Robust cross-validation of linear regression QSAR models.

Area of Science:

• Computational chemistry
• Cheminformatics
• Quantitative structure-activity relationship (QSAR) modeling

Background:

• QSAR models predict biochemical activity from molecular structures.
• Blind prediction is the gold standard for QSAR model validation.
• Cross-validation (CV) is commonly used to estimate predictive power during QSAR model development.

Purpose of the Study:

• To examine predictive power and fitting ability of multiple linear regression (MLR) QSAR models within a CV context, accounting for outliers.
• To assess commonly used predictive and fitting statistics using Monte Carlo cross-validation.
• To evaluate QSAR data sets including human intestinal absorption, blood-brain partition coefficient, and toxicity values.

Main Methods:

• Monte Carlo cross-validation was employed to assess statistical methods.
• Analysis included saxitoxin QSAR data sets and benchmark data sets with known outliers.
• Robust statistics were specifically evaluated for their performance in the presence of outliers.

Main Results:

• Robust MLR is consistently superior to ordinary-least-squares MLR, irrespective of outlier presence.
• Model predictive power must be assessed using robust statistical measures.
• Outlier contamination does not diminish the advantage of robust methods.

Conclusions:

• Robust statistical methods are crucial for reliable QSAR model validation.
• The choice of statistical approach significantly impacts the assessment of predictive power.
• Software and source code are available for academic use to reproduce results and apply methods to new data.