Advancements in Ligand-Based Virtual Screening through the Synergistic Integration of Graph Neural Networks and...

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Ligand-based virtual screening is crucial for drug discovery.
• Graph neural networks (GNNs) are increasingly used in cheminformatics.
• Traditional chemical descriptors provide valuable molecular information.

Purpose of the Study:

• To evaluate the impact of integrating traditional chemical descriptors with various GNN architectures for ligand-based virtual screening.
• To compare the performance of different GNNs (GCN, SchNet, SphereNet) when combined with descriptors.
• To assess the robustness of descriptors and GNN-descriptor models in scaffold-split scenarios.

Main Methods:

• Implemented and evaluated GCN, SchNet, and SphereNet models.
• Integrated traditional chemical descriptors with each GNN architecture.
• Performed virtual screening experiments using benchmark datasets.
• Analyzed model performance, particularly in scaffold-split validation settings.

Main Results:

• The integration of chemical descriptors significantly improved performance for GCN and SchNet, but only marginally for SphereNet.
• All evaluated GNNs achieved comparable performance when augmented with descriptors.
• Expert-crafted descriptors alone demonstrated strong performance, even outperforming combined GNN-descriptor models in scaffold-split tests.
• The effectiveness of descriptor integration varied across different GNN architectures.

Conclusions:

• Sophisticated GNN architectures are not always necessary; simpler GNNs enhanced with descriptors can achieve similar efficacy.
• Expert-crafted descriptors are robust and highly effective, especially in challenging scaffold-split scenarios.
• Future GNN research should focus on developing models adept at handling scaffold diversity for real-world drug discovery applications.