MIT Team Uses AI to Discover 9 Antibiotic Candidates From 700B Molecules

MIT researchers Regina Barzilay and James J. Collins in Cambridge, Massachusetts, have received the 2026 Pritzker Prize for AI Research in Science for their work with the Antibiotics-AI Project, which has used artificial intelligence to screen more than 70 billion molecules and identify or design nine novel antibiotic candidates. The prize, established by the University of Chicago in 2024, recognizes AI-driven advances in science and engineering and, for the first time, was awarded to a collaborative team. The project addresses one of the most serious challenges in modern medicine: antimicrobial resistance. As bacteria evolve, existing antibiotics lose effectiveness, while new antibiotic development has lagged because of scientific difficulty, limited commercial incentives, and lengthy validation requirements. Barzilay, an MIT professor focused on machine learning in health, and Collins, a pioneer in synthetic biology and biomedical engineering, have worked to accelerate the earliest and most resource-intensive stage of discovery: finding molecules that can kill resistant pathogens. Their initial approach used deep learning to screen existing chemical libraries. In 2020, the team reported in Cell that AI had identified Halicin, a compound originally unrelated to antibiotics but active against multiple bacteria, including in mouse infection models. In 2023, the group reported Abaucin, a candidate targeting Acinetobacter baumannii, a major cause of hospital infections. By 2024, the team had also developed interpretable AI models that connected molecular structures to antibacterial activity, including candidates active against methicillin-resistant Staphylococcus aureus, or MRSA. The field advanced further in a 2025 study in Cell, where the researchers moved from filtering known compounds to using generative AI to design new antibiotic structures. In one method, they began with antibacterial chemical fragments for Neisseria gonorrhoeae and used a generative model to build about 7 million candidate molecules. After computational filtering, they synthesized 80 compounds and obtained two successes, including NG1, which showed activity against resistant gonorrhea bacteria and acted on the LptA protein involved in outer membrane formation. In a second approach, the model generated more than 29 million molecules without being restricted to predefined chemical fragments. The team selected compounds based on predicted activity, toxicity, and synthesizability, ultimately testing 22 molecules. Six showed strong activity against multidrug-resistant Staphylococcus aureus, and one, DN1, worked in a mouse skin infection model by disrupting bacterial cell membranes. Across both strategies, the models produced more than 36 million candidate molecules, but only 24 were successfully synthesized and tested, and seven showed selective antibacterial activity. The team also reported in Science Translational Medicine that specialized graph neural networks outperformed large language models when screening 6 million compounds for activity against resistant gonorrhea. The models identified 83 active molecules among 213 tested, with about two showing promising results in animal and organ-chip models, including a candidate linked to disruption of bacterial cell wall synthesis. The Antibiotics-AI Project has screened tens of thousands of molecules in antimicrobial and human-cell assays and now focuses on seven priority pathogens: Acinetobacter baumannii, Pseudomonas aeruginosa, Klebsiella pneumoniae, Escherichia coli, Staphylococcus aureus, Neisseria gonorrhoeae, and Mycobacterium tuberculosis. It has also partnered with Phare Bio to help move candidates toward practical drug development. The work demonstrates that AI can greatly expand the searchable chemical space and speed early discovery. However, successful drug development still requires reliable data, accurate model evaluation, safe synthesis, toxicity testing, pharmacokinetic analysis, and clinical trials. For antibiotics, where resistance develops quickly and commercial returns are limited, AI can reduce search time, but turning predicted molecules into approved medicines remains a long and resource-intensive process.