By 2050, experts predict antibiotic-resistant infections could cause over 8 million deaths worldwide each year. These infections occur when bacteria resist standard drugs such as penicillin. They arise from contaminated food, wounds or surgery, with E. coli strains showing strong resistance. Secondary infections like pneumonia can also develop. Creating new antibiotics is slow and costly, often taking 10 years and over $1 billion per drug. Ten of 13 antibiotics introduced since 2017 already fail against some bacteria. Generative AI models, directed by scientists, offer a way to propose new molecules. Physics-based simulations then test these designs quickly and cheaply by mimicking real-world rules. Peptides serve as a useful starting point because they are short proteins with varied roles, including insulin for diabetes and vancomycin as an antibiotic. AI systems combine a generator to create many designs with a recommender to select the next simulation. Tests showed that limited but relevant data trains generators better than large amounts of mixed information. Recommenders were also refined to map search paths more clearly. Physics simulations validate suggestions by modeling how peptides change shape near different membranes. In these models, atoms act as soft spheres in water, allowing observation of interactions similar to a video game engine. This approach distinguishes harmless dances near mammal cells from lethal ones near bacterial cells.
