Benchmarking Variable Selection in QSAR.

Area of Science:

• * Cheminformatics
• * Computational Chemistry
• * Machine Learning in Drug Discovery

Background:

• * Variable selection enhances QSAR model interpretability, performance, and computational efficiency.
• * Optimal variable selection strategies for QSAR remain largely uncharacterized.
• * Random forest models are widely used in QSAR but require careful variable selection.

Purpose of the Study:

• * To benchmark the performance of eight variable selection methods in QSAR modeling.
• * To identify methods that improve predictive accuracy and transparency of random forest models.
• * To assess the impact of variable selection on computational cost in QSAR.

Main Methods:

• * Conducted 1728 benchmarking experiments using seven diverse QSAR datasets.
• * Evaluated eight distinct variable selection techniques.
• * Utilized random forest as the modeling algorithm.
• * Assessed model performance, transparency, and variable reduction efficiency.

Main Results:

• * Univariate variable selection methods demonstrated suboptimal performance in QSAR.
• * Multivariate adaptive regression splines (MARS) and forward selection significantly reduced dataset dimensionality.
• * Approximately 60% variable reduction was achieved without compromising predictive performance.
• * Enhanced model transparency and reduced computational costs were observed with effective methods.

Conclusions:

• * MARS and forward selection are recommended for variable selection in QSAR modeling.
• * Significant variable reduction is feasible, improving QSAR model efficiency.
• * Careful method selection is key to optimizing QSAR model performance and interpretability.