4.In silico tox prediction-Data sources.pdf

REVIEW...

REVIEW published: 20 February 2018 doi: 10.3389/fchem.2018.00030 In Silico Prediction of Chemical Toxicity for Drug Design Using Machine Learning Methods and Structural Alerts Hongbin Yang, Lixia Sun, Weihua Li, Guixia Liu and Yun Tang* Shanghai Key Laboratory of New Drug Design, School of Pharmacy, East China University of Science and Technology, Shanghai, China During drug development, safety is always the most important issue, including a variety of toxicities and adverse drug effects, which should be evaluated in preclinical and clinical trial phases. This review article at first simply introduced the computational methods used in prediction of chemical toxicity for drug design, including machine learning methods and structural alerts. Machine learning methods have been widely applied in qualitative classification and quantitative regression studies, while structural alerts can be regarded Edited by: as a complementary tool for lead optimization. The emphasis of this article was put on Daniela Schuster, the recent progress of predictive models built for various toxicities. Available databases Paracelsus Private Medical University and web servers were also provided. Though the methods and models are very helpful of Salzburg, Austria for drug design, there are still some challenges and limitations to be improved for drug Reviewed by: Huixiao Hong, safety assessment in the future. United States Food and Drug Keywords: drug safety, chemical toxicity, drug design, machine learning, structural alerts Administration, United States Heebeom Koo, Catholic University of Korea, South Korea INTRODUCTION *Correspondence: Yun Tang Drug discovery and development is a long journey full of high risk. It is estimated that the attrition [email protected] rate of drug candidates is up to 96% (Paul et al., 2010) and the average cost to develop a new drug reaches to 2.6 billion U.S. dollars in recent years (PhRMA, 2015). One of the major causes for the Specialty section: high attrition rate is drug safety, which accounts for 30% of drug failures (Giri and Bader, 2015). This article was submitted to Even if a drug is approved in market, it could be withdrawn due to safety problems. Therefore, drug Medicinal and Pharmaceutical safety should be evaluated extensively as early as possible. Chemistry, Usually, in vitro and in vivo tests are performed to investigate drug safety, including a variety a section of the journal of toxicities and adverse drug effects. In recent years, there are also some efforts to develop Frontiers in Chemistry in vitro models such as “organ on a chip” to reduce cost (Huh et al., 2010, 2011). However, Received: 11 January 2018 those approaches are still costly and time-consuming. In comparison of experimental approaches, Accepted: 05 February 2018 computational methods have shown great advantages since they are green, fast, cheap, accurate, Published: 20 February 2018 and most importantly they could be done before a compound is synthesized (Segall and Barber, Citation: 2014). Yang H, Sun L, Li W, Liu G and Tang Y Till now, many computational models have been developed for drug safety assessment, which (2018) In Silico Prediction of Chemical Toxicity for Drug Design Using could be generally divided into three categories: qualitative classification, quantitative regression Machine Learning Methods and and read-across. As the first step of drug safety assessment, we only need to know a compound is Structural Alerts. Front. Chem. 6:30. toxic or non-toxic, highly toxic or slightly toxic, rather than its exact toxicity value, so classification doi: 10.3389/fchem.2018.00030 models can be used. For a small number of chemical analogs, quantitative structure-toxicity Frontiers in Chemistry | www.frontiersin.org 1 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity relationship (QSTR) models can be derived for prediction of In addition to the phenotype data that are directly relevant to exact toxicity values. For those unique compounds, read-across toxicity, databases on bioactivity, pathway and side effects are also is also a feasible approach to deduce certain toxicity endpoint important to toxicity prediction. Several bioactivity databases are from their similar structures with experimental toxicity values. free available, such as PubChem (Wang et al., 2009), ChEMBL These models have high accuracies especially in a local chemical (Gaulton et al., 2017), and BindingDB (Gilson et al., 2016). We space, and sometimes they can replace in vitro or in vivo assays developed a web server named MetaADEDB that integrates CTD for certain endpoints. Furthermore, structural alerts (SAs) can (Davis et al., 2017), SIDER (Kuhn et al., 2010), and OFFSIDES be derived from the models as keys for a compound to cause (Tatonetti et al., 2012) with regard to the ADE of drugs (Cheng adverse effects on organs (Pizzo et al., 2015), which can be used et al., 