Gene Expression Data Based Deep Learning Model for Accurate

Article pubs.acs.org/jcim

Cite This: J. Chem. Inf. Model. XXXX, XXX, XXX−XXX

Gene Expression Data Based Deep Learning Model for Accurate Prediction of Drug-Induced Liver Injury in Advance Chunlai Feng,*,# Hengwei Chen,# Xianqin Yuan, Mengqiu Sun, Kexin Chu, Hanqin Liu, and Mengjie Rui* Department of Pharmaceutics, School of Pharmacy, Jiangsu University, Zhenjiang 212013, PR China

Downloaded via NOTTINGHAM TRENT UNIV on July 19, 2019 at 17:47:51 (UTC). See https://pubs.acs.org/sharingguidelines for options on how to legitimately share published articles.

S Supporting Information *

ABSTRACT: Drug-induced liver injury (DILI), one of the most common adverse effects, leads to drug development failure or withdrawal from the market in most cases, showing an emerging challenge that is to accurately predict DILI in the early stage. Recently, the vast amount of gene expression data provides us valuable information for distinguishing DILI on a genomic scale. Moreover, the deep learning algorithm is a powerful strategy to automatically learn important features from raw and noisy data and shows great success in the field of medical diagnosis. In this study, a gene expression data based deep learning model was developed to predict DILI in advance by using gene expression data associated with DILI collected from ArrayExpress and then optimized by feature gene selection and parameters optimization. In addition, the previous machine learning algorithm support vector machine (SVM) was also used to construct another prediction model based on the same data sets, comparing the model performance with the optimal DL model. Finally, the evaluation test using 198 randomly selected samples showed that the optimal DL model achieved 97.1% accuracy, 97.4% sensitivity, 96.8% specificity, 0.942 matthews correlation coefficient, and 0.989 area under the ROC curve, while the performance of SVM model only reached 88.9% accuracy, 78.8% sensitivity, 99.0% specificity, 0.794 matthews correlation coefficient, and 0.901 area under the ROC curve. Furthermore, external data sets verification and animal experiments were conducted to assess the optimal DL model performance. Finally, the predicted results of the optimal DL model were almost consistent with experiment results. These results indicated that our gene expression data based deep learning model could systematically and accurately predict DILI in advance. It could be a useful tool to provide safety information for drug discovery and clinical rational drug use in early stage and become an important part of drug safety assessment.

■

INTRODUCTION The liver represents a critical player and therapeutic target in a variety of diseases, because of its major metabolic and excretory functions, central role in biotransformation, and anatomic and physiologic structure.1,2 Therefore, liver would suffer from many commonly used drugs that are capable of evoking some degree of hepatotoxicity.3,4 As an uncommon idiosyncratic side effect,5 severe hepatotoxicity could not be obviously detected during preclinical phases or early clinical phases of drug development.6 Even highly regulated and rigorous testing practices still fail in making clear the toxicological properties, especially for late-onset drug-induced liver injury (DILI).7 Consequently, DILI remains a serious issue in clinic treatment, and has also been a common cause of drug development failure or withdrawal from the market. Thus, reliable early prediction and accurate characterization of DILI are urgently needed by clinicians, the pharmaceutical industries and regulators.8 Up to now, various approaches have been applied to predict DILI in vitro, using some hepatocyte-like cell lines.9,10 It is showed that various cell lines were successfully applied for the prediction of several adverse drug reactions, however, in vitro © XXXX American Chemical Society

cell lines cannot completely reflect the sophisticated traits of human DILI.11 Also, compared with stem cells, hepatocyte-like cells lack the ability to represent the full mature hepatocellular phenotype. Thus, more complex multicellular models of the liver should be developed for the DILI risk assessment, which is still challenging. In another aspect, in silico approaches have been applied into DILI risk prediction, for example, pharmacophore models.12−14 On basis of approved mechanism of toxicity, the related signaling pathways were regarded as crucial targets and the effective ligands of targets were searched to construct binding modes. Then the binding modes as the pharmacophore models were applied to screen a series of compounds with special chemical structures. Obviously, this approach heavily relies on approved mechanism of toxicity; however, most of mechanisms underlying human DILI are complicated and still undefined. Thus, the lack of mechanism information limits the further development of pharmacophore models for DILI prediction. On the other hand, in silico prediction models that combined machine learning with Received: February 16, 2019 Published: June 12, 2019 A

DOI: 10.1021/acs.jcim.9b00143 J. Chem. Inf. Model. XXXX, XXX, XXX−XXX

Journal of Chemical Information and Modeling

Article

■

MATERIALS AND METHODS Data Collection. The gene expression data corresponding to diverse categories of hepatotoxicity of 143 compounds measured in vivo based on GeneChip Rat Genome 230 2.0 which is known to be a good proxy to predict hepatotoxicity in humans (TG-GATEs data set)23,29 were carefully retrieved from the ArrayExpress database.47 The TG-GATEs data set included gene expression data of rat liver at 4, 8, 15, and 29 days treated by various dosages (low, middle and high dose) of compounds. After the removal of the missing data and unclear pathologic results of experimental subjects treated by some compounds in the downloaded data set, 988 samples associated with 87 compounds were selected to construct prediction models finally (Supporting Information, Table S1). Furthermore, according to the lesions range of rats’ liver, the liver injury induced by drugs was classified into five classes: severe (lesion range of 76∼100%), moderately severe (lesion range of 51∼75%), moderate (lesion range of 26∼50%), slight (lesion range of 1∼25%), and minimal (lesion range of 0∼1%). In this study, if one drug could induce liver injury with lesion range of 1∼100%, this drug was labeled as DILI positive. Accordingly, the samples related to DILI positive drugs were defined as DILI positive samples. Then the remaining samples and control groups were labeled as DILI negative samples. Finally, 496 DILI positive samples and 492 DILI negative samples were obtained. Data Preprocessing. The data preprocessing was conducted based on the platform R v3.5.0.48 First, the downloaded raw gene expression data (CEL files) were converted into standard gene expression profile by oligopackage.49 Since invalid and missing values among raw data may provide wrong information for model training and also lead to unreliable predicted results, these values were replaced and filled using the impute 1.52.0 package from Bioconductor Library.50 Then the refilled data sets were normalized using the Robust Multichip Average (RMA) algorithm from the oligopackage in order to avoid inaccurate results induced by maxima and minima as well as to save computing cost. Next, each probe ID was matched with its corresponding gene symbol, according to the annotation file of gene chip. Then the mean expression value of the multiple probe IDs matched an official gene symbol was calculated to represent the expression intensity. Finally, 15406 genes with expression value were remained for each sample. Model Construction. 80% of positive samples and 80% of negative samples were randomly selected as training set and the remaining 20% of the samples were used as testing set to develop prediction models. As a result, 397 DILI positive samples and 393 DILI negative samples made up the training set of 790 samples, 99 DILI positive samples and 99 DILI negative samples constituted the testing set of 198 samples (Table 1). Then deep learning prediction model and Support Vector Machine (SVM) prediction model were constructed and optimized respectively based on the training set and

