The effect of leverage and/or influential on structure-activity relationships

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are crucial for drug discovery and development.
• Assessing model reliability and identifying influential data points are essential for robust QSAR development.
• Leverage (Hat matrix, h(i)) and influence (Cook's distance, D(i)) are key metrics for evaluating QSAR model components.

Purpose of the Study:

• To evaluate the impact of leverage and influential compounds on QSAR model reliability.
• To compare the performance of QSAR models after removing high-leverage or influential compounds.
• To determine the optimal strategy for identifying and removing problematic compounds in QSAR modeling.

Main Methods:

• Analysis of seven QSAR datasets from previously published studies.
• Development of three models per dataset: full-model, h(i)-model (leverage-based removal), and D(i)-model (influence-based removal).
• Application of statistical validation criteria and comparison of model performance metrics.

Main Results:

• Correlation coefficients improved in 5 out of 7 datasets after removing compounds with high h(i) or D(i).
• Removal of 1-13 influential compounds (D(i)-models) or 2-4 leverage compounds (h(i)-models) was performed.
• D(i)-models demonstrated systematically better agreement compared to full-models and h(i)-models.

Conclusions:

• Removing influential compounds significantly enhances QSAR model performance and reliability.
• Cook's distance analysis, combined with Hat matrix analysis, is recommended for identifying compounds for removal.
• This approach is crucial for developing robust and dependable Quantitative Structure-Activity Relationship models.