The Brief
Science & Discovery 4 min read

From Years to Weeks: What AI-Discovered Antibiotics Reveal

NAVION

Share

Two MIT professors have won the 2026 Pritzker Prize for AI in Science Research Excellence. Regina Barzilay and James Collins were recognized for work that sits at the intersection of machine learning and one of medicine’s most urgent problems: the global antimicrobial resistance crisis. Their research did not just produce new drugs. It demonstrated that the process of finding them could be fundamentally accelerated.

The Discovery That Started With 2,500 Molecules

The collaboration between Barzilay and Collins produced its first major result in 2020, in a paper published in Cell with then-postdoc Jonathan Stokes as first author. The subject was halicin, the first antibiotic identified through artificial intelligence. The name is a nod to HAL 9000, the fictional AI from Stanley Kubrick’s 1968 film 2001: A Space Odyssey, a choice that signals something about how the researchers understood the moment they were in.

Halicin was not designed from scratch. It was originally studied as a potential diabetes drug. A deep learning model, trained on 2,500 molecules, identified it as a candidate worth testing. When the team ran those tests, halicin proved effective against several antibiotic-resistant bacteria, including Clostridium difficile, Acinetobacter baumannii, and Mycobacterium tuberculosis. These are not obscure pathogens. They are among the organisms that make drug-resistant infections so dangerous in clinical settings.

The speed of the discovery is what most coverage underplays. A process that would traditionally have taken years was completed in a matter of weeks. That compression is not a footnote. It is the core of what makes this approach significant.

Narrow Beats Broad: The Logic Behind Abaucin

The second major result came in 2023. Working again with Stokes, now an assistant professor at McMaster University, Barzilay and Collins used machine learning to identify abaucin. Unlike halicin, abaucin is a narrow-spectrum antibiotic: it targets a single species, A. baumannii, rather than a broad range of bacteria.

This specificity is not a limitation. It is a design advantage. Broad-spectrum antibiotics, while powerful, create conditions that encourage resistance to develop and can disrupt beneficial bacteria in the gut. A narrow-spectrum drug that works precisely against one resistant organism reduces both of those risks. A. baumannii is a particularly relevant target: it is commonly found on hospital surfaces and has been documented in the wounds of soldiers stationed in conflict zones.

The shift from halicin to abaucin also illustrates something about how this research program is evolving. The first discovery proved the concept. The second began to refine it, moving toward more targeted interventions rather than simply replicating what broad-spectrum antibiotics already do.

What This Means Beyond the Lab

The Pritzker Prize, established in 2024 by the University of Chicago in partnership with Caltech, was created specifically to recognize work that advances both AI and the natural sciences or engineering together. That framing matters. It reflects a growing recognition that the most consequential applications of AI are not happening inside the technology sector alone. They are happening in biology, chemistry, and medicine, fields where the problems are harder and the stakes are higher.

Collins has described the historical model of antibiotic discovery as slow, costly, and unpredictable, a trial-and-error process that yields few viable candidates. The shift he points to is from AI as a discovery tool, scanning existing molecules for hidden potential, toward AI as a design tool, capable of generating new candidates rather than just evaluating known ones. That distinction matters for what comes next.

Barzilay offers a more measured framing. In her view, the field is still in early stages. AI can automate parts of the drug discovery pipeline, but not the whole of it. Building realistic AI models of human biology remains an open and difficult problem. This is not pessimism. It is precision. The research has demonstrated genuine capability while being honest about where the gaps remain.

Here is what that combination of ambition and caution actually signals: the work is credible precisely because it does not overclaim. The results are real and documented. The limitations are acknowledged. That combination is rarer than it should be in discussions of AI and medicine.

In Short

Barzilay and Collins used deep learning to identify two antibiotics, halicin in 2020 and abaucin in 2023, that conventional methods would have taken years to find. The speed, the specificity, and the focus on drug-resistant pathogens make this work directly relevant to one of global medicine’s most serious challenges. The 2026 Pritzker Prize recognizes not just the results but the approach: using AI to augment scientific discovery in a domain where human effort alone has not been enough. The researchers themselves are clear that this is a beginning, not a solution. Building the AI models that could genuinely transform drug discovery is still the work ahead.

Based on reporting from MIT News.

Read the original study.

Written by

NAVION