A generative AI system built at the University of Pennsylvania is tackling one of drug discovery’s most stubborn problems: taking antibiotic candidates that show promise in the lab but fail to become usable drugs, and rewriting them into something stronger. The tool, called ApexGO, redesigned peptide antibiotics so effectively that two of its AI-generated molecules matched the potency of polymyxin B, a last-resort antibiotic used against drug-resistant infections, when tested in mice.
The research, led by César de la Fuente, a Presidential Associate Professor at Penn, along with Jacob R. Gardner of Penn’s Department of Computer and Information Science and Marcelo Torres, a research assistant professor of psychiatry, was published in Nature Machine Intelligence on May 13, 2026. The paper, titled ‘A generative artificial intelligence approach for peptide antibiotic optimization,’ describes a transformer variational autoencoder that embeds peptide sequences into a continuous latent space, paired with Bayesian optimization that proposes targeted sequence edits to boost antimicrobial potency.
Why Imperfect Candidates Matter
Antibiotic peptides are short protein chains that occur naturally or can be synthesized, and they represent one of the most active areas of antimicrobial research because bacteria have a harder time evolving resistance to them compared with conventional small-molecule drugs. The catch is that most candidate peptides identified through screening are mediocre: potent enough to show activity in a petri dish, but too weak, too toxic, or too unstable to become an actual medicine. Historically, improving those candidates meant slow, manual rounds of chemical tweaking guided largely by trial and error.
De la Fuente’s lab set out to automate that optimization step. Rather than generating antibiotics from scratch, ApexGO takes an existing imperfect peptide and searches computationally for edits that increase its ability to kill bacteria while limiting toxicity to human cells.
What the Lab Results Showed
According to the Nature Machine Intelligence paper and reporting from Phys.org, the Penn team generated hundreds of antibiotic candidates over months of computational runs, then synthesized and tested a batch of them directly against clinically relevant, drug-resistant bacterial strains. Of the AI-optimized peptides that were synthesized and tested in vitro, 85% halted bacterial growth, and 72% outperformed the original peptide sequences they were derived from. The team also ran mechanism-of-action studies and cytotoxicity evaluations to check that the optimized peptides weren’t simply becoming more effective by becoming more toxic.
The most striking result came from live-animal testing: two of the AI-designed antimicrobial peptides matched the efficacy of polymyxin B in mouse infection models. Polymyxin B is a decades-old antibiotic still used clinically as a treatment of last resort for infections caused by multidrug-resistant gram-negative bacteria, precisely because so few newer drugs can match it.
Quotes From the Research Team
‘ApexGO gives us a way to navigate that space with far more direction,’ de la Fuente said, referring to the vast and largely unexplored landscape of possible peptide sequences that traditional chemistry could never fully search. Gardner, whose lab focuses on machine learning methods including Bayesian optimization, said the real test was whether the model’s predictions would survive contact with biology: ‘What is striking is that ApexGO’s predictions held up in the real world.’
A Field Crowded With Competing Approaches
ApexGO enters a drug-discovery landscape where multiple AI approaches to antibiotic design are advancing in parallel, from large-scale generative models that design entirely novel molecules from scratch to models trained on genomic data mined from ancient or environmental samples. Public health researchers have repeatedly cautioned that computational hits, however promising, still face the same bottleneck that has throttled antibiotic development for decades: getting a molecule through toxicology studies, manufacturing scale-up, and human clinical trials costs far more than any AI model can shortcut. Antimicrobial resistance researchers have also warned that AI-generated candidates require the same rigorous, multi-year safety and efficacy testing as any other drug, meaning computational speed gains at the design stage don’t necessarily translate into faster approvals.
The Broader Antibiotic Resistance Problem
The urgency behind this work stems from a well-documented crisis: bacteria are evolving resistance to existing antibiotics faster than pharmaceutical companies are bringing new ones to market, in part because antibiotics are far less profitable than chronic-disease drugs. Public health researchers have flagged drug-resistant infections as a growing cause of preventable deaths worldwide, and the pipeline of genuinely new antibiotic classes has been thin for years. Tools like ApexGO are aimed less at replacing the drug-development pipeline than at making the early design phase faster and less wasteful, so more candidates that are actually viable make it into that pipeline in the first place.
What Comes Next
The Penn team says its next steps involve testing ApexGO-optimized peptides against a broader range of pathogens and refining the toxicity-prediction side of the model, which remains a harder computational problem than potency prediction. De la Fuente’s lab has previously worked on AI-driven antibiotic discovery projects mining unconventional data sources, and ApexGO is described as one piece of a broader effort to industrialize antibiotic design using machine learning. Whether any ApexGO-derived candidate advances toward human trials will depend on additional toxicology work and, eventually, outside pharmaceutical partners willing to fund that expensive next phase.
