SS-GNN: A Simple-Structured Graph Neural Network for Affinity Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• Accurate drug-target binding affinity (DTBA) prediction is vital for efficient drug screening but is computationally demanding.
• Graph neural networks (GNNs) offer powerful representation capabilities for complex molecular interactions.

Purpose of the Study:

• To develop a computationally efficient and accurate GNN-based model for predicting DTBA.
• To simplify the graph representation of protein-ligand interactions for reduced computational cost.

Main Methods:

• Proposed SS-GNN, a simple-structured GNN model utilizing a single undirected graph representation with a distance threshold.
• Ignored covalent bonds in protein representation to further decrease computational expense.
• Employed independent GNN-MLP modules for atom and edge feature extraction, alongside edge-based atom-pair feature aggregation and graph pooling.

Main Results:

• Achieved state-of-the-art prediction performance with a model of only 0.6 million parameters.
• Obtained Pearson's R = 0.853 on the PDBbind v2016 core set, surpassing existing GNN methods by 5.2%.
• Demonstrated high prediction efficiency, with affinity prediction for a typical complex taking only 0.2 ms.

Conclusions:

• SS-GNN provides a simplified yet highly effective approach for DTBA prediction.
• The model's efficiency and accuracy make it a valuable tool for accelerating drug screening processes.
• The open-source availability of the code facilitates further research and application.