The flexibility of a protein is an essential contributor to its function and affects properties such as stability, catalysis, and molecular interactions. Experiments can be used to study side-chain flexibility, but cannot be applied at scale, and most structure prediction methods focus on reconstructing a single configuration. Here, we use the internal side-chain representations of AlphaFold2 to develop AF2chi to generate conformational ensembles representing the dynamics of protein side chains. We benchmark AF2chi predictions on a broad set of proteins using NMR 3 J-couplings, S 2 order parameters, and dihedral-angle distributions from collections of experimental structures, demonstrating the high accuracy of AF2chi. We also compare the accuracy of AF2chi with molecular dynamics simulations and machine learning models that generate conformational ensembles and show that AF2chi provides state-of-the-art accuracy. The speed and accuracy of AF2chi open up for a wide range of applications such as in protein design, ligand docking, and interpretation of biophysical experiments such as from X-ray crystallography.
Cagiada et al. (Sun,) studied this question.