P03-03
A deep learning model for predicting chemical-induced rat hepatocellular necrosis using transcriptome data.
Kouki MAEBARA *1, Kyoko ONDO2, Tomoaki TOCHITANI2, Toru USUI2, Izuru MIYAWAKI2, Kaori AMBE1
1Department of Regulatory Science, Graduate School of Pharmaceutical Sciences, Nagoya City University
2Preclinical Research Unit, Sumitomo Pharma Co., Ltd. Kiyoshi HASEGAWA *, Yuya SEKI, Yu LIU, Yukiyo ITO
Division of Informatics Promotion, TECHNOPRO R&D company
( * E-mail: c202050@ed.nagoya-cu.ac.jp )
Open TG-GATEs is a toxicity database which includes transcriptome data from animal experiments on chemicals [1]. Recently, transcriptome data is expected to be utilized for toxicity evaluation of chemicals including pharmaceuticals. However, since transcriptome data contains a large amount of information, appropriate data pre-processing is required to utilize it. The DeepInsight method is capable of converting high-dimensional data into images [2], and it could be used for processing transcriptome data. In this study, we tried to construct a deep learning model to predict chemicals that cause hepatocellular necrosis in rats, an important histopathological finding when evaluating hepatotoxicity, from imaged liver transcriptome data.
This study used animal study data and transcriptome data published on Open TG-GATEs. Chemicals that showed histopathological findings related to hepatocellular necrosis in a 28-day repeated-dose rat study were designated as positive, and chemicals that did not show hepatocellular necrosis were designated as negative. The expression data of 31,099 genes observed in the liver of rats in the single-dose study of these chemicals were imaged using the DeepInsight method and used as explanatory variables in the prediction model. A pre-trained convolutional neural network (CNN) model was constructed to determine whether or not each compound induced hepatocellular necrosis in rats.
For model construction, 127 chemicals (25 positive, 102 negative) were randomly split 4:1 into training data and validation data, and their prediction performance was evaluated using the hold-out method. After five trials with different splitting patterns in the validation data, the mean and standard deviation of the evaluation indices ROC-AUC, f1 score, sensitivity, and specificity were 0.752 (0.052), 0.531 (0.069), 0.800 (0.112), and 0.705 (0.084), respectively. These results suggest that the liver transcriptome data obtained from a single-dose study in rats might predict hepatocellular necrosis induced in a 28-day repeated-dose study.
[1] https://dbarchive.biosciencedbc.jp/en/open-tggates/desc.html
[2] Sharma, et al., Sci Rep., 9, 11399 (2019).