Temperature-Dependent Small-Molecule Solubility Prediction Using MoE-Enhanced Directed Message Passing Neural...

Area of Science:

• Computational Chemistry
• Machine Learning in Materials Science
• Drug Discovery Informatics

Background:

• Accurate solubility prediction is vital for pharmaceutical development and materials science.
• Existing models often lack generalizability across varied solvents and temperatures.
• Challenges include capturing complex molecular interactions and structural nuances.

Purpose of the Study:

• To develop a novel, generalizable solubility prediction model, DMPNN-MoE.
• To enhance the prediction of temperature-dependent solubility for diverse chemical systems.
• To provide an interpretable model for understanding key molecular drivers of solubility.

Main Methods:

• Integration of a directed message passing neural network (DMPNN) for feature extraction.
• Utilization of a mixture-of-experts (MoE) algorithm for adaptive learning.
• Training and validation on a large dataset of 56,945 experimental solubility values.

Main Results:

• DMPNN-MoE achieved superior performance (MAE = 0.256, R² = 0.863) via 10-fold cross-validation.
• Outperformed existing graph-based and descriptor-based models by up to 35%.
• Demonstrated robust generalization to unseen rare solvents and solutes with high accuracy.

Conclusions:

• The DMPNN-MoE model offers a significant advancement in solubility prediction accuracy and generalizability.
• The model provides insights into critical molecular features influencing solubility, such as ring structures.
• This interpretable model has broad applicability in pharmaceutical discovery and chemical engineering.