-
From the home page, click Models.
-
From the left panel, click a model that you created.
A page opens with Quality statistics for the
model.
-
Review the statistics for the model:
| Model Types | Description |
|---|
| Bayesian Classification, Random Forest Classification, and Random Forest
Regression |
Values marked with “test” are from k-fold cross-validation—the remaining values
are for the complete training set. |
| Bayesian Classification, Pharmacophore Classification Ensemble, and Random
Forest Classification |
ROC AUC values close to 1 are good, and close to 0.5 are bad. Few off-diagonal
elements in a confusion matrix are good. Many off-diagonal elements are bad. |
| Pharmacophore Classification Ensemble |
Pharmacophore models are treated like the other classification models. However,
because they take longer to train, GTD uses one-pass test set validation rather than k-fold cross-validation to assess the
model quality:
- Input data are randomly divided into training and test sets, with 75% of data in
training, and 25% in test.
- A pharmacophore model is built from the training data and evaluated against the
test data, computing ROC curves, ROC AUC values, and confusion matrices.
- The model is rebuilt with 100% of the input data and saved in the GTD system.
|
| Random Forest Regression |
Rsquared and Qsquared values
closer to 1 are good, and closer to 0 are bad. The smaller the root-mean-square error
(RMSE) value the better, especially relative to the null
model value. If the test RMSE is close to the null model RMSE, that is a sign of a
poor model. Qsquared (Q2) is the Rsquared value that you get from applying the model
to the test set instead of the training set. For GTD, the test set is the data that was held-out from cross-validation. |
|