ABOVE: Generative artificial intelligence can design new drugs including antibiotics, but these models often produce chemicals in silico that are difficult to synthesize in the laboratory. ©istock, Just_Super

Clinicians routinely administer antibiotics before surgeries to reduce the risk of infection.1 However, with the rise of antimicrobial resistance worldwide and the lack of new antibiotics, bacterial infections are becoming a growing challenge for the medical field. “We have a super lean antibiotic pipeline now that is being populated [mostly] by analogs of existing drug classes,” said Jon Stokes, a biochemist at McMaster University. Without the development of structurally and functionally novel antibacterial drugs, researchers predict that the mortality rate associated with antimicrobial-resistant bacteria will continue to rise reaching up to 10 million deaths annually by the year 2050.2  

In a recently published Nature Machine Intelligence paper, Stokes and his team enlisted the help of artificial intelligence (AI) to design structurally novel antibiotics that they could easily synthesize in the laboratory.3 This approach could help accelerate both antibiotic development and drug discovery.

Scientists are particularly concerned about six highly virulent and drug-resistant bacterial species, Enterococcus faecium, Staphylococcus aureus, Klebsiella pneumoniae, Acinetobacter baumannii, Pseudomonas aeruginosa, and Enterobacter species, known as the ESKAPE pathogens. One of these pathogens, A. baumannii, is a desiccation- and disinfectant-resistant microbe that is responsible for life-threatening, hospital-acquired infections of the skin, lungs, urinary tract, brain, bloodstream, and soft tissues.4 Because of the gravity of its threat to global health, the World Health Organization ranked this pathogen as a critical priority for new treatment and diagnostic tool development.5

To address this need, Stokes and his team decided to use AI to discover novel small molecules with antibacterial activity against A. baumannii. Traditionally used property prediction AI models forecast the characteristics of chemicals to allow researchers to find synthesizable compounds with the desired properties, but they screen only existing chemical libraries. On the other hand, new generative AI models allow scientists to produce novel chemical compounds with the required qualities, but this approach is not without its own pitfalls. “A lot of generative algorithms for de novo molecular design exist. The problem is they tend to build molecules, atom by atom, which means in a computer you can draw amazing, beautiful compounds, but you cannot bring them into the laboratory because you cannot make them. They are synthetically intractable. I always compare it to kids when they draw wild stuff, like a giraffe flying a spaceship. It is a really cool picture, but we cannot make that happen,” said Stokes.

Photos of Jon Stokes (left) and members of his team (right).
Jon Stokes and his team developed SyntheMol, a generative artificial intelligence model that they used to create novel antibiotics with predicted efficacy against the ESKAPE pathogen, Acinetobacter baumannii.
McMaster University

To overcome these challenges, the researchers developed a generative AI model called SyntheMol, which uses a property prediction model to assess the new chemical’s potential bioactivity. They first tested the ability of molecules from three chemical libraries to inhibit or not inhibit the growth of A. baumannii in culture. This binarized information allowed the researchers to train the property prediction model to identify important chemical features potentially contributing to their antibacterial activity. To ensure SyntheMol produces synthesizable compounds, Stokes and colleagues allowed the generative algorithm to use approximately 132,000 molecular building blocks and 13 established chemical reactions to build the molecules. Following in silico generation, the property prediction model evaluates each molecule, and that information informs future chemicals that the algorithm designs.

After filtering the small molecules generated to select structurally novel chemicals, Stokes and his team identified 58 potential compounds and worked with their colleagues at Enamine to synthesize them in the laboratory. They exposed A. baumannii cultures to these chemicals and found that six of them inhibited the bacterium’s growth. Additionally, these same six molecules also showed antibacterial activity against other ESKAPE species, suggesting that these compounds are good antibiotic candidates for future development.

“[This paper] is a great example of using generative AI in the field of antibiotics,” said César de la Fuente, a bioengineer at the University of Pennsylvania, who also uses AI for antibiotic discovery but was not involved in this study. “I am excited to see what they do next with some of the molecules that they have generated, how they optimize them, and how they improve them in the future,” de la Fuente said. “The more groups working in this area, the better for humanity.”

While SyntheMol yielded new molecules that displayed antibacterial activity against A. baumannii in vitro, it is unknown how well these chemicals will work in the clinic. “Many things are good antibiotics in a dish, like fire and bleach, but are not good antibiotics to give to patients,” Stokes said. “We are thinking about this problem [beyond] how do I find a new molecule that can kill bacteria in a dish, which is a single property. We are now developing algorithms that are much more interesting in that they consider many, many properties that are required of human medicines.” Stokes and his team are now refining SyntheMol and other generative AI models to contemplate characteristics, such as solubility, toxicity, and drug localization, which will help these algorithms output antibiotics optimized for clinical use. But the knowledge they gained from this study goes beyond antibiotic development. The researchers are also adapting SyntheMol to discover novel antiviral, antifungal, and anticancer drugs and Stokes hopes that other scientists will leverage this approach for other therapeutic domains.