Penn Engineers Develop AI Method to Refine Promising Antibiotic Peptides

University of Pennsylvania engineers have developed an artificial intelligence method that takes promising but imperfect antibiotic peptide candidates and improves them through successive rounds of optimization. The work is detailed in a study titled "A generative artificial intelligence approach for peptide antibiotic optimization," published in the journal Nature Machine Intelligence. Rather than screening large, fixed libraries of compounds, the approach is designed to navigate the vast molecular space of drug discovery with more precision by starting from known leads and refining their structures. Laboratory testing found that 85 percent of the AI-generated molecules halted bacterial growth, and 72 percent outperformed the original peptides from which they were derived.

The Details

The new method is called ApexGO. It was developed by researchers at the University of Pennsylvania and represents an advance beyond APEX, an earlier artificial intelligence model from the same laboratory that predicts whether peptides possess antimicrobial properties. While APEX is used to identify promising starting compounds, ApexGO is built to refine those compounds after they have been identified. This distinction shifts the focus from discovery alone to the improvement of leads that already show some potential but require further enhancement before they can be considered viable drug candidates.

The study, titled "A generative artificial intelligence approach for peptide antibiotic optimization," appears in the journal Nature Machine Intelligence. It lays out a workflow in which AI does not merely filter a static collection of compounds, but instead proposes structural changes to existing leads and learns from the results of each round. This iterative loop is central to how ApexGO operates.

According to the study, the researchers began with ten peptide templates. From these templates, ApexGO generated optimized derivative structures. The team then moved these computer-generated designs into the laboratory, chemically synthesizing one hundred compounds for in vitro characterization. This step was necessary to determine whether the AI-optimized molecules retained activity against bacteria and whether they performed better than the original templates from which they were drawn. This approach of starting from a small set of known templates and generating derivatives contrasts with the conventional practice of screening large, fixed libraries without subsequent structural refinement.

The laboratory results were measured against bacterial cultures. Of the one hundred synthesized compounds, 85 percent halted bacterial growth. This hit rate indicates that the vast majority of the computationally designed molecules retained biological activity after optimization. In subsequent tests against Gram-negative pathogens, 72 percent of the optimized peptides demonstrated improved antimicrobial activity relative to the peptides they were derived from. These figures represent the proportion of AI-designed molecules that not only worked, but worked better than their starting points. The 85 percent hit rate applied to bacterial growth inhibition broadly, while the 72 percent improvement rate was specific to Gram-negative pathogen assays.

The evaluation extended beyond cell cultures. The researchers tested the optimized molecules in two mouse models of infection caused by Acinetobacter baumannii, a drug-resistant pathogen frequently encountered in healthcare environments. In these animal models, the AI-optimized molecules exhibited anti-infective activity that was superior to the template controls. Their effectiveness was described as comparable to or better than polymyxin B, a last-resort antibiotic that the Penn release identified as the comparator in the study.

César de la Fuente, who leads the laboratory where ApexGO was developed, described antibiotic discovery as fundamentally a search problem across an enormous molecular space. "ApexGO gives us a way to navigate that space with far more direction," he said.

Jacob R. Gardner, a researcher involved in the project, commented on the transition from computational prediction to laboratory validation. "What is striking is that ApexGO's predictions held up in the real world," Gardner said.

Yimeng Zeng, another member of the research team, described the optimization mechanism. According to Zeng, the model relies on Bayesian optimization to guide its choices. "Bayesian optimization helps the model make informed choices about what to try next, balancing candidates that look promising with candidates that could teach the model something new," Zeng said.

Context

Antibiotic resistance is described in the study as a growing global problem. The development of faster and more reliable pathways for antimicrobial design is therefore considered strategically important. Peptide antibiotics represent one area of active investigation, particularly for pathogens that have developed resistance to conventional small-molecule drugs. The study focuses on Gram-negative pathogens, a category that includes Acinetobacter baumannii. These organisms are frequently associated with hospital-acquired infections and are described as a major drug-resistant threat in healthcare settings. They are often difficult to treat because of their complex cellular envelopes and their ability to resist multiple classes of antibiotics.

The distinction between identifying a candidate and improving it is significant in drug development. A lead compound that shows initial activity may still be too weak, too unstable, or too toxic for therapeutic use. ApexGO is aimed at the refinement stage, attempting to lift an imperfect molecule toward viability without starting over with a new chemical structure.

The de la Fuente laboratory at Penn previously released APEX, an artificial intelligence model designed to predict antimicrobial properties in peptides. ApexGO is presented as the next step in that workflow. Rather than only identifying whether a peptide might have antimicrobial properties, the new method actively improves imperfect candidates through iterative structural refinement. APEX and ApexGO are presented as sequential tools within the same research program: APEX generates candidate peptides, and ApexGO then optimizes their structures to improve antimicrobial performance.

Phys.org, in its republication of the Penn announcement, independently noted the step-by-step nature of the work. The outlet reported that the method improved imperfect antibiotic candidates incrementally and that the majority of AI-designed molecules worked in laboratory validation. CBS News was also consulted during the research phase as an adjacent source.

What's Next

The study demonstrates the viability of using iterative artificial intelligence optimization to improve peptide antibiotics at the preclinical stage. Because the validation was conducted in laboratory cultures and in mouse infection models, the findings represent an intermediate step between computational design and clinical application. The researchers did not indicate whether any of the optimized molecules are being prepared for human trials. Further preclinical testing would be required before regulatory review could begin. For now, the method adds a refinement step to the existing APEX pipeline, giving the laboratory a way to move from identification to improvement without relying solely on large-scale screening. The researchers have positioned ApexGO as a complement to APEX, linking identification with subsequent structural improvement. No commercial partnerships were mentioned in the published findings or in the accompanying university announcement.