Evaluation of QSAR Equations for Virtual Screening.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Quantitative Structure Activity Relationship (QSAR) models correlate molecular structure with activity.
• Traditional QSAR metrics (e.g., R2, QF1/F2/F32) are context-ignorant and may not reflect performance in virtual screening (VS).
• Virtual screening requires models optimized for identifying active compounds within large datasets.

Purpose of the Study:

• To propose and validate a virtual screening-aware metric for QSAR model development.
• To demonstrate the limitations of classical QSAR metrics in predicting VS performance.
• To introduce a new algorithm optimizing an enrichment-based metric for QSAR model derivation.

Main Methods:

• Developed 21 Multiple Linear Regression (MLR) models using classical metrics for seven targets.
• Tested MLR models on datasets mimicking virtual screening campaigns.
• Developed a new algorithm optimizing an enrichment-based metric for MLR model derivation and compared it with classical MLR, Random Forest (RF), and Support Vector Machine (SVM) models.

Main Results:

• Classical QSAR metrics showed no correlation with virtual screening success (enrichment).
• Models optimized with an enrichment-based metric consistently outperformed classical MLR models in virtual screening tests.
• The new algorithm-derived models outperformed RF and SVM models in most virtual screening tests, attributed to better handling of inactive compounds.

Conclusions:

• Virtual screening-aware metrics are crucial for developing effective QSAR models for VS.
• Optimizing for enrichment provides a promising strategy for QSAR model derivation in classification and virtual screening.
• The developed Enrichment Optimizer Algorithm (EOA) offers improved performance in virtual screening tasks.