Generative artificial intelligence (AI) can now design antibiotics. Scientists at the University of Pennsylvania developed AMP-Diffusion, a tool that invents antimicrobial peptides, or AMPs. These are short chains of amino acids, the basic units of proteins, that can kill bacteria. In tests on animals, the best AMPs worked as well as drugs approved by the FDA, with no noticeable side effects.
Earlier work used AI to find antibiotics in data from nature, like ancient animals or microbes. But this study shows AI can make new ones not found in nature. The scientists adapted image-making AI to build molecules. They say nature has limits, but AI can try new designs.
The tool uses a diffusion model, which starts with random noise, like static, and refines it step by step into a clear output. Here, it refines amino acid sequences instead of pictures. AMP-Diffusion builds on ESM-2, a protein language model trained on millions of real proteins. This model acts like a map of protein structures, helping the tool create realistic peptides quickly without relearning basics.
Testing the designs
The researchers made about 50,000 candidate sequences with the tool. They used another AI, APEX 1.1, to pick the best ones based on bacteria-killing power, newness, and variety. From these, they made 46 peptides in the lab and tested them. In mice with skin infections, two peptides matched the effectiveness of levofloxacin and polymyxin B, drugs for resistant bacteria, without harm.
In the future, the scientists plan to improve the tool to target specific infections and add useful drug features. This could shorten the time to find new antibiotics from years to days, helping fight rising resistance where bacteria outpace discoveries.
“It’s exciting to see that our AI-generated molecules actually worked,” reads a press release. “This shows that generative AI can help combat antibiotic resistance.”
The scientists have described the methods and results of this study in a paper published in Cell Biomaterials.