DynamicGT: A dynamic-aware geometric transformer model to predict protein-binding interfaces in flexible and...

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning for protein structure analysis

Background:

• Protein-protein interactions are crucial for cellular functions.
• Current deep learning models for binding site prediction use static protein structures, limiting accuracy for flexible or disordered regions.
• Accurate prediction of binding sites is essential for understanding biological processes and drug discovery.

Purpose of the Study:

• To develop a dynamic-aware deep learning model for improved protein binding site prediction.
• To address the limitations of static structure-based approaches by incorporating conformational dynamics.
• To enhance the prediction accuracy for disordered, transient, and unbound protein structures.

Main Methods:

• Introduction of Dynamic Geometric Transformer (DynamicGT), a dynamic-aware model.
• Integration of conformational dynamics using a cooperative graph neural network (Co-GNN) and a Geometric Transformer (GT).
• Encoding of dynamic features at node (atom) and edge (interaction) levels, considering bound and unbound states.
• Dynamic regulation of message passing between core and surface residues.

Main Results:

• DynamicGT was trained on a 1-ms molecular dynamics simulation dataset and augmented with AlphaFlow-generated conformations.
• Extensive benchmarking demonstrated significant improvement in prediction accuracy for flexible protein regions.
• The model showed superior performance on diverse datasets including disordered, transient, and unbound structures.
• DynamicGT requires substantially less data compared to leading static approaches.

Conclusions:

• Incorporating conformational dynamics into a cooperative architecture significantly enhances protein binding site prediction accuracy, especially for flexible regions.
• DynamicGT offers a more robust and generalizable approach compared to static models.
• This dynamic-aware method holds promise for advancing drug discovery and understanding complex biological interactions.