Generative AI Designs Novel Antibiotics That Defeat Defiant Drug-Resistant Superbugs

Harnessing generative AI, MIT scientists have created groundbreaking antibiotics with unique membrane-targeting mechanisms, offering fresh hope against two of the world’s most formidable drug-resistant pathogens.

Research: A generative deep learning approach to de novo antibiotic design. Image Credit: Dabarti CGI / Shutterstock

With the help of artificial intelligence, MIT researchers have designed entirely new antibiotics capable of tackling two of today’s toughest bacterial threats: drug-resistant Neisseria gonorrhoeae and multi-drug-resistant Staphylococcus aureus (MRSA).

Using generative AI, the team explored an enormous chemical universe, designing more than 36 million hypothetical compounds and screening them computationally for antimicrobial potential. The most promising candidates turned out to be structurally unlike any existing antibiotic and appear to attack bacteria by novel mechanisms, chiefly by disrupting their protective cell membranes.

“We’re excited about the new possibilities that this project opens up for antibiotics development,” says James Collins, senior author of the study and the Termeer Professor of Medical Engineering and Science at MIT. “Our work shows the power of AI from a drug design standpoint, and enables us to exploit much larger chemical spaces that were previously inaccessible.” The results are published in the journal Cell, with MIT postdoc Aarti Krishnan, former postdoc Melis Anahtar ’08, and Jacqueline Valeri, PhD ’23, as lead authors.

Expanding the search

For decades, new antibiotics have largely been minor variations on old ones. In the past 45 years, only a few dozen have been approved by the U.S. Food and Drug Administration, and resistance to many of them is rising fast. Globally, drug-resistant bacterial infections are estimated to contribute to nearly 5 million deaths annually.

Collins and his colleagues at MIT’s Antibiotics-AI Project have already made headlines by using AI to screen existing chemical libraries, discovering candidates such as halicin and abaucin. This time, they pushed further, tasking AI with inventing entirely new molecules that don’t yet exist in any database.

The researchers used two strategies. In one, they began with a known chemical fragment that had antimicrobial activity and asked their algorithms to build full molecules around it. In the other, they let the AI generate plausible molecules from scratch, guided only by chemical rules rather than any specific starting point.

Targeting N. gonorrhoeae

The fragment-based search began with a massive library of about 45 million possible chemical fragments, made from combinations of carbon, nitrogen, oxygen, fluorine, chlorine, and sulfur, plus options from Enamine’s REadily AccessibLe (REAL) space. A machine-learning model previously trained to spot antibacterial activity against N. gonorrhoeae narrowed this pool to 4 million. Filtering out toxic, unstable, or already-known antibiotic-like structures left about 1 million candidates.

Further screening led to a fragment called F1, which the team fed into two generative AI systems. One, chemically reasonable mutations (CReM), tweak a starting molecule by adding, swapping, or removing atoms and groups. The other, a fragment-based variational autoencoder (F-VAE), builds complete molecules by learning how fragments are typically combined, based on over 1 million examples from the ChEMBL database.

These algorithms produced about 7 million F1-containing candidates, which were whittled down to 1,000 and then to 80, which were considered suitable for synthesis. Only two could be made by chemical vendors, and one, dubbed NG1, proved highly effective against N. gonorrhoeae in both lab tests and a mouse model of drug-resistant gonorrhea. NG1 works by interfering with LptA, a protein critical for constructing the bacterium’s outer membrane, fatally compromising the cell.

Designing without constraints

The second approach targeted S. aureus, this time with no predefined fragment. Again, using CReM and a variational autoencoder, the AI generated over 29 million chemically plausible molecules. After applying the same filters, about 90 remained. Twenty-two of these were synthesized, and six showed potent activity against multidrug-resistant S. aureus in lab tests. The most promising, DN1, cleared MRSA skin infections in mice. Like NG1, DN1 appears to damage bacterial membranes, but through broader mechanisms not tied to a single protein.

Next steps

Phare Bio, a nonprofit partner in the Antibiotics-AI Project, is now refining NG1 and DN1 to prepare them for more advanced testing. “We’re exploring analogs and advancing the best candidates preclinically, through medicinal chemistry work,” Collins says. “We’re also excited about applying these platforms toward other bacterial pathogens, notably Mycobacterium tuberculosis and Pseudomonas aeruginosa.”

For a field where resistance often outpaces discovery, the ability to rapidly explore vast, uncharted chemical space offers a fresh advantage. By combining computational muscle with medicinal chemistry, the MIT team hopes to stay ahead in the race against antibiotic resistance and perhaps rewrite the rulebook for how new drugs are found.