The openOCHEM consensus model is the best-performing open-source predictive model in the First EUOS/SLAS joint...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in chemistry

Background:

• Accurate prediction of aqueous solubility is crucial for drug discovery and development.
• The EUOS/SLAS challenge provided a dataset of 100,000 compounds to spur algorithm development.
• Nephelometry assay was used to categorize compounds into low, medium, and high solubility.

Purpose of the Study:

• To describe the winning model developed for the EUOS/SLAS challenge on aqueous solubility prediction.
• To detail the methodology, including data selection, descriptor choice, and modeling strategy.
• To highlight the effectiveness of consensus modeling and advanced machine learning techniques.

Main Methods:

• Utilized the Online CHEmical database and Modeling environment (OCHEM) for model development.
• Employed a consensus approach combining 28 individual models.
• Incorporated both descriptor-based and representation learning methods, including Transformer Convolutional Neural Networks (CNNs).

Main Results:

• The consensus model significantly outperformed individual approaches and single-method consensus models.
• The Transformer CNN model yielded the highest individual prediction score, demonstrating the power of Natural Language Processing (NLP) methods.
• A combination of diverse models effectively reduced bias and variance, leading to superior performance.

Conclusions:

• Consensus modeling, particularly with diverse methods, is a highly effective strategy for accurate aqueous solubility prediction.
• Natural Language Processing (NLP) based models show significant promise in cheminformatics tasks.
• Incorporating aleatoric uncertainty estimation is recommended for future challenges to improve data interpretation.