Apo-Holo: a neural network to predict residue-level backbone and side-chain displacements during the formation of CAZyme-sugar complexes.

Predicting local structural rearrangements of enzymes upon ligand binding remains a major challenge. We present Apo-Holo (pronounced Apolo), a dataset and neural-network models that predict per-residue backbone and side-chain displacements (RMSD) and binary mobility labels for carbohydrate-processing enzymes (CAZymes). We constructed a curated dataset of experimentally solved Apo/Holo Protein Data Bank (PDB) pairs, extracted residue-level physicochemical and structural descriptors, and computed the per-residue RMSD for both the backbone and side chain after backbone superposition. We trained classical machine learning models and deep neural networks for regression and classification tasks. Classification networks reached accuracies above 85% in held-out tests, while regression networks reproduced RMSD trends. We discuss biological patterns, limitations, and how Apo-holo can guide experimental prioritization. Code and data are available on GitHub.