Artificial intelligence is becoming a more practical tool in the search for new medicines. In a study focused on methicillin-resistant Staphylococcus aureus, or MRSA, MIT researchers identified a new class of compounds that could become antibiotic candidates against a dangerous drug-resistant bacterium.
Why MRSA remains a hard target
MRSA is a drug-resistant bacterium linked to more than 10,000 deaths a year in the US. That makes it a major challenge for medicine, especially because multiresistant pathogens can limit the usefulness of existing treatments.
The MIT work does not describe an approved drug. Instead, it identifies compounds with features that make them worth deeper study: they can kill MRSA, and they appear to have low toxicity to human cells.
That combination matters. A potential antibiotic has to do more than damage bacteria. It also has to avoid causing unacceptable harm to the patient, which is why the researchers included toxicity prediction as a central part of the selection process.
How the AI search worked
The team used deep learning models to search for compounds with likely antimicrobial activity. A key part of the method was making those models easier to interpret, rather than treating their outputs as a black box.
To do that, the researchers used the Monte Carlo tree search algorithm. This helped them examine how the models reached their predictions and identify molecular substructures that were likely connected to antimicrobial activity.
That interpretability is important for drug discovery. If researchers can see which parts of a molecule may be driving the effect, they gain a clearer path for deciding what to study next and how future compounds might be designed.
The team then added another filter: safety. They trained three additional models to predict toxicity across three types of human cells. By combining those toxicity predictions with antimicrobial predictions, they narrowed the search toward compounds that could attack microbes while having minimal negative effects on humans.
From millions of compounds to a smaller shortlist
The researchers screened around 12 million commercially available compounds. From that large pool, they identified five different classes with likely activity against MRSA.
Two compounds stood out as especially promising antibiotic candidates. In two mouse models, those compounds reduced the MRSA population by a factor of 10.
The source article also describes a possible mechanism. The compounds appear to disrupt the bacteria's ability to maintain an electrochemical gradient across their cell membranes. That gradient is essential for key cellular functions, so interfering with it could weaken or kill the bacteria.
The result is not just a list of molecules. It is also a demonstration of a workflow:
- Use AI models to predict antimicrobial activity.
- Apply interpretability methods to understand which molecular features may matter.
- Predict toxicity to human cells before prioritizing compounds.
- Test the most promising candidates in biological models.
Each step reduces the field of possibilities. Together, they turn a huge chemical search space into a more focused set of candidates for further research.
What happens next
The research has been shared with Phare Bio, a non-profit organization founded by some researchers involved in the work. Phare Bio will carry out more detailed analyses of the compounds' chemical properties and potential clinical applications.
At the same time, the laboratory is working to develop further drug candidates based on the study results. That means the work is moving in two directions: deeper analysis of the identified compounds and continued design of additional candidates.
The researchers also believe the same AI-based approach can be applied beyond MRSA. Felix Wong, a postdoctoral fellow at the Institute for Medical Engineering and Science (IMES) at MIT and Harvard University, said, "We are already leveraging similar approaches based on chemical substructures to design compounds de novo, and of course, we can readily adopt this approach out of the box to discover new classes of antibiotics against different pathogens."
The study included researchers from MIT and Harvard University, as well as the Leibniz Institute for Polymer Research and the Max Bergmann Center of Biomaterials in Germany.
Why this matters for AI in medicine
This work shows one of the more grounded uses of AI in medicine: narrowing a difficult search. Antibiotic discovery requires finding compounds that affect bacteria, avoiding compounds likely to harm human cells, and understanding enough about the chemistry to guide the next round of research.
The MRSA project addresses all three. The AI models helped search across around 12 million compounds. The toxicity models helped screen for safer profiles. The interpretability step gave the researchers a way to connect predictions to chemical substructures.
The result is a set of promising antibiotic candidates, not a finished therapy. But for a field facing multiresistant pathogens, even a credible new class of antibacterials is meaningful. It gives researchers something concrete to analyze, refine, and test further.