Elucidating bacterial therapeutic mechanisms against c. Difficile via physics-informed active machine learning - living bacterial therapeutics (LBT) are an emerging class of drugs that have the…

The National Institute of Allergy and Infectious Diseases awarded Duke University $609,185 on September 1, 2026, under the Allergy and Infectious Diseases Research program (CFDA 93.855) to develop a physics-informed machine learning model for elucidating the therapeutic mechanisms of bacterial therapeutics against Clostridium difficile infection.

The funded work centers on creating a Neural Species Mediator model to systematically understand how living bacterial therapeutics function in the human gut microbiome and to predict their mechanisms of action. The project addresses a critical gap: while fecal microbiota transplantation shows clinical success against recurrent C. difficile infection, the underlying inter-species interactions remain poorly understood, and the approach lacks scalability and carries risks of transferring antibiotic-resistant pathogens. The proposed computational framework, integrating Bayesian optimization with machine learning, will enable efficient navigation of the vast experimental design space for bacterial therapeutic candidates, capturing complex interaction modalities and revealing metabolites and compounds critical to community function.

Performance occurs in Durham, North Carolina. The award runs from September 1, 2026, through August 31, 2027. This is a Project Grant, the standard assistance type for investigator-initiated research under the program.

Elucidating bacterial therapeutic mechanisms against c. Difficile via physics-informed active machine learning - living bacterial therapeutics (LBT) are an emerging class of drugs that have the potential to treat or prevent a wide range of diseases involving substantial alterations of the human GUT microbiome. However, we lack an understanding of their mechanisms of action. As a testbed for developing and applying this framework, clostridium difficile (c. Difficile) is a leading cause of antibiotic associated diarrhea and colitis. The notable success of fecal microbiota transplantation (FMT) at treating recurrent c. Difficile infection (rcdi) indicates that GUT microbiota inter-species interactions are major variables impacting c. Difficile's fitness in the GUT. However, FMT suffers from critical limitations including unknown mechanisms of action, lack of scalability, transfer of potential antibiotic-resistant pathogens and potential negative health outcomes. While LBT hold tremendous promise as next-generation microbiome therapies for multiple human diseases, we lack a systematic and quantitative framework for understanding the networks driving their target functions, key inputs and outputs and potential failure modes. Developing this framework is critical for realizing their full potential for precision microbiome modulation of the human GUT microbiome. There are an astronomical number of LBT that could be constructed using commensal human GUT bacteria. Therefore, computational modeling approaches are essential for predicting and uncovering the factors shaping their functions. We propose to develop a novel physics-informed machine learning model (neural species mediator or NSM) that can be used to efficiently navigate this large experimental design landscape with bayesian optimization. Mediators represent metabolites or other compounds critical to the function of the community. In addition to capturing complex interaction modalities, the NSM can accurately reveal the metabolite-species interaction networks driving community dynamics and functions. Major advantages of this framework include the ability to capture the physical laws, allowing flexibility to adapt to experimental data and enabling interpretation to uncover key molecular interactions driving system functions. For the first time, we will develop a flexible computational model that can capture the metabolite-mediated regulation of c. Difficile toxin production. We will combine high-throughput bottom-up experiments, EXO-metabolomics, transcriptional profiling, advanced computational modeling and bayesian optimization. We will investigate the interactions between model-optimized communities and c. Difficile using a germ-free mouse model FED distinct diets. Targeted and untargeted metabolomics from our mouse experiments will guide further in vitro experiments to understand differences in in vitro and in vivo outcomes. In SUM, this generalizable framework for understanding the networks shaping LBT functions and their robustness to key sources of environmental variability can be applied more generally to other human pathogens, human diseases and microbiomes.