A differentiable representation of solvent-solute interface

The success of an implicit solvent model hinges on the accurate representation of the solvent-solute interface. However, current models primarily depend on specific geometric criteria, making the calculation of interface derivatives an arduous task.

In response to this challenge, we have introduced an innovative solution using a Transformer-like neural network to depict the solvent-solute interface. This network, trained on an extensive dataset of 1251 all-atom solvated protein models, exhibits robust generalization capabilities, showing excellent performance even with protein structures absent from the training set.

An additional advantage of our approach lies in the auto-differentiable nature of the neural network model. It enables the efficient computation of hydration energy estimations and the precise calculation of forces, both critical components for understanding biophysical interactions within protein processes.

Furthermore, this differentiable representation also stands as a promising candidate for capturing the variation in relative permittivity, potentially offering a more precise implementation of the Poisson-Boltzmann equation on researching the protein's hydration. This innovation underscores the potential of our model as a comprehensive tool for advancing our understanding of protein biophysics.