Iktos ADMET Models are global predictive models for ADMET values, trained on a compilation of large datasets from public sources and some specific client collaborations. Our ADMET models are designed to give a first insight on some important ADMET properties in your chemical space.
Iktos ADMET Models are available as a Makya add-on. If your software license does not include ADMET Models and you are interested in getting a demonstration, do not hesitate to contact your regional Application Scientists.
Which models are available?
The full list of available models is given here: Which ADMET Models are available?.
What type of models are Iktos ADMET Models?
There are two categories of models, the categories impacting the interpretability of predictions:
-
Regressor:
- Regressors are models that predict an interpretable value for each molecule (for example, a model that predicts the LogD)
-
Scorer:
- Scorers are models that predict a number correlated with the ADMET property. Let's consider the PPB Scorer for example: the higher the predicted score, the higher the PPB rate. But the predicted score is not a direct estimation of the PPB rate. This type of model does not return interpretable scores but is designed to rank a list of molecules, and select the top ones, with respect to an ADMET property.
- To help you select good molecules based on Scorer models's predictions, Makya tags the molecules as good / medium / bad, based on thresholds we estimated on real drug discovery projects (see more details in: Which ADMET Models are available?).
Illustration of the data transformation performed in Scorer models.
When to use the ADMET Models?
Our ADMET Models are meant to be used when no data points (or few data points, i.e., < 30) are available in your chemical space. In the early stages of your project, you might not have yet enough ADMET data to train your own QSAR models; in that case, Iktos ADMET Models can help you select and prioritize generated molecules.
If some ADMET data is available for your project, we recommend testing the ADMET Models on your data and comparing it with the experimental data (using performance metrics such as Spearman correlation coefficient, for example), so as to check if there is a strong correlation.
Tools that you can use to compare ADMET Models predictions and your data are:
- The Spearman correlation coefficient, which measures the ranking power (SciPy implementation; Excel implementation)
- The Matthews correlation coefficient (MCC), which measures the classification power, for Scorer models; you can use your TPP threshold for your data, and calculate the MCC on a range of possible thresholds on the ADMET predictions (Scikit-learn implementation; R example)
- The Mean Absolute Error, which measures the regression power, for Regressor models (Scikit-learn implementation; Excel example)
What is the applicability domain of Iktos ADMET Models?
Because Iktos ADMET models are trained on very diverse molecules, they have a large applicability domain and a good ranking power in diverse chemical spaces.
However, it is not possible to predict if the models will have a strong predictive power on a given and specific chemical space. Moreover, these Models are not expected to be the top performers on a specific chemical space. Models trained specifically on your project data (such as QSAR models) will generally perform better.
If you have your own project data, we highly recommend that you (1) check the accuracy of the ADMET Models on your data as described above, and (2) train your own models (using Makya’s QSAR module): these “local” models will likely be stronger on your chemical series.