Prediction of Hydration Free Energy Distributions around Proteins using Deep Learning

While hydration is an important factor in various biological processes involving ligand binding, its explicit consideration has been limited due to the huge computational cost of the hydration structures around proteins. In the present study, we propose a deep-learning model for the fast computation of grid inhomogeneous solvation theory (GIST), a method for calculating the distribution function of the hydration energy and entropy [1]. The network architecture of our deep-learning model is essentially the same as that employed in the gr Predictor [2], enabling a fast and accurate computation of the water-oxygen (or water-hydrogen) site distribution function. Our deep-learning model, referred to as the “Deep GIST”, enables the computation of the distribution function of the hydration free energy for tens of seconds. The R2 score values for the proteins used for the test were between 0.75 and 0.83, indicating high prediction performance. We apply Deep GIST for the computation of the hydration free energy of a water molecule upon the replacement of the ligand, ΔG_Replace. It is found that the correlation coefficient of ΔG_Replace obtained using Deep GIST and that obtained using GIST is 0.78, indicating high prediction performance. It is also shown that Deep GIST can be used for discussing whether a water molecule at the ligand-binding site can be replaced or not upon ligand binding.

[1] Nguyen, C. N. et al., J. Chem. Phys. 2012, 137, 044101.

[2] Kawama, K. et al., J. Chem. Inf. Model. 2022, 62, 4460.