Transparency in Modeling through Careful Application of OECD's QSAR/QSPR Principles via a Curated Water Solubility...

Area of Science:

• * Computational Chemistry
• * Cheminformatics
• * Environmental Science

Background:

• * Increasing data size and machine learning accessibility necessitate robust validation for Quantitative Structure-Activity/Property Models (QSAR/QSPR).
• * Regulatory bodies like the US Environmental Protection Agency require stringent scrutiny of QSAR/QSPR models for environmental risk assessment.
• * The Organisation for Economic Co-operation and Development (OECD) provides established principles for validating structure-activity models.

Purpose of the Study:

• * To apply OECD validation principles to a machine learning-based QSAR/QSPR model.
• * To demonstrate the application of these principles to a random forest regression model for predicting water solubility.
• * To ensure QSAR/QSPR models derived using advanced methods are suitable for regulatory consideration.

Main Methods:

• * Assembly and curation of a dataset of 10,200 unique chemical structures with water solubility measurements from public sources.
• * Development of a water solubility prediction model using random forest regression.
• * Systematic application and discussion of OECD QSAR/QSPR validation principles to the developed model, including descriptor selection guidance.

Main Results:

• * A QSAR/QSPR model for predicting water solubility was developed using random forest regression.
• * The model achieved a 5-fold cross-validated performance of 0.81 R² and 0.98 RMSE.
• * Performance was comparable to previously published models, demonstrating the efficacy of the applied validation principles.

Conclusions:

• * Explicitly leveraging and modernizing OECD principles is crucial for developing reliable QSAR/QSPR models with machine learning.
• * Rigorous validation ensures QSAR/QSPR models meet regulatory standards for environmental assessment.
• * This work encourages a dialogue on integrating advanced machine learning with established validation frameworks for regulatory science.