Protein Docking Model Evaluation by Graph Neural Networks.

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Protein complex structures are vital for understanding cellular functions.
• Experimental methods for determining these structures are resource-intensive.
• Computational approaches are essential for predicting protein complex structures.

Purpose of the Study:

• To develop an accurate computational method for evaluating protein docking models.
• To improve the identification of near-native structures from large sets of predicted models.

Main Methods:

• Developed Graph Neural Network-based DOcking decoy eValuation scorE (GNN-DOVE), a deep learning approach.
• Represented protein interfaces as graphs, using atomic properties and distances as node and edge features.
• Trained and validated GNN-DOVE on diverse protein docking datasets, including Dockground, ZDOCK, and CAPRI.

Main Results:

• GNN-DOVE demonstrated superior performance in evaluating protein docking models compared to existing methods.
• The graph neural network approach effectively captured interface characteristics for accurate scoring.
• Outperformed previous convolutional neural network-based methods like DOVE.

Conclusions:

• GNN-DOVE offers a powerful new tool for computational protein structure prediction.
• Deep learning applied to graph representations of protein interfaces enhances decoy evaluation.
• This method can accelerate the study of molecular mechanisms involving protein complexes.