Mechanistic interpretation of artificial neural network‐based QSAR model for prediction of cathepsin K inhibition potency

Abstract

Cathepsin K (Cat K) is a lysosomal cysteine protease that plays an important role in many bone diseases including osteoporosis, which makes Cat K an interesting drug target. Several compounds, including balicatib, were tested in clinical trials for their potency as Cat K inhibitors. Balicatib was found as a potent inhibitor; however, the side effects caused by the lysosomotropic drug candidate behavior as a consequence of its basic nature prevented further drug development. Here, we present a counter‐propagation artificial neural network (CP‐ANN) quantitative structure–activity relationship model of benzamide‐containing aminonitriles with good predictive ability. The quality of all models developed was evaluated internally by leave‐one‐out cross‐validation (LOOCV) on the training set and externally by an independent validation set. The best model performed with the LOOCV and external validation squared correlation coefficient of 0.81 and 0.84, respectively. In order to interpret the selection of variables and consequently discuss the mechanism of inhibition, the layer of the CP‐ANN model representing the distribution of individual molecular descriptors was compared with the output layer representing the response surface. The measures indicating the overlap/similarity of the response surface with selected levels in the input layer were introduced. The results signify not only the importance of the covalent bonding parameters, which are responsible for the S1 binding pocket of the enzyme, but also the impact of 3D shapes of the molecules on the inhibitor–enzyme interactions implying the stabilization of the inhibitors poses within the S2 and S3 binding sites. Copyright © 2014 John Wiley & Sons, Ltd.