Three machine learning models for the 2019 Solubility Challenge.

Area of Science:

• Computational chemistry
• Machine learning
• Predictive modeling

Background:

• The 2019 Solubility Challenge aimed to advance accurate prediction of molecular solubility.
• Machine learning models, particularly tree-based classifiers, are increasingly used in cheminformatics.

Purpose of the Study:

• To describe and evaluate three machine learning models for predicting intrinsic aqueous solubility.
• To compare the performance of Random Forest, Extra Trees, and a consensus model (Vox Machinarum).
• To assess the impact of data variance on model performance.

Main Methods:

• Development of three tree-like classification models: Random Forest, Extra Trees, and a consensus model (Vox Machinarum) using Bagging.
• Evaluation of models on two test sets from the 2019 Solubility Challenge: one with low variance (100 molecules) and one with high variance (32 molecules).
• Comparison of predicted solubilities with gold standard and literature-based solubility values.

Main Results:

• Extra Trees performed best on the low-variance solubility dataset.
• Vox Machinarum and Random Forest showed slightly superior performance on the high-variance dataset compared to Extra Trees.
• The consensus model, Vox Machinarum, demonstrated the benefits of the 'Wisdom of Crowds' approach.

Conclusions:

• Tree-based machine learning models, including ensemble methods, are effective for predicting molecular solubility.
• Model performance can vary depending on the variance within the dataset.
• Consensus models offer a robust approach to solubility prediction.