McMaster’s generative AI model accelerates antibiotic discovery with novel water-soluble candidate

Researchers have developed a generative artificial intelligence model that has designed a clinically promising, water-soluble antibiotic candidate and demonstrated efficacy against drug-resistant infections in preclinical testing

Researchers at McMaster University, Hamilton, Ontario, Canada, have developed a generative artificial intelligence (AI) model that has demonstrated the capacity to accelerate drug discovery and has already produced a novel antibiotic candidate in early-stage testing. The work addresses a long-standing bottleneck in antimicrobial research where the pace of discovery has failed to keep up with the rapid emergence of drug-resistant pathogens.

The model – called SyntheMol-RL – has been designed to navigate an immense chemical landscape comprising up to 46 billion potential compounds, a scale that far exceeds conventional laboratory screening approaches, which typically can assess no more than around one million molecules.

By integrating approximately 150,000 molecular building blocks with a defined set of 50 chemical synthesis reactions, the system has enabled the generation of structurally diverse antibiotic candidates with a high degree of novelty.

“In the lab, we can build chemical compounds using a set of smaller chemical fragments, which can be stuck together like molecular Lego blocks,” said Dr. Jon Stokes, assistant professor and was lead investigator on the study.

“SyntheMol-RL configures those fragments in different ways, faster than humans ever could, to create new, larger chemical compounds that should – based on its knowledge – be antibacterial,” he said.

Stokes, who is affiliated with the Michael G. DeGroote Institute for Infectious Disease Research, also located in Hamilton, noted that although generative AI has become increasingly adept at proposing candidate molecules, the translation of those candidates into viable medicines has remained constrained by complex pharmacological requirements. He emphasised that properties such as solubility, toxicity and metabolic clearance are essential determinants of clinical success and cannot be inferred solely from antibacterial activity.

“It doesn’t matter if you find a chemical that’s antibacterial in the lab if it can’t dissolve inside the body, if it’s toxic to human cells, or if it can’t be metabolised and expelled after it has done its job,” he said.

“Bleach is antibacterial – [and] so is fire. But they obviously don’t tick those other boxes. Good drug candidates must meet several different criteria, otherwise they’ll never become actual medicine,” he added.

Earlier iterations of SyntheMol had focused primarily on antibacterial efficacy, without systematic incorporation of these additional constraints. Over the past two years, however, Stokes’ team, in collaboration with researchers at Stanford University, California, USA, has refined the system to embed key developability criteria directly within the generative process. This refinement has enabled the model to prioritise compounds that are not only antibacterial but also amenable to synthesis and likely to exhibit favourable solubility profiles.

“There is a lot of conflict between compounds that are antibacterial and compounds that are water soluble,” said Gary Liu, a doctoral candidate in Stokes’ laboratory and lead developer of the model.

“In previous studies, filtering for compounds that were both antibacterial and soluble after our prompt often left us with significantly fewer viable drug candidates, so we built solubility right into the generation process and now the model can efficiently design antibiotic candidates with greater clinical promise,” he said.

In this work, the research team tasked the model with generating water-soluble antibiotic candidates capable of treating infections caused by Staphylococcus aureus, a common and sometimes drug-resistant pathogen responsible for a range of clinical conditions.

From a set of 79 candidate molecules proposed by the model, the researchers identified a particularly promising compound, which they named ‘synthecin’. This molecule exhibited both predicted antibacterial activity and favourable solubility characteristics. Subsequent experimental validation involved formulation as a topical cream and evaluation in a mouse model of drug-resistant wound infection.

“Synthecin was highly effective at controlling the infection,” said Denise Catacutan, a doctoral candidate who led the experimental component of the study.

“It worked extremely well as a topical drug and also shows early promise as something that could be applied or optimized for systemic use in the future,” she said.

Despite these encouraging results, the precise mechanism by which synthecin exerts its antibacterial effect has yet to be established. Stokes indicated that elucidation of this mechanism represents a critical next step, as it will inform both safety assessment and the likelihood of successful progression towards clinical application. His laboratory has initiated detailed mechanism-of-action studies to address this question.

The researchers have framed the identification of synthecin as a proof of principle for the broader capability of their AI framework. Rather than relying on extensive screening to identify rare viable compounds, the model has demonstrated the capacity to generate high-quality candidates at the outset, thereby shifting the focus of drug discovery towards rational design and optimisation.

“We used our model to design antibiotics, but it’s capable of so much more,” said Stokes.

“We built it to be disease agnostic, meaning it could just as easily generate novel drug candidates for diabetes or cancer or other indications,” he concluded.

Stokes, who is also a faculty member at the Marnix E. Heersink School of Biomedical Innovation and Entrepreneurship, at the University of Alabama at Birmingham, and an executive member of NexusHealth, has indicated that further refinements to the model are underway. His team has anticipated the release of an enhanced version later this year, with the aim to extend its applicability across multiple therapeutic domains and to improve predictive accuracy for clinically relevant properties.

Taken together, the study has illustrated how generative AI can begin to address fundamental inefficiencies in drug discovery. If such approaches continue to mature, they may help to restore momentum in antibiotic development, an area that has struggled to keep pace with the evolving threat of antimicrobial resistance.

For further reading please visit: 10.1038/s44320-026-00206-9