structural properties of compounds have also been greatly applied into DILI risk assessment.15−18 Nevertheless, the prediction models completely based on molecular descriptors did not take genetic and environmental factors into consideration,6 and compounds with diverse structures could not well describe the complex biology processes,19 especially the mechanism of toxicity. Recently, as the rapid development of toxicogenomics, it has been identified that gene expression level could be altered either directly or indirectly in response to exposure to a toxic compound.20−22 To be specific, toxic compounds would bind to target proteins to interfere their functions, and as a result, these compound-bound proteins might have the ability to influence coding genes of other proteins that are related to liver activity, resulting in human DILI.23 Therefore, toxicogenomics data could provide valuable information for toxicity prediction.7,24,25 Compared with the prediction model based on chemical structures, toxicogenomics data could systematically describe the effect induced by compounds at gene expression level.26−28 On the other hand, DILI prediction based on toxicogenomics data could apply anomalies at gene expression level to predict drug adverse effect in later stages.23 Pessiot et al. predicted DILI using available in vivo and in vitro microarray data from TG-GATEs by constructing a support vector machine (SVM) prediction model, and results showed that even though the goal was to predict human DILI, using rat in vivo data was more informative than using human in vitro data.29 In 2017, Rueda-ZÃ rate. combined SVM with empirical Bayes statistic to develop prediction model trained by toxicogenomics data, identifying the ability of gene expression profiles to distinguish subtypes of hepatotoxicity and label DILI compounds.23 However, most DILI prediction models based on genomics data were developed using conventional machine learning algorithms which need extensive manual interventions for data preprocess and feature selection. In this way, the model performance heavily depends on artificial extraction of features from raw data. As the rapid generation of genomics data, the prediction models based on conventional machine learning algorithms often have poor performance due to useful information being obscured by too many extracted features.30 Recently, as a state-of-the-art technology, deep learning (DL) architecture greatly improved the performance in many aspects of computer vision,31,32 speech recognition,33,34 natural language processing,35,36 and even medical diagnosis.37,38 These great successes of DL model have been mainly attributed to its powerful ability to automatically extract important features from raw and noisy data. With more hidden layers and more neurons in each layer, DL architecture could be capable of handling extensive volume of raw data using its strong computing ability.39,40,42,44,46 Compared with conventional machine learning algorithms, recent applications of DL could also benefit biomedical research, such as pharmacological properties of drugs prediction, protein torsion angle prediction,43 and anticancer drug synergy prediction.45 In this study, we try to combine the DL algorithm with genomics data to develop a more stable and highly accurate prediction model, which could systematically predict DILI in advance. Meanwhile, the conventional machine learning algorithms SVM was also used to compare the prediction performance with that of DL model. Furthermore, animal experiments were conducted to verify the predicted results of our model.

Table 1. Gene Expression Samples for Model Construction DILI labels

B

category

DILI-positive

DILI-negative

total number

training set testing set

397 99

393 99

790 198


Article


Figure 1. Basic architecture of deep learning prediction model.

where x was the actual label of input data in input layer, z was the predicted label of output layer, and d was the epoch number. In addition, the optimization function, Root Mean Square prop (RMSprop), was used to compile model. During the training process, the above processes were repeated to update the weights and bias until the optimal weight matrix W and bias b were obtained. Optimization of Deep Learning prediction Model. The performance of DL prediction model is not only determined by the architecture of DL prediction model but also by hyperparameters. In order to improve the performance of DL prediction model, feature gene selection and parameter optimization were carried out. Feature Gene Selection. When dealing with the large size of gene expression data, a common problem is the so-called “curse of dimensionality”.41 Thus, it is necessary to conduct feature selection to reduce the feature dimensionality. Here, two methods were used to select feature genes from the whole 15406 genes: (1) differential genes expression analysis and (2) feature selection based on weights values of feature vectors which were calculated by DL algorithm. Differentially expressed genes were analyzed using the empirical Bayes algorithm in the limma package53 from Bioconductor Library. Genes were considered to be the differentially expressed genes if the absolute value of fold change value was more than 2 and adjusted P-value was less than 0.05. In addition, since DL algorithm enabled the automatic extraction of important features from sample information, it was applied to calculate the weight values of each feature and rank the importance of each gene. Then proper threshold value was set to filter the inessential genes. Ultimately, 375 feature genes were chosen by analysis of gene differential expression and 1574 feature genes were selected by DL algorithm (Supporting Information, Tables S2 and S3). Parameter Optimization. The optimal combination of parameters for DL was usually determined by personal experience; obviously, it was not accurate and objective. To address this issue, grid search algorithm was applied to seek the best combination from the parameter space including epoch number, batch size, learning rate, dropout rate and node numbers of hidden layers. Here, the two data sets based on selected 375 feature genes and 1574 feature genes were applied to develop the optimal prediction model, respectively, resulting in 900 (5 × 4 × 3 × 3 × 5) models in each data set.

testing set, and their performances were compared with each other. Deep Learning Prediction Model. The deep learning (DL) prediction model was developed in the way of sequential mode using Python (version 3.6) based on the Keras platform, which is a high-level neural networks API running on top of Theano.51 The basic architecture of DL model was showed in Figure 1. First, each gene symbol as one feature was loaded in the nodes (also called neurons) of the input layer. Then the loaded information from input layer was propagated through the neighboring hidden layers, which included dense layer and dropout layer. Finally, the output layer could provide the classification results of each sample. To deal with the sophisticated classification problems, each layer among the DL architecture was followed by nonlinear activation functions. In this work, the Rectified Linear Unit (ReLU) activation function was applied to activate the input layer and hidden layer due to its ability to decrease the vanishing gradient problem as well as its fast computational speed (eq 1). Then for the output layer, Sigmoid activation function was applied to produce the classification labels (eq 2), because Sigmoid function can perform better on binary classification where it could map the field of real numbers to the range from 0 to 1, representing the probability of positive samples. y = ReLU(Wx + b)

(1)

where y was the activation value of the hidden layer, x was the input data, W was weight matrix, and b was bias. z = sigmoid(W ′y + b′)

(2)

where z was the classification labels, y was the activation value of the hidden layer, W′ was transposed weight matrix and b’ was transposed bias. During the model training process, the model performance was estimated by comparing the difference between the actual label of input data in input layer (x) and the predicted label of output layer (z) via the loss function binary cross-entropy, where stochastic gradient descent was applied to search the optimal parameters.52 (eq 3) d

LH (x , z) = − ∑ [xk log zk + (1 − xk)log(1 − zk)] k=1

(3) C


Article


for model based on 375 genes were 5.11 and 0.10, and those of model based on 1574 genes were 3.47 and 0.10 respectively. After model training, 198 testing samples were used to evaluate the prediction ability of all optimal DL and SVM models. The model performance was assessed using these criteria including sensitivity (SEN, eq 4), specificity (SPE, eq 5), accuracy (ACC, eq 6), and Matthews correlation coefficient (MCC, eq 7).

