AI maps how more than 18,000 human proteins move in under two days
Proteins are not frozen shapes. They flex and shift, and some disease mutations or drug binding sites only show up when a protein moves into a particular pose. Capturing that motion has traditionally required huge molecular dynamics simulations that can run for weeks or months. Researchers at the University of Illinois Urbana-Champaign built ESMDynamic, a deep learning model that predicts which contacts inside a protein are likely to change, and roughly how often and how fast. Using it, the team mapped the moving parts of more than 18,000 human proteins in less than two days. The study, in Nature Communications, found the model matched or beat leading methods while running much faster.