Evaluation of Deep Learning Architectures for Aqueous Solubility Prediction

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Materials science

Background:

• Accurate prediction of aqueous solubility is essential for pharmaceutical development, environmental science, and energy storage.
• Existing solubility prediction models face accuracy challenges across diverse molecular structures.
• Deep learning approaches offer potential for improved solubility prediction accuracy.

Purpose of the Study:

• To evaluate deep learning methods for predicting molecular aqueous solubility.
• To develop a generalizable model for predicting the solubility of various organic molecules.
• To analyze the influence of data characteristics, molecular representations, and model architectures on predictive performance.

Main Methods:

• Utilized the largest available solubility dataset for training and validation.
• Implemented deep learning models including fully connected neural networks, recurrent neural networks, graph neural networks (GNNs), and SchNet.
• Explored molecular representations: molecular descriptors, simplified molecular-input line-entry system (SMILES) strings, molecular graphs, and 3D atomic coordinates.

Main Results:

• Models employing molecular descriptors demonstrated the highest predictive accuracy.
• Graph neural networks (GNNs) also exhibited strong performance in solubility prediction.
• Error and feature analyses identified key molecular properties and structural information influencing model accuracy.

Conclusions:

• Molecular descriptors combined with deep learning provide a robust approach for accurate aqueous solubility prediction.
• Further research can optimize GNNs and explore transfer learning for enhanced performance with limited data.
• This work advances the development of reliable computational tools for solubility assessment.