Enhancing Predictions of Drug Solubility Through Multidimensional Structural Characterization Exploitation

Area of Science:

• Computational chemistry
• Drug discovery
• Molecular modeling

Background:

• Solubility is a critical physical property influencing small-molecule drug development.
• Experimental solubility determination is resource-intensive, requiring specialized equipment and controlled conditions.
• Accurate computational prediction of drug solubility is a long-standing research objective.

Purpose of the Study:

• To introduce MSCSol, a novel computational model for predicting drug solubility.
• To leverage multidimensional molecular structure information for enhanced prediction accuracy.
• To develop a more efficient and accessible method for assessing drug solubility.

Main Methods:

• Integration of a graph neural network with geometric vector perceptrons (GVP-GNN) to encode 3D molecular structures.
• Application of Selective Kernel Convolution with Global and Local attention mechanisms for multi-scale feature extraction.
• Utilization of various molecular descriptors and data augmentation strategies for 2D and 3D structural data.

Main Results:

• MSCSol demonstrated superior performance on benchmark and independent datasets.
• Ablation studies validated the effectiveness of individual model components.
• Interpretability analysis identified key atomic groups and substructures influencing solubility.

Conclusions:

• MSCSol offers a powerful and accurate computational approach to drug solubility prediction.
• The model effectively captures complex molecular structural information and higher-order knowledge.
• MSCSol has the potential to streamline the drug development process by reducing experimental costs and time.