2013b,c). in structural modification to reduce the risk by chemists. In recent years, we have worked on drug safety assessment Data Description and developed a lot of predictive models for chemical toxicity There are two ways to represent chemical structures as numeric with machine learning methods and structural alerts. A web features which can be processed by machine learning methods. server named admetSAR was also developed for publicly free One way is to use molecular descriptors, which can be calculated access (Cheng et al., 2012b). In a previous paper published from chemical structures, physicochemical or topological in 2013, we reviewed the advances and challenges of in silico properties. Currently thousands of continuous and discrete prediction of chemical toxicity together with pharmacokinetic molecular descriptors can be obtained via chemoinformatics properties (Cheng et al., 2013a). Here, we would like to review the toolkits such as PaDEL-Descriptor (Yap, 2011), OpenBabel progress of in silico chemical toxicity prediction in recent 5 years, (O’Boyle et al., 2011), CDKit (Steinbeck et al., 2003), RDKit including methodologies of machine learning and structural (Landrum, 2017), or web servers like E-Dragon (Tetko et al., alerts, and major toxicity endpoints in drug discovery and 2005), ChemBCPP (Dong et al., 2017a), and ChemDes (Dong development (Figure 1). Available data sources and web servers et al., 2015). Using numeric features may result in overfitting were also mentioned. At last, challenges and future directions in when the size of training set is small (Xue et al., 2004). Hence, this field were provided. feature selection should be done before model building, to reduce the risk of overfitting and enhance the performance of model (Sun et al., 2017). MODEL BUILDING WITH MACHINE The other way is to use molecular fingerprints, which LEARNING METHODS represent a molecule as a binary string, such as MACCS, PubChemFP, and KRFP (Klekota and Roth, 2008). In a molecular The general procedure to build a predictive model contains fingerprint, lists of substructures or other kinds of patterns are roughly four steps: data collection, data description, model predefined. If a specified pattern presents in a molecule, the building, and model evaluation. Each step has its own corresponding bit in the binary string is set to “1,” otherwise it will requirements to guarantee the reliability and accuracy of the be set to “0.” Comparing to molecular descriptors, these binary models. features are more interpretable because each bit corresponds to a specific substructure. In addition to the common fingerprints, Data Collection custom patterns can also be used to enhance the predictability of The quality of experimental data is the most important in the models (Yang et al., 2017b). model building. Currently, there are numerous well-defined data available online, which greatly facilitates the construction of Single-Label Model Building computational models by machine learning methods. In Table 1, Machine learning methods are usually used to build the we listed some widely used databases, including those linking predictive models. There are many free and open access tools chemical structures with safety outcomes, protein targets and/or and development kits to fulfill this task. For example, Scikit- biological pathways. learn (Pedregosa et al., 2011) is a popular python toolkit for TOXNET is a comprehensive source that integrates several machine learning and TensorFlow (https://www.tensorflow.org) toxicity databases such as ToxLine and ChemIDplus (Fowler and is a widely used python library for deep learning. WEKA (Frank Schnall, 2014). ACToR is a large database that aggregates data et al., 2004), Orange (Demsar et al., 2013) and RapidMiner from thousands of public sources (Judson et al., 2008). DSSTox, (https://rapidminer.com/) are machine learning toolboxes with a subset of ACToR, provides a high quality resource for toxicity GUI (Graph user interface). prediction, including ToxCast and Tox21 data (Williams-DeVane Support vector machine (SVM), Random forest (RF), boost et al., 2009). OECD established eChemPortal to provide chemical tree (BT), and k-nearest neighbor (kNN) are classic machine information including physicochemical properties, and toxicity. learning methods that are widely used in classification and Many databases are contained in eChemPortal, such as ACToR regression models. SVM, also known as support vector classifier and HSDB (Fonger et al., 2014). Some other toxicity databases (SVC) or support vector regression (SVR) in particular tasks, is include SuperToxic (Schmidt et al., 2009), T3DB (Wishart et al., well-known for its high predictive performance and less risk of 2015), and ToxBank (http://www.toxbank.net). We previously overfitting (Cortes and Vapnik, 1995). The basic idea of SVM developed a web server admetSAR, which also contains toxicity is to construct a hyperplane in