AI Designs Better Peptide Antibiotics in Landmark UPenn Study

Researchers at the University of Pennsylvania have developed ApexGO — an artificial intelligence system that designs better peptide antibiotics by starting with weak candidates and improving them. In lab tests, 85% of the AI-generated molecules halted bacterial growth, and 72% outperformed the peptides from which they were derived.

The findings, published in Nature Machine Intelligence on May 13, represent a significant advance in the application of generative AI to antimicrobial discovery — a field racing to address the global crisis of antibiotic-resistant infections.

How ApexGO Works

ApexGO is built on two complementary AI components. The first, building on the lab’s earlier APEX model published two years ago, predicts whether a given peptide sequence is likely to have antimicrobial activity. The second — the generative engine — proposes molecular tweaks to existing peptide candidates, creating new sequences that the prediction model then scores.

“It’s fundamentally a search problem across an enormous molecular space,” said César de la Fuente, PhD, presidential associate professor at the University of Pennsylvania and senior author of the study. “ApexGO gives us a way to navigate that space with far more direction.”

The system works iteratively: generate a variant, predict its activity, rank the results, and feed the best performers back into the generative loop. After multiple cycles, ApexGO converges on peptide sequences that are computationally predicted to be more potent than any of the starting candidates.

Real-World Results

What distinguishes this study from purely computational work is the lab validation. The researchers synthesized the top AI-generated peptides and tested them against real bacterial pathogens.

“What is striking is that ApexGO’s predictions held up in the real world,” said Jacob R. Gardner, PhD, assistant professor in computer and information science at Penn and co-author. “ApexGO was optimizing against a computational prediction of antimicrobial activity, but the resulting molecules actually worked.”

In mouse models, two antimicrobial peptides created by ApexGO reduced bacterial infection levels — a critical proof point for a field where many computational predictions fail to translate to in vivo efficacy.

Marcelo Torres, PhD, research assistant professor of psychiatry in the Perelman School of Medicine and lead author, noted the efficiency gain: “APEX helped us find promising antibiotic candidates in enormous biological datasets. ApexGO helps us improve them.”

The Antibiotic Resistance Context

The study arrives as antimicrobial resistance (AMR) is recognized as one of the most pressing public health threats of the 21st century. The World Health Organization has declared AMR a top global health priority, warning that routine infections could become untreatable without new classes of antibiotics.

Traditional antibiotic discovery has slowed dramatically, stymied by the high cost of drug development pipelines and the complexity of bacterial defense mechanisms. Peptide-based antibiotics — short chains of amino acids that mimic natural immune defenses — have long been considered promising alternatives because bacteria find it harder to develop resistance to them. But designing effective peptide antibiotics has remained challenging due to the vast chemical space of possible sequences. The broader context — where peptide hype has been outpacing the evidence in consumer wellness markets — makes rigorous, peer-reviewed advances like this one particularly important.

ApexGO demonstrates one way AI can compress that search from years to days.

What This Means for Peptide Research

The study is significant for the broader peptide field for several reasons. First, it validates that AI-driven optimization can work on peptides — molecules that are increasingly central to drug development pipelines across oncology, metabolic disease, and infectious disease. As the BPC-157 compound guide illustrates, even widely used research peptides operate in a regulatory gray area where clinical evidence remains thin. Second, it shows that generative AI can not only discover but actively improve existing candidates, a capability with implications for peptide design well beyond antibiotics.

The approach could eventually extend beyond antimicrobials. “ApexGO shows that AI can do more than predict which molecules might work,” de la Fuente said. “It can help us improve them.”

Caveats and Next Steps

The researchers emphasize that even the best-performing ApexGO peptides remain early-stage candidates. None are ready for clinical trials; additional optimization and safety testing will be required before any could be evaluated in humans. Still, the study suggests a future in which AI-guided optimization becomes a standard step in peptide drug development.

“The approach could eventually extend beyond antibiotics,” the authors wrote, pointing to potential applications in oncology, metabolic disorders, and other therapeutic areas where peptide-based treatments are under development.

The UPenn lab — which has previously identified antibiotic candidates from frog secretions, ancient microbes, and computational models of extinct proteins — continues to expand its toolkit. ApexGO, the researchers said, is now being applied to additional therapeutic targets.

PeptidesBeat is an independent editorial publication covering peptide policy, research, and industry developments. We do not sell peptides, recommend dosing, or provide medical advice. All content is informational. Peptides referenced may be subject to FDA restrictions; consult a licensed healthcare provider for any therapeutic question.