MGraphDTA: deep multiscale graph neural network for explainable drug-target binding affinity prediction.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Accurate drug-target affinity (DTA) prediction accelerates drug discovery.
• Graph neural networks (GNNs) are used for DTA prediction, but shallow models struggle with global compound structures.
• Existing GNN interpretability methods, like graph attention, lack comprehensive global relationship insights.

Purpose of the Study:

• To develop a deep multiscale graph neural network (MGraphDTA) for enhanced DTA prediction.
• To improve the capture of both local and global compound structural information.
• To introduce a novel, chemically intuitive visual explanation method (Grad-AAM) for deep learning models in DTA prediction.

Main Methods:

• Proposed MGraphDTA, a super-deep GNN with 27 graph convolutional layers and dense connections.
• Integrated chemical intuition into the GNN architecture for simultaneous local and global structure analysis.
• Developed Grad-AAM for visual analysis of DTA prediction models from a chemical perspective.

Main Results:

• MGraphDTA significantly outperformed state-of-the-art deep learning models across seven benchmark datasets.
• The Grad-AAM method generated explanations consistent with pharmacological knowledge.
• The approach demonstrated improved generalization and interpretation capabilities in DTA prediction.

Conclusions:

• The proposed MGraphDTA model offers superior performance in drug-target affinity prediction.
• Grad-AAM provides valuable chemical insights, aiding in understanding model predictions beyond human perception.
• This integrated approach enhances both the predictive power and interpretability of computational drug discovery models.