a high dimensional space with data (Cheng et al., 2012b). the largest distance to the nearest training data points (support Frontiers in Chemistry | www.frontiersin.org 2 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity FIGURE 1 | The roadmap of in silico prediction of chemical toxicity with machine learning methods and structural alerts. (A) Examples of available data and web servers. (B) The state-of-the-art machine learning algorithms. (C) Scheme of building QSAR or structural alerts models for prediction of chemical toxicity. vectors). RF and BT are derived from decision tree (Breiman, recognition (Deng et al., 2013; LeCun et al., 2015). Multilayer 2001; Elith et al., 2008). RF can be viewed as bagging many neural network (MNN) is one of the DL techniques. Different decision trees that use a random subset of features and combine from common ANN that only has three layers including them via a voting system. Different from RF, in which each input layer, hidden layer and output layer (Shen et al., 2004), tree is equal, BT dynamically adjusts the weight of each tree MNN contains more than one hidden layers and thus is according to the mean error of prediction. kNN is one of the more competent in large toxicological data with complex simplest algorithms (Cover and Hart, 1967). The creed of kNN mechanisms. When the training set is large, it can perform is that compounds with similar structures have similar biological better than ANN and above-mentioned classic machine learning properties. In kNN, a sample is classified by the votes of the methods (Mayr et al., 2016). However, more complex network categories of its neighbors. indicates more weights to fit and more likely to be overfitting. Sometimes, to enhance performance of prediction models, Graph-convolutional networks (Duvenaud et al., 2015) and combination of these algorithms is applied. We developed a long short-term memory architectures (Altae-Tran et al., 2017) combined method using an artificial neural network (ANN) are recently developed to extract features from molecules model to generate the final combination decision probability, based on atom features and show better performance in which showed that the combined methods would be superior to handling thousands of compounds or even more (Goh et al., “single” methods (Cheng et al., 2011b; Du et al., 2017; Sun et al., 2017). DeepChem (https://deepchem.io) is an open source 2017). python library devoted to providing a high quality toolchain Recently, deep learning (DL) has been applied in solving to facilitate the use of DL in drug discovery and other such challenging problems as computer vision and speech fields. Frontiers in Chemistry | www.frontiersin.org 3 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity TABLE 1 | Data sources for prediction of chemical toxicity. Database name Typea URL TOXNET CTA https://toxnet.nlm.nih.gov/ ToxBank Data Warehouse CTA http://www.toxbank.net/data-warehouse admetSAR CTA http://lmmd.ecust.edu.cn/admetsar1/ Pharmaco Kinetics Knowledge Base (PKKB) CTA http://cadd.zju.edu.cn/pkkb/ ToxCast CTA https://www.epa.gov/chemical-research/toxicity-forecasting Tox21 CTA https://tripod.nih.gov/tox21 CTD (Comparative Toxicogenomics Database) CTA http://ctdbase.org/ ECOTOX CTA https://cfpub.epa.gov/ecotox/ SuperToxic CTA http://bioinformatics.charite.de/supertoxic/ DSSTox CTA https://www.epa.gov/chemical-research/distributed-structure-searchable-toxicity-dsstox-database ACToR CTA https://actor.epa.gov/actor/home.xhtml T3DB CTA http://www.t3db.ca eChemPortal CTA https://www.echemportal.org/echemportal/index.action PubChem CPI http://pubchem.ncbi.nlm.nih.gov/ ChEMBLdb CPI https://www.ebi.ac.uk/chembldb/ BindingDB CPI http://www.bindingdb.org/bind/index.jsp ChemProt CPI http://potentia.cbs.dtu.dk/ChemProt/ STITCH CPI http://stitch.embl.de/ DrugBank CPI http://www.drugbank.ca/ TTD CPI http://bidd.nus.edu.sg/group/cjttd/ IntAct MI http://www.ebi.ac.uk/intact/ SIDER SE http://sideeffects.embl.de/ MetaADEDB SE http://lmmd.ecust.edu.cn/online_services/metaadedb/ OFFSIDES SE http://www.pharmgkb.org Chemical Effects in Biological Systems (CEBS) SE http://tools.niehs.nih.gov/cebs3/ui/ IntSide SE http://intside.irbbarcelona.org Reactome Pathway http://www.reactome.org/ Pathway Commons Pathway http://www.pathwaycommons.org/ KEGG Pathway http://www.kanehisa.jp/ PharmGKB Pathway https://www.pharmgkb.org/ a CTA, compound-toxicity association; MI, molecular interaction; SE, side effect; CPI, compound-protein interaction. Multi-Label Model Building multiclass data can be utilized as the underlying base classifier Unlike aforementioned single-label classification or regression including SVM, ANN, decision tree, kNN, and so on. The models, multi-label classification (MLC) is a data mining second alternative aims for adapting existent algorithms to deal approach in which each data instance can be assigned to with multi-label data, such as multi-label C4.5 (Al-Otaibi