Enhancing protein-ligand binding affinity prediction through sequential fusion of graph and convolutional neural...

Area of Science:

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

Background:

• Accurate prediction of protein-ligand binding affinity is essential for structure-based drug discovery.
• Deep learning models show promise but can be further optimized for enhanced predictive power.

Purpose of the Study:

• To develop and evaluate a novel fusion model integrating Graph Neural Networks (GNNs) and Convolutional Neural Networks (CNNs) for predicting protein-ligand binding affinity.
• To assess the model's performance, generalization ability, and utility in virtual screening applications.

Main Methods:

• A sequential fusion model combining GNN and CNN architectures was developed.
• Intermediate GNN outputs were concatenated with CNN input features to capture atomic chemical environments.
• Model performance was evaluated on the CASF-2016 benchmark and a virtual screening task for PI5P4Kα.

Main Results:

• The GNN-CNN fusion model demonstrated improved prediction accuracy compared to standalone CNN models on the CASF-2016 benchmark.
• Generalization ability was confirmed through rigorous testing with varying similarity thresholds.
• Masking experiments indicated the model effectively identifies critical interaction regions.
• The fusion model significantly enhanced virtual screening performance for the PI5P4Kα target.

Conclusions:

• The proposed GNN-CNN fusion strategy offers a novel and effective approach to enhance deep learning model accuracy for protein-ligand binding affinity prediction.
• This method holds potential for advancing structure-based drug discovery and virtual screening applications.