Some today fear that artificial intelligence will one day destroy humanity. But if the rise of the machines doesn’t get us, drug-resistant bacteria just might. These microscopic killers already claim millions of lives each year worldwide, and the world’s arsenal of effective antibiotics is dwindling.
But could one threat be trained perhaps to help stave off the other? A study published today in the journal Cell certainly suggests the possibility.
A team led by Jim Collins, MIT professor of biological engineering, showed how generative AI algorithms trained on vast datasets of antibacterial substances could dream up millions of previously unimagined molecules with predicted microbe-killing power—some of which proved potent in mouse experiments.
The researchers synthesized a small subset of these AI-designed molecules and found them lethal to superbugs responsible for drug-resistant gonorrhea and stubborn staphylococcus skin infections.
“It’s a great addition to this emerging field of using AI for antibiotic discovery,” says César de la Fuente, a synthetic biologist at the University of Pennsylvania who was not involved in the research.
“It shows quite well how generative AI can produce molecules with real-world activity,” he adds. “It’s elegant and potentially clinically meaningful.”
A social-enterprise non-profit created by Collins, called Phare Bio, now plans to advance these and other AI-discovered antibiotics toward clinical development.
The candidate antibiotics build on earlier finds from Collins’ lab—including halicin, a potent broad-spectrum antibiotic identified in 2020; a more targeted agent called abaucin with activity against Acinetobacter baumannii, a major cause of hospital-acquired infections; and a novel structural class of molecules described last year that proved effective against the superbugs MRSA and VRE.
With the team’s earlier discoveries, however, Collins and his colleagues were still mining existing chemical libraries, using deep-learning models to spot overlooked compounds with antibacterial potential.
The new work sets down a new path altogether: rather than searching for hidden gems in familiar territory, the generative AI platform starts from scratch, conjuring entirely new molecular structures absent from any database.
“This is moving from using AI as a discovery tool to using AI as a design tool,” Collins says. The shift, he adds, opens new frontiers in antibiotic discovery—unexplored territory that could harbor the next generation of lifesaving drugs.
Explore: Best Free AI Tools One Click
Anti-Germ Intelligence Proves Its Mettle
To train their generative AI model, Collins and his colleagues first used a neural network framework to virtually screen more than 45 million chemical fragments—the building blocks of would-be drugs—looking for pieces predicted to have activity against Neisseria gonorrhoeae (the cause of sexually transmitted gonorrhea infections) and Staphylococcus aureus (the germ behind deadly bloodstream infections, pneumonia, and flesh-eating skin disease).
Two algorithms then went to work: one assembling the fragments into complete molecular structures, the other predicting which of those structures would pack the strongest antibacterial punch.
Together, the algorithms generated more than 10 million candidate molecules, none of which had ever existed before. But then came what MIT study author and computational biologist Aarti Krishnan describes as “a massive bottleneck”: very few of these prophesied antibiotics could actually be made in the lab.
The researchers manually sifted through the AI hits, filtering for properties suggestive of drug-likeness and synthetic feasibility. They ultimately arrived at a shortlist of around 200 promising designs, 24 of which could be successfully generated.
Seven proved to be bona fide antimicrobial agents, as confirmed by laboratory tests, with two showing particularly strong efficacy in mouse models of gonorrhea and staph infections. Notably, each seems to work through a distinct and novel mechanism of action not exploited by existing antibiotics.
Explore: Best Free AI Tools One Click
“That’s pretty cool,” says Phare co-founder Jonathan Stokes, an antimicrobial chemical biologist at Canada’s McMaster University in Hamilton, Ontario. He praises Collins’ team for unearthing two highly promising antibiotic leads but notes that the labor-intensive trial-and-error process underscores how far the technology still has to go in producing compounds that can be readily synthesized.
“It’s a bit of an elephant in the room,” he says of synthetic tractability in GenAI drug discovery. “Antibiotics, because of the financial disincentives in this space, have to be cheap,” Stokes, who was not involved in the research, says.
“They have to be cheap to discover, cheap to develop, and cheap to make. So if there are opportunities to avoid all of these issues with synthetic feasibility, I feel like that is a major advantage.”
Moving From Model to Molecule
To tackle that challenge, Stokes and his colleagues developed a generative AI tool that designs antibiotic candidates with chemical blueprints tailored for real-world manufacturing, not just computer screens.
This tool, called SyntheMol, operates within a more limited chemical space than Collins’ GenAI model, choosing only molecules whose building blocks can be synthesized with known, lab-proven reaction steps.
Explore: Best Free AI Tools One Click
That narrows the search parameters to tens of billions of molecules, compared to the 1060 possible structures that Collins’ model explored.
It’s enough, however, for SyntheMol to have already yielded several drug candidates that Stokes and his colleagues, through a startup called Stoked Bio, hope to develop into treatments for bacteria linked to Crohn’s disease and other hard-to-treat conditions.
The team aims to balance the sheer breadth of biochemical possibilities the models can explore with crucial metrics like drug potency, safety, low toxicity, and ease of synthesis.
“It’s a multi-objective optimization problem,” says de la Fuente, who advises Phare and builds his own generative AI models to design antimicrobial peptide drugs.
But for now, the tools powering Phare’s discovery efforts—rooted in Collins’ approaches—are already delivering early wins, says Akhila Kosaraju, Phare Bio’s CEO and president.
“We are getting substantially more potent and less toxic initial compounds,” she notes. And backed by the U.S. government’s Advanced Research Projects Agency for Health (ARPA-H), along with the philanthropic arm of Google—which is funding Phare to build open-source infrastructure around AI-guided antibiotic design— Kosaraju and her colleagues aim to move the most promising candidates into human trials.
“We are building what we think is the most novel and robust pipeline of antibiotics in the world,” she says.
Explore: Best Free AI Tools One Click
