Multivariate linear QSPR/QSAR models: Rigorous evaluation of variable selection for PLS.

Area of Science:

• Computational Chemistry
• Chemoinformatics
• Quantitative Structure-Property Relationships (QSPR)

Background:

• Quantitative structure-property relationship (QSPR) models are crucial for predicting chemical compound properties.
• Chemometric methods, particularly Partial Least Squares (PLS) regression, offer powerful tools for developing empirical regression models.
• Robust model evaluation is essential for reliable QSPR predictions.

Purpose of the Study:

• To describe basic chemometric methods for QSPR model development from a user's perspective.
• To demonstrate the application of PLS regression and variable selection for predicting gas chromatographic retention indices.
• To emphasize the importance of rigorous model performance evaluation using repeated double cross-validation (rdCV).

Main Methods:

• Utilized Partial Least Squares (PLS) regression for building empirical QSPR models.
• Employed stepwise variable selection to reduce a large set of molecular descriptors (2688) to a smaller, informative subset (22).
• Applied repeated double cross-validation (rdCV) for careful and cautious evaluation of PLS model performance.

Main Results:

• Developed QSPR models to predict gas chromatographic retention indices for 209 polycyclic aromatic compounds (PACs).
• Achieved the most favorable models using descriptors from 3D structures including all hydrogen atoms.
• The final QSPR model demonstrated a typical prediction error of ±12 units for the retention index on test set objects.

Conclusions:

• Chemometric methods, especially PLS regression combined with effective variable selection and rdCV, provide reliable QSPR models.
• The developed QSPR models accurately predict gas chromatographic retention indices for PACs.
• Open-source R software environment programs and data are provided for reproducibility and further research.