AI tool boosts 85% of imperfect antibiotic candidates in lab

Researchers at the University of Pennsylvania have developed ApexGO, a new artificial intelligence method designed to transform imperfect antibiotic candidates into highly effective treatments. Unlike traditional AI approaches that scan vast databases for existing promising molecules, ApexGO starts with a small set of flawed candidates and iteratively improves them. The system uses a predictive algorithm to evaluate each proposed modification and guide the next step in the optimization process, effectively navigating the enormous molecular space where antibiotic discovery typically occurs. The study, detailed in Nature Machine Intelligence, was led by senior co-authors César de la Fuente and Jacob R. Gardner. The approach combines a previously published model called APEX, which predicts antimicrobial properties, with Bayesian optimization techniques that allow the system to efficiently explore possible solutions. This hybrid model suggests specific molecular tweaks, predicts their potential to enhance activity, and uses those predictions to direct subsequent edits. As doctoral student Yimeng Zeng explained, the system balances exploring promising areas of the chemical space with investigating less certain regions where hidden improvements might exist. Laboratory results validated the AI's predictive capabilities. When tested against disease-causing bacteria, 85% of the AI-generated molecules successfully halted bacterial growth. Furthermore, 72% of these new variants outperformed the original peptide candidates from which they were derived. In animal trials using mice, two of the antimicrobial peptides produced by ApexGO reduced bacterial counts to levels comparable to polymyxin B, an FDA-approved antibiotic currently used as a last-resort treatment for drug-resistant infections. Jacob R. Gardner noted that the success was particularly significant because the AI was optimizing based on its own computer models, which often risk creating molecules that perform well in simulations but fail in physical testing. The fact that the majority of ApexGO's designs worked in the real world demonstrates the tool's reliability. Until now, antibiotic discovery has often relied on serendipity, such as the accidental discovery of penicillin. ApexGO offers a systematic alternative to this trial-and-error method, potentially allowing researchers to identify hundreds of candidates in a matter of months rather than years. While the findings mark a major advancement, researchers caution that these peptides are still early-stage candidates. Before they can be used to treat human infections, further development is required to ensure safety, stability, and efficacy within the human body. However, the study suggests that AI tools like ApexGO can significantly narrow the search for viable therapies, reducing the time and cost associated with traditional discovery methods. The team believes this technology could eventually extend beyond antibiotics to optimize peptides for other biological functions, such as immune modulation or tumor targeting. As antibiotic resistance continues to rise globally, ApexGO represents a critical step toward accelerating the transition from theoretical molecular design to practical therapeutic solutions.