AI Model Designs Synthetic Antibiotics With in vivo Efficacy

Researchers at the University of Pennsylvania have developed a generative artificial intelligence (AI) model that can create synthetic antimicrobial peptides (AMPs) with bacteria-killing potential. The tool, named AMP-Diffusion, produced tens of thousands of candidate peptides, two of which demonstrated efficacy in treating infections in mice comparable to existing antibiotic drugs.

The work, published in Cell Biomaterials, presents a proof-of-concept that generative AI can be used not just to mine nature’s existing molecules, but to create novel compounds that follow the same biological rules. The study was conducted by researchers in the labs of Pranam Chatterjee and César de la Fuente.

Leveraging image-generation models for biology

AMP-Diffusion is built on a type of generative AI known as a diffusion model. These models typically begin with random inputs and iteratively refine them toward an intended output – a method widely used in AI tools for image creation. Instead of refining pixels, AMP-Diffusion outputs strings of amino acids, the building blocks of peptides.

Unlike other tools that train from scratch, AMP-Diffusion incorporates ESM-2, a protein language model trained by Meta on hundreds of millions of protein sequences. This integration allows the AI to generate sequences with biologically plausible structures and properties without relearning foundational protein rules. The model also references ESM-2’s internal rules during generation, a process that keeps the outputs within known biological constraints.

Filtering high volumes of synthetic candidates

The generative process produced approximately 50,000 potential AMP sequences. Researchers then applied a second AI model, APEX 1.1, to rank and filter these candidates. This model prioritized sequences with predicted antimicrobial activity, while removing those that were either redundant or too similar to known peptides.

From this pool, 46 top-ranked peptides were synthesized and tested in both human cell cultures and animal models. In mice, two of the peptides were effective against skin infections. Their efficacy was comparable to levofloxacin and polymyxin B – two US Food and Drug Administration (FDA)-approved antibiotics – without observable adverse effects.

Focusing future iterations on therapeutic goals

Although diffusion models have been explored previously in the context of AMP generation, this study introduces several key methodological improvements. By relying on a pre-trained protein language model, the team bypassed the need for creating a new “latent space” for peptide structure, accelerating the generation process. The tool’s structure also enables future modifications to steer the AI toward producing peptides for specific therapeutic targets or improved drug-like properties.

While the current work focuses on demonstrating feasibility, researchers aim to develop AMP-Diffusion further to reduce the timeline for antibiotic discovery and allow more precise tuning of peptide characteristics.

Reference: Torres MDT, Chen LT, Wan F, et al. Generative latent diffusion language modeling yields anti-infective synthetic peptides. Cell Mat. 2025. doi: 10.1016/j.celbio.2025.100183 

This article has been republished from the following materials. Note: material may have been edited for length and content. For further information, please contact the cited source. Our press release publishing policy can be accessed here.

This content includes text that has been generated with the assistance of AI. Technology Networks' AI policy can be found here.