The first antibiotic discovered by artificial intelligence

Researchers at MIT trained a model on 2,500 molecules and asked it to search 100 million more; it returned halicin, an antibiotic that kills bacteria resistant to everything else.

A pipeline that had stopped

In a 2011 review in Clinical Microbiology Reviews, Silver wrote that the discovery of novel small-molecule antibacterial drugs "has been stalled for many years." She identified two reasons: the difficulty of choosing molecular targets that bacteria cannot quickly evolve around, and chemical libraries that lacked the kind of molecules able to enter bacterial cells, particularly Gram-negative ones, and stay there without being pumped out. The compounds that pharmaceutical screening had on its shelves were, in the main, the wrong shapes.

The cost of the stall was measured a decade later. In 2022 the Antimicrobial Resistance Collaborators, led by Christopher Murray, estimated in The Lancet that in 2019 there were 4.95 million deaths associated with bacterial antimicrobial resistance worldwide, including 1.27 million deaths directly attributable to it. Six pathogens accounted for 929,000 of the attributable deaths: Escherichia coli, Staphylococcus aureus, Klebsiella pneumoniae, Streptococcus pneumoniae, Acinetobacter baumannii and Pseudomonas aeruginosa.

Training on 2,335 molecules

Jonathan Stokes, Regina Barzilay, James Collins and colleagues at MIT published their approach in Cell on February 20, 2020. They began with a training set of 2,335 molecules, drawn from approved drugs and natural products, which they screened against E. coli in the laboratory. Of these, 120, or 5.14 percent, inhibited growth, defined as an optical density below 0.2 at the end of the assay. Each molecule was then converted into a graph of atoms and bonds, and a deep neural network using a directed message passing architecture was trained to predict, from structure alone, whether a molecule would inhibit E. coli. The model was augmented with computed molecular features and tuned by training on twenty different random splits of the data and combining the results.

A diabetes candidate that killed bacteria

The trained model was then pointed at the Drug Repurposing Hub, a library of 6,111 molecules at various stages of investigation for human diseases. After removing compounds that overlapped the training set, the 99 molecules with the highest predicted scores were tested in the laboratory. Fifty-one of them inhibited E. coli. Among the 63 lowest-ranked molecules tested as a control, only two did.

The team then filtered the hits for those in preclinical or clinical development, structurally distant from known antibiotics and predicted to have low toxicity. One compound met every criterion: SU3327, a nitrothiazole inhibitor of c-Jun N-terminal kinase under investigation as a treatment for diabetes. They renamed it halicin. Its nearest neighbour in the training set, by Tanimoto similarity, was an antiprotozoal drug at a score of about 0.37; the nearest antibiotic was metronidazole at about 0.21. It inhibited E. coli at a minimum inhibitory concentration of 2 micrograms per millilitre. According to MIT's announcement of the work, the name honours HAL, the computer in 2001: A Space Odyssey.

How it kills, and why resistance did not appear

Halicin killed E. coli in rich medium and continued to kill cells in nutrient-free buffer, where most antibiotics fail because the cells are not growing. To find the mechanism, the investigators first tried to breed resistant mutants and could not: no spontaneous suppressors emerged after 30 days of serial passaging in liquid medium, over which time E. coli readily evolved resistance to ciprofloxacin, or after seven days of continuous exposure on solid medium. RNA sequencing of treated cells showed a rapid response consistent with disruption of the proton motive force. Further experiments established that halicin selectively dissipates the pH component of the gradient across the bacterial membrane, a mechanism the authors described as unconventional. Collapsing that gradient kills cells whether or not they are dividing.

What it killed and what it did not

Halicin was rapidly bactericidal against Mycobacterium tuberculosis. Against panels of 36 multidrug-resistant clinical isolates each of carbapenem-resistant Enterobacteriaceae, A. baumannii and P. aeruginosa, three pathogen groups the World Health Organization has placed at the top of its priority list, it strongly inhibited the first two. It did not inhibit P. aeruginosa, which the authors attributed to insufficient permeability of that organism's membrane.

Two mouse models followed. In neutropenic Balb/c mice, the team created a 2 square centimetre wound on the back and infected it with about 250,000 colony-forming units of A. baumannii strain 288, an isolate from the Centers for Disease Control and Prevention that is not susceptible to the antibiotics generally used against it. One hour after infection the wound was treated with a cream containing either vehicle or 0.5 percent halicin, with repeat applications at 4, 8, 12, 20 and 24 hours. At 25 hours the vehicle-treated wounds carried about 100 million colony-forming units per gram of tissue. Five of the six halicin-treated mice carried fewer than 1,000 per gram, below the limit of detection, and the sixth about 100,000.

In mice infected with Clostridioides difficile strain 630, halicin at 15 milligrams per kilogram in four mice began to sterilize the gut at 72 hours and had cleared the infection in all four by 96 hours after treatment. Metronidazole at 50 milligrams per kilogram, in six mice, did not clear the infection faster than vehicle in seven controls.

107 million molecules in four days

The model was then run across a far larger space. From the ZINC15 database of about 1.5 billion molecules, the team selected the subsets whose physical and chemical properties resembled those of antibiotics, which came to 107,349,233 molecules. The in silico screen took four days, a scale two orders of magnitude beyond what laboratory screening permits. From the predictions, 23 molecules structurally dissimilar to known antibiotics were purchased and tested. Eight showed activity against at least one of the tested pathogens, and two of the eight were broad-spectrum.

What came after

The same approach has since produced two more candidates from the Collins and Stokes laboratories. In 2023 Liu and colleagues screened about 7,500 molecules against A. baumannii, trained a model on the results and discovered abaucin, a narrow-spectrum compound active against that organism alone, which acts on lipoprotein trafficking. In 2024 Wong and colleagues determined antibiotic activity and human cell toxicity for 39,312 compounds, trained ensembles of graph neural networks, screened 12,076,365 molecules and tested 283; among the hits was a structural class active against methicillin-resistant S. aureus and vancomycin-resistant enterococci that reduced bacterial counts in mouse infection models. That model, unlike the halicin model, was built to show which chemical substructures drove its predictions.

What the story does and does not show

Halicin was tested in bacteria in culture and in mice, in groups of four to seven animals, over periods of one to four days. It has not been given to a human, and the 2020 paper makes no claim about safety, dosing or efficacy in people. A compound that kills bacteria by collapsing a membrane gradient must be shown not to do the same to human mitochondria, and that work was not part of this study. The mouse wound model used immunosuppressed animals and a topical cream, which is a narrower test than a systemic infection treated by injection. Thirty days without resistance in a flask is a laboratory result, not a prediction about hospitals.

What the paper does show is a method. A model trained on 2,335 measured molecules ranked 6,111 others well enough that 51 of its top 99 picks inhibited bacteria, found a compound with no structural relationship to any antibiotic in use, and then searched 107 million molecules in four days to find eight more. The rate of discovery changed before any drug did.

No new class in decades. A machine found one in days.