et al., multiple categories at once (Tsoumakas et al., 2010; Zhang and 2014), multi-label back-propagation (Zhang and Zhou, 2006), Zhou, 2014; Gibaja and Ventura, 2015). The demand for multi- Rank-SVM (Wang et al., 2014), and multi-label kNN (Zhang label techniques is constantly growing in biology and genomics and Zhou, 2007). Finally, the classification ensemble is also a (Diplaris et al., 2005; Avila et al., 2009). The current algorithms widespread technique in multi-label field. For example, Ensemble used for this task are pretty new and many of them are still in an of Classifier Chain (ECC) (Read et al., 2011), which consists of early stage of development. a set of CC with diverse label orders and then votes for the There are three major approaches for multi-label learning: final prediction, is proposed to allow for the effect of chain data transformation, method adaptation and ensembles of order. Some other MLC methods based on the ensemble of classifiers. The first one, including Binary Relevance (BR) multi-class classifiers were also proposed, such as EPS (Read et al., (Godbole and Sarawagi, 2004), classifier chains (CC) (Read et al., 2008), RAkEL (Tsoumakas and Vlahavas, 2007), and HOMER 2011), and Label Powerset (LP) (Boutell et al., 2004), is to (Tsoumakas et al., 2008). transform original multi-label dataset (MLD) to a set of binary datasets (BIDs) or one multi-class dataset (MCD) first, and then Model Evaluation process them with traditional classification algorithms (Barot For regression models, three evaluation metrics, namely Pearson and Panchal, 2014). With the development of these frameworks product moment correlation coefficient (R2 ), mean absolute error for MLC, classification algorithms available for binary and (MAE) and root mean squared error (RMSE) are frequently used Frontiers in Chemistry | www.frontiersin.org 4 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity to estimate the performance of models. These parameters are where Yi represents the real label-set of the ith instance, and defined as following: Zi the predicted one. n is the number of instances and k is the number of labels.  PN
2 Furthermore, another example-based metric named ranking (xi −1 x) y i − y loss can be used. The ranking loss metric portrays how many R2 =  qP  (1) N 2 PN 2 times an irrelevant label is ranked above a relevant one according 1 (xi − x) 1 (y i − y) PN to their probabilities belonging to each label. As for label-based 1 xi − yi metrics, micro-AUC is the most commonly used one. It is also a MAE = (2) ranking based metric similar to ranking loss. However, different N s PN
2 from the ranking loss that compares the ranks for each example, 1 xi − yi micros-AUC counts the number of all the relevant-irrelevant RMSE = (3) N pairs meeting the condition that the relevant label is ranked above irrelevant one (in which the labels are not necessarily for the same where xi is the experimental value, yi is the predicted value, example). x, y are their corresponding means and N is the number of samples. For traditional single-label binary or multiple classification METHODS FOR DETECTING models, most of the performance metrics are calculated based on STRUCTURAL ALERTS the count of true positive (TP), true negative (TN), false positive (FP), and false negative (FN). Accuracy, sensitivity and specificity Structural alerts (SAs) are key substructures responsible for metrics can be calculated as the following equations to represent certain toxicity. They are directly connected to toxicity and the overall predictive ability, the predictive accuracy for positive hence could be used for structural optimization by medicinal samples and the predictive ability for negative ones: chemists to reduce the risk. In 1985, Ashby found strong associations between occurrence of some substructures or TP + TN patterns and chemical mutagenicity to Salmonella, which was Accuracy = (4) the first appearance of the concept of SA (Ashby and Tennant, TP + TN + FP + FN TP 1988). Sensitivity = (5) Till now, many methods and software have been developed TP + FN for detecting SAs, such as SARpy (Ferrari et al., 2013), MoSS, TN Specificity = (6) Gaston, and MolFea. ToxAlerts is a web server that collects SAs TN + FP defined by experts or identified by computational tools. It can predict toxicity according to the appearance of SAs (Sushko et al., In addition to these computed from binary partition of labels, 2012). Automatic detection of SAs by computational tools now metrics these calculated from a confidence degree of being becomes a hotspot as the development of cheminformatics and positive are also used like area under the receiver operating the explosion of available data (Lepailleur et al., 2013; Floris et al., characteristic curve (AUC). 