Scientists Develop AI Tool to Predict Protein Binding Sites

Researchers at EPFL have created PeSTo, an AI model that predicts protein binding interfaces with high accuracy. The tool identifies sites on a protein’s surface where it may interact with other proteins, nucleic acids, lipids, ions, and small molecules. PeSTo’s efficient design enables large-scale analysis of structural data, revealing new insights into protein interactions. The work is published in Nature Communications.

Proteins carry out most of the biological functions in living organisms by interacting with various molecules. Understanding these interactions is key to elucidating cellular processes.

PeSTo utilizes a neural network architecture called a transformer to analyze protein structures. It focuses on chemically significant atoms using a self-attention mechanism. This captures the complex atomic interactions within proteins to make accurate interface predictions.

Unlike other methods, PeSTo works directly from the positions and types of atoms in a structure. It does not require preprocessing steps to calculate molecular surfaces or properties. This makes PeSTo very fast, robust, and widely applicable compared to existing tools.

The researchers applied PeSTo to the human foldome, producing a detailed map of interactions across the human proteome. They provide the tool through an easy-to-use web server, enabling scientists to rapidly analyze protein structures.

PeSTo demonstrates improved performance over previous methods and can reliably predict diverse interaction types. Its efficient design and general applicability make it a valuable new tool for elucidating protein interfaces and molecular interactions.

More information: Lucien F. Krapp et al, PeSTo: parameter-free geometric deep learning for accurate prediction of protein binding interfaces, Nature Communications (2023). DOI: 10.1038/s41467-023-37701-8