Water position prediction with SE(3)-Graph Neural Network

Abstract

Most protein molecules exist in a water medium and interact with numerous water molecules. Consideration of interactions between protein molecules and water molecules is essential to understanding the functions of the protein. In computational studies on protein functions, either implicit solvation or explicit solvation methods are used to consider the effect of water on the protein. Implicit solvation methods consider water as a continuous solvent and have lower computational costs than explicit methods that consider water as a collection of individual water molecules. However, some water molecules have specific interactions with protein molecules, which are critical to protein function and require explicit treatment to consider these specific interactions. Thus, as a compromise between computational cost and consideration of specific interactions, hybrid methods use explicit consideration of water molecules with specific interaction with protein molecules while considering other water molecules implicitly. Prediction of the water positions having specific interaction is required to perform such hybrid methods, where various water position prediction methods have been developed. However, currently developed water position prediction methods still require considerable computational cost. Here, we present a water position prediction method with low computational cost and state-of-the-art prediction performance by utilizing SE(3)-an equivariant graph neural network. The introduction of a graph neural network enabled the consideration of the atom as a single data point, which makes computational costs less than our previous water prediction method using a convolutional neural network, which considers an atom as multiple data points. Our new water position prediction method, WatGNN, showed an average computation time of 1.86 seconds while maintaining state-of-the-art prediction performance. The source code of this water prediction method is freely available at https://github.com/shadow1229/WatGNN.