An efficient piecewise linear model for predicting activity of caspase-3 inhibitors

Background and purpose of the study: Multimodal distribution of descriptors makes it more difficult to fit a single global model to model the entire data set in quantitative structure activity relationship (QSAR) studies. Methods: The linear (Multiple linear regression; MLR), non-linear (Artificial neural network; ANN), and an approach based on “Extended Classifier System in Function approximation” (XCSF) were applied herein to model the biological activity of 658 caspase-3 inhibitors. Results: Various kinds of molecular descriptors were calculated to represent the molecular structures of the compounds. The original data set was partitioned into the training and test sets by the K-means classification method. Prediction error on the test data set indicated that the XCSF as a local model estimates caspase-3 inhibition activity, better than the global models such as MLR and ANN. The atom-centered fragment type CR2X2, electronegativity, polarizability, and atomic radius and also the lipophilicity of the molecule, were the main independent factors contributing to the caspase-3 inhibition activity. Conclusions: The results of this study may be exploited for further design of novel caspase-3 inhibitors