Prediction of Protein Hydration using Deep Learning

Hydration is necessary for protein folding, stability, and functions. The hydration structure of proteins is formed inside proteins and at the interfaces between bulk solvent and protein, and has been visualized mostly using cryogenic X-ray crystallography. From massive structural data of proteins with hydration structures, we developed a three-dimensional convolutional network to generate distributions of the existence probability of hydration water molecules on protein surfaces and in protein cavities. We also devised a positional search method of hydration water molecules based on the probability maps. The predicted hydration sites are located on an average within 0.3 Å from the experimentally identified sites in crystal structures.