Descriptor Selection Improvements for Quantitative Structure-Activity Relationships.

Area of Science:

• * Cheminformatics and computational chemistry.
• * Development of predictive modeling techniques.

Background:

• * Quantitative structure-activity relationship (QSAR) models require careful molecular descriptor selection to enhance predictive performance.
• * Existing QSAR models often suffer from redundant, noisy, or irrelevant descriptors, leading to suboptimal results and potential overfitting.
• * Efficient descriptor selection is crucial for building robust and accurate QSAR models.

Purpose of the Study:

• * To propose a novel descriptor selection framework, termed self-paced learning (SPL) via sparse logistic regression (LR) with Logsum penalty (SPL-Logsum).
• * To simultaneously identify simple and complex samples adaptively and prevent model overfitting.
• * To enhance QSAR model performance by selecting a minimal yet significant subset of molecular descriptors.

Main Methods:

• * Implementation of a self-paced learning (SPL) strategy, mimicking gradual learning from simple to complex samples.
• * Utilization of sparse logistic regression (LR) with a Logsum penalty to promote the selection of a parsimonious set of descriptors.
• * Validation of the proposed SPL-Logsum framework through simulations and application to three public QSAR datasets.

Main Results:

• * The SPL-Logsum framework demonstrated superior performance compared to existing sparse methods.
• * Key performance metrics including area under the curve, sensitivity, specificity, and accuracy were significantly improved.
• * The method effectively identified and utilized significant molecular descriptors while mitigating the impact of irrelevant ones.

Conclusions:

• * The proposed SPL-Logsum framework offers an effective approach for molecular descriptor selection in QSAR modeling.
• * This method enhances predictive accuracy and model robustness by adaptively handling sample complexity and descriptor sparsity.
• * SPL-Logsum provides a valuable tool for advancing the field of cheminformatics and predictive modeling.