Multi-label classification and label dependence in in silico toxicity prediction

Abstract

Most computational predictive models are specifically trained for a single toxicity endpoint and lack the ability to learn dependencies between endpoints, such as those targeting similar biological pathways. In this study, we compare the performance of 3 multi-label classification (MLC) models, namely Classifier Chains (CC), Label Powersets (LP) and Stacking (SBR), against independent classifiers (Binary Relevance) on Tox21 challenge data. Also, we develop a novel label dependence measure that shows full range of values, even at low prior probabilities, for the purpose of data-driven label partitioning.Using Logistic Regression as the base classifier and random label partitioning (k = 3), CC show statistically significant improvements in model performance using Hamming and multi-label accuracy scores (p<0.05), while SBR show significant improvements in multi-label accuracy scores. The weights in the Logistic Regression and Stacking models are positively associated with label dependencies, suggesting that learning label dependence is a key contributor to improving model performance.An original quantitative measure of label dependency is combined with the Louvain community detection method to learn label partitioning using a data-driven process. The resulting MLCs with learned label partitioning were generally found to be non-inferior to their corresponding random or no label partitioning counterparts. Additionally, using the Random Forest classifier in a 10-fold stratified cross validation Stacking model, we find that the top-performing stacking model out-performs the corresponding base model in 11 out of 12 Tox21 labels. Taken together, these results suggest that MLC models could potentially boost the performance of current single-endpoint predictive models and that label partitioning learning may be used in place of random label partitionings.