A team of Penn engineers says a homegrown artificial intelligence model is turning so-so peptide molecules into serious antibiotic contenders, with results that already hold up in lab dishes and infected mice. Working with an AI tool dubbed ApexGO, the researchers report that out of 100 optimized peptides they synthesized, about 85% stopped bacteria from growing in vitro and roughly 72% outperformed the original “template” molecules they started from. Several of the ApexGO-designed candidates also cleared infections in Acinetobacter baumannii mouse models at levels comparable to last-resort drugs.
How ApexGO works
ApexGO combines a transformer-based variational autoencoder with Bayesian optimization to make targeted, rule-following edits to an existing peptide template and rank new variants that are predicted to have stronger antimicrobial punch, according to a paper in Nature Machine Intelligence. Instead of blindly trawling through an enormous, fixed menu of possible molecules, the system builds on the lab’s earlier APEX predictor and focuses on tuning and upgrading existing peptide scaffolds.
Lab and animal results
In lab assays and two preclinical mouse infection models, several peptides proposed by ApexGO cut bacterial counts and performed on par with polymyxin B, a last-resort antibiotic, as reported by Penn Today. The team began with ten template peptides that had been pulled from ancient proteomes, then leaned on the AI’s built-in uncertainty estimates to select a relatively small, focused group of candidates for chemical synthesis and testing. The outcome, the researchers say, is a surprisingly high hit rate for a generative optimization pipeline that is still in its early days.
Voices from the lab
Presidential Associate Professor César de la Fuente has his own catchy tagline for the tool. ApexGO is “a way of improving lousy antimicrobials,” he told The Daily Pennsylvanian. In that interview, he noted that about 85% of the peptides they actually synthesized killed bacteria and that roughly 72% did better than the initial templates. Project co-worker Natalie Maus said the model helps shrink the number of molecules that have to be made and tested in the wet lab, so bench time can be spent on the most promising shots on goal.
Why this matters
The stakes here are not abstract. Drug-resistant infections are already cutting lives short around the world, with a 2022 analysis in The Lancet estimating that about 4.95 million deaths in 2019 were associated with bacterial antimicrobial resistance. Anything that can speed up the risky, failure-heavy early stages of antibiotic discovery could help widen the pipeline of candidates that move into full-on drug development and clinical testing.
Limits and next steps
The Penn team is not pretending ApexGO is a magic “click and cure” button. The authors emphasize that a more potent peptide is only the start of a viable medicine. A drug has to survive long enough in the body, reach the right tissues, and do its job without causing too much collateral damage. That means stability, bioavailability, and safety all have to be engineered in, and future versions of ApexGO will need to juggle several design constraints at once while also taking into account the specific pathogens being targeted, according to the Nature Machine Intelligence paper. Even so, the study offers a working blueprint for using generative AI to steer the vast space of possible peptide sequences toward molecules that can actually show measurable, in vivo efficacy.
The work, published in Nature Machine Intelligence in May 2026, marks an early but concrete advance out of a Philadelphia lab that is trying to chip away at a global health crisis. If ApexGO can scale up and the familiar obstacles of toxicity and drug delivery can be solved, the tool could help shorten the long, expensive path from a promising peptide to an antibiotic that physicians can actually prescribe.