2017). Comparing to the single-label classification patterns, multi- In a previous paper, we evaluated several methods for label classifiers can have multiple outputs for an instance, identification of SAs (Yang et al., 2017a). At present, the of which the predictions can be fully or partially correct. methods can be divided into three categories: fragment-based, The multi-label performance metrics introduced there can graph-based, and fingerprint-based. Fragment-based methods, be classified into two groups, i.e., example-based and label- such as SARpy (Ferrari et al., 2013), cut the bonds of the based metrics (Tsoumakas et al., 2007; Zhang and Zhou, molecules in dataset first to get all possible fragments. Then 2014). Here, five example-based metrics (subset accuracy, each fragment is evaluated according to their occurrence in Jaccard similarity coefficient, hamming-loss, micro-precision, toxic and non-toxic compounds. These methods have been used micro-recall) are described with mathematical formulations in detecting SAs for carcinogenicity (Golbamaki and Benfenati, below. 2016; Golbamaki et al., 2016). Graph-based approaches use subgraph searching algorithms, treating molecules as graphs 1 Xn SubsetAccuracy = Yi = Zi (7) that consist of a set of vertices and edges, to find the frequent n i=1 patterns. MoSS uses depth-first search association rules to 1 Xn |Yi ∩ Zi | mine substructures (Borgelt and Berthold, 2002). Gaston is a Jaccard Similarity Coefficient = (8) n i = 1 |Y i ∪ Zi | stand-alone tool that uses a graph-based approach to obtain 1 1 Xn substructures from dataset (Kazius et al., 2006). Another graph- Hamming Loss = |Yi 1Zi | (9) nk i=1 based method proposed by Ahlberg (Ahlberg et al., 2014) 1 Xn |Yi ∩ Zi | uses Atom Signature, a linear expression of a compound, to Recallmicro = (10) n i = 1 |Yi | mined sub-signature as SAs. Fingerprint-based approaches do 1 Xn |Yi ∩ Zi | not obtain fragments from the dataset. Instead, the fragments Precisionmicro = (11) are defined by different molecular fingerprints such as MACCS n i=1 |Zi | Frontiers in Chemistry | www.frontiersin.org 5 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity and SubFP (Shen et al., 2010). The selection of fingerprints may toxicity (Zhu et al., 2009). Based on the data set, several machine affect the final results of the identified SAs. Fingerprints such as learning methods were developed and applied to construct Morgan, used by Bioalerts (Cortes-Ciriano, 2016) might lead to classifiers and regression models to predict LD50 or their toxic redundant SAs which are very similar and related to the same categories (Li et al., 2014; Lei et al., 2016; Xu et al., 2017). mechanism. Noticeably, the models built by Xu et al. have high performance Information gain (IG) can also be used to evaluate the in two test sets, more than 95% of accuracy for classification significance of a substructure. Compounds containing the and 0.861 of R2 for regression, and the model is free available substructure are categorized as toxic and others are categorized in web server (http://www.pkumdl.cn/DLAOT/DLAOThome. as non-toxic. IG is defined as the difference between the php). information entropy of original dataset and the weighted average information entropies of two datasets separated by a substructure Cardiotoxicity (Sokolova and Szpakowicz, 2010). We previously used IG to Blockade of the hERG (human ether-a-go-go related gene) detect privileged substructures whose occurrences have strong potassium channel is the main adverse effect with regard to relevance to some endpoints (Shen et al., 2010). cardiotoxicity (Gintant et al., 2016). Several in silico models were developed according to the in vitro hERG blockage test PROGRESS IN TOXICITY PREDICTION in early screening assays. Our group recently developed an in silico model that used chemical category approaches to Carcinogenicity and Mutagenicity predict hERG blockage (Zhang et al., 2016b), in which 1,570 Chemical carcinogenesis is of increasing importance in drug unique compounds were collected from ChEMBL database and discovery for its serious effect on human health. Most of the early studies (Doddareddy et al., 2010; Wang et al., 2012). predictive models use Carcinogenic Potency Database (CPDB) as In addition to machine learning methods, combination with the data source, which contains more than 1,500 chemicals with multiple pharmacophores can improve the predictive capabilities their labels (carcinogen or non-carcinogen) according to their and the model would be more interpretable (Wang et al., TD50 values (Gold et al., 2005). Recently several publications 2016). shared their protocols to construct models to predict chemical However, as the simplified in vitro approaches for detection of carcinogenesis, including Naïve Bayes, kNN, probabilistic neural cardiac safety are less specific, the in silico models will also output network, and SVM (Singh et al., 2013; Tanabe et al., 2013; Li et al., the false-positive predictions that may result in unwarranted 2015; Zhang H. et al., 2016). Zhang et al. developed a web server, attribution of novel drug