Graph Neural Networks Model Based on Atomic Hybridization for Predicting Drug Targets

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Traditional quantitative structure-activity relationship (QSAR) models struggle to integrate local structural patterns and global physicochemical properties for accurate bioactivity prediction.
• Predicting half-maximal inhibitory concentration (IC50) values is crucial for drug discovery but remains challenging.

Purpose of the Study:

• To develop a hybrid deep learning framework combining graph neural networks (GNNs) and molecular descriptors for improved IC50 prediction.
• To enhance the accuracy and interpretability of structure-activity relationship (SAR) modeling.

Main Methods:

• Developed a hybrid deep learning framework integrating GNNs with explicit molecular descriptors.
• Incorporated atomic/bond features from molecular graphs and interpretable physicochemical properties/fingerprints.
• Trained and validated the model on 14,316 compounds across nine diverse biological targets.

Main Results:

• Achieved a high test R-squared (R²) of 0.87, outperforming existing methods by 6-42%.
• Demonstrated robust generalization with comparable training and test performance.
• Maintained partial interpretability via descriptor contributions and attention mechanisms.

Conclusions:

• The hybrid framework synergistically combines data-driven learning with domain knowledge for superior SAR modeling.
• Offers improved accuracy and interpretability, facilitating efficient compound prioritization in early drug discovery.
• Represents a significant advancement in computational approaches for accelerating drug development.