Researchers have tapped into the power of generative artificial intelligence to aid them in the fight against one of humanity’s most pernicious foes: antibiotic-resistant bacteria. Using a model trained on a library of about 40,000 chemicals, scientists were able to build never-before-seen antibiotics that killed two of the most notorious multidrug-resistant bacteria on earth.
Two lead compounds, NG1 and DN1, both were able to eliminate multidrug-resistant gonorrhea, while DN1 could also kill methicillin-resistant Staphylococcus aureus (MRSA). MRSA is “probably the most famous of the resistant pathogens,” James Collins, Ph.D., a synthetic biologist at the Broad Institute of MIT and Harvard and leader of the new study, told Fierce Biotech.
Details of the model and its chemical creations were published in Cell on Aug. 14.
Collins is now working with Phare Bio, a nonprofit antibiotic developer he co-founded in 2020, to move the two candidates toward the clinic, he said.
Collins’ team has been trying to understand how antibiotics work and how to make new ones for more than two decades. He launched a collaboration in 2018 with two AI experts at MIT to develop deep learning models that can screen chemicals for antibiotic activity. This led to the 2020 discovery of a new antibiotic called halicin and, in late 2023, an improved model revealed an entirely new class of bacteria-killing compounds hidden within a vast chemical library.
After these successes, the team decided to aim higher: using models not to discover antibiotics but to design entirely new ones.
“We developed two different generative AI platforms that enabled us to build out molecules with antibacterial activity, non-toxicity and drug-like properties in mind,” Collins said.
The first of these models, used to build NG1, was trained on compounds known to be effective against gonorrhea. The model analyzed fragments of the chemicals and rated their antibacterial activity, finding that one structure, called F1, seemed particularly potent. Using F1 as a starting point, the model iteratively added elements and judged the resulting compounds’ potential to subdue bacteria as it went.
The second model, which made DN1, did something similar, but instead of starting from a fragment of a molecule, it built new compounds starting from a single atom.
The models essentially “decorate” the starting fragment or atom, Collins said, “meaning you're adding bonds and additional elements to it" and at each step predicting the compound's antibiotic potential and toxicity.
Of several hundred candidates the models identified, Ukrainian manufacturing partner Enamine was able to synthesize 24 of them. Of that bunch, seven showed antibiotic effects, including the two most promising: NG1 and DN1.
Being able to synthesize an AI-designed compound is the main hurdle to using the technique to develop new antibiotics, Collins said. His team’s models did estimate the likelihood that a generated chemical could actually be made, using techniques developed by MIT chemical engineer Connor Coley, Ph.D.
“The real stopgap in all of this is not computation time or computation resources,” Collins said. “It's really, can you synthesize the molecules you come up with?”
The code for the new models is publicly available, and Collins hopes that groups with access to more money and larger chemical libraries—like pharma companies and national governments—will adopt the technique to develop sorely needed novel antibiotics.
“I think if we could develop 10 to 15 new antibiotics, we could be in a much better state,” Collins said. “It's not that expensive for a nation-state, or even one of our more wealthy philanthropists, to address that.”