Refinement of Boosting-ERDA for ordered categorical data and its application to QSAR analysis of multi-drug resistance agents

Abstract

The attrition of new drugs in development due to toxicity remains high. Therefore, it becomes more important that prediction of toxicity at an early stage for avoiding attrition during drug development. For rough toxicity prediction, it is efficient to use quantitative structure-activity relationship for ordered categorical data. However it is known that analyzing ordered categorical data in accurately is difficult with existing methods because they treat their objective parameters as nominal instead of ordered. We have already reported the development of the Boosting-ERDA (Evolved Regression Disciminant Analysis) method, which was developed by combining the ERDA method with AdaBoost. The ERDA method is used to generate models for ordered categorical data and AdaBoost is an ensemble learning. Our method has both accuracy and generalization capability at the same time. In this study, we have made an improvement in the Boosting-ERDA method to acquire information by calculating coefficients of explanatory variable. To evaluate performance of our new method, we have applied the Boosting-ERDA method for QSAR analysis of multi-drug resistance reversal agents using a data set of 609 compounds.