying bacterial species associated with bile acid epimerization independent of genomic annotation. A consortium of candidate species produced immunomodulatory 3-epiDCA and its intermediate 3-oxoDCA despite lacking predicted hydroxysteroid dehydrogenase (HSDH) genes required for this transformation. Structure-based screening and heterologous expression identified divergent 3b-HSDHs sharing less than 40% amino acid sequence identity with characterized reference enzymes. The discovered HSDHs were detected in nearly 20% of human gut metagenomes, with distinct enrichment patterns observed across inflammatory bowel disease and cancer.

CONCLUSION:
Mass spectrometry-derived metabolic phenotypes, analyzed through a phenotype-driven GNN framework, enable functional discovery of microbial enzymes that systematically evade homology-based detection. The approach revealed novel bile acid epimerization capacity associated with chronic diseases, including inflammatory disease and cancer. This framework is generalizable to other microbial biotransformations and represents a scalable strategy for extracting mechanistic functional insight from existing clinical and population-level metabolomics datasets.