Finding Gut Partners Among Trillions of Combinations: AI and Robotic Experiments Design Butyrate-Producing Microbial Communities

When probiotics enter the gut, they do not encounter an empty space, but rather a diverse microbial ecosystem that varies from person to person. These differences mean that the same probiotic or dietary fiber may not produce the same effect in everyone. A Duke University team has now combined machine learning with robotic experiments in an effort to identify microbial and fiber pairings that are less susceptible to interference from existing microbial communities within a vast combination space.

The research was published in *Nature Chemical Biology*. The team selected 15 gut microorganisms and six dietary fibers, using beneficial metabolites such as butyrate and community characteristics as design targets. Even with only 21 variables, the possible combinations of microbial species and diets can exceed one trillion, making exhaustive testing impossible.

To narrow the search space, the researchers established a “design–test–learn” cycle: the model first selected conditions that could best fill knowledge gaps while also showing promise for achieving the targets. An automated platform then cultured and analyzed the microbial communities, and the resulting data were fed back into the model for the next round. The system completed five batches of high-throughput experiments, testing up to 390 conditions in parallel per batch and concentrating the search on areas with higher information value.

The analysis identified an ecological combination centered on inulin. When *Bacteroides uniformis* and *Anaerostipes caccae* were both present, together with a higher-order interaction involving *Prevotella copri*, the community produced butyrate more consistently. Butyrate is a short-chain fatty acid formed when gut bacteria break down fiber and is associated with energy supply to colon cells and immune regulation. However, increased production alone does not mean that efficacy against disease has been demonstrated.

Validation of the work was not limited to synthetic communities consisting of only a few microbial species. After the researchers added the model-designed combinations of microbial species and fibers to human fecal microbial communities, they still observed predictable beneficial metabolic outputs, indicating that the combinations retained a degree of resilience in a more complex ecological setting. The paper also made publicly available the data used for model training and experimental design, sequencing data, and analysis code, enabling other teams to evaluate the method.

At this stage, the findings remain an in vitro proof of concept. The tested microbial species and fibers represent only a small fraction of the real gut ecosystem, and the study has not yet answered whether the strains can colonize over the long term, how different diets and medications may alter the effects, or whether increasing butyrate will translate into measurable clinical benefits. The team is evaluating candidate combinations in a mouse model of inflammatory bowel disease. A university report described the early results as promising, but complete animal data have not yet been published, so no conclusions about efficacy in humans can be drawn from them. For now, the platform’s most practical value lies in shifting the development of probiotics and prebiotics from empirical formulations toward ecological engineering that can be tested repeatedly.