Hybrid Approach to Protein-Protein Complex Affinity Prediction Based on Language Models and Molecular Dynamics.

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Protein-protein and protein-peptide interactions are vital for biological functions.
• Accurate prediction of binding affinity is essential for drug design and understanding mutations.
• Existing methods often struggle with static structures or lack multimodal integration.

Purpose of the Study:

• To develop HyBind-NN, a novel multimodal graph neural network for predicting protein-protein and protein-peptide binding affinity.
• To integrate protein language models (PLMs) with 3D structural and dynamic information for enhanced prediction accuracy.
• To outperform existing sequence-based and structure-based algorithms in binding affinity prediction.

Main Methods:

• Developed HyBind-NN, a multimodal graph neural network architecture.
• Integrated ESM-2 sequence embeddings with 3D Voronoi spatial geometry.
• Employed a multi-task learning framework using residue-level root mean square fluctuations (RMSF) from molecular dynamics (MD) for dynamic regularization.
• Benchmarked against state-of-the-art algorithms on diverse structural datasets.

Main Results:

• HyBind-NN achieved high accuracy in predicting binding affinity, with a mean absolute error of 0.89 for pK_D (1.12 kcal/mol for ∆G).
• The multimodal approach, combining PLMs with geometric and dynamic features, outperformed purely sequence-based or structure-based methods.
• Ablation studies confirmed the dominant contribution of PLMs, with geometric and dynamic regularization being crucial for capturing conformational changes.

Conclusions:

• HyBind-NN offers a robust and accurate method for predicting intermolecular binding affinity.
• The synergistic integration of PLMs with physics-aware architectures enhances predictive power.
• This approach holds significant potential for advancing drug design and biological mechanism studies.