A Linear Prediction Rule Based on Ensemble Classifiers for Non-Genotoxic Carcinogenicity

Abstract

This article presents a novel approach for developing a linear prediction rule to predict the non-genotoxic carcinogenicity potential of a new compound in the drug development pipeline. We construct the approach using data from 24-hour microarray experiments on rats treated with the compound. This method was developed to address an actual problem that we were presented with by scientists in mechanistic toxicology. Short-term, preclinical assays for non-genotoxic carcinogenicity, a toxicity commonly observed in long-term rodent carcinogenicity studies, have proven difficult to develop. A quick and early preclinical assay, such as this, is of particular interest and urgency. The linear prediction rule is derived using an Ensemble Linear Discriminant classifier. This is a hybrid approach which leverages the advantages of ensemble approaches for addressing over-fitting (a problem endemic to microarray data), and that of LDA for interpretability. In a limited comparison with some other classifiers, including random forest, we show that our approach has good predictive performance in addition to being interpretable.

Keywords:

• Classification
• Gene signature
• Linear discriminant analysis (LDA)
• Microarray
• Toxicogenomics

Acknowledgments

The authors would like to thank James Colaianne and Peter Lord for their support of this work.