candidates (Gintant et al., 2016). Other CarcinoPred-EL, for chemists to predict carcinogenicity online, categories such as contractile and structural cardiotoxicity should in which Ensemble XGBoost was used to build the model (Zhang be considered and more in vitro or in vivo data should be used to et al., 2017). construct sophisticated models. Due to its complicated mechanism and less available data, the predictive models based on phenotypic assays are Hepatotoxicity not precise and reliable enough. It is an alternative to Chemical hepatotoxicity in drug discovery, also termed “drug construct models based on in vitro assays. The mechanisms induced liver injury (DILI),” is the leading cause for drug of carcinogenesis of chemicals can be categorized into: (1) failure or withdrawn from the market (Schuster et al., 2005). genotoxicity, which are primarily caused by the mutagenicity Due to its complicated mechanism and inconsistency in diverse of chemicals damaging DNA (Fan et al., in press); (2) patients, experimental detection of hepatotoxicity in preclinical non-genotoxic carcinogens acting through different specific and clinical trials is difficult. mechanisms, which are more complicated (Golbamaki and Computational approaches to predict DILI of compounds Benfenati, 2016). Ames test devised by Bruce Ames is a well- are widely applied for their low cost and high efficiency. known in vitro assay to detect mutagenic effects of chemicals. Hewitt reviewed the in silico models on DILI prediction from Currently more than 8,000 compounds with Ames mutagenicity 2000 to 2015, including statistics-based methods and expert are available. Both predictive models and structural alerts were systems (Hewitt and Przybylak, 2016). Chemical or hybrid promoted with these toxicity data in recent years (Kazius descriptors as features, and different machine learning methods et al., 2005; Hansen et al., 2009; Xu et al., 2012; Yang et al., such as linear discriminant analysis and ANN were used in 2017a). these models to predict general or specific endpoints related to hepatotoxicity (Hewitt and Przybylak, 2016). Zhu constructed Acute Oral Toxicity a human hepatotoxicity database for QSTR models using post- According to the exposure routes of chemicals, acute toxicity market safety data originated from FDA adverse event reporting can be divided into oral, dermal and inhalation, among which system (Zhu and Kruhlak, 2014). Our group previously used acute oral toxicity is the most widely studied in computational molecular fingerprints and machine learning methods to build prediction. It is often the first performed endpoint in drug classification models with a data set containing 1,317 diverse discovery because any compounds causing acute toxicity will compounds (Zhang et al., 2016a). Xu et al. used a deep learning not be further considered for its strong hazardous to human method called undirected graph recursive neural networks health. Zhu et al. collected 7,385 compounds with LD50 values (UGRNN) that encodes molecules into an undirected graph to and built several models for prediction of chemical acute oral build QSTR models (Xu et al., 2015). The performance was Frontiers in Chemistry | www.frontiersin.org 6 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity excellent compared to other models, up to 0.955 of AUC. More proliferation, reproductive disorders, metabolic disorders, or recently, Mulliner et al. classified the complex pathology of even cancers (Colborn, 1995; Chawla et al., 2001; Grün and hepatotoxicity into 21 endpoints at three levels, with a large Blumberg, 2007). data set comprising 3,712 compounds. Then the specific models For the specific mechanisms such as binding to ER, using were combined into an optimized global human hepatotoxicity in silico models to predict the bioactivity of chemicals and that has high sensitivity of 68% and excellent specificity of 95% evaluate their risk of being EDCs is preferred for its high (Mulliner et al., 2016). accuracy and less cost. We previously built in silico models for AR and ER binding using molecular fingerprints and machine Respiratory Toxicity learning methods and the best performance in the test set Respiratory toxicity is another toxicity category with complicated was 0.84 and 0.79, respectively (Chen et al., 2014). The Tox21 mechanisms. The most concerned endpoint is drug-induced project also includes nuclear receptors assays which involve interstitial lung disease (DILD), which can be classified into more diverse compounds (Hsieh et al., 2015). DeepTox, the two categories in terms of their mechanisms: (1) cytotoxic lung winner of the “Tox21 Data Challenge,” used deep neural network injury and (2) immune-mediated (Matsuno, 2012). Another type and obtained an excellent performance against other machine of respiratory toxicity is respiratory sensitization, of which the learning methods such as SVM (Mayr et al., 2016). mechanism is more complicated. There