Fragment-Based Prediction of Cytochromes P450 2D6 and 1A2 Inhibition by Recursive Partitioning

The evaluation of the ADME (absorption, distribution, metabolism, and excretion) properties of drug candidates is an important stage in drug discovery. To speed up the numerous tests carried out on large databases of compounds, the help of robust and accurate in silico filters is increasingly required. We propose here a method to build predictive and interpretable models for the prediction of cytochrome P450 (CYP) 1A2 and 2D6 inhibition using recursive partitioning (RP), a well-known technique for the construction of decision trees. The originality of the work is the use of several descriptions of the molecules in terms of fragments, i.e. the MACCS keys and five in-house fingerprints based on the electron density properties of fragments, employed to draw easily understandable structure-activity models. The classifiers reached performances of 87.5% and 76.5% of prediction on a validation set for CYP1A2 and CYP2D6, respectively. The analysis of the first nodes of the RP trees permits us to highlight some relations between the structural fragments and the inhibition of CYPs.