Grid-Context Convolutional Model for Efficient Molecular Surface Construction from Point Clouds

Accurate and efficient molecular surface representation is essential in computational chemistry, impacting applications such as enzymology, rational drug design, and molecular recognition. Traditional approaches, including solvent-accessible surface (SAS) and solvent-excluded surface (SES), are widely used but often suffer from computational inefficiencies and limited adaptability to complex molecular structures. More recently, grid-independent machine learning methods have accelerated surface construction; however, these approaches rely solely on local single-grid information and fail to capture grid-context interactions, leading to inaccuracies in modeling complex interior surface regions. In this work, we introduce Con2SES, a novel framework that explicitly incorporates grid-context interactions to improve the accuracy and robustness of molecular surface construction. By leveraging learnable convolutional operators in both two-dimensional (2D) and three-dimensional (3D) space, Con2SES mitigates the limitations of grid-independent methods and ensures a more continuous and accurate surface construction. Extensive evaluations on protein molecules and diverse molecular collections demonstrate that Con2SES significantly outperforms grid-independent approaches in accuracy while maintaining computational efficiency comparable to that of the fastest grid-independent method, achieving approximately 99% accuracy and a 28-fold speedup compared to the classical SES implementation in the AMBER software package. Furthermore, Con2SES has been integrated into the AMBER platform and released as an open-source tool, providing a high-performance solution for large-scale molecular surface analysis.