SAR based on self consistent classifier.

Area of Science:

• * Cheminformatics
• * Computational chemistry
• * Predictive modeling

Background:

• * Quantitative structure-activity relationship (QSAR) model performance relies heavily on training data quality and descriptor selection.
• * Publicly available datasets often present challenges like diverse chemical classes and imbalanced activity ratios.
• * Current descriptor selection methods for QSAR models may lack robust mathematical justification, impacting model stability.

Purpose of the Study:

• * To introduce a novel self-consistent classifier approach for improved QSAR analysis.
• * To enhance classification capabilities through Logistic (SCLC) and Extreme (SCEC) extensions of self-consistent regression (SCR).
• * To address issues of model stability and generalizability in QSAR by refining feature selection.

Main Methods:

• * Implementation of Self-Consistent Logistic Classification (SCLC) and Self-Consistent Extreme Classification (SCEC) models.
• * Application of the developed models to predict HIV-1 activity and toxicity endpoints.
• * Performance evaluation using fivefold cross-validation and comparison with original SCR and support vector machine (SVM).

Main Results:

• * The proposed SCLC and SCEC models demonstrated comparable accuracy to existing SCR and SVM methods.
• * The new approach offers superior feature selection, leading to more generalizable QSAR models.
• * Crucial factors influencing activity and toxicity endpoints were unambiguously identified.

Conclusions:

• * The developed self-consistent classifier extensions (SCLC, SCEC) provide a robust framework for QSAR modeling.
• * Enhanced feature selection capabilities contribute to the development of more stable and reliable predictive models.
• * This methodology facilitates a clearer understanding of the key molecular descriptors driving biological activity and toxicity.