AI designs antibiotics, solves math
- Cell published a generative antibiotic-design study after researchers used deep learning to create two lead compounds active against drug-resistant gonorrhea and MRSA in mice. - The paper’s lead molecules, NG1 and DN1, were designed de novo, showed distinct mechanisms, and worked in vivo against multidrug-resistant Neisseria gonorrhoeae and Staphylococcus aureus. - Separately, Google DeepMind’s Aletheia reported autonomous solutions to four open questions and led on First Proof, a new research-math test. (arxiv.org)
Antibiotics are drugs that kill bacteria, but researchers first have to find molecules that hit microbes harder than human cells. A Cell paper reported an artificial-intelligence system that designed two new lead antibiotics for drug-resistant gonorrhea and methicillin-resistant Staphylococcus aureus, or MRSA. (cell.com) (nature.com) The team’s platform combined fragment screening with generative design, which means the model proposed new molecules instead of only ranking old ones from a library. The resulting leads, called NG1 and DN1, showed selective antibacterial activity and distinct mechanisms of action. (cell.com) (nature.com) NG1 worked against multidrug-resistant Neisseria gonorrhoeae in vivo, and DN1 lowered bacterial burden in a mouse model of drug-resistant skin infection caused by Staphylococcus aureus. Nature Biotechnology said the study addressed a central problem in antibiotic discovery: finding compounds that are both novel and testable. (cell.com) (nature.com 1) (nature.com 2) A separate March 2026 paper pushed the same idea into peptide antibiotics, which are short protein-like chains that can punch holes in bacterial membranes. In that study, a system called CAMPER designed a 12-amino-acid peptide, WP-CAMPER1, against persistent MRSA cells that survive standard treatment. (nature.com) The peptide killed S. aureus MW2 at a minimum inhibitory concentration of 4 micrograms per milliliter. A 2% topical formulation cut bacterial burden by 2.5 log10 in a mouse skin model, and a related version reduced established biofilm infection by 1.37 log10. (nature.com) Math research asks for proofs, which are step-by-step arguments that other mathematicians can check line by line. In February 2026, the First Proof project released 10 unpublished research-level problems, with official solutions withheld until the deadline so AI systems could not train on them in advance. (arxiv.org) (1stproof.org) OpenAI said on February 20 that its internal model produced proof attempts for all 10 First Proof problems and that at least five attempts had a high chance of being correct after expert feedback. The company later said its earlier confidence in problem 2 had dropped after official commentary and community review. (openai.com) Google DeepMind’s March 9 paper described Aletheia, a research agent built on Gemini Deep Think that generates, checks, and revises proofs in natural language. The paper said Aletheia produced multiple publication-grade papers, including one with no human intervention, and autonomously solved four open questions from a 700-problem evaluation set. (arxiv.org) The same paper said Aletheia led on First Proof, a benchmark built to test research mathematics rather than contest tricks. That puts the antibiotic work and the math work in the same lane: models are moving from sorting known options to proposing new molecules and new arguments that experts can test. (arxiv.org) (cell.com)
