Variable selection using pair-correlation method. Environmental applications

Area of Science:

• Quantitative Structure-Activity Relationships (QSAR)
• Cheminformatics
• Statistical Modeling

Background:

• Selecting optimal descriptor variables is crucial for accurate QSAR models.
• Correlated descriptors can complicate model development and interpretation.
• Existing methods may not adequately address correlated descriptor selection.

Purpose of the Study:

• To introduce and evaluate the Pair-Correlation Method (PCM) for selecting between correlated descriptor variables in QSAR.
• To compare the performance of different statistical tests within the PCM framework.
• To establish methods for ordering multiple descriptor variables.

Main Methods:

• Development and adaptation of the Pair-Correlation Method (PCM).
• Utilization of a 2x2 contingency table for data ordering.
• Application and comparison of statistical tests: Conditional Fisher's exact test (CE), McNemar's test (MN), Chi-square test, and Williams' t-test (Wt).
• Implementation of three ordering methods for multiple variables: simple ordering, ordering by win-loss differences, and probability-weighted ordering.

Main Results:

• PCM effectively discriminates between correlated descriptor variables.
• The tested statistical methods (CE, MN, Chi-square, Wt) show varying sensitivities in descriptor selection.
• The developed ordering methods allow for robust ranking of multiple descriptors.
• PCM demonstrated utility in case studies involving flavone inhibition, chlorobenzene toxicity, and aromatic amine mutagenicity.

Conclusions:

• The Pair-Correlation Method (PCM) provides a systematic approach for selecting and ordering correlated descriptor variables in QSAR.
• PCM enhances the reliability of QSAR model development by identifying superior descriptors.
• The method is applicable to diverse chemical and biological datasets, improving predictive accuracy.