Energy-based graph convolutional networks for scoring protein docking models.

Area of Science:

• Computational Biology
• Structural Biology
• Machine Learning

Background:

• Understanding protein-protein interactions is crucial for cell mechanisms and drug discovery.
• Protein docking computationally predicts these interactions but faces challenges in ranking models (scoring) and assessing their quality.
• Existing methods often struggle with both ranking accuracy and quality estimation.

Purpose of the Study:

• To develop a unified deep learning framework for protein docking that addresses both model ranking and quality assessment.
• To introduce a novel graph convolutional network capable of learning interaction energies directly from 3D structural data.

Main Methods:

• Representing protein structures as residue contact graphs with atom-resolution features.
• Proposing a novel graph convolutional kernel for aggregating node features to learn generalized interaction energies.
• Developing Energy-based Graph Convolutional Networks (EGCN) with multihead attention to predict energies, binding affinities, and quality measures (interface RMSD).

Main Results:

• EGCN significantly improved model ranking on a Critical Assessment of Predicted Interactions (CAPRI) homology docking test set.
• EGCN performance was comparable or slightly better than state-of-the-art methods on the Score_set benchmark.
• EGCN demonstrated a 27% improvement in quality assessment for the Score_set compared to previous methods.

Conclusions:

• EGCN is the first successful application of graph convolutional networks to protein docking, directly learning from 3D structure data.
• The framework offers a powerful approach for both scoring and quality assessment in protein docking.
• This advancement has implications for accelerating the discovery of novel therapeutics through improved protein interaction prediction.