Antibiotics present a significant challenge for patients with inflammatory bowel disease (IBD). While frequently prescribed for gut flare-ups, these broad-spectrum medications are known to eliminate beneficial microbes alongside harmful ones. This indiscriminate action can, paradoxically, intensify symptoms over time, underscoring the need for more targeted therapeutic approaches in managing gut inflammation.
Researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), in collaboration with McMaster University, have pinpointed a novel compound, enterololin, that offers a highly targeted intervention for Crohn’s disease. This molecule uniquely suppresses specific bacterial groups associated with disease flare-ups, crucially leaving the broader gut microbiome largely undisturbed. The team also leveraged a generative AI model to rapidly decipher the compound’s mechanism of action, accelerating a process that typically spans years into just a few months.
A critical challenge in antibiotic development lies not in identifying compounds that eradicate bacteria in a laboratory setting, but in deciphering their exact mechanisms of action within bacterial cells, explained Jon Stokes. Stokes, a senior author of new research, assistant professor of biochemistry and biomedical sciences at McMaster University, and a research affiliate at MIT’s Abdul Latif Jameel Clinic for Machine Learning in Health, emphasized that this detailed understanding is essential for transforming early-stage antibiotic discoveries into safe and reliable treatments for patients.
Enterololin marks a significant step toward precision antibiotics, offering treatments engineered to eliminate only the problematic bacteria. In studies using mouse models of Crohn’s-like inflammation, the drug effectively targeted *Escherichia coli*, a gut bacterium known to exacerbate flare-ups, while largely preserving the majority of other microbial residents. Mice administered enterololin demonstrated quicker recovery and maintained a healthier microbiome compared to those treated with vancomycin, a conventional antibiotic.
Uncovering a drug’s precise mechanism of action—identifying the specific molecular target it binds within bacterial cells—traditionally necessitates years of painstaking experiments. While Stokes’ laboratory successfully discovered enterololin using a high-throughput screening approach, the subsequent task of determining its target was expected to be the primary bottleneck. To expedite this crucial step, the research team employed DiffDock, a generative AI model pioneered at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) by PhD student Gabriele Corso and Professor Regina Barzilay.
DiffDock was engineered to tackle a formidable challenge in structural biology: predicting the precise fit of small molecules within protein binding pockets. While conventional docking algorithms traditionally navigate potential molecular orientations using scoring rules—a method often fraught with imprecise outcomes—DiffDock introduces a fundamentally different paradigm. It reframes the docking problem as one of probabilistic reasoning, employing a diffusion model that incrementally refines initial approximations until it converges on the most probable binding configuration.
Within minutes, a computational model accurately predicted that enterololin binds to the LolCDE protein complex. This complex is vital for transporting lipoproteins in specific bacteria, a discovery Barzilay, who also co-leads the Jameel Clinic, characterized as a “very concrete lead.” He underscored that such a finding provides a clear direction for experiments rather than a substitute for them.
Stokes’ research group subsequently embarked on the experimental validation of their prediction. Guided by DiffDock’s computational insights, they first cultured enterololin-resistant *E. coli* mutants. Genomic analysis of these resistant strains pinpointed DNA alterations within the lolCDE region, corroborating DiffDock’s precise forecast of enterololin’s binding location.
The team also conducted RNA sequencing to map changes in bacterial gene expression upon drug exposure and employed CRISPR gene-editing to selectively silence the predicted target. Collectively, these diverse laboratory experiments consistently unveiled significant disruptions in lipoprotein transport pathways, aligning precisely with DiffDock’s initial predictions.
According to Stokes, a robust understanding of a scientific mechanism solidifies when both computational models and experimental (‘wet-lab’) data independently corroborate the same underlying process.
The project signifies a notable shift in artificial intelligence’s application within the life sciences, according to Barzilay. She explains that while AI in drug discovery has largely concentrated on exploring chemical spaces to pinpoint new, potentially active molecules, this current work showcases AI’s capacity to deliver crucial mechanistic explanations. Such insights are considered indispensable for successfully navigating a molecule through the complex development pipeline.
Understanding a drug’s mechanism of action is a critical, often rate-limiting stage in pharmaceutical development. Traditional studies for this process typically span 18 months to two years, or even longer, and carry costs running into millions of dollars. However, the MIT-McMaster team significantly streamlined this timeline, completing the work in approximately six months and at a mere fraction of the conventional expenditure.
Enterololin, a compound currently in its early stages of development, is already undergoing active translation. Stokes’ spinout company, Stoked Bio, has licensed the compound and is now focused on optimizing its properties for eventual human application. Concurrently, preliminary research is exploring derivatives of Enterololin to combat other resistant pathogens, including Klebsiella pneumoniae. Should development proceed favorably, clinical trials for the compound are anticipated to commence within the next few years.
The research carries broader implications, suggesting a significant shift in antimicrobial development. Scientists have long sought narrow-spectrum antibiotics, which treat infections without the collateral damage to the microbiome often associated with broader treatments. Yet, the discovery and validation of these targeted drugs have proven exceptionally difficult. AI tools like DiffDock could make this process far more practical, potentially accelerating the creation of a new generation of precise antimicrobials.
The development of a medication that can alleviate symptoms for patients with Crohn’s disease and other inflammatory bowel conditions, without disturbing the delicate balance of their microbiome, holds the potential for a significant improvement in their daily lives. On a larger scale, such precision antibiotics are poised to become a vital tool in confronting the escalating global threat of antimicrobial resistance.
Stokes articulated his enthusiasm, stating it extends beyond any single compound to encompass the revolutionary potential of rapidly deciphering a drug’s mechanism of action. He suggests this crucial understanding can be significantly accelerated by judiciously blending artificial intelligence, human insight, and rigorous laboratory experiments. This innovative paradigm, Stokes believes, could fundamentally transform the landscape of drug discovery for a multitude of diseases, reaching far beyond conditions like Crohn’s.
The escalating challenge of antimicrobial-resistant bacteria, capable of circumventing even our most potent antibiotics, represents one of the greatest threats to public health, according to Yves Brun. A professor at the University of Montreal and distinguished professor emeritus at Indiana University Bloomington, Brun, who was not directly involved in the research, highlighted the growing role of artificial intelligence (AI) in this critical fight. He lauded the study for its “powerful and elegant combination of AI methods” to pinpoint the mechanism of action of a promising new antibiotic candidate, calling it “an important step in its potential development as a therapeutic.”
The research paper was co-authored by Corso, Barzilay, and Stokes, alongside a team of McMaster University researchers and professors, including Denise B. Catacutan, Vian Tran, Jeremie Alexander, Yeganeh Yousefi, Megan Tu, Stewart McLellan, Dominique Tertigas, Jakob Magolan, Michael Surette, Eric Brown, and Brian Coombes. Financial support for the study was provided by numerous organizations, notably the Weston Family Foundation, the David Braley Centre for Antibiotic Discovery, the Canadian Institutes of Health Research, and the U.S. Defense Threat Reduction Agency.
To ensure broad access, the research team deposited their sequencing information into public databases and made the DiffDock-L software code publicly available on GitHub.