Subsequently, 10-fold cross validation was used to evaluate the model performance. Results revealed the best combination of parameters for the model based on 375 feature genes, which was 2 hidden layers with 100 nodes in each one. The dropout layer of 0.2 rejection ratio was used to avoid the overfitting problems. In addition, the number of nodes for output layer was equal to the number of classification label. Also, a smaller learning rate of 10−5, batch size of 32 and epoch number of 50 were essential for this optimal prediction model. Taken together, the optimal model based on 1574 feature genes has 2 hidden layers with 50 nodes in each one, a dropout rate of 0.5, a learning rate of 10−3, a batch size of 128 and an epoch number of 50 (Table 2, Supporting Information, Tables S4 and S5).

optimal value of parameters parameters epoch number batch size learning rate dropout rate node number

model of 375 genes 50

50

10, 32, 64, 128 0.01, 0.001, 0.00001 0, 0.2, 0.5 50, 100, 300, 500, 1000

32 0.00001 0.2 100

128 0.001 0.5 50

TN FP + TN

(5)

TP + TN TP + FN + FP + TN

(6)

(TP + FN)*(TP + FP)*(TN + FN)*(TN + FP) (7)

where TP meant true positive, TN meant true negative, FP meant false positive, and FN meant false negative. Furthermore, the area under the receiver operating characteristic (ROC) curve (AUC) was also used to evaluate the model performance. Specifically, the best possible prediction was 100% sensitivity and 100% specificity with area under the curve (AUC) of 1, while an AUC value of ≤0.5 represented random selection. Literature Validation. To further verify the performance of constructed models above, seven approved DILI positive compounds which were excluded from the compounds for model training were retrieved from literatures.55−61 Then the gene expression data corresponding to the seven compounds were downloaded from ArrayExpress database and processed via the same procedures in the section of data preprocessing (Table 4). In sum, 46 gene expression samples stimulated by seven compounds were collected for validating the predictive performance of optimal model. Meanwhile, a larger toxicogenomics data set from DrugMatrix of National Toxicology Program was applied to conduct further model validations. In this data set, gene expression profiles of liver samples were assayed by whole genome RG230_2.0 rat GeneChip arrays after exposure of rats to 200 different toxicants. Specifically, liver samples were obtained from rats that were treated with either test compound or vehicle control for 0.25, 1, 3, and 5 days, respectively. In a few samples, rats were exposed with target compounds for 7 days instead of 5 days. Then the reference information on DILI related to 200 compounds was searched and verified from Pubchem database, LiverTox database, Hazardous Substances Data Bank (HSDB) and literatures. As a result, 159 out of 200 compounds have been already reported with definite DILI or hepatotoxicity and 1 out of 200 compounds has been verified with no significant hepatotoxicity while the rest 40 compounds have no definite reference information on DILI so far (Supporting Information, Table S8). Thus, after profiling these samples with the same procedures that were described above, a total of 1579 gene expression samples of 160 compounds with positive or negative DILI information were constructed as another external validation data set for model performance testing.

Table 3. Parameters for SVM Prediction Model Optimization optimal value of parameters

K-fold cross validation type of SVM kernel function

SPE =

TP*TN − FP*FN

Support Vector Machine Prediction Model. To compare with the performance of DL prediction model, support vector machine (SVM) as one of highly robust conventional machine learning algorithms was selected to develop another prediction model. Based on the platform of MATLAB version R2016b, the SVM model was developed using libsvm version 3.2.2 toolbox.54 Then SVM models with 375 feature genes and 1574 feature genes were also respectively optimized by tuning several decisive parameters, including the type of cross validation mode, SVM and kernel function in the value space. The results showed that the optimal model based on 375 feature genes consisted of 3-fold cross validation, nu-SVC and radial basis function (RBF). The optimal model based on 1574 feature genes was set as 7-fold cross validation, nu-SVC and RBF function (Table 3,

parameters

(4)

MCC =

model of 1574 genes

10, 50, 100, 200, 500

TP TP + FN

ACC =

Table 2. Parameters for DL Prediction Model Optimization value space for optimization

SEN =

value space for optimization

model of 375 genes

model of 1574 genes

3, 4, 5, 6, 7, 8, 9, 10

3

7

C-SVC, nu-SVC linear, polynomial, RBF, sigmoid

nu-SVC RBF

nu-SVC RBF

Supporting Information, Tables S6 and S7). After determining the above parameters, other two critical parameters (the parameter g of RBF and penalty factor c) were optimized to further improve the performance of prediction model. Thus, particle swarm optimization (PSO) was utilized to obtain the best combination of the above two parameters. According to the final optimization results, the best combination of c and g D


Article

Journal of Chemical Information and Modeling Table 4. Compounds and Their Corresponding Gene Expression Data for Literature Validation compound name

CAS number

acarbose 2-acetylaminofluorene acetaminophen

56180-94-0 53-96-3 103-90-2

acetazolamide

59-66-5

doxepin hydrochloride

1229-29-4

ajmaline

4360-12-7

allopurinol

315-30-0

gene chip serial number 3017395027.CEL 3017564004.CEL 3017643003.CEL 3017643006.CEL 3017643009.CEL 3017271024.CEL 3017271029.CEL 3017711019.CEL 3017711022.CEL 3017711025.CEL 3017381014.CEL 3017381017.CEL 3017381020.CEL 3017748012.CEL 3017748015.CEL 3017748018.CEL

Experiment Validation. To further evaluate the accuracy and robustness of optimal model, in vivo hepatotoxicity experiments that stimulated by the compound without the report of obvious induced hepatotoxicity were carried out. Here, Vinblastine (CAS number: 865-21-4; molecular weight: 810.97g/mol; chemical structure see Figure 2) is one of the

3017395028.CEL 3017564005.CEL 3017643004.CEL 3017643007.CEL 3017643011.CEL 3017271026.CEL 3017271030.CEL 3017711020.CEL 3017711023.CEL 3017711026.CEL 3017381015.CEL 3017381018.CEL 3017381021.CEL 3017748013.CEL 3017748016.CEL 3017748019.CEL

3017395029.CEL 3017564006.CEL 3017643005.CEL 3017643008.CEL 3017271027.CEL 3017711021.CEL 3017711024.CEL 3017711027.CEL 3017381016.CEL 3017381019.CEL 3017381022.CEL 3017748014.CEL 3017748017.CEL 3017748020.CEL

were conducted under the operating conditions, which were same as the treatment of above gene expression data measurement. In order to validate whether the DL model can predict DILI in advance, a time series experiment of 1d, 3d, 5d and 9d was designed. The experiment instruments and reagents used for animal experiments were listed in Table 5. Table 5. List of Experiment Instruments and Reagents Used for Animal Experiments experiment instruments and reagents vinblastine (CAS: 865-21-4; HPLC purity ≥98%; batch no. CCJ18J7B8690) NaCl (CAS: 7647-14-5; AR; batch no. 2017021401) double distilled water AST diagnostic kit ALT diagnostic kit microplate spectrophotometer centrifuge Thermostat water bath SPF SD male rats

Figure 2. Chemical structure of vinblastine.

source Nanjing spring and autumn biological, Inc. Chengdu kelong chemical reagent, Inc. Made in our lab Nanjing Jiancheng Bioengineering Institute Nanjing Jiancheng Bioengineering Institute BioTek Instruments, Inc. Beckman Coulter, Inc. Jiangsu jinyi instrument technology, Inc. Laboratory animal center in Jiangsu University

