Novel Solubility Prediction Models: Molecular Fingerprints and Physicochemical Features vs Graph Convolutional Neural...

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Drug discovery and development

Background:

• Accurate prediction of molecular solubility is essential for chemical process design and drug development.
• Traditional methods for solubility prediction are often time-consuming and resource-intensive.
• Developing robust predictive models remains a significant challenge in cheminformatics.

Purpose of the Study:

• To develop and evaluate machine learning and deep learning models for accurate solubility prediction.
• To compare the performance of molecular fingerprint-based methods with graph convolutional network (GCN) approaches.
• To assess the utility of physicochemical properties and graph representations for solubility prediction.

Main Methods:

• Utilized machine learning and deep learning techniques for solubility prediction.
• Employed two primary methods: (1) molecular fingerprints augmented with physicochemical properties, and (2) graph convolutional networks (GCNs) for graph-based molecular representation.
• Conducted both regression (solubility value) and classification (solubility class) tasks for each method.

Main Results:

• The molecular fingerprint-based method demonstrated high performance when incorporating key physicochemical descriptors.
• The GCN method proved effective in predicting chemical compound properties using simplified graph representations.
• Both methods provided reliable predictions for solubility values and classes.

Conclusions:

• The study successfully developed effective surrogated model-based methods for predicting molecular solubility.
• Incorporating physicochemical properties with molecular fingerprints enhances prediction accuracy.
• GCN models offer a powerful approach for predicting chemical properties from graph representations, aiding in the selection of suitable solutes and solvents.