Predicting side chain conformations in folded proteins by AlphaFold: Perspective and challenges

AlphaFold has revolutionized protein structure prediction by accurately creating 3D structures from just the amino acid sequence. However, a key question important for the molecular modeling field remains: can AlphaFold predict the conformations of individual amino acid residue side chains within a folded protein? Herein, we investigate the ability of ColabFold, an online implementation of AlphaFold2, and AlphaFold3 to predict the side chain conformations in folded proteins. We find that over a set of 10 benchmark proteins (set A) representing several different highly populated fold families, which are included in the AlphaFold protein structure database, the side chain conformation prediction error of ColabFold is ∼14% for χ₁ dihedral angles, and it increases to ∼48% for χ₃ dihedral angles. Prediction error is smaller for nonpolar side chains and is somewhat improved using structural templates. ColabFold demonstrates a bias toward the most prevalent rotamer states in the Protein Data Bank, potentially limiting its ability to capture rare side chain conformations effectively. Additionally, for 10 recently released protein structures, which were not employed in the training of AlphaFold2, we show that ColabFold predicts side chain conformations with almost the same accuracy as for set A. Also, we demonstrate the side chain prediction accuracy by AlphaFold3 is slightly better than by ColabFold. As an application of AlphaFold to explore the structural consequences of strongly cooperative mutations on side chain rearrangements, we employ a Potts sequence-based statistical energy model to perform large-scale mutational scans of two proteins, ABL1 and PIM1 kinase, searching for the most strongly cooperative mutational pairs, and then, we use ColabFold to predict the structural signatures of this cooperativity on the interacting side chains. Our results demonstrate that integration of the sequence-based Potts model with AlphaFold into a single pipeline provides a new tool that can be used to explore the fundamental relationship between protein mutations, cooperative changes in structure, and fitness.