PSMa: Learning Protein Surface Representations with Physicochemical Masked Autoencoders

Protein modeling has largely centered on sequence representations, leaving the geometry and physicochemical context of molecular surfaces underused. We introduce PSMa (Protein Surface Masked autoencoder), a self-supervised framework that operates directly on molecular-surface point clouds with geometric and physicochemical attributes, reconstructing masked surface patches to learn transferable representations at scale from ∼220,000 PDB structures. On held-out proteins, PSMa outperforms sequence-only and structure-only baselines on two downstream tasks─(i) protein-protein interface (PPI) binding-site identification and (ii) protein thermal-stability prediction─matching state-of-the-art AUROC and AUPRC while substantially improving threshold-dependent classification performance. The learned representations generalize across protein-length ranges and across experimental as well as predicted structures, with decision thresholds closer to the neutral prior; these findings show that surface-centric pretraining provides a strong, generalizable foundation for structure-aware protein machine learning.