Researchers at McMaster University and Stanford have used an AI model called SyntheMol-RL to design a new antibiotic candidate, named synthecin, that controlled drug-resistant Staphylococcus aureus wound infections in mice. The model explores a chemical space of up to 46 billion possible compounds, compared with around a million molecules in a conventional lab screen.
What happened
The work is published in the journal Molecular Systems Biology, and McMaster announced it on 23 April. The team is led by Jon Stokes at McMaster.
The researchers describe the model as exploring “a vast chemical space of up to 46 billion possible compounds.” To build that space, it draws on about 150,000 molecular building blocks and 50 chemical reactions.
From that search came synthecin. In mice with wound infections caused by drug-resistant Staphylococcus aureus, often shortened to staph, the university says the compound “was highly effective at controlling the infection.”
Before anything else, the limits. Synthecin has been tested only in mice. It is not a drug for people, it has not been approved by any regulator, and it should not be read as a new antibiotic for patients. It is an early candidate that showed promise in an animal model.
How it works
The classic way to find a new antibiotic is to screen a library of existing molecules: test each one against bacteria and see which ones kill them. Conventional lab screens of this kind cover around a million molecules, according to McMaster. That sounds like a lot until you compare it with the number of molecules chemistry could, in principle, make.
SyntheMol-RL flips the process. Instead of choosing from a fixed shelf, it designs molecules by combining building blocks through known chemical reactions. A good analogy is a set of construction bricks. A conventional screen is like walking through a toy shop and picking from the models already assembled on the shelves. SyntheMol-RL is handed about 150,000 kinds of bricks and 50 known ways of joining them, and it works out which new models are worth building.
That design choice has a practical advantage. Because every candidate is assembled from real building blocks using real reactions, the molecules it suggests should be ones chemists have a recipe for. That matters, because a structure that looks promising on screen is of little use if nobody can make it in a lab. Tying the search to known reactions keeps it grounded.
The scale comes from combinations. A modest number of bricks and joining rules can produce an enormous number of possible structures, which is how the model reaches a space of up to 46 billion compounds.
By the numbers
| Item | Figure | Source |
|---|---|---|
| Chemical space explored | Up to 46 billion compounds | McMaster University |
| Molecular building blocks | About 150,000 | McMaster University |
| Chemical reactions used | 50 | McMaster University |
| Typical conventional lab screen | Around 1 million molecules | McMaster University |
| Test setting for synthecin | Wound infections in mice | McMaster via EurekAlert |
| Journal | Molecular Systems Biology | McMaster University |
Why it matters
Drug-resistant infections are rising faster than new antibiotics arrive. Generative AI that can search a far bigger space of molecules, and propose ones that can actually be made, could shorten the early search from years to weeks.
What caught my attention is the gap between the two numbers: around a million molecules in a lab screen against up to 46 billion in the model’s search space. Even if only a tiny share of those are worth testing, the model is looking in places a physical screen never reaches.
In my view, the mouse result is best read as a proof of concept for the method rather than for the molecule. It shows the pipeline can go from an AI design to a compound that works in a living animal. Whether synthecin itself ever becomes a medicine is a separate and much longer question, and animal results often do not carry over to humans.
What comes next
The paper is now public, with the DOI 10.1038/s44320-026-00206-9, so other researchers can examine the method and its results. Any path toward human use would require much more testing, and nothing announced so far goes beyond the mouse studies. For now, the clearest takeaway is about the tool: a way of searching chemical space that is larger and more practical than a shelf of existing molecules.
Sources
- McMaster University, news release on its AI model that speeds up drug discovery and designs a new antibiotic candidate, 23 April 2026, https://news.mcmaster.ca/mcmaster-built-ai-model-speeds-up-drug-discovery-designs-new-antibiotic/
- EurekAlert, McMaster University news release on SyntheMol-RL and synthecin, 23 April 2026, https://www.eurekalert.org/news-releases/1125168
- News-Medical, report on the AI model designing an antibiotic candidate against resistant bacteria, 23 April 2026, https://www.news-medical.net/news/20260423/AI-model-rapidly-designs-new-antibiotic-to-fight-resistant-bacteria.aspx