SurfSol: A Multimodal Surface-Based Deep Learning Framework for Protein Solubility Prediction

Protein solubility is a critical determinant for the application of recombinant proteins in biotechnology and pharmaceuticals. Although experimental techniques have advanced, accurate prediction of solubility remains a persistent challenge. This work introduces SurfSol, a novel computational approach that predicts protein solubility by leveraging explicit surface representations. The method integrates surface geometry and physicochemical properties, such as electrostatics, hydropathy, and hydrogen-bonding potential, within an E(3)-equivariant graph neural network, and combines them with sequence embeddings from ESM-2 and structural features extracted via a TransformerConv architecture. Evaluated on the processed eSOL data set, SurfSol achieves an <i>R</i>² of 0.555 and an AUC of 0.895, outperforming existing predictors. Ablation studies confirm the complementary contributions of the surface, sequence, and structural modalities. SurfSol demonstrates the importance of explicit surface modeling for solubility prediction and provides a generalizable framework for other protein property prediction tasks.