3. ADMET Modeler™

The ADMET Modeler™ module ("Modeler") is the part of Predictor used to construct structure-property and structure-activity models from experimental data. It makes use of molecular descriptors generated by ADMET Predictor as well as any external descriptors you supply to build a predictive model, and then lets you export the model back to ADMET Predictor for incorporation as a new predicted property.

ADMET Modeler automates each of the following steps necessary to build high quality predictive models:

Filtering descriptors to eliminate those that are underrepresented, those with very small variance, and those that are highly correlated with other descriptors.
Selecting the method by which data will be partitioned between the test set and the training pool. Doing so may involve clustering compounds based on descriptor values to identify similar structures and provide for optimal sampling.
Ranking descriptors by their effect on the predicted property (sensitivity analysis).
Training multiple sets of models (ensembles) to allow selection of the most appropriate model architecture.
Selecting the best ensemble to use as the final predictive model.

The mechanics of building is covered in 3.2: Running ADMET Modeler, whereas detailed discussion of the theory behind various steps in the process is presented in the Modeler Theory section.

Definitions of the descriptors used in Modeler are provided in the ADMET Properties chapter on theory and details.