P01-07
Deep-Learning model for Predicting the Replacement of Water Molecule upon Ligand Binding
Yuki ITO1, Masateru OHTA2, Mitsunori IKEGUCHI2, 3, Takashi YOSHIDOME *1
1Department of Applied Physics, Graduate School of Engineering, Tohoku University
2AI-Driven Drug Discovery Collaborative Unit, HPC- and AI-Driven Drug Development Platform Division, Center for Computational Science, RIKEN
3Graduate School of Medical Life Science, Yokohama City University
( * E-mail: takashi.yoshidome.b1@tohoku.ac.jp )
In drug discovery, the performance of docking software is often limited due to the exclusion of water molecules located at the interface between proteins and ligands from the input data. A proposed solution involves incorporating only those water molecules that remain bound during the protein-ligand binding. Although molecular dynamics (MD) simulations can in principle be possible to incorporate the water molecules, they are notoriously time-consuming. Thus, a fast and accurate method is required for predicting the water molecules that should be incorporated into the drug discovery.
To address this challenge, here the following protocol for predicting the water molecules that should remain in the binding pocket is proposed. First, the hydration structure around a protein is computed using our deep-learning model “gr Predictor” [1] enabling the prediction of the hydration structure in approximately a minute while MD requires a few hours to obtain the hydration structure. Next water molecules are placed using the obtained hydration structure and the program suite “placevent” [2]. Finally, a convolutional neural network (CNN) is implemented to predict each water molecule as either “displaceable” or “non-displaceable”. Upon testing the model on unknown data, it achieved an accuracy of 0.6971 and a recall of 0.584. We will also show a prediction result of replaced and non-replaced water molecules in a horo structure using the apo structure.
[1] K. Kawama, Y. Fukushima, M. Ikeguchi, M. Ohta, and T. Yoshidome, J. Chem. Inf. Model., 62, 4460 (2022).
[2] D.J. Sindhikara, N. Yoshida, F. Hirata, J. Comput. Chem., 33, 1536 (2012).