Generative AI Designs Novel Antibiotics Against MRSA and Drug-Resistant Gonorrhea
Researchers at MIT and the Broad Institute have reported generative artificial intelligence systems capable of designing entirely new antibiotics, including a compound active against methicillin-resistant Staphylococcus aureus (MRSA). The study, published August 14 in Cell, describes two lead molecules—NG1 and DN1—produced by deep learning models trained on a dataset of ~40,000 small molecules with known antibacterial profiles.
Model Design
Two separate generative platforms were used:
- Fragment-based model (NG1): Trained on compounds effective against Neisseria gonorrhoeae. The model decomposed molecules into fragments, scored antibacterial activity, and iteratively expanded a particularly active fragment (F1) by adding bonds and substituents, predicting antibacterial potential and toxicity at each step.
- Atom-seeded model (DN1): Started from a single atom rather than a fragment. The model incrementally added chemical groups, evaluating each intermediate for antibacterial activity, predicted toxicity, drug-likeness, and synthetic feasibility.
Both models employed predictive filters developed at MIT to estimate whether proposed molecules could be synthesized. The code for the generative models has been released openly.
From several hundred computational candidates, Ukrainian contract manufacturer Enamine synthesized 24 molecules. Laboratory testing found seven with antibacterial activity, with NG1 demonstrating activity against multidrug-resistant gonorrhea and DN1 being active against both multidrug-resistant gonorrhea and MRSA.
James Collins, Ph.D., Broad Institute, who co-led the project, has previously worked on AI-guided antibiotic discovery, including the 2020 identification of halicin. Unlike prior work, which focused on screening libraries for hidden antibacterial activity, the present study demonstrates de novo antibiotic design through generative methods.
The two lead molecules are being advanced toward preclinical development through Phare Bio, a nonprofit co-founded by Collins in 2020. Phare Bio partnered with the Collins Lab at MIT to develop what it calls the first generative AI platform for antibiotic discovery, named AIBiotics. The nonprofit is supported by TED’s Audacious Project, ARPA-H, and grant funding from Google.org through its Generative AI Accelerator. Its model is structured around donor-funded preclinical development and later-stage partnerships or spin-outs to bring candidates through clinical testing.
Antimicrobial resistance (AMR) is projected to cause up to 10 million deaths annually by 2050, with the CDC estimating a resistant infection occurs every 11 seconds in the U.S. and a death every 15 minutes. While many drugmakers have exited antibiotic development due to weak market incentives, Phare Bio is targeting high-priority pathogens including Acinetobacter baumannii, Pseudomonas aeruginosa, and Klebsiella pneumoniae. All three are listed by the World Health Organization as critical-priority pathogens for new antibiotic development.
Image credit: Deaths attributable to AMR by pathogen, global, 1990–2021, for people 5 years and older and those younger than 5 years
Phare’s platform applies generative AI to explore chemical space more extensively and at lower computational cost than traditional methods, as claimed by the company. Compounds that meet initial computational criteria are tested for pharmacokinetics, toxicity, and efficacy, then advanced into preclinical development. Candidates that progress further are taken into clinical testing through partnerships with larger pharmaceutical companies or spin-out ventures.
Topics: AI & Digital