Male Sprague−Dawley (SD) rats weighting 160−180 g were feed under temperature 22 ± 3 °C, humidity 30∼70%, photoperiod 12 h: 12 h and SPF bedding. All rats were allowed free access to SPF food and water. Studies were conducted in accordance with the Guide for the Care and Use of Laboratory Animals, and were approved by the Animal Care and Use Committee of Jiangsu University. After 1 week of acclimation, 24 rats were randomly divided into four experimental groups and four corresponding control groups (n = 3 for each group). Four experimental groups were administered with vinblastine at a dose of 0.3 mg/kg every day for 1 day, 3 days, 5 days and 9 days, respectively. In the contrast, rats in control groups were injected with normal saline instead of vinblastine. Rats in both control and treatment groups were injected via tail vein. After vinblastine injection, rats were closely observed, and behavioral and physiological changes were recorded. Rat blood samples were collected from retro-orbital plexus, followed by centrifugation at 3500 rpm for

alkaloids from catharanthus roseus62 and it has been widely used to treat Hodgkin’s lymphoma, nonsmall cell lung cancer, bladder cancer, brain cancer, melanoma, and testicular cancer in clinic treatment.63 According to a variety of clinical reports and related literatures, the toxic and side effects of vinblastine mainly focus on the decrease of leucocyte level, gastrointestinal reaction and neurotoxicity.64 Nevertheless, reports about vinblastine-induced liver toxicity and damage were rare in clinical treatment and medical research. In this study, the gene expression data of Sprague−Dawley (SD) rats which were repeatedly injected with vinblastine by the tail vein for 5 days were downloaded from Gene Expression Omnibus (GEO) database (GSM1389259, GSM1389373, GSM1389467, GSM1393535, GSM1393542, and GSM1393549). These gene expression data were processed to be predicted by optimal model and then corresponding animal experiments E


Article


Figure 3. Predicted results of the test data set based on the four optimal models. (A) The comparison between true and predicted results of SVM model based on 375 feature genes. The red pentacles represented predicted results and the blue squares represented true results covering the same parts between true and predicted results. (B), (C) and (D) indicated the predicted results of DL models based on 375 feature genes, SVM model based on 1574 feature genes and DL model based on 1574 feature genes, respectively.

10 min. Alanine aminotransferase (ALT) and aspartate aminotransferase (AST) activities in the obtained plasma were determined using ELISA kits. All the measured results were analyzed through statistical test to estimate whether there was significant difference between experimental and control groups. After the withdrawal of blood, rats were sacrificed and liver specimens were obtained, sliced and immediately fixed in 10% neutral buffered formalin for 24 h. Dehydration of liver specimens in ascending grades of alcohol was subsequently performed. Tissues were cleaned using three changes of xyline 1 h each was carried out, followed by the embedment with paraffin wax. Paraffin-embedded blocks were sectioned and stained with hematoxylin and eosin (H&E). Histopathological examination was performed using a microscopy. Gene Ontology Analysis of Feature Genes. After the validation of model performance, the 1574 feature genes (Supporting Information, Table S3) of modeling data for the optimal DL model could be also deemed as the informative genes related to DILI. In order to identify the biological roles of these genes, gene ontology (GO) enrichment analyses were carried out.

■

Table 6. Overall Performance of DL and SVM Prediction Model feature genes 375 1574

model

SEN (%)

SPE (%)

MCC

ACC (%)

AUC

SVM DL SVM DL

77.8 92.0 78.8 97.4

98.0 94.9 99.0 96.8

0.774 0.868 0.794 0.942

87.9 93.4 88.9 97.1

0.892 0.971 0.901 0.989

than that of SVM models, our DL model could still show a stable predictive ability in both positive and negative samples, compared with the significant difference between SEN and SPE of SVM models. Therefore, it could be identified that the DL prediction model in this work could greatly improve accuracy of DILI prediction, compared with the previous models based on conventional machine learning algorithms. In another aspect, both the DL and SVM model trained by data sets of 1574 feature genes achieved a significant improvement in prediction performance than models based on data sets of 375 feature genes. Meanwhile, the data sets based on whole 15406 genes without feature selection were also applied to train DL model, but it did not perform well, coming up with only 78.9% for ACC, 0.808 for AUC, 0.621 for MCC, 96.4% for SEN and 61.3% for SPE. Thus, it was greatly essential to conduct feature selection before model construction, and this method revealed the promising capability of DL model in feature selection, indicating that it could be applied to deal with other similar problems. Furthermore, the selected 1574 genes were considered as important feature genes with regard to hepatotoxicity. Eventually, the optimal DL model trained by 1574 feature genes (Final_DL model) was applied to conduct further validation.

RESULTS AND DISCUSSION

Model Performance. Four optimized models (SVM models based on 375 and 1574 feature genes, DL models based on 375 and 1574 feature genes) were respectively evaluated using test data set and the results were shown in Figure 3 and Supporting Information, Table S9. The evaluation indexes showed that no matter trained by data set consisting of either 375 feature genes or 1574 feature genes, both DL models performed much better than SVM models (Table 6). Although the SPE of DL models was a bit lower F


Article


the samples were treated with various experimental conditions (administered doses and treatment time points), its samples in high administered dose group or long treatment time group were predicted as DILI positive and (2) for the compounds with the samples were treated with the same experimental condition, the number of DILI positive samples was larger than that of DILI negative samples. All other cases were considered as negative. According to the above criteria, among the 159 compounds with definite DILI, 149 compounds were predicted as DILI positive while 10 compounds were predicted as DILI negative (Supporting Information, Table S10). In addition, the rest one (tosufloxacin, CAS: 100490-36-6) which has been verified with no significant hepatotoxicity was predicted as DILI negative due to its all samples were predicted as DILI negative (successive administration for 3d and 5d at a daily dose of 2000 mg/kg). The previous toxicity studies of tosufloxacin65−67 demonstrated that it had no significant toxicity, especially in liver tissue, in the acute toxicity and chronic toxicity studies. For the acute toxicity study, the rats were orally administrated with tosufloxacin at a dose of 10 g/ kg for successive 14 days. Additionally, compared with control groups, there was no significant difference observed from pathological sections of liver tissue. With regard to the studies of chronic toxicity, the rats were orally administrated with tosufloxacin at a daily dose of 80, 400, and 2000 mg/kg for 6 months, respectively. As a result, rats in the group of 80 mg/kg showed no pathological changes compared with control groups. However, for rats in the groups of 400 and 2000 mg/kg, no significant pathological changes were observed, except for the degenerated renal tubular epithelial cell, cell infiltration in tissue interspace and drug crystal formed in kidney tubules. Based on these results of toxicological studies, it is revealed that the predicted results of tosufloxacin are in accordance with true results. Furthermore, the considerable reason for false negative results of 10 compounds was that the gene expression samples were obtained after exposed at a low dose for a short time (0.25 and 1 d). Since DILI occurs in a dose and time dependent manner, the chronic liver injury has been attributed to either drug overdose or long-term drug use. Thus, these 10 compounds with reported hepatotoxicity may not lead to DILI in short-term use at low doses, and they should be predicted as DILI negative. On the other hand, it could be also deemed that the DL model showed a sensitive and stable predictive ability in both positive and negative samples. Therefore, as a well-performed prediction model, the final_DL could be applied to predict DILI using gene expression data from various resources. Experiment Validation. The gene expression data after 5 days of repeat administration of vinblastine was used to predict whether vinblastine could trigger DILI by final_DL model, and the results showed that all samples were predicted as DILI positive (Figure 5). Thus, it meant 5 days of repeat administration of vinblastine would damage liver to varied extent. To verify the predicted results, the in vivo animal experiments were carried out in terms of physical signs, serum levels of ALT and AST and pathological outcome. During the treatment, rats begun to shed hair along with loss of appetite after repeat administration with vinblastine for 5 days, compared with control group. With regard to levels of ALT and AST in rats, no significant difference was observed between the treatment and control group (P > 0.05) after

