DeepSES: Learning solvent-excluded surfaces via neural signed distance fields

The solvent-excluded surface (SES) is essential for revealing molecular shape and solvent accessibility in applications such as molecular modeling, drug discovery, and protein folding. Its signed distance field (SDF) delivers a continuous, differentiable surface representation that enables efficient rendering, analysis, and interaction in volumetric visualization frameworks. However, analytic methods that compute the SDF of the SES cannot run at interactive rates on large biomolecular complexes, and grid-based methods tend to result in significant approximation errors, depending on molecular size and grid resolution. We address these limitations with DeepSES, a neural inference pipeline that predicts the SES SDF directly from the computationally simpler van der Waals (vdW) SDF on a fixed high-resolution grid. By employing an adaptive volume-filtering scheme that directs processing only to visible regions near the molecular surface, DeepSES yields interactive frame rates irrespective of molecule size. By offering multiple network configurations, DeepSES enables practitioners to balance inference time against prediction accuracy. In benchmarks on molecules ranging from one thousand to nearly four million atoms, our fastest configuration achieves real-time frame rates with a sub-angstrom mean error, while our highest-accuracy variant sustains interactive performance and outperforms state-of-the-art methods in terms of surface quality. By replacing costly algorithmic solvers with selective neural prediction, DeepSES provides a scalable, high-resolution solution for interactive biomolecular visualization. • Grid-based signed distance field representation of solvent-excluded surfaces. • Neural inference from the van der Waals surface to the solvent-excluded surface. • Low prediction errors compared to other grid-based methods. • Real-time capable 3D U-Net architectures for signed distance field prediction.