ProtCross: Bridging the PDB-AlphaFold Gap for Binding Site Prediction with Protein Point Clouds

AlphaFold2 (AF2) has greatly increased the availability of predicted protein structures, but many binding-site prediction methods trained on experimentally determined Protein Data Bank (PDB) complexes perform less well when applied to AF2 models. This loss of accuracy reflects differences between ligand-bound experimental structures and predominantly apo-like predicted models, as well as nonuniform local structural reliability indicated by pLDDT scores. To address this, we introduce <b>ProtCross</b>, a confidence-aware domain adaptation framework for residue-level binding-site prediction on predicted protein structures. Proteins are represented as residue point clouds and encoded using a hierarchical PointNet++ architecture, with ESM-C protein language model embeddings providing physicochemical and evolutionary information. To transfer models trained on labeled PDB structures to unlabeled AF2 models, ProtCross employs adversarial domain adaptation with a gradient reversal layer, while weighting the domain-adversarial loss by pLDDT to reduce the influence of low-confidence regions. On an AF2 test set derived from the PDBbind v2020 Refined Set, ProtCross shows improved performance relative to existing binding-site predictors and an architecture-matched geometric baseline, as measured by area under the ROC curve and segmentation accuracy. Ablation analyses indicate that pLDDT-guided weighting mitigates negative transfer observed with standard domain-adversarial training.