AI-Designed Antibiotics Show Promise Against Drug-Resistant Bacteria
The findings demonstrates how AI models can generate millions of theoretical molecular structures.
Complete the form below to unlock access to ALL audio articles.
Screening millions of hypothetical molecules
Researchers at the Massachusetts Institute of Technology have developed novel antibiotics using generative artificial intelligence (AI), targeting two drug-resistant bacterial pathogens: Neisseria gonorrhoeae and methicillin-resistant Staphylococcus aureus (MRSA). The study, published in Cell, demonstrates how AI models can generate millions of theoretical molecular structures that are computationally screened for antibacterial activity.
To identify potential treatments for N. gonorrhoeae, the team first curated a set of around 45 million known chemical fragments, combining them with compounds from Enamine’s REadily AccessibLe (REAL) space. Machine-learning models were then applied to narrow this pool to 4 million fragments with predicted activity. After removing cytotoxic, chemically unstable or structurally redundant candidates, around 1 million novel fragments remained.
From these, the researchers identified a promising fragment named F1. Two separate generative models – Chemically Reasonable Mutations (CReM) and a fragment-based variational autoencoder (F-VAE) – were then used to produce approximately 7 million F1-containing candidates. After a final screening phase, 1,000 molecules were shortlisted, and 80 were selected for potential synthesis. Of these, two were successfully synthesized, and one candidate, NG1, showed strong activity against drug-resistant N. gonorrhoeae in both cell cultures and a mouse model.
Membrane disruption as a mechanism
Initial studies suggest that NG1 targets LptA, a protein involved in bacterial outer membrane synthesis. By disrupting this membrane, the compound exerts its bactericidal effects. LptA has not previously been targeted by existing antibiotics, indicating that NG1 may act through a novel mechanism.
Expanding into unconstrained compound generation
The team also applied generative AI models to design antibiotics without fragment constraints, aiming to target S. aureus, a Gram-positive bacterium. CReM and F-VAE models were again used, producing over 29 million unique compounds. After applying similar filters, around 90 compounds were shortlisted for synthesis, and 22 were successfully produced. Six of these showed strong antibacterial activity in laboratory tests.
Among them, one compound, DN1, cleared MRSA infections in a mouse skin infection model. DN1 also appears to disrupt bacterial membranes but with less specificity, indicating a broader mechanism of action.
Next steps in development
The newly formed compounds, NG1 and DN1, are being developed further by Phare Bio, a nonprofit affiliated with the Antibiotics-AI Project. Efforts are now focused on modifying these molecules to improve their pharmacological properties and expanding the AI platform to design candidates for other pathogens, including Mycobacterium tuberculosis and Pseudomonas aeruginosa.
Reference: Krishnan A, Anahtar MN, Valeri JA, et al. A generative deep learning approach to de novo antibiotic design. Cell. 2025. doi: 10.1016/j.cell.2025.07.033
This content includes text that has been generated with the assistance of AI. Technology Networks' AI policy can be found here.
