Predicting drug solubility in binary solvent mixtures using graph convolutional networks: a comprehensive deep...

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Drug solubility prediction is crucial for pharmaceutical development but traditionally relies on time-consuming experimental methods.
• Accurate solubility prediction accelerates formulation design and reduces development costs.

Purpose of the Study:

• To evaluate Graph Convolutional Networks (GCNs) for predicting drug solubility in binary solvent mixtures across various temperatures.
• To assess the performance of GCNs compared to traditional machine learning methods for solubility prediction.

Main Methods:

• Utilized an extensive dataset of 27,000 solubility measurements for small molecules, solvents, and binary mixtures.
• Developed a GCN architecture with multi-head attention, hierarchical learning, and advanced pooling for molecular interaction analysis.
• Performed ablation studies and attention visualization to understand model behavior and identify key structure-solubility relationships.

Main Results:

• The GCN model achieved a mean absolute error (MAE) of 0.28 solubility units, outperforming traditional methods by 15%.
• GCNs demonstrated particular strength in modeling structure-solubility relationships for pharmaceutically relevant compounds.
• Prospective validation confirmed predictive reliability, with experimental verification yielding MAE < 0.5 solubility units for similar compounds.

Conclusions:

• GCNs are powerful tools for accelerating pharmaceutical formulation development, potentially reducing experimental needs by 60-80%.
• The study highlights the effectiveness of integrating graph neural networks for computational solubility prediction in binary solvent systems.
• Attention mechanisms in GCNs provide valuable, interpretable molecular insights for drug development.