are still no good models Previous studies on EDCs mainly focused on nuclear for identification of respiratory sensitization (Mekenyan et al., receptors. However, chemicals that do not directly interact 2014; Dik et al., 2015). The current QSTR studies tend to use with these receptors may also interfere through the pathway. phenotype data such as LD50 , LC50 or symptoms such as asthma For instance, aromatase (CYP19A1) is an important enzyme as endpoints to represent the respiratory toxicity of a chemical, affecting the biosynthesis of estrogen and plays a key role in and the built models performed well enough (Jarvis et al., 2015; maintaining the balance between estrogen and androgen in many Lei et al., 2017). of the EDC-sensitive organs (Sonnet et al., 1998). Therefore, we recently built in silico models for prediction of aromatase Irritation and Corrosion inhibitors as potential EDCs using machine learning methods Risk assessment of eye and skin irritation/corrosion (EI/EC, with molecular fingerprints (Du et al., 2017). The data used for SI/SC) is of importance in pharmaceutical and cosmetics training and test were collected from Tox21 and the best model industries. Though these endpoints might not be directly had 0.84 of accuracy for the test set and 0.91 for the external considered in drug discovery stage, in silico models for these validation set. endpoints are yet required since a lot of substances may cause irritation and corrosion and should be assessed, including the Eco-Toxicity ocular and dermal pharmaceuticals and final products used in Pharmaceuticals and their metabolites exposed to the manufacturing, agriculture, and warfare (Wilhelmus, 2001; Kolle environment may affect the ecosystem since they are designed et al., 2017). to be bioactive to creature (Halling-Sørensen et al., 1998). For Verheyen et al. evaluated the existing QSTR models in Derek instance, chemicals with binding affinities to hormone receptors Nexus, Toxtree and Case Ultra for the prediction of skin and may be EDCs of fishes or concentrate in fish body and finally eye irritation/corrosion, and found that the performance of reach to high-level animal bodies (He et al., 2017). To evaluate those models is unsatisfactory because of narrow applicability the environmental persistence of a chemical, biodegradation domain and low accuracy (Verheyen et al., 2017). However, using half-life is widely used as a common criterion (Raymond machine learning methods to predict eye injury was reported et al., 2001). We previously categorized chemicals as ready having high performance. For instance, Verma et al. build biodegradability and not ready biodegradability according to combined QSTR models by ANN and got 88% of sensitivity and their biological oxygen demand (BOD) with a threshold of 60% 82% of specificity for EI (Verma and Matthews, 2015a), 96% of and built several classification models. The best model used kNN sensitivity and 91% of specificity for EC (Verma and Matthews, with molecular descriptors and had a AUC of 0.873 in test set 2015b). Our group recently developed in silico models for EI/EC (Cheng et al., 2012a). using machine learning methods and molecular fingerprints Fishes are usually used as model species to evaluate aquatic (Wang et al., 2017). In the paper, more positive data were toxicity and avian species are widely used as model species to manually collected from X-Mol (http://www.x-mol.com) and evaluate the terrestrial toxicity. Our group previously collected ChemIDplus and the performance is excellent, 94.6% of overall LC50 data of three fish species from ECOTOX database and built accuracy for EI and 95.9% for EC. several local and global models (Sun et al., 2015). Recently, we reported a model focusing on the aquatic toxicity of pesticides Endocrine Disruption and found that the molecule fingerprints performed different Chemicals interacting with nuclear receptors such as estrogen between local and global models (Li et al., 2017). For the and androgen receptors (ER and AR) as off-targets or exposed in avian species, several in silico models were developed including environment may cause endocrine disruption. These chemicals, classification (Zhang et al., 2015) and regression (Mazzatorta called endocrine disrupting chemicals (EDCs), may interfere et al., 2006; Toropov and Benfenati, 2006). In addition to the with the normal functions of these endogenous steroid hormones endpoints mentioned above, another commonly used model and lead to adverse health consequences such as tissue or organ species for eco-toxicology is Tetrahymena pyriformis (Sauvant Frontiers in Chemistry | www.frontiersin.org 7 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity et al., 1999). Cheng et al. collected 1,571 unique chemicals with limitations to be improved. At first, data quality is still a big issue. toxicity to Tetrahymena pyriformis and built several models of Currently many toxicity data are obtained from