Development of <i>in silico</i> predictive classification models for chemical-induced hepatocellular hypertrophy based on molecular descriptors

Abstract

Introduction: Because a lot of cost, time, and laboratory animals are required in the repeated dose toxicity tests, in silico prediction of toxicity is desired. Chemical-induced hepatocellular hypertrophy often influences the No-Observed-Adverse-Effect Level in repeated dose toxicity tests. Since hepatocellular hypertrophy is caused by various chemicals and the mechanisms are mostly unknown, it is necessary to develop a prediction method which does not require comprehensive understanding of mechanism. In this study, we developed predictive classification models of hepatocellular hypertrophy in rats using machine learning methods. Methods: Hepatocellular hypertrophy data was extracted from two reliable databases. One is our original database from risk assessment reports such as pesticides (ORAD: Original Risk Assessment reports Database). The other is the Hazard Evaluation Support System Integrated Platform (HESS). We constructed predictive classification models based on molecular descriptors calculated by Dragon 6 software using machine learning methods known as deep learning (DL), random forest (RF), and support vector machine (SVM), respectively. Model performance were evaluated using test chemicals independent from training chemicals used to build the model. Further, we also developed consensus models, which accepted matching results for more than two machine learning models (DL, RF, and SVM). Results: Among DL, RF, and SVM, the SVM model showed the highest performance with a prediction accuracy of 0.75 or more both in the ORAD dataset and the HESS dataset. Although the prediction accuracy was similar between the ORAD dataset and the HESS dataset in any model, the sensitivity was higher in the HESS dataset. The consensus model which is based on the combined three models could not increase the prediction accuracy or sensitivity from the SVM model in either the ORAD dataset or the HESS dataset. In the consensus model, the sensitivity was higher in the HESS dataset than the ORAD dataset as same as the SVM model. Conclusions:We have developed in silico predictive classification models of hepatocellular hypertrophy of diverse chemicals. The best model was SVM regardless the dataset. These our models could be faster and more cost-effective assessments tools for hepatocellular hypertrophy.