New AI Model Designs Therapeutic Peptides Beyond AlphaFold's Scope

A multi-institutional team from McMaster University, Duke University, and Cornell University has developed an artificial intelligence tool that can create small, drug-like peptides designed to bind to and degrade harmful proteins—without requiring those proteins’ 3D structures. The study, published in Nature Biotechnology, introduces PepMLM, an AI model based on an algorithm originally built to understand human language but subsequently trained to understand the “language” of proteins from their amino acid sequences.

This approach marks a fundamental departure from AlphaFold, the 2024 Nobel Prize–winning AI developed by Google DeepMind, which predicts 3D protein structures to guide drug discovery. While AlphaFold transformed understanding of structured proteins, many critical disease-associated proteins, including those linked to cancer, neurodegenerative disorders, and certain viral infections, lack stable structures. Such intrinsically disordered or unstable proteins have been considered “undruggable” by traditional structure-based methods.

“Most drug design tools rely on knowing the 3D structure of a protein, but many of the most important disease targets don’t have stable structures,” said senior author Pranam Chatterjee. “PepMLM changes the game by designing peptide binders using only the protein’s amino acid sequence.”

In lab tests, the team showed that PepMLM could design peptides that stick to disease-related proteins and, in some cases, help destroy them. These included proteins involved in cancer, reproductive disorders, Huntington’s disease, and even live viral infections.

The team is now advancing the technology with next-generation AI models such as PepTune and MOG-DFM to improve stability, targeting, and delivery of peptides in the body. Chatterjee described the ultimate goal as “a general-purpose, programmable peptide therapeutic platform–one that starts with a sequence and ends with a real-world drug.”