Quantitative structure–activity relationships of dihydrofolatereductase inhibitors

Abstract

Pneumocystis carinii is a nonpathogenic fungus found in respiratory tract of healthy humans. It is a major cause of death in AIDS patients. P. carinii dihydrofolatereductase inhibitors such as trimetrexate are currently used to treat P. carinii pneumonia. These drugs have several adverse side effects and, therefore, there is a critical need for the identification of novel dihydrofolatereductase inhibitors. In this work, quantitative structure – activity relationship models were derived for 395 stucturally diverse P. carinii dihydrofolatereductase inhibitors. A wide variety of molecular descriptors belonging to various structural properties were calculated for each molecule. Both linear (multiple linear regression) and non-linear (generalized regression neural network) models were developed to link the calculated descriptors as independent variables to their reported biological activity as dependent variable. Different variable transformation was performed on the both dependent and independent variables to obtain better multiple linear regression models. At first glance, it seemed applying variable transformation on the dependent variable was necessary to improve quality of the quantitative structure–activity relationships models because it made the model more predictive. Upon further inspection, the results were not necessarily surprising. In the other words, back-transforming predictions did not give unbiased estimations of pIC50. At the final trial, generalized neural network was employed to develop predictive quantitative structure–activity relationships model.