Scientists at the National Centre for Biological Sciences, Bengaluru, built Disobind, a deep-learning tool that predicts how intrinsically disordered proteins attach to partner molecules. These proteins have no fixed 3D shape. The tool learns from millions of protein sequences and does not need structural data. It outperformed tools such as AlphaFold-Multimer and could help research on immune signalling, cancer and brain diseases, and on drug design.