MSc Computational Chemistry ยท Cheminformatics
QSAR Model Validation Lab
Evaluate QSAR model quality with external prediction, cross-validation, Y-randomization, leverage, and residual analysis.
Theory
QSAR relates molecular descriptors to biological activity. Robust validation checks predictive performance, chance correlation, and applicability domain.
Procedure
- Split data into training and external test sets.
- Fit the descriptor model.
- Check cross-validation and external prediction.
- Inspect residuals and leverage warnings.
Observations
- Training fit: Training R2 is high at 0.91. R2 measures fitted variance on compounds used for calibration. Good fit is necessary but does not prove predictive power.
- Cross-validation: Leave-one-out Q2 remains above 0.7. Predicted residual sums remain substantially below total variance. Internal validation supports moderate robustness.
- External test set: External test compounds are predicted with R2pred = 0.68. External residuals are larger than training residuals but still structured around zero. The model has usable but not excellent predictive performance.
- Leverage outlier: One compound has high leverage and a large standardized residual. Williams plot places it outside the applicability domain. Do not trust this prediction without additional similar training data.
Interpretation rule
Accept a QSAR model only when internal and external validation agree and predictions stay inside the applicability domain.
Validation
Y-randomization and external test performance help detect chance correlations and overfitting.
This page runs an interactive simulation of the experiment; enable JavaScript to use it.