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

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in drug discovery

Background:

• AlphaFold2 (AF2) significantly increases protein structure availability but poses challenges for binding-site prediction models trained on experimental data.
• Existing methods struggle with AF2 models due to differences from experimental structures (apo-like) and variable local reliability (pLDDT scores).

Purpose of the Study:

• To develop ProtCross, a confidence-aware domain adaptation framework for accurate residue-level binding-site prediction on AF2 models.
• To address the performance gap between models trained on experimental Protein Data Bank (PDB) data and predicted protein structures.

Main Methods:

• Proteins represented as residue point clouds, encoded with hierarchical PointNet++ architecture.
• ESM-C protein language model embeddings used for physicochemical and evolutionary information.
• Adversarial domain adaptation with gradient reversal layer, weighted by pLDDT to mitigate low-confidence regions.

Main Results:

• ProtCross demonstrated improved binding-site prediction performance on an AF2 test set compared to existing methods and a geometric baseline.
• Performance measured by area under the ROC curve and segmentation accuracy.
• pLDDT-guided weighting effectively reduced negative transfer common in standard domain-adversarial training.

Conclusions:

• ProtCross offers a robust solution for binding-site prediction on large-scale predicted protein structures from AF2.
• Confidence-aware domain adaptation is crucial for adapting models trained on high-quality experimental data to lower-confidence predicted structures.
• The framework enhances the utility of AF2 models for downstream applications like drug discovery.