Credit: Chris Ratcliffe/Bloomberg via Getty
Researchers are turning to artificial intelligence (AI) to help design the next generation of antibiotics to combat rising antimicrobial resistance. In minutes, AI can design thousands of chemical compounds with potential antibacterial properties, although there are hurdles to overcome before the first of these medicines can be tested in people.
Last week, the US Centers for Disease Control reported that rates of dangerous bacterial infections surged by 69% between 2019 and 2023. Enterobacterales bacteria, also called ‘nightmare bacteria’, are particularly difficult to treat with existing antibiotics. Globally, 1.1 million deaths a year are linked to bacterial resistance to antimicrobial drugs.
The standard method of antibiotic discovery involves going into nature and sifting through dirt to find antibacterial compounds, says César de la Fuente, a machine biologist at the University of Pennsylvania, Philadelphia. “That’s really painstaking work that relies on trial and error, and it can take many years,” he adds. His team has been using AI to discover antibiotics for about a decade. The whole process of discovering a candidate, creating it in the laboratory and testing it in cells can be done “within a week or two”, he adds.
Most AI-designed antibiotics are still in early development. None has yet been tested in humans.
AI-designed drugs
Machine learning and generative AI (genAI) can speed up the process. De la Fuente and his team train machine-learning algorithms by showing them compounds that can harm bacteria, and others that cannot. The AI designs antibiotics by looking for fragments of proteins with antibacterial properties in data sets that it has not seen before. These include the proteomes — the complete set of proteins that an organism can express — of animals, plants and bacteria.
GenAI algorithms — similar to the AI used in chatbots or image generators — are trained on the same data but designed to create new compounds. Earlier this month, de la Fuente and his team reported that their genAI model designed 50,000 peptides, or short chains of amino acids, that have antimicrobial properties and can destroy pathogens1. A deep-learning model then ranked these on the basis of how effective it thinks they might be at killing several bacteria types. Of the top 46 synthesized peptides, about 35 killed at least one bacterial strain in a dish, and most were not toxic to human embryonic kidney cells. The top two candidates were then tested and found to be effective against Acinetobacter baumannii in mouse models.
Challenges ahead
But making AI-designed antibiotics in the lab can be challenging. For example, bioengineer Jim Collins, at the Massachusetts Institute of Technology in Cambridge, has found that some AI-designed antibiotics are chemically unstable and cannot be synthesized. Others take too many steps to make and would be too costly and time-consuming to produce commercially.
