Applicability domain-expansion studies for machine learning models reveal new inhibitors of CYP2B6.

CYP2B6 is an important enzyme in the phase 1 metabolism of key pharmaceuticals, and inhibition of this enzyme can lead to adverse drug events. Machine learning models can potentially predict interactions with CYP2B6; however, there is limited data with which to train these models in the public domain. We proposed enhancing the applicability domain and improving the predictive capability of our CYP2B6 inhibition model by selecting a small, diverse set of compounds to test in vitro and adding the results to our model training set. We used a distance-based approach to define the applicability domain of the model and then measured the chemical diversity by creating t-distributed stochastic neighbor embedding plots to represent the chemical space of our model. After comparing this chemical space with a 49-plate drug-repurposing library, we were able to identify a plate with the highest average minimum Euclidean distance from the model training set. We then performed in vitro testing of this plate for CYP2B6 inhibition activity at 10 μM and added this new data to our machine learning model. A one-class classification approach was used to evaluate the efficacy of our applicability domain-expansion technique. The results showed that this method did not appreciably increase the performance of the model or the applicability domain, but we did increase the diversity of the training set. Additionally, the in vitro experiments identified vilanterol and allylestrenol as inhibitors of CYP2B6 with IC