Researchers Utilize Generative AI to Create Compounds Effective Against Drug-Resistant Bacteria | MIT News

Leveraging artificial intelligence, researchers at MIT have developed innovative antibiotics that target two challenging infections: drug-resistant Neisseria gonorrhoeae and multi-drug-resistant Staphylococcus aureus (MRSA).

Utilizing generative AI algorithms, the research team created over 36 million potential compounds, screening them computationally for antimicrobial activity. The most promising candidates they identified are structurally unique compared to existing antibiotics and seem to operate through novel mechanisms that disrupt bacterial cell membranes.

This strategy enabled the researchers to generate and assess theoretical compounds that have never been previously encountered — a method they aspire to use for identifying and designing compounds effective against other bacterial species.

“We’re thrilled about the new avenues this project opens for antibiotic development. Our work illustrates the power of AI in drug design and allows us to explore significant chemical spaces previously thought to be out of reach,” says James Collins, the Termeer Professor of Medical Engineering and Science at MIT’s Institute for Medical Engineering and Science (IMES) and the Department of Biological Engineering.

Collins serves as the senior author of the study, which is published today in Cell. The lead authors of the paper include MIT postdoc Aarti Krishnan, former postdoc Melis Anahtar ’08, and Jacqueline Valeri PhD ’23.

Exploring chemical space

In the last 45 years, only a few dozen new antibiotics have gained FDA approval, most being variations of existing antibiotics. Concurrently, bacterial resistance to many of these medications has been on the rise. It is estimated that drug-resistant bacterial infections contribute to nearly 5 million deaths globally each year.

To address this growing issue by identifying new antibiotics, Collins and his colleagues at MIT’s Antibiotics-AI Project have leveraged AI to screen vast libraries of existing chemical compounds. This initiative has led to several promising candidates, including halicin and abaucin.

To enhance this momentum, Collins and his team opted to explore molecules that are absent from any chemical libraries. By employing AI to generate hypothetically possible molecules that are either undiscovered or non-existent, they aimed to investigate a greater diversity of potential drug compounds.

In their recent study, the researchers utilized two distinct methodologies: Firstly, they instructed generative AI algorithms to design molecules based on a specific chemical fragment exhibiting antimicrobial properties, and secondly, they allowed the algorithms to generate molecules freely, without any constraints.

For the fragment-based strategy, the researchers aimed to identify molecules capable of annihilating N. gonorrhoeae, a Gram-negative bacterium responsible for gonorrhea. They began by compiling a library of approximately 45 million known chemical fragments, formed from all permissible combinations of 11 atoms: carbon, nitrogen, oxygen, fluorine, chlorine, and sulfur, as well as fragments from Enamine’s REadily AccessibLe (REAL) space.

Next, they screened the library utilizing machine-learning models that Collins’ lab previously trained to forecast antibacterial activity against N. gonorrhoeae. This process resulted in nearly 4 million fragments, which they refined by eliminating any fragments predicted to be cytotoxic to human cells, displayed chemical liabilities, or were similar to existing antibiotics. This refinement left around 1 million candidates.

“Our aim was to eliminate anything resembling an established antibiotic to address the antimicrobial resistance crisis fundamentally. By exploring underrepresented areas of chemical space, we sought to uncover novel mechanisms of action,” explains Krishnan.

Through several rounds of experimental and computational analysis, the researchers identified a fragment referred to as F1, which showed promising activity against N. gonorrhoeae. They utilized this fragment as a basis for generating additional compounds via two different generative AI algorithms.

One of the algorithms, known as chemically reasonable mutations (CReM), begins with a particular molecule containing F1 and subsequently generates new molecules by adding, replacing, or removing atoms and chemical groups. The second algorithm, F-VAE (fragment-based variational autoencoder), takes a chemical fragment and constructs it into a complete molecule by learning patterns of fragment modifications based on its training with over 1 million molecules from the ChEMBL database.

These two algorithms produced around 7 million candidates containing F1, which were then screened computationally for activity against N. gonorrhoeae. This screening yielded approximately 1,000 compounds, from which 80 were selected for potential synthesis by chemical vendors. Of these, only two could be synthesized, with one, named NG1, proving to be highly effective at eliminating N. gonorrhoeae in both lab dishes and mouse models of drug-resistant gonorrhea infection.

Further experiments demonstrated that NG1 interacts with a protein known as LptA, a novel drug target involved in the synthesis of the bacterial outer membrane. The drug appears to disrupt membrane synthesis, which is lethal to cells.

Unconstrained design

In a subsequent round of investigations, the researchers delved into the potential of generative AI for unrestricted molecule design, using the Gram-positive bacterium S. aureus as their target.

Once again, the researchers employed CReM and VAE to develop molecules, albeit this time without any constraints except for the general principles of atom bonding that form chemically feasible molecules. Collectively, the models generated over 29 million compounds, which were filtered using the same criteria applied to the N. gonorrhoeae candidates but focusing on S. aureus, eventually narrowing the pool to about 90 compounds.

They successfully synthesized and tested 22 of these molecules, with six demonstrating significant antibacterial activity against multi-drug-resistant S. aureus in laboratory dishes. The top candidate, named DN1, was capable of clearing a methicillin-resistant S. aureus (MRSA) skin infection in a mouse model. These molecules also seem to affect bacterial cell membranes, but with broader implications beyond interacting with a single protein.

Phare Bio, a nonprofit organization involved in the Antibiotics-AI Project, is currently working on further modifications of NG1 and DN1 to prepare them for additional testing.

“In collaboration with Phare Bio, we are exploring analogs and advancing the best candidates through preclinical testing, aided by medicinal chemistry work,” Collins remarks. “We are also keen on applying the platforms developed by Aarti and the team to other bacterial pathogens of concern, particularly Mycobacterium tuberculosis and Pseudomonas aeruginosa.”

The research received funding from various sources, including the U.S. Defense Threat Reduction Agency, the National Institutes of Health, the Audacious Project, Flu Lab, the Sea Grape Foundation, Rosamund Zander and Hansjorg Wyss for the Wyss Foundation, and an anonymous donor.