Researchers at The University of Texas at Austin have leveraged artificial intelligence to develop a new antibiotic showing promise in animal trials. Published today in Nature Biomedical Engineering, the study details how scientists used a large language model (LLM) to engineer a previously toxic bacteria-killing drug to be safe for human use.
The rise of antibiotic-resistant bacterial strains has worsened the prognosis for patients with dangerous infections, as new treatment options have stalled. However, UT researchers believe AI tools can change this scenario.
“We have found that large language models are a major step forward for machine learning applications in protein and peptide engineering,” said Claus Wilke, professor of integrative biology and statistics and data sciences, and co-senior author of the paper. “Many use cases that weren’t feasible with prior approaches are now starting to work. I foresee that these and similar approaches are going to be used widely for developing therapeutics or drugs going forward.”
Originally designed to generate text sequences, LLMs are being creatively applied in other domains. Proteins, like sentences, consist of sequences—in this case, amino acids. LLMs cluster words sharing common attributes into an "embedding space" with thousands of dimensions. Similarly, proteins with functions such as fighting bacteria without harming their hosts may cluster together in an AI embedding space.
“The space containing all molecules is enormous,” said Bryan Davies, co-senior author of the paper. “Machine learning allows us to find the areas of chemical space that have the properties we’re interested in, and it can do it so much more quickly and thoroughly than standard one-at-a-time lab approaches.”
For this project, researchers employed AI to reengineer Protegrin-1—an antibiotic effective against bacteria but toxic to humans. Naturally produced by pigs, Protegrin-1 belongs to antimicrobial peptides (AMPs), which generally kill bacteria by disrupting cell membranes but often target both bacterial and human cells.
Initially, researchers used a high-throughput method they developed earlier to create over 7,000 variations of Protegrin-1 and identify modifiable areas without losing its antibiotic activity. They then trained a protein LLM on these results to evaluate millions of possible variations based on three features: targeting bacterial membranes selectively, killing bacteria effectively, and not harming human red blood cells. This led them to develop bacterially selective Protegrin-1.2 (bsPG-1.2).
Mice infected with multidrug-resistant bacteria treated with bsPG-1.2 were less likely to have detectable bacteria six hours post-infection compared with untreated mice. If further testing yields positive results, researchers hope eventually to bring an AI-informed antibiotic drug into human trials.
“Machine learning’s impact is twofold,” Davies said. “It’s going to point out new molecules that could have potential to help people, and it’s going to show us how we can take those existing antibiotic molecules and make them better and focus our work more quickly get those to clinical practice.”
This project underscores how academic researchers are advancing artificial intelligence for societal needs—a key theme at UT during its Year of AI declaration for 2024.
The study's other authors include research associate Justin Randall and graduate student Luiz Vieira from UT.
Funding was provided by the National Institutes of Health, The Welch Foundation, the Defense Threat Reduction Agency, and Tito’s Handmade Vodka.