Building on earlier AI work, scientists from one university use neural networks and Bayesian optimization to design new antibiotics for validation.

Antibiotic resistance has become one of the most pressing public health threats worldwide, yet the pipeline for new drugs remains alarmingly thin.

Over the past few decades, the development of novel antibiotics has slowed significantly, with many pharmaceutical firms scaling back or abandoning research in this area due to high costs and limited financial returns. As a result, clinicians are increasingly forced to rely on a small and aging arsenal of treatments, some of which are losing effectiveness as bacteria evolve resistance.

Now, researchers at the University of Pennsylvania have developed a new AI system that could transform how antibiotics are discovered and improved. Instead of relying on traditional methods that screen large libraries of molecules, the system focuses on refining and optimizing antimicrobial peptides through an iterative design process. Unveiled in Nature Machine Intelligence in May 2026, the system works as follows:

  • Starting with an initial peptide that may only have modest antibacterial activity, the system then repeatedly suggests modifications, predicts how those changes will affect performance, and selects the most promising candidates for further refinement.
  • This cycle continues in a loop, allowing the AI to efficiently explore the vast space of possible peptide designs. The framework combines advanced neural network models with Bayesian optimization, enabling it to balance experimentation with informed decision-making.
  • In laboratory tests, the results were striking. A large majority of the peptides generated by the system had successfully inhibited bacterial growth, and many showed improved effectiveness compared to their original versions. The system was particularly successful against Gram-negative bacteria, which are often difficult to treat due to their strong defenses against antibiotics.

To further validate the approach, researchers tested two AI-designed peptides in mice. These candidates reduced bacterial levels to a degree comparable to polymyxin B, a powerful antibiotic typically reserved for severe, drug-resistant infections. This suggests that the AI-generated compounds have real therapeutic potential, not just theoretical promise.

Building on earlier innovation

The current system builds on earlier work from the same research group, which previously focused on identifying antimicrobial peptides from natural and even extinct sources. The new system goes a step further by actively designing and improving molecules rather than simply discovering them.

Although the findings are promising, the researchers caution that these peptides are still in early stages of development. Additional testing is needed to ensure safety, stability, and effectiveness in humans.

Looking ahead, the team believes this technology could also be applied beyond antibiotics, potentially aiding in the design of treatments for cancer or immune-related diseases.