Protein-ligand binding affinity prediction model based on graph attention network.

Area of Science:

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

Background:

• Accurate estimation of protein-ligand binding affinity is crucial for structure-based drug design.
• Machine learning models offer potential to enhance classical scoring functions for binding affinity prediction.

Purpose of the Study:

• To develop a novel graph attention network (GAT) based model, GAT-Score, for predicting protein-ligand binding affinity.
• To improve upon existing machine learning methods for binding affinity estimation in drug design.

Main Methods:

• Representing protein-ligand complexes as graph structures where atoms are nodes.
• Implementing a dynamic feature mechanism to incorporate bond information.
• Introducing a virtual super node for aggregating node-level features into graph-level representations for regression tasks.
• Training the model on the PDBbind database v.2018.

Main Results:

• GAT-Score demonstrated superior performance compared to traditional machine learning models using molecular descriptors.
• The model effectively handles graph-level regression problems for binding affinity prediction.
• Validation using the $C_s$ (Core set) and Cross-Validation (CV) schemes confirmed the model's efficacy.

Conclusions:

• The GAT-Score model offers a promising advancement in predicting protein-ligand binding affinity.
• The proposed architectural improvements enhance the capability of graph attention networks for molecular modeling.
• This method has significant implications for accelerating structure-based drug design and discovery.