MIT researchers use explainable AI to identify new structural class of antibiotics against MRSA

Researchers at the Massachusetts Institute of Technology (MIT), the Broad Institute of MIT and Harvard, and collaborating institutions have used artificial intelligence to identify a new structural class of antibiotic compounds effective against methicillin-resistant Staphylococcus aureus (MRSA). The findings were published in Nature on December 20, 2023 (nature.com).

MRSA infects more than 80,000 people in the United States each year and causes more than 10,000 deaths annually, often leading to skin infections, pneumonia, or life-threatening sepsis, according to the MIT News report on the study (news.mit.edu). The work forms part of the Antibiotics-AI Project led by James Collins, the Termeer Professor of Medical Engineering and Science at MIT.

The team trained deep learning models, specifically ensembles of graph neural networks, on experimental data from testing approximately 39,312 compounds for antibiotic activity against MRSA, along with their chemical structures. Additional models predicted toxicity against three types of human cells. These models then screened roughly 12 million commercially available compounds (broadinstitute.org).

A key advance was making the models explainable. The researchers adapted a Monte Carlo tree search algorithm to identify specific chemical substructures, or rationales, that the models associated with antibiotic activity and low cytotoxicity. This revealed compounds belonging to five predicted structural classes. Of about 283 purchased and tested compounds, two from the same novel class showed strong activity (eurekalert.org).

In laboratory tests, the compounds killed MRSA. In two mouse models—one of MRSA skin infection and one of systemic thigh infection—each compound reduced the bacterial population by a factor of about 10. They also proved effective against vancomycin-resistant enterococci and showed very low toxicity to human cells. Experiments indicated the compounds act by selectively dissipating the proton motive force across bacterial cell membranes, disrupting the electrochemical gradient essential for energy production (ATP) and other functions, without substantially damaging human cell membranes (news.mit.edu).

Lead authors Felix Wong, a postdoc at MIT’s Institute for Medical Engineering and Science and the Broad Institute, and Erica Zheng, then a Harvard Medical School graduate student, noted that the explainable approach provides chemical insights previously lacking in black-box AI models for drug discovery. “Our work provides a framework that is time-efficient, resource-efficient, and mechanistically insightful, from a chemical-structure standpoint,” Collins said in the MIT announcement.

The compounds have been shared with Phare Bio, a nonprofit affiliated with the Antibiotics-AI Project, for further analysis of chemical properties and potential clinical development. Collins’ laboratory is using similar methods to design additional candidates and target other bacterial pathogens. The research received funding from sources including the National Institute of Allergy and Infectious Diseases, the James S. McDonnell Foundation, and the Audacious Project (nature.com).

This discovery adds to prior AI-assisted antibiotic finds by the same group, such as halicin in 2020, and demonstrates how explainable deep learning can accelerate identification of novel structural classes amid rising antimicrobial resistance.