For patients with IBD, antibiotics can be a double-edged sword. Broad-spectrum medications, often prescribed for gut flare-ups, can kill both beneficial and harmful microbes, sometimes making symptoms worse over time. When fighting IBD, you don’t always want to exploit a hammer in a knife fight.
Scientists from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and McMaster University did it identified a new compound this requires a more targeted approach. A molecule called enterololin suppresses a group of bacteria associated with flare-ups of Crohn’s disease, leaving the rest of the microbiome largely intact. Using a generative artificial intelligence model, the team mapped how the compound worked. This process usually takes years, but in this case it was sped up to just months.
“This discovery highlights a major challenge facing antibiotic development,” says Jon Stokes, senior author of the book a new article about workassistant professor of biochemistry and biomedical sciences at McMaster and research associate at the Abdul Latif Jameel Machine Learning Clinic at MIT. “The problem isn’t finding bacteria-killing molecules in a dish – we’ve been able to do that for a long time. The main hurdle is figuring out what these molecules actually do in the bacteria. Without this detailed understanding, early-stage antibiotics can’t be developed into safe and effective treatments for patients.”
Enterololin is a step towards precision antibiotics: treatments designed to eliminate only the bacteria causing problems. In mouse models of Crohn’s disease-like inflammation, the drug targeted a gut-dwelling bacterium that can worsen flares while leaving most other microbes untouched. Mice given enterololin recovered faster and maintained a healthier microbiome than those given vancomycin, a common antibiotic.
Determining a drug’s mechanism of action, i.e. the molecular target to which it binds inside bacterial cells, usually requires years of painstaking experimentation. Stokes’ lab discovered enterololin using a high-throughput screening method, but the bottleneck would be determining its target. In this case, the team turned to DiffDock, a generative artificial intelligence model developed at CSAIL by MIT graduate student Gabriele Corso and MIT professor Regina Barzilay.
DiffDock aimed to predict how petite molecules fit into protein binding pockets, an extremely hard problem in structural biology. Conventional docking algorithms search through possible orientations using scoring rules, often producing loud results. Instead, DiffDock views docking as a probabilistic reasoning problem: the diffusion model iteratively refines its guesses until it converges on the most likely binding mode.
“In just a few minutes, the model predicted that enterololin binds to a protein complex called LolCDE, which is necessary for the transport of lipoproteins in some bacteria,” says Barzilay, who is also co-director of the Jameel Clinic. “It was a very specific lead – one that could guide experiments, not replace them.”
Stokes’ group then put this prediction to the test. Using DiffDock’s predictions as an experimental GPS, they first evolved enterololin-resistant mutants in the lab, which revealed that changes in the mutant’s DNA mapped to lolCDE, exactly where DiffDock predicted enterololin would bind. They also performed RNA sequencing to see which bacterial genes were turned on and off by the drug, and used CRISPR technology to selectively downregulate the expression of an expected target. Together, these laboratory experiments revealed disruptions in pathways related to lipoprotein transport, just as DiffDock predicted.
“When you see a computational model and wet lab data pointing to the same mechanism, that’s when you start to believe you’ve discovered something,” Stokes says.
According to Barzilay, the project highlights a shift in the way artificial intelligence is used in the life sciences. “Many applications of artificial intelligence in drug discovery involve searching the chemical space and identifying new molecules that may be active,” he says. “Here we show that AI can also provide mechanistic explanations that are crucial for moving a molecule through the development process.”
This distinction is critical because mechanism of action studies are often the main rate-limiting step in drug development. Conventional methods can take anywhere from 18 months to two years or longer and cost millions of dollars. In this case, the MIT–McMaster team reduced the time to about six months at a fraction of the cost.
Enterololin is still in its early stages of development, but translation is already underway. Stokes’ spinout company, Stoked Bio, has licensed the compound and is optimizing its properties for potential exploit in humans. Early work also includes testing derivatives of this molecule against other resistant pathogens such as . If all goes well, clinical trials could begin within the next few years.
Scientists also see broader implications. Narrow-spectrum antibiotics have long been sought as a way to treat infections without collateral damage to the microbiome, but they have been hard to discover and validate. Artificial intelligence tools such as DiffDock could make this process more practical, enabling the rapid development of a fresh generation of targeted antimicrobials.
For patients with Crohn’s disease and other inflammatory bowel diseases, the prospect of a drug that relieves symptoms without destabilizing the microbiome could mean a significant improvement in quality of life. More broadly, precision antibiotics can assist combat the growing threat of antimicrobial resistance.
“I’m excited not only by this connection, but also by the idea that we can start to think about elucidating the mechanism of action as something we can do faster, with the right combination of artificial intelligence, human intuition and laboratory experiments,” Stokes says. “This could change the way we approach drug discovery for many diseases, not just Crohn’s disease.”
“One of the greatest challenges to our health is the rise of antimicrobial-resistant bacteria, which evade even the best antibiotics,” adds Yves Brun, a professor at the University of Montreal and distinguished professor emeritus at Indiana University Bloomington, who was not involved in the work on the publication. “Artificial intelligence is emerging as an important tool in our fight against these bacteria. This study uses a powerful and elegant combination of artificial intelligence methods to determine the mechanism of action of a new antibiotic candidate, an important step in its potential development as a therapeutic agent.”
Corso, Barzilay and Stokes wrote the paper with McMaster researchers Denise B. Catacutan, Vian Tran, Jeremie Alexander, Yeganeh Yousefi, Megan Tu, Stewart McLellan and Dominique Tertigas, and professors Jakob Magolan, Michael Surette, Eric Brown and Brian Coombes. Their research was supported in part by the Weston Family Foundation; David Braley Antibiotic Discovery Center; Canadian Institutes of Health Research; Natural Sciences and Engineering Research Council of Canada; M. and M. Heersink; Canadian Institutes of Health Research; Ontario Graduate Scholarship Award; Jameel Clinic; and the U.S. Defense Threat Reduction Agency’s program to discover medical countermeasures against fresh and emerging threats.
The researchers published the sequencing data in public repositories and openly published the DiffDock-L code on GitHub.