7.7. Building a DELTA Model™

7.7. Building a DELTA Model™

In this tutorial, we use rat liver microsomal (RLM) clearance for a particular data series as an example of how to build a DELTA Model. The procedure can be applied to any ADMET property model that you wish to augment with new data or which you want to extend to a closely related endpoint - e.g., where a somewhat different assay protocol was used. The example described here uses ANNE regression models throughout. It has worked well in our hands to date but the technique may be more broadly applicable.

You will be guided through the process of building a model for such in-house experimental CLint data using a modest number of structurally related compounds to build both a Local_RLM_CLint model and a Delta_CLint model. We will then compare the model's predictive performance to each other as well as to that of a global GLB_RLM_CLint model. Global RLM CLint predictions are provided as part of the supporting data for the tutorial so it can be run without having a license for the Metabolism Module.

The local CLint data is taken from Gomtsyan et al. (2016), who reported rat and human liver microsomal (RLM and HLM, respectively) intrinsic clearance data for 51 selective antagonists of the Transient Receptor Potential Vanilloid 3 (TRPV3) receptor. The tabulated CLint values were corrected for nonspecific microsomal binding using experimental binding data, just as the CYP_RLM_CLint and CYP_HLM_CLint predictions in ADMET Predictor are. 

You do need to have an ADMET Modeler license to run this tutorial, and it is recommended that you work through the other Modeler tutorials provided before running this one.

The tutorial is divided into five parts:

1. Loading the local data.

2. Inspecting the predictions made by the global model for the local data.

3. Building a local ANNE model that just makes use of the local data.

4. Building a DELTA Model that adjusts the global predictions to take local deviations into account.

5. Comparing the predictive performance of the three models.

QSAR models are generally built on property data for molecular structures that have already been made and characterized in hopes of being able to predict the properties of compounds that have not been made or tested yet - new molecules that may or may not be structurally similar to those in the training set. Here we will build a model using 31 randomly selected compounds and reserve 20 compounds as a test set to assess the predictive power of GLB_RLM_CLint and the newly constructed models.