New AI model tailors phage treatment for resistant bacterial infections

With the rapid development of antibiotics in the 1930s, phage therapy – using viruses known as bacteriophages or phages to tackle bacterial infections – fell into oblivion. But as the current rise in antibiotic resistance is making it increasingly difficult to treat bacterial infections, phage therapy is once again sparking interest among physicians and scientists – although it remains complex in practice because of the great diversity and specificity of phages.

Against this backdrop, scientists from the Institut Pasteur, Inserm, the Paris Public Hospital Network (AP-HP) and Université Paris Cité have developed a simple and effective new tool that recommends the best possible phage cocktail for a given patient. They did so by developing and training an artificial intelligence model capable of making a tailored selection of phages based solely on the genome of the targeted bacteria. The results of this research, published on October 31, 2024 in the journal Nature Microbiology, pave the way for personalized phage therapies to treat antibiotic-resistant bacterial infections.

Aude Bernheim, last author of the study and Head of the Institut Pasteur's Molecular Diversity of Microbes laboratoryWe put phages in contact with bacteria in culture and observed which bacteria were killed. We studied 350,000 interactions and successfully identified the characteristics in the bacterial genome likely to predict phage efficacy."

"Contrary to what we initially thought, the ability of phages to infect bacteria, which indicates their efficacy, is determined by receptors at the bacterial surface rather than bacterial anti-viral defense mechanisms," continues Florian Tesson, co-first author of the paper and a PhD student in the Molecular Diversity of Microbes laboratory at the Institut Pasteur and the IAME laboratory at Université Paris Cité-Inserm.

This precise, comprehensive analysis of the interaction mechanisms between bacteria and phages enabled the bioinformaticians in the team to design an optimized, effective artificial intelligence program. The program is based on an analysis of the bacterial genome, especially the regions involved in coding bacterial membrane receptors – the gateway for phages. "We are not dealing with a "black box," and that's what makes our AI model so effective. We know exactly how it works, and that helps us to improve its performance," says Hugo Vaysset, co-first author of the paper and a PhD student in the Institut Pasteur's Molecular Diversity of Microbes laboratory. After more than two years of development and training, the AI model was able to correctly predict the efficacy of phages in treating the E. coli bacteria in the dataset in 85% of cases, simply by analyzing the bacterial DNA. "This result exceeded our expectations," says Aude Bernheim. To take their research further, the scientists tested the model on a new collection of E. coli bacterial strains responsible for pneumonia and selected a tailored "cocktail" of three phages for each of them. In 90% of cases, the phages specifically chosen by AI were successful and destroyed the bacteria. This method, which can easily be transferred to hospital laboratories, paves the way in the coming years for a strategy whereby a personalized selection of phage treatments can be made rapidly if bacterial infection with highly antibiotic-resistant Escherichia coli is diagnosed. "We still need to test the effect of phages in different environments, but proof of concept has been established. We hope to be able to extend it to other pathogenic bacteria, since our AI model has been designed to adapt easily to other scenarios with the aim of offering personalized phage therapy treatments in future," concludes Aude Bernheim.

Source:

Institut Pasteur

Journal reference: