Publicly Accessible AI Tool Could Speed Antibiotic Discovery

A tool that uses artificial intelligence (AI) to screen for molecules with potential therapeutic value against drug-resistant bacteria has been released for free to accelerate new antibiotic discovery globally.

Called the ESKAPE Model, the tool was built specifically to identify candidate antibiotics targeting the ESKAPE pathogens — Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter spp. These pathogens are responsible for a large proportion of the resistant infections identified clinically and have been identified by the World Health Organization as priority pathogens for the development of new treatments.

Rather than waiting to publish a paper detailing the development of the tool, Jonathan Stokes, PhD, assistant professor of biochemistry and biomedical sciences at McMaster University in Hamilton, Ontario, Canada, decided to release it in its current state on January 27.

“Our paper will be ready to submit in about a month, and then it will take another 8 months or more to undergo peer review and do revisions,” Stokes told Medscape Medical News. “That’s just too long. The tool is built. I know it’s not perfect, but it’s good enough to get into users’ hands for feedback. So, I decided to just get it out there so people can use it.

“The economic model to support antibiotic development is fundamentally broken,” he added. Large pharmaceutical companies are no longer interested in antibiotics because “there’s no clear path to get a return on investment,” he said. “But that doesn’t mean everybody has exited this space. Many talented people working in academia and in small biotech companies are doing everything they can to ensure that we have antibiotics in development. Our tool could accelerate their efforts.”

In fact, using an approach like that of the ESKAPE Model led to the team’s earlier discovery of abaucin, an antibiotic targeting A baumannii.

‘Easy to Use’

Half the people in Stokes’s lab are computer scientists, and half are wet-lab microbiologists and biochemists, he said. “Together, we had the idea to train AI models to help us predict antibiotics against all the ESKAPE pathogens. So, we screened our chemical libraries, which include small molecules for which we know the structure and the associated antibiotic activity, against all six to develop training sets.”

The team trained AI models for each ESKAPE pathogen on the basis of the structures of existing molecules that have antibacterial activity against each pathogen, as well as those that don’t. Now, when shown structures of unfamiliar chemicals, the models can make predictions about the molecules’ antibacterial activity and about whether they might make good candidates for further study as potential antibiotics.

Investigators from anywhere in the world can upload their own molecules to the ESKAPE Model website, and the models will run predictions on their chemicals. “This [tool] will help other people make decisions about what compounds they should be pursuing and which ones they should perhaps ignore,” Stokes said.

The ESKAPE model was designed to be easy to use, even by people who have no AI experience, he noted. It simply asks users to input “SMILES codes,” which are strings of alphanumeric characters that represent the structures of various chemical compounds. After copying and pasting the code into the appropriate space on the website, users quickly receive AI-guided predictions about whether their chemicals have therapeutic potential against any of the ESKAPE pathogens.

Users can enter as many as 100 molecules in each session. Predictions on one molecule take ~2 minutes. Predictions on 100 molecules take ~3.5 minutes.

Overall, the researchers say that ESKAPE Model users can currently screen upward of 20,000 chemicals in an average workday, whereas the same output using a traditional wet-lab approach would take several weeks, cost thousands of dollars, and may not yield any positive results.

The team is updating the user interface based on feedback from current users, Stokes noted. “Some folks wanted to be able to draw the chemical structures instead of inputting the SMILES codes, and so we’re currently coding something to enable them to do that.

“Over time,” he added, “as we gather more training data, train additional models, and get more user feedback, we’re going to be continuously refining and updating all of these models so the tool becomes more and more robust.”

‘Arms Race Against Bacteria’

Amesh Adalja, MD, a spokesperson for the Infectious Diseases Society of America and senior scholar at Johns Hopkins Center for Health Security in Baltimore, commented on the ESKAPE Model and its release for Medscape Medical News.

“New tools are needed for antibiotic discovery because it’s harder and harder to find new antibiotics, and there is a great potential to use AI to speed that process up and to make it less costly,” he said. “We’re still in an arms race with bacteria, and the barriers to new antibiotic discovery are fairly high. So, anything that makes it easier, cheaper, and faster is welcome.”

The tool’s release will have “scale effects” because it will enable more people to screen virtually and efficiently than could be done by a single company or academic institution, he added. “For clinicians, this means that people are still working, striving, and using new technologies to develop new antibiotics. And the more people do that, the more tools we will eventually have.”

No funding for the development of the ESKAPE Model was disclosed, and Stokes and Adalja reported no relevant financial relationships.

Marilynn Larkin, MA, is an award-winning medical writer and editor whose work has appeared in numerous publications, including Medscape Medical News and its sister publication MDedge, The Lancet (where she was a contributing editor), and Reuters Health.