high-throughput which the best performance was 92.6% for validation set (Cheng in vitro assays or in vivo tests on animals. For example, Tox21 et al., 2011a). and ToxCast provide the activity data of thousands of chemicals against hundreds of assays (Huang et al., 2016). While false SOFTWARE AND WEB SERVERS positive and false negative data are inevitable in those assays, in vivo data from animals are also questionable to be used directly Currently many software and web servers can predict chemical on humans. Therefore, more data from drug clinical trials and toxicity before synthesis. Drug design software suites such as clinic applications are highly demanded. Discovery Studio and Pipeline Pilot integrate toxicity prediction Secondly, more computational methods should be developed models to help filter compounds with risk of toxicity. But the to enhance the accuracy of the predictive models. For instance, endpoints are not as diverse as that in some toxicity-oriented read-across has gained wide attention recently because it can fill commercial software including ADMET Predictor, Leadscope the gap of missing data (Shah et al., 2016). Meanwhile, some and Lhasa Derek, which take efforts primarily on predicting and endpoints have complex mechanisms such as hepatotoxicity alerting molecules with potential toxicity. and respiratory toxicity, computational systems toxicology has Free software or web servers are more preferred by academia, emerged to use comprehensive data sources from gene to organ which can promote the development of high quality models to understand the mechanisms of toxicity (Jack et al., 2013; and algorithms, and their applications in various fields including Sauer et al., 2015). With the help of machine learning methods drug discovery. OCED Toolbox is an official suite for toxicity and cheminformatics techniques, more accurate models could be prediction and modeling using QSTR. Web servers are easier developed for toxicity prediction. and lighter to use and will be preferred by outsiders of Thirdly, medicinal chemists are more interested in the computational toxicology, such as medicinal chemists. Lazar is relationship between substructures and chemical toxicity, such a tool that can predict several toxicity endpoints with a which can guide the optimization of lead compounds. Using user interface of drawing chemical structures (Maunz et al., computational tools to identify SAs is a promising way. Current 2013). ToxTree is an open source application that estimates approaches of SA identification can only generate numerous toxic hazard by applying a decision tree approach (Patlewicz but redundant substructures in terms of their frequency of et al., 2008). Compared to QSTR-like models, ToxTree is occurrence, disregarding the chemical or biological mechanisms more interpretable and the fragments (SAs) can guide the (Yang et al., 2017a). It is not difficult to obtain “potential” SAs chemists in modification of the molecules. The performance for almost every endpoint with support of assay results, yet of ToxTree, OECD Toolbox, and other commercial tools innovative protocol or framework is still required to further were compared in literature (Devillers and Mombelli, 2010; refine these substructures and explore the chemical mechanisms Mombelli and Devillers, 2010; Bhatia et al., 2015; Bhhatarai of toxicity. et al., 2016). Our group developed admetSAR that can also predict toxicity of compounds in SMILES format (Cheng et al., AUTHOR CONTRIBUTIONS 2012b). Web servers such as ChemSAR (Dong et al., 2017b) and YT, GL, and WL contributed conception and design of the study; ChemBench (Capuzzi et al., 2017) enable users to build custom HY wrote the first draft of the manuscript; HY and LS wrote models for particular use with machine learning methods and sections of the manuscript. All authors contributed to manuscript molecular descriptors. For chemists who have in-house data for revision, read and approved the submitted version. some particular endpoints, it will be convenient to use these web servers to build predictive models to prioritize or substitute ACKNOWLEDGMENTS in vitro or in vivo tests. This work was supported by the National Key Research and PERSPECTIVES Development Program of China (Grant 2016YFA0502304), the National Natural Science Foundation of China (Grants Though in silico prediction of chemical toxicity has made a 81373329 and 81673356) and the 863 Project (Grant good progress in recent years, there are still some challenges and 2012AA020308). REFERENCES Altae-Tran, H., Ramsundar, B., Pappu, A. S., and Pande, V. (2017). Low data drug discovery with one-shot learning. Ahlberg, E., Carlsson, L., and Boyer, S. (2014). Computational derivation of ACS Cent. Sci. 3, 283–293. doi: 10.1021/acscentsci.6b structural alerts from large toxicology data sets. J. Chem. Inf. Model. 54, 00367 2945–2952. doi: 10.1021/ci500314a Ashby, J., and Tennant, R. W. (1988). Chemical structure