Optibrium Community

Downloads

HLM Stability Models

Dec 17, 2012

HLM Stability Models

Dec 17, 2012

Alexey Zakharov of the National Cancer Institute, National Institutes of Health has developed models of stability in Human liver Microsomes (HLM). Alexey presented details of these models and their validation at the American Chemical Society 2012 Fall National Meeting in Philadelphia and his slides are available in an article on this community site. The methods have also been published in Future Med. Chem. (2012) 4(15), pp. 1933–1944. Alexey has kindly shared three of the models built with StarDrop so that all StarDrop users can access them free-of-charge.

There are three models available, built using the decision tree, random forest and radial basis function (RBF) methods. These classify the in vitro half-life of compounds in HLM,  as ‘stable’ or ‘unstable’, based on a cut-off of 15 minutes.

The two classification models and the associated detailed Auto-Modeller output including descriptors and validation results for each model can be downloaded from the links in the table below:

Validation Set

Model

Accuracy

k

Model file

Details

Decision Tree

0.80

0.45

DT_T_Half_Life.aim

Model_Details_DT.pdf

Random Forest

0.84

0.57

RF_T_Half_Life.aim

Model_Details_RF.pdf

Based on these results, we would suggest that the decision tree model is not sufficiently robust, but that the random forest model appears to be more reliable. Unfortunately, the random forest model was not included in the original study, so data on the performance on the external sets are not available.

The final model is a RBF model that outputs a continuous value which, in turn, can be used to generate a classification. This model produced the best results of the three StarDrop models on the three external sets studied with, accuracies ranging from 75% to 87%. The model placed second overall of the models built by Alexey using various methods and described in his presentation. The model file and detailed output from the Auto-Modeller can be downloaded from the following links:

RBF_T_Half_Life.aim

Model_Details_RBF.pdf

To use the models within StarDrop, download and save these files in a convenient place. Load them into StarDrop using the Load Model button on the Models tab. Alternatively, the directory in which the model files have been saved can be added to the paths from which models are automatically loaded when StarDrop starts by selecting the File->Preference menu option and adding the directory under Models in the File Locations tab.

In order to use the RBF model to classify compounds, Alexey suggests that a cut-off of 0.5 in the output of the model should be used. This can easily be achieved in StarDrop by using the mathematical function tool Function Editor and entering the following function:

if({RBF_T_Half_Life}>0.5, ‘stable’, ‘unstable’)

This will report the class ‘stable’ for those compounds predicted to have a half-life in HLM of >15 minutes and ‘unstable’ for those predicted to have a half-life of less than or equal to 15 minutes.

There is an important caveat to the use of this model. We have noted that many drug-like compounds lie outside of the domain of applicability of this model and therefore the reported uncertainty is “inf” (i.e. infinite). For these compounds, the probabilities of each class will be equal and the reported class will be ‘stable’. When looking at the results of this model, we recommend that you view the statistics by selecting the statistics button Statistics Button from the toolbar.