MIT Researchers Develop New Antibiotics with AI Assistance
Researchers at MIT have made strides in creating innovative antibiotics using artificial intelligence, targeting two particularly challenging infections: drug-resistant Neisseria gonorrhoeae and multi-drug-resistant Staphylococcus aureus (MRSA).
The team utilized generative AI algorithms to design over 36 million potential compounds, which they then screened for their antimicrobial properties. Notably, their top discoveries are structurally unique compared to existing antibiotics and seem to disrupt bacterial cell membranes through new mechanisms.
This strategy permitted researchers to analyze theoretical compounds that had never been encountered before—a method they aim to apply to discover and create treatments for other bacterial strains.
James Collins, a leading figure in the study and the Termeer Professor of Medical Engineering and Science, expressed enthusiasm about the new antibiotic potentials this project may unveil. “Our work illustrates AI’s power in drug design and allows us to explore vastly larger chemical spaces than ever before,” he noted.
The findings were published today in Cell, with MIT postdoc Aarti Krishnan, former postdoc Melis Anahtar, and Jacqueline Valeri contributing as lead authors.
The Challenge of Antibiotic Resistance
In the past 45 years, the FDA has only approved a handful of new antibiotics, mostly variations of those already available. Meanwhile, resistance to these drugs has become a significant global problem, contributing to an estimated 5 million deaths annually from drug-resistant infections.
In hopes of discovering new antibiotics, Collins and his MIT colleagues are tapping into AI to sift through extensive libraries of existing chemical compounds. Their efforts have already uncovered notable candidates like halicin and abaucin.
Building on previous successes, the researchers decided to cast a wider net by exploring unexplored molecular possibilities. By generating hypothetical molecules that are not present in established chemical libraries, they believe they can discover a broader array of potential drugs.
In this study, they employed two strategies: one where generative AI designed molecules based on a specific fragment known for antimicrobial activity, and another where the AI could freely create molecules without the need for a designated fragment.
The fragment-based method focused on identifying compounds effective against N. gonorrhoeae. The team compiled a library of approximately 45 million known chemical fragments from various atom combinations and then screened them using predictive machine-learning models. They filtered this initial pool of nearly 4 million fragments down to about 1 million, excluding anything resembling existing antibiotics.
“We aimed to avoid existing antibiotics to tackle antimicrobial resistance differently,” Krishnan expressed. “By exploring less-charted chemical spaces, we hoped to uncover innovative mechanisms of action.”
After further experiments and computational work, researchers pinpointed a fragment, referred to as F1, which displayed promising activity against N. gonorrhoeae. This fragment served as a foundation for generating more compounds using two distinct AI algorithms.
One algorithm, called chemically reasonable mutations (CReM), modifies a specific molecule with F1 by adding, replacing, or removing atoms. The other, F-VAE (fragment-based variational autoencoder), constructs complete molecules from a fragment by learning how fragments are typically altered from prior knowledge of over 1 million molecules.
This process generated about 7 million candidates containing F1, which were then screened for their effectiveness against N. gonorrhoeae. The screening narrowed the selection to roughly 1,000 compounds, with 80 chosen for potential synthesis. Ultimately, only two were successfully produced, and one, NG1, showed significant effectiveness against N. gonorrhoeae in laboratory and mouse models of drug-resistant infections.
Additional studies indicated that NG1 interacts with a protein named LptA, a novel target involved in bacterial outer membrane synthesis. It appears this drug works by disrupting membrane synthesis, which proves lethal for the bacteria.
Open-Ended Design Exploration
In another phase of the research, the team employed generative AI to explore freely without predefined constraints, focusing on S. aureus as the target.
Again utilizing CReM and VAE, the researchers generated over 29 million compounds, filtering them similarly to those targeted at N. gonorrhoeae, which eventually reduced the selection to about 90 compounds.
They managed to synthesize and evaluate 22 of these molecules, with six demonstrating strong antibacterial effects against multi-drug-resistant S. aureus. Notably, the leading candidate, DN1, successfully cleared a methicillin-resistant S. aureus (MRSA) skin infection in mouse models. These molecules also seem to affect bacterial cell membranes but through a broader range of mechanisms, rather than just one specific target.
Phare Bio, another partner in the Antibiotics-AI Project, is currently working on further modifications of NG1 and DN1 to prepare them for additional testing.
Collins remarked, “In collaboration with Phare Bio, we are developing analogs and advancing our best candidates through medicinal chemistry.” He also expressed excitement about applying their platforms to other significant bacterial threats, like Mycobacterium tuberculosis and Pseudomonas aeruginosa.
This research received funding from various sources, including the U.S. Defense Threat Reduction Agency, the National Institutes of Health, the Audacious Project, and others.
