Random forest models to predict aqueous solubility

Area of Science:

• Computational chemistry
• Cheminformatics
• Quantitative Structure-Property Relationship (QSPR) modeling

Background:

• Accurate prediction of aqueous solubility is crucial for drug discovery and environmental fate assessment.
• Existing QSPR models have limitations in accuracy and interpretability.

Purpose of the Study:

• To develop and compare QSPR models for predicting aqueous solubility using various machine learning algorithms.
• To identify the most accurate and robust model for solubility prediction.

Main Methods:

• Employed Random Forest (RF) regression, Partial Least Squares (PLS) regression, Support Vector Machines (SVM), and Artificial Neural Networks (ANN).
• Utilized experimental data for 988 organic molecules to train and validate models.
• Performed external validation on a set of 330 molecules.

Main Results:

• The RF model demonstrated superior predictive accuracy for aqueous solubility compared to PLS, SVM, and ANN.
• RF provided automatic descriptor selection, importance assessment, and in-parallel evaluation of predictive ability.
• External test set prediction yielded r² = 0.89 and RMSE = 0.69 log S units for log molar solubility.

Conclusions:

• Random Forest regression is a highly effective method for QSPR modeling of aqueous solubility.
• The RF model's features enhance model interpretability and reliability.
• The developed model performs competitively with existing literature methods and covers relevant chemical space.