How AI designed two antibiotic leads from millions of possible molecules

An AI-generated molecule becomes scientifically useful when someone can make it and test what it does. In research reported in August 2025, an MIT-led team turned a search across millions of possible structures into two antibiotic leads, NG1 and DN1. Both reached experiments in infected mice, giving researchers physical compounds to investigate further.[1]

The earlier AI-discovered candidate abaucin illustrates a different starting point. In that 2023 project, researchers trained a model on the effects of thousands of compounds and used it to screen an existing chemical collection for activity against Acinetobacter baumannii. The newer work asks AI to design structures beyond the molecules already sitting in such collections.[5]

Two routes to a new chemical structure

One route kept a promising chemical fragment, called F1, inside the generated molecules. The other allowed the system to design molecules without a required antibacterial fragment. Both used variations of a mutation-based generator and a variational autoencoder, which learns a representation of molecular structures and generates new candidates from it.[1]

The project’s published software documentation shows how generation was paired with scoring. Separate models estimated antibacterial activity and toxicity to human cell lines. Further filters checked problematic chemical groups, similarity to 559 known antibiotics and predicted ease of synthesis. A molecule could therefore receive a strong antibacterial score and still be removed for another property.[3]

Synthesis was an experiment in its own right

The paper gives an instructive intermediate number: vendors attempted synthesis of 27 of the 80 shortlisted fragment-based designs, and obtained two. Both successful molecules came from the variational autoencoder. A proposed structure and an accessible route to making it are different outputs; computational predictions of synthesizability did not remove that gap.[2]

The researchers’ code also exposes differences between the generators. Its unconstrained mutation route starts from very small molecules such as water, ammonia or methane. The variational autoencoder samples a learned molecular space. These procedures create different candidate pools before the antibacterial models score them. Researchers can inspect the filtering notebooks, generation settings and released datasets to see how those choices shaped the search.[3]

The funnel is useful because it locates the scarce work. Expanding a virtual collection makes more designs available, while synthesis and biological experiments determine which designs deserve another round of effort. Improving the generator could mean producing more chemical variety, increasing the proportion that can be made, or producing molecules with better measured activity. A larger generation count alone does not distinguish those improvements.

What the mouse experiments showed

NG1 reduced bacterial counts in a mouse model of gonorrhoea following local treatment. The experiment included nine NG1-treated mice, eight vehicle controls and five receiving ceftriaxone. Experiments implicated LptA, a protein involved in the bacterium’s outer-membrane machinery, as its main target.[2]

DN1 was tested against methicillin-resistant Staphylococcus aureus, or MRSA, in a skin-infection experiment using mice with reduced infection-fighting white blood cells. Six treated mice and six vehicle controls were compared. At 25 hours after infection, the mean bacterial load was roughly tenfold lower with topical DN1 than with vehicle.[2]

Designing the next compounds

In a February 2026 account of the programme, senior researcher James Collins described the next emphasis as combining deep learning with high-throughput biological testing to obtain molecules with stronger drug-like properties. He also described collaboration on organ-chip systems, which allow drug behaviour to be studied in tissue-like environments alongside conventional animal work.[4]

Phare Bio, the nonprofit involved in advancing the programme, is intended to connect academic discovery with further development. Collins said an ARPA-H-funded project aims to design 15 antibiotics and develop them as preclinical candidates. That is a development objective; it is distinct from a count of approved treatments.[4]

NG1 and DN1 give that work a concrete starting point: identifiable molecules, experimental results and structures that chemists can modify. The next useful advance will be a better compound with an improved combination of potency, exposure, tolerability and manufacturability. AI can propose changes quickly; the value of those changes becomes visible in the next set of measurements.

Report an error or suggest a correction