In further analysis, we attempted to identify whether the sample information from different groups (low, middle, and high dose; 4, 8, 15, and 29 d treatment time) would have influence on model performance. Thus, the test data set was divided into seven groups according to administered dose and treatment time, and they were respectively used to evaluate performances of Final_DL model (Table 7). In terms of Table 7. Model Performances of Final_DL Model Based on Samples in Various Groupsa sample group

SEN (%)

SPE (%)

MCC

ACC (%)

AUC

low dose middle dose high dose 4d 8d 15d 29d

94.1 96.9 98.0 96.0 95.5 96.6 100.0

nan nan nan 95.7 100.0 100.0 87.5

nan nan nan 0.917 0.953 0.967 0.882

94.1 96.9 98.0 95.8 97.6 98.3 93.8

nan nan nan 0.988 0.991 0.996 0.976

a

Note: due to no negative samples included in low dose, middle dose, and high dose groups of test data set, the values of the three evaluation indexes (SPE, MCC, and AUC) in the groups were replaced by nan.

various dose groups, results demonstrated that the model performance in high dose group was much better than that in middle and low dose groups. Thus, in certain extent, a dose− response trend was observed for the model performance. Besides, the model performances became better as the treatment time increased in most of the cases. However, only the 29d group showed a relatively low specificity, which might be due to several control samples that were predicted as DILI positive. Overall, it is reckoned that the gene expression samples with high administered dose and long treatment time might provide more accurate information for model construction. Literature Validation. 46 gene expression samples stimulated by seven compounds were used to evaluate the performance of final_DL model. All the 46 samples were predicted as DILI positive and the results were completely consistent with the approved results of literatures (Figure 4).

Figure 4. Predicted results of literature validation.

In addition, with regard to the structural diversity of these compounds, the final_DL showed robust prediction ability without the limitation of conventional structure-based models. In addition, 1579 gene expression samples of 160 compounds from DrugMatrix were also predicted by final_DL model to test model performance. Since a compound corresponds to multiple samples, if the prediction results of the samples of a compound included both positive and negative, this compound needed to meet the following criteria to be considered as DILI positive: (1) for the compounds with G


Article


Gene Ontology Analysis of Feature Genes. In order to identify the biological roles of 1574 feature genes used for constructing the optimal DL model, Gene Ontology (GO) enrichment analyses were carried out and their biological processes were obtained based on p value cutoff = 0.01, (Supporting Information, Table S11). The enriched biological processes were mostly related to metabolic process of biomoleculars and compounds. The top20 enriched biological processes included metabolism of monocarboxylic acid, lipids, fatty acid, organic acid, small molecule and enzyme activity (Figure 7). In addition, 112 genes related to a variety of biological processes, including “liver development”, “hepaticobiliary system development”, “response to toxic substance” and “cellular response to toxic substance” (Supporting Information, Table S12). It suggested that metabolic abnormalities caused by compounds are highly correlated with DILI, and these feature genes are informative to predict DILI. In sum, after the evaluation of model performance, both literatures and in vivo animal experiments indicated that the DL model based on gene expression data could show a great performance in accurate DILI prediction indeed. Although the animal experiments treated with a single compound served to a concept level to show how the prediction model could be used in a preclinical drug safety testing; a considerable number of external test data sets in literature validation have also evaluated and verified the high performance of the DL model. Most important of all, the DL model could be efficiently applied to preclinical drug safety assessment. DILI that tends to be late-onset and severe hepatotoxic could not be clearly detected by either routine blood tests or pathological findings during preclinical phases or early clinical phases of drug development. As a result, the assessment of false DILI negative caused by traditional animal testing methods may lead to serious public health problems and financial loss. To avoid the undetected DILI in preclinical drug safety assessment, the DL model combined with toxicogenomics could detect the anomalies induced by drug on a genomic scale, thereby predicting and warning liver injury in advance. Additionally, as an alternative method of animal testing in DILI detection, the method we developed not only could reduce the time and cost of drug development, but also contribute to the 3R (replacement, reduction and refinement) principles of animal research.

Figure 5. Predicted results of gene expression samples treated with vinblastine.

repeat administration for 1, 3, and 5 days, respectively. However, after 9 days of repeat administration, level of ALT was increased up to 149.25 ± 7.1 in treatment group compared with 33.17 ± 2.50 in control group. Also, level of AST was changed from 41.46 ± 2.17 in control group up to 205.13 ± 4.78 in treatment group (Table 8). The significant differences (P < 0.01) between treatment group and control group indicated a obvious liver injury with a repeat administration for 9 days. Table 8. Serum Levels of ALT and ASTa group 1 1 3 3 5 5 9 9

d d d d d d d d

(control) (treatment) (control) (treatment) (control) (treatment) (control) (treatment)

number of rats

ALT (KarU)

AST (KarU)

3 3 3 3 3 3 3 3

24.82 ± 1.28 22.31 ± 2.06 27.09 ± 4.81 32.12 ± 1.08 15.60 ± 5.26 18.96 ± 7.18 33.17 ± 2.50 149.25 ± 7.13b

45.69 ± 3.02 49.75 ± 2.16 52.95 ± 5.06 50.71 ± 1.84 31.20 ± 6.96 32.40 ± 3.45 41.46 ± 2.17 205.13 ± 4.78b

a

Measured results were conducted independent-samples t tests. Means P < 0.05.

b

Moreover, histological evaluation of rat livers was performed. In group that was treated for only day, the histological features of liver injury were not observed, in comparison with that of control group. Similarly, no significant liver injury was identified in group that was treated with vinblastine for 3 days, only showing vacuolar degeneration in partial hepatic cell. After further administration for next 2 days, hyperemia in tissue interspace and cell shrinkage were also observed in rat livers. However, some typical histological features were observed in samples that were repeat administration for 9 days, showing necrosis and a large area of pathological lesions. Obviously, results of 9-day repeat administration indicated the onset of liver injury, in accordance with results of changes in ALT and AST levels (Figure 6). With regard to the predicated livery injury onset date, the time of onset in in vivo results seems slightly later. This discrepancy might be explained that a variety of potential factors which would trigger obvious hepatic lesions already occurred in 5-day treatment of vinblastine, for instance, hepatic vein thrombosis and hepatic veno-occlusive disease in partial liver tissue, the cytopathy as well as inflammation onset. Although the predicted results were not exactly consistent to the experiment results, the model accurately predicted that vinblastine could trigger liver injury. In other words, the final_DL model could predict tardive DILI in early stage.

■

CONCLUSION In conclusion, a gene expression data based deep learning prediction model for drug-induced liver injury was successfully developed through feature genes selection and parameters optimization, followed by validation by literatures and animal experiments. It was of high performance to detect the anomalies induced by drug on a genomic scale, thereby predicting and warning liver injury in early stage. Meanwhile, the roles of important feature genes related to hepatotoxicity were analyzed via GO enrichment analyses. Our model could be a very useful tool to provide safety information for drug discovery and clinical rational drug use in early stage and was expected to become an important part of drug safety assessment. Furthermore, with more available biomedical data updating, this approach of combining gene expression data with deep learning could be also applied to predict other drug induced adverse effects, such as neurotoxicity and cardiotoxicity. H


Article


Figure 6. Histological evaluation of liver tissue at different treatment time points. (A) Histological sections of liver tissue with one-day administration of vinblastine (1-a) and control groups (1-b) based on 10 × 10 microscopic magnification and 10 × 40 microscopic magnification, respectively. (2-a), (3-a), and (4-a) belonged to experimental groups of three different rats based on 10 × 10 microscopic magnification and (2-b), (3-b), and (4-b) belonged to their corresponding partial enlarged photos based on 10 × 40 microscopic magnification. Histological sections of liver tissue with 3-day (B), 5-day (C), and 9-day (D) administration of vinblastine were observed, respectively. Each photo was labeled as the same way described in (A).

Figure 7. GO enrichment analysis of the 1574 feature genes.

■

ASSOCIATED CONTENT

* Supporting Information S

The Supporting Information is available free of charge on the ACS Publications website at DOI: 10.1021/acs.jcim.9b00143. Details of the supporting tables (PDF) Table S1, 988 samples with related information (XLSX) Table S2, selected 375 feature genes (XLSX) Table S3, selected 1574 feature genes (XLSX) Table S4, the results of parameter optimization for DL model based on 375 genes (XLSX) Table S5, the results of parameter optimization for DL model based on 1574 genes (XLSX) Table S6, the results of parameter optimization for SVM model based on 375 genes (XLSX) Table S7, the results of parameter optimization for SVM model based on 1574 genes (XLSX)

■

Table S8, reference information on compounds from NTP drugmatrix (XLSX) Table S9, the predicted results of test data sets (XLSX) Table S10, predicted results of NTP drugmatrix data sets (XLSX) Table S11, the results of GO enrichment analysis of 1574 genes (XLSX) Table S12, the selected 112 important genes for DILI prediction (XLSX)

AUTHOR INFORMATION

Corresponding Authors

*E-mail: [email protected]. Phone: +86-511-8879-1526. *E-mail: [email protected]. Phone: +86-511-8879-1526. ORCID

Chunlai Feng: 0000-0002-9628-1331 I


Article


Understanding and Prediction of Human Drug-induced Liver Injury. Hepatology 2017, 65, 710−721. (12) Ho, S.; Mclachlan, A.; Chen, T.; Hibbs, D.; Fois, R. Relationships Between Pharmacovigilance, Molecular, Structural, and Pathway Data: Revealing Mechanisms for Immune-Mediated Drug-Induced Liver Injury. CPT: Pharmacometrics Syst. Pharmacol. 2015, 4, 426−441. (13) Welch, M. A.; Köck, K.; Urban, T. J.; Brouwer, K. L.; Swaan, P. W. Toward Predicting Drug-Induced Liver Injury: Parallel Computational Approaches to Identify Multidrug Resistance Protein 4 and Bile Salt Export Pump Inhibitors. Drug Metab. Dispos. 2015, 43, 725−734. (14) Kotsampasakou, E.; Ecker, G. F. Predicting Drug-Induced Cholestasis with the Helpof Hepatic TransportersAn in Silico Modeling Approach. J. Chem. Inf. Model. 2017, 57, 608−615. (15) Xu, Y.; Dai, Z.; Chen, F.; Gao, S.; Pei, J.; Lai, L. Deep Learning for Drug-Induced Liver Injury. J. Chem. Inf. Model. 2015, 55, 2085− 2093. (16) Wu, L.; Liu, Z.; Auerbach, S. S.; Huang, R.; Chen, M.; Mceuen, K.; Xu, J. Z.; Fang, H.; Tong, W. Integrating Drug’s Mode of Action into Quantitative Structure-Activity Relationships for Improved Prediction of Drug-Induced Liver Injury. J. Chem. Inf. Model. 2017, 57, 1000−1006. (17) Kim, E.; Nam, H. Prediction Models for Drug-Induced Hepatotoxicity by Using Weighted Molecular Fingerprints. BMC Bioinf. 2017, 18, 227. (18) Hong, H.; Thakkar, S.; Chen, M.; Tong, W. Development of Decision Forest Models for Prediction of Drug-Induced Liver Injury in Humans Using A Large Set of FDA-Approved Drugs. Sci. Rep. 2017, 7, 17311. (19) Ran, S.; Xiong, S.; Zink, D.; Loo, L. H. High-Throughput Imaging-Based Nephrotoxicity Prediction for Xenobiotics with Diverse Chemical Structures. Arch. Toxicol. 2016, 90, 2793−2808. (20) Steiner, S.; Anderson, N. L. Expression Profiling in Toxicology– Potentials and Limitations. Toxicol. Lett. 2000, 112, 467−471. (21) Guerreiro, N.; Staedtler, F.; Grenet, O.; Kehren, J.; Chibout, S. D. Toxicogenomics in Drug Development. Toxicol. Pathol. 2003, 31, 471−479. (22) Lord, P. G.; Nie, A.; McMillian, M. Application of Genomics in Preclinical Drug Safety Evaluation. Basic Clin. Pharmacol. Toxicol. 2006, 98, 537−546. (23) Rueda-ZÃ rate, H. A.; Imaz-Rosshandler, I.; CÃ rdenas-Ovando, R. A.; Castillo-FernÃ ndez, J. E.; Noguez-Monroy, J.; Rangel-Escareno, C. A Computational Toxicogenomics Approach Identifies a List of Highly Hepatotoxic Compounds from a Large Microarray Database. PLoS One 2017, 12, No. e0176284. (24) Otava, M.; Shkedy, Z.; Kasim, A. Prediction of Gene Expression in Human Using Rat in Vivo Gene Expression in Japanese Toxicogenomics Project. Syst. Biomed. 2014, 2, 8−15. (25) Rodrigues, R. M.; Heymans, A.; De Boe, V.; Sachinidis, A.; Chaudhari, U.; Govaere, O.; Roskams, T.; Vanhaecke, T.; Rogiers, V.; De Kock, J. Toxicogenomics-Based Prediction of AcetaminophenInduced Liver Injury Using Human Hepatic Cell Systems. Toxicol. Lett. 2016, 240, 50−59. (26) Jennen, D.; Polman, J.; Bessem, M.; Coonen, M.; van Delft, J.; Kleinjans, J. Drug-Induced Liver Injury Classification Model Based on in Vitro Human Transcriptomics and in Vivo Rat Clinical Chemistry Data. Syst. Biomed. 2014, 2, 63−70. (27) Krauskopf, J.; Caiment, F.; Claessen, S. M.; Johnson, K. J.; Warner, R. L.; Schomaker, S. J.; Burt, D. A.; Aubrecht, J.; Kleinjans, J. C. Application of High-Throughput Sequencing to Circulating MicroRNAs Reveals Novel Biomarkers for Drug-Induced Liver Injury. Toxicol. Sci. 2015, 143, 268−276. (28) Kawamoto, T.; Ito, Y.; Morita, O.; Honda, H. MechanismBased Risk Assessment Strategy for Drug-Induced Cholestasis Using the Transcriptional Benchmark Dose Derived by Toxicogenomics. J. Toxicol. Sci. 2017, 42, 427−436. (29) Pessiot, J. F.; Wong, P. S.; Maruyama, T.; Morioka, R.; Aburatani, S.; Tanaka, M.; Fujibuchi, W. The Impact of Collapsing

Hanqin Liu: 0000-0001-7984-6305 Mengjie Rui: 0000-0001-5566-817X Author Contributions

C.F. and H.C. contributed equally to this work. C.F. and H.C. wrote the main manuscript. C.F. and M.R. designed the research. H.C. and X.Y. constructed the prediction models and statistical analysis. X.Y. and M.S. collected the data and conducted animal experiments. H.L. conducted the pathological section. All authors have read and given approval to the final version of the manuscript. Author Contributions #

C.F. and H.C. contributed equally to this work.

Notes

The authors declare no competing financial interest.

■

ACKNOWLEDGMENTS The authors thank Professor Min Xu at Jiangsu University for his kind help with pathological section. This work was supported, in part, by the National Natural Science Foundation of China (Nos. 81373897 and 81672582), Natural Science Foundation of Jiangsu Province (No. BK20181445), Six Talent Peak Project from Government of Jiangsu Province (Nos. SWYY-013 and 2015-SWYY-019), Natural Science Foundation of Jiangsu Province for Distinguished Young Scholars (No. BK20160013), Postdoctoral Science Foundation of Jiangsu Province (No. 1402174C), and the Scientific Research Foundation of Jiangsu University (No. 12JDG034 and 14JDG163). The code with examples of input files and outputs is freely available at DOI: 10.5281/zenodo.2639910.

■

REFERENCES

(1) Chandra, P.; Brouwer, K. L. R. The Complexities of Hepatic Drug Transport: Current Knowledge and Emerging Concepts. Pharm. Res. 2004, 21, 719−735. (2) Almazroo, O. A.; Miah, M. K.; Venkataramanan, R. Drug Metabolism in the Liver. Clin Liver Dis 2017, 21, 1−20. (3) Jaeschke, H.; Gores, G. J.; Cederbaum, A. I.; Hinson, J. A.; Pessayre, D.; Lemasters, J. J. Mechanisms of Hepatotoxicity. Toxicol. Sci. 2002, 65, 166−176. (4) Shehu, A. I.; Ma, X.; Venkataramanan, R. Mechanisms of DrugInduced Hepatotoxicity. Clin Liver Dis 2017, 21, 35−54. (5) Kaplowitz, N. Idiosyncratic Drug Hepatotoxicity. Nat. Rev. Drug Discovery 2005, 4, 489−499. (6) Regev, A. Drug-Induced Liver Injury and Drug Development: Industry Perspective. Semin. Liver Dis. 2014, 34, 227−239. (7) Kohonen, P.; Parkkinen, J. A.; Willighagen, E. L.; Ceder, R.; Wennerberg, K.; Kaski, S.; Grafström, R. C. A Transcriptomics DataDriven Gene Space Accurately Predicts Liver Cytopathology and Drug-Induced Liver Injury. Nat. Commun. 2017, 8, 15932. (8) Brinker, A. D.; Lyndly, J.; Tonning, J.; Moeny, D.; Levine, J. G.; Avigan, M. I. Profiling Cumulative Proportional Reporting Ratios of Drug-Induced Liver Injury in the FDA Adverse Event Reporting System (FAERS) Database. Drug Saf. 2013, 36, 1169−1178. (9) Kim, D. E.; Jang, M. J.; Kim, Y. R.; Lee, J. Y.; Cho, E. B.; Kim, E.; Kim, Y.; Kim, M. Y.; Jeong, W. I.; Kim, S. Prediction of Drug-Induced Immune-Mediated Hepatotoxicity Using Hepatocyte-like Cells Derived from Human Embryonic Stem Cells. Toxicology 2017, 387, 1−9. (10) Yu, K. N.; Nadanaciva, S.; Rana, P.; Dong, W. L.; Ku, B.; Roth, A. D.; Dordick, J. S.; Will, Y.; Lee, M. Y. Prediction of MetabolismInduced Hepatotoxicity on Three-Dimensional Hepatic Cell Culture and Enzyme Microarrays. Arch. Toxicol. 2018, 92, 1295−1310. (11) Goldring, C.; Antoine, D. J.; Bonner, F.; Crozier, J.; Denning, C.; Fontana, R. J.; Hanley, N. A.; Hay, D. C.; Ingelman-Sundberg, M.; Juhila, S. Stem Cell-Derived Models to Improve Mechanistic J


Article

Journal of Chemical Information and Modeling Data on Microarray Analysis and DILI Prediction. Syst. Biomed. 2013, 1, 137−143. (30) Kalinin, A. A.; Higgins, G. A.; Reamaroon, N.; Soroushmehr, S.; Allynfeuer, A.; Dinov, I. D.; Najarian, K.; Athey, B. D. Deep Learning in Pharmacogenomics: From Gene Regulation to Patient Stratification. Pharmacogenomics 2018, 19, 629−650. (31) Krizhevsky, A.; Sutskever, I.; Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. In International Conference on Neural Information Processing Systems; 2012; pp 1097−1105. (32) Nie, S.; Zheng, M.; Ji, Q. The Deep Regression Bayesian Network and Its Applications: Probabilistic Deep Learning for Computer Vision. IEEE Signal Processing Magazine 2018, 35, 101− 111. (33) Huang, P. S.; Kumar, K.; Liu, C.; Gong, Y.; Deng, L. Predicting Speech Recognition Confidence Using Deep Learning with Word Identity and Score Features. In IEEE International Conference on Acoustics, Speech and Signal Processing 2013, 7413−7417. (34) Noda, K.; Yamaguchi, Y.; Nakadai, K.; Okuno, H. G.; Ogata, T. Audio-Visual Speech Recognition Using Deep Learning. Applied Intelligence 2015, 42, 722−737. (35) Collobert, R.; Weston, J. A Unified Architecture for Natural Language Processing:Deep Neural Networks with Multitask Learning. In International Conference on Machine Learning 2008, 160−167. (36) Sarikaya, R.; Hinton, G. E.; Deoras, A. Application of Deep Belief Networks for Natural Language Understanding. IEEE/ACM Transactions on Audio Speech & Language Processing 2014, 22, 778− 784. (37) Hazlett, H. C.; Gu, H.; Munsell, B. C.; Kim, S. H.; Styner, M.; Wolff, J. J.; Elison, J. T.; Swanson, M. R.; Zhu, H.; Botteron, K. N.; Collins, D. L.; Constantino, J. N.; Dager, S. R.; Estes, A. M.; Evans, A. C.; Fonov, V. S.; Gerig, G.; Kostopoulos, P.; McKinstry, R. C.; Pandey, J.; Paterson, S.; Pruett, J. R.; Schultz, R. T.; Shaw, D. W.; Zwaigenbaum, L.; Piven, J. Early Brain Development in Infants at High Risk for Autism Spectrum Disorder. Nature 2017, 542, 348− 351. (38) Esteva, A.; Kuprel, B.; Novoa, R. A.; Ko, J.; Swetter, S. M.; Blau, H. M.; Thrun, S. Dermatologist-Level Classification of Skin Cancer with Deep Neural Networks. Nature 2017, 542, 115−118. (39) Lusci, A.; Pollastri, G.; Baldi, P. Deep Architectures and Deep Learning in Chemoinformatics: The Prediction of Aqueous Solubility for Drug-like Molecules. J. Chem. Inf. Model. 2013, 53, 1563−1575. (40) Ma, J.; Sheridan, R. P.; Liaw, A.; Dahl, G. E.; Svetnik, V. Deep Neural Nets as a Method for Quantitative Structure-Activity Relationships. J. Chem. Inf. Model. 2015, 55, 263−274. (41) Aliper, A.; Plis, S.; Artemov, A.; Ulloa, A.; Mamoshina, P.; Zhavoronkov, A. Deep Learning Applications for Predicting Pharmacological Properties of Drugs and Drug Repurposing Using Transcriptomic Data. Mol. Pharmaceutics 2016, 13, 2524−2530. (42) Pereira, J. C.; Caffarena, E. R.; Dos Santos, C. N. Boosting Docking-Based Virtual Screening with Deep Learning. J. Chem. Inf. Model. 2016, 56, 2495−2506. (43) Li, H.; Hou, J.; Adhikari, B.; Qiang, L.; Cheng, J. Deep Learning Methods for Protein Torsion Angle Prediction. BMC Bioinf. 2017, 18, 417. (44) Xu, Y.; Pei, J.; Lai, L. Deep Learning Based Regression and Multiclass Models for Acute Oral Toxicity Prediction with Automatic Chemical Feature Extraction. J. Chem. Inf. Model. 2017, 57, 2672− 2685. (45) Preuer, K.; Rpi, L.; Hochreiter, S.; Bender, A.; Bulusu, K. C.; Klambauer, G. DeepSynergy: Predicting Anti-Cancer Drug Synergy with Deep Learning. Bioinformatics 2018, 34, 1538−1546. (46) Fernandez, M.; Ban, F.; Woo, G.; Hsing, M.; Yamazaki, T.; LeBlanc, E.; Rennie, P. S.; Welch, W. J.; Cherkasov, A. Toxic Colors: The Use of Deep Learning for Predicting Toxicity of Compounds Merely from Their Graphic Images. J. Chem. Inf. Model. 2018, 58, 1533−1543. (47) Brazma, A.; Parkinson, H.; Sarkans, U.; Shojatalab, M.; Vilo, J.; Abeygunawardena, N.; Holloway, E.; Kapushesky, M.; Kemmeren, P.;

Lara, G. G. Array Express–a Public Repository for Microarray Gene Expression Data at the EBI. Nucleic Acids Res. 2003, 31, 68−71. (48) Ihaka, R.; Gentleman, R. R: A Language for Data Analysis and Graphics. Journal of Computational and Graphical Statistics 1996, 5, 299−314. (49) Carvalho, B. S.; Irizarry, R. A. A Framework for Oligonucleotide Microarray Preprocessing. Bioinformatics 2010, 26, 2363−2367. (50) Gentleman, R. C; Carey, V. J; Bates, D. M; Bolstad, B.; Dettling, M.; Dudoit, S.; Ellis, B.; Gautier, L.; Ge, Y.; Gentry, J. Bioconductor: Open Software Development for Computational Biology and Bioinformatics. Genome Biol. 2004, 5, 1−16. (51) Team, T. D.; Alrfou, R.; Alain, G.; Almahairi, A.; Angermueller, C.; Bahdanau, D.; Ballas, N.; Bastien, F.; Bayer, J.; Belikov, A. Theano: A Python Framework for Fast Computation of Mathematical Expressions. 2016. (52) Alkawaa, F. M.; Chaudhary, K.; Garmire, L. X. Deep Learning Accurately Predicts Estrogen Receptor Status in Breast Cancer Metabolomics Data. J. Proteome Res. 2018, 17, 337−347. (53) Smyth, G. K. Limma: Linear Models for Microarray Data; Springer: New York, 2005. (54) Chang, C. C.; Lin, C. J. LIBSVM: A Library for Support Vector Machines. 2011, 2, 1−27. (55) Thorgeirsson, S. S.; Felton, J. S.; Nebert, D. W. Genetic Differences in the Aromatic Hydrocarbon-Inducible n-Hydroxylation of 2-Acetylaminofluorene and Acetaminophen-Produced Hepatotoxicity in Mice. Mol. Pharmacol. 1975, 11, 159−165. (56) Monges, B.; Monges, G.; Salducci, J. [Ajmaline-Induced Hepatitis. A Case Report with Ultrastructural Study]. Gastroenterol. Clin. Biol. 1983, 7, 540−544. (57) Maren, T. H. Acetazolamide and Advanced Liver Disease. Am. J. Ophthalmol. 1986, 102, 672−672. (58) Carrascosa, M.; Pascual, F.; Aresti, S. Acarbose-Induced Acute Severe Hepatotoxicity. Lancet 1997, 349, 698−699. (59) Cheo, D. L.; Burns, D. K.; Meira, L. B.; Houle, J. F.; Friedberg, E. C. Mutational Inactivation of the Xeroderma Pigmentosum Group C Gene Confers Predisposition to 2-Acetylaminofluorene-Induced Liver and Lung Cancer and to Spontaneous Testicular Cancer in Trp53−/− Mice. Cancer Res. 1999, 59, 771−775. (60) Yoon, J. Y.; Min, S. Y.; Park, J. Y.; Hong, S. G.; Park, S. J.; Paik, S. Y.; Park, Y. M. [A Case of Allopurinol-Induced Granulomatous Hepatitis with Ductopenia and Cholestasis]. Korean. J. Hepatol 2008, 14, 97−101. (61) Keegan, A. D. Doxepin-Induced Recurrent Acute Hepatitis. Aust. N. Z. J. Med. 2010, 23, 523−523. (62) Guéritte, F.; Bac, N. V.; Langlois, Y.; Potier, P. Biosynthesis of Antitumour Alkaloids from Catharanthus Roseus. Conversion of 20′Deoxyleurosidine into Vinblastine. J. Chem. Soc., Chem. Commun. 1980, 10, 452−453. (63) Zelnak, A. B. Clinical Pharmacology and Use of MicrotubuleTargeting Agents in Cancer Therapy. Methods Mol. Med. 2007, 137, 209−234. (64) Suresh, P.; Kapoor, R.; Kapur, B. N. Severe Neurotoxicity Due to Vinblastine in Hodgkin Lymphoma. South Asian J. Cancer 2014, 3, 147−148. (65) Niki, Y. Pharmacokinetics and Safety Assessment of Tosufloxacin Tosilate. J. Infect. Chemother. 2002, 8, 1−18. (66) Mei, L. Pharmacology and Clinical Effect of the Tosufloxacin. Journal of Pharmaceutical Practice 1995, 13, 145−149. (67) Xian, G.; Wuqing, O.; Mengyun, L.; Yue, L.; Bozhen, W.; Yin, Z.; Mingqi, Y. Study on the Preparation and Acute Toxicity of Tosufloxacin Nanoemulsion. Chinese Journal of Animal and Veterinary Sciences 2015, 46, 2069−2077.

K


Gene Expression Data Based Deep Learning Model for Accurate

Recommend Documents