Accurate Prediction of Hydration Sites of Proteins Using Energy Model With Atom Embedding

Abstract

We propose a method based on neural networks to accurately predict hydration sites in proteins. In our approach, high-quality data of protein structures are used to parametrize our neural network model, which is a differentiable score function that can evaluate an arbitrary position in 3D structures on proteins and predict the nearest water molecule that is not present. The score function is further integrated into our water placement algorithm to generate explicit hydration sites. In experiments on the OppA protein dataset used in previous studies and our selection of protein structures, our method achieves the highest model quality in terms of F1 score, compared to several previous studies.

Prediction of hydrophilic and hydrophobic hydration structure of protein by neural network optimized using experimental data

“…Here, we compare the performance of our NN with that of two other NN-based hydration prediction methods, GalaxyWater-CNN_42 30 and Accutar 31 , with respect to the distributions of predicted hydration sites (Fig. 5 and SI Appendix, Fig.…”

“…In this study, as an alternative computational approach, we constructed a neural network (NN) for predicting the hydration probability distribution over the surfaces and the interior cavities of proteins. The constructed NN was optimized for experimentally identified hydration structures from protein crystallography as recently reported NN-based hydration prediction methods 30 , 31 , rather than the method trained by the hydration structures predicted by MD and theoretical calculations 32 . Here, we describe the NN architecture details and demonstrate the performance of predicting hydration structures in the interior, hydrophilic, and hydrophobic surfaces of proteins.…”

Prediction of Water Distributions and Displacement at Protein–Ligand Interfaces

“…Nittinger et al reviewed the performance of these algorithms and characterized the methods as broadly knowledge-based, such as AcquaAlta, Dowser++, HydraMap, and GalaxyWater-wKGB and simulation-based, such as WATsite and GIST . Fusani et al have recently proposed a genetic algorithm method (Gasol) based on the 3D-RISM approach, and Huang et al used a neural network to predict solvation sites. A comparison with several of these algorithms indicated broad agreement with the 83% accuracy in WATGEN for prediction of water positions within 1.5 Å.…”

“…24 Fusani et al 25 have recently proposed a genetic algorithm method (Gasol) based on the 3D-RISM approach, and Huang et al 26 used a neural network to predict solvation sites. A comparison with several of these algorithms 26 indicated broad agreement with the 83% accuracy in WATGEN for prediction of water positions within 1.5 Å. Other methods 42,43 have been described to calculate the energetics of water molecules that are already known to be displaced upon ligand binding.…”

“…As a first step to evaluating the roles of water molecules at protein–ligand interfaces, we have developed an algorithm for computing and classifying water displacement upon ligand binding. This analysis requires prediction of water networks, in part because of the limitations in detection of water molecules in X-ray structures. , There are several algorithms that have been described for placement of water molecules at protein–ligand interfaces. − We have previously shown that our algorithm, WATGEN, can predict water molecules with accuracy at protein–peptide and protein–RNA interfaces. Here, we report use of WATGEN as a method for evaluating water displacement upon ligand binding to a protein.…”