Implicit Solvation Improvements

GNN-Based Implicit Solvent (GNNIS) Model

The accuracy of MD simulations for studying biological systems depends also on the description of the water environment around the molecules. Until today the most accurate approach involves the calculation of each water molecule explicitly, which is associated with considerable computational effort, while widely applied models using an implicit solvent description cannot account well for local solvent effects. To address this issue, we developed the concept of the graph neural networt (GNN)-based implicit-solvent model GNNIS. Keeping the physical basis of a generalized Born (GB) implicit-solvent model, the GNN is used to correct for the local effects of solvation by scaling the effective Born radii to reproduce average explicit-solvent forces. In a first proof-of-concept [1], we evaluated the accuracy and transferability of the model by varying the central residue of a series of pentapeptides. Encouraged by the findings, a general GNNIS water model for organic molecules was developed [2]. Using average explicit-solvent forces for nine conformers of approx. 370’000 diverse organic molecules, the resulting GNNIS model reproduced the free-energy landscapes of unseen molecules. Next, we extended GNNIS to 39 organic solvents (including water) using a multi-task model [3]. Based on this, we devised a pipeline starting from a fast in silico conformer generator followed by minimization involving GNNIS to provides rapid access to Boltzmann-weighted solvent-dependent conformational ensembles. These ensembles agreed with conformational preferences in different solvent measured in NMR experiments.

[1] external page Katzberger et al.,J. Chem. Phys. (2023), 158,204101.
[2] external page Katzberger et al., Chem. Sci. (2024), 15, 10794.
[3] external page Katzberger et al., J. Am. Chem. Sci. (2025), 147 ,13264.

While the GNNIS model was developed for classical force fields, we found that the learned correction (i.e., the difference between the underlying standard implicit-solvent model and explicit solvation) can be transferred to QM calculations carried out with the same or similar implicit-solvent model [4]. Despite the simplicity and assumptions of this approach, QM-GNNIS outperformed the conventional implicit-solvent models for QM when compared against NMR and IR data.

[4] external page Katzberger et al., J. Chem. Theory Comput. (2025), 21, 7450.