Coping with Unbalanced Class Data Sets in Oral Absorption Models

Abstract

Class imbalance occurs frequently in drug discovery data sets. In oral absorption data sets, in the literature, there are considerably more highly absorbed compounds compared to poorly absorbed compounds. This produces models that are biased toward highly absorbed compounds which lack generalization to industry settings where more early stage drug candidates are poorly absorbed. This paper presents two strategies to cope with unbalanced class data sets: undersampling the majority high absorption class and misclassification costs using classification decision trees. The published data set by Hou et al. [J. Chem. Inf Model. 2007, 47, 208-218], which contained percentage human intestinal absorption of 645 drug and drug-like compounds, was used for the development and validation of classification trees using classification and regression tree (C&RT) analysis. The results indicate that undersampling the majority class, highly absorbed compounds, leads to a balanced distribution (50:50) training set which can achieve better accuracies for poorly absorbed compounds, whereas the biased training set achieved higher accuracies for highly absorbed compounds. The use of misclassification costs resulted in improved class predictions, when applied to reduce false positives or false negatives. Moreover, it was shown that the classical overall accuracy measure used in many publications is particularly misleading in the case of unbalanced data sets and more appropriate measures presented here may be used for a more realistic assessment of the classification models' performance. Thus, these strategies offer improvements to cope with unbalanced class data sets to obtain classification models applicable in industry.