Development of a Predictive Model for Chemically Induced Rat Liver Cell Necrosis Using Visualized Transcriptome Data
Kouki MAEBARA *1, Kei KINOSHITA3, Kyoko ONDO3, Tomoaki TOCHITANI3, Toru USUI3, Izuru MIYAWAKI3, Kaori AMBE1, 2, Hiroshi ARAKAWA1
1Department of Regulatory Science, Graduate School of Pharmaceutical Sciences, Nagoya City University
2Graduate School of Data Science, Nagoya City University
3Preclinical Research Unit, Sumitomo Pharma Co., Ltd.
【Background】The toxicity database Open TG-GATEs contains rat toxicity test results for chemicals, including transcriptome data. This transcriptome data is expected to be useful for toxicity evaluation of chemicals such as pharmaceuticals, and various analytical approaches have been proposed. However, the enormous amount of data is an issue, and no standardized method for its utilization has been established. Recently, visualization methods of high-dimensional data have gained attention in data analysis, due to its compatibility with deep learning, enabling efficient analysis of necessary information from high-dimensional data.
【Purpose】The aim of this study was to develop a deep learning model to predict chemicals that induce hepatic necrosis, a representative type of liver damage, in rats, using transcriptome data from rat livers that were converted into low-dimensional data for model construction by visualization using the DeepInsight method.
【Methods】This study utilized data publicly available on Open TG-GATEs. Chemicals that showed pathological findings associated with hepatocyte necrosis in 28-day repeated dose studies in rats were classified as positive, while those that did not were classified as negative. Additionally, the expression levels of 31,099 genes in the livers of rats in single-dose studies for these chemicals were visualized using the DeepInsight method and used as explanatory variables for the prediction model. Using these data and a convolutional neural network (CNN) model, a model was constructed to distinguish whether each chemical induces hepatocyte necrosis in rats, and the model was evaluated using external data. Subsequently, class activation map (CAM) was applied to interpret the model and extract the genes that contributed to the prediction.
【Results】The 126 substances (24 positive and 102 negative) that could be used for model construction were divided into training and validation datasets, and the predictive performance was evaluated using the cross-validation method. In the validation data, the average evaluation metrics were ROC-AUC 0.896, f1-score 0.733, sensitivity 0.96, and specificity 0.832. These values were superior to the model constructed without visualization. In addition, all positive compounds were predicted as positive in the external data. In conclusion, it was suggested to be possible to predict liver cell necrosis induced in 28-day repeated-dose studies in rats by visualization of transcriptome data obtained from single-dose rat studies and applying it to a deep learning model. This demonstrates the potential for a new approach that could serve as an alternative to long-term toxicity testing and improve prediction accuracy.