A dual-branch graph neural network architecture for drug-target binding affinity prediction

Area of Science:

• Cheminformatics
• Computational Drug Discovery
• Artificial Intelligence in Medicine

Background:

• Traditional drug discovery is time-consuming and resource-intensive.
• Computational strategies, including Graph Neural Networks (GNNs), offer accelerated target identification and candidate prioritization.
• Integrating FDA-approved drug libraries enhances computational approaches.

Purpose of the Study:

• Introduce a novel dual-branch GNN architecture for enhanced molecular representation learning.
• Improve accuracy and robustness in drug candidate screening and prioritization.
• Establish a new benchmark for cheminformatics tasks.

Main Methods:

• Developed a dual-branch GNN combining Graph Convolutional Neural Networks (GCN), GraphSage, and Jumping-Knowledge modules.
• Jointly encoded molecular graph structural topology and functional attributes for enriched embeddings.
• Evaluated the model against 45 state-of-the-art baselines on Davis and KIBA datasets.

Main Results:

• The proposed GNN model demonstrated quantitative improvements over the GCN model.
• Achieved a reduction in Mean Squared Error (MSE) from 35.24 to 33.98.
• Showcased superior performance with higher Pearson (76.49 vs. 76.19) and Concordance indices (85.41 vs. 84.41).

Conclusions:

• The dual-branch GNN offers superior accuracy and robustness for molecular candidate screening.
• The model establishes a new reference point for cheminformatics tasks.
• A COVID-19 drug repurposing case study identified potential antiviral drugs, including Imunovir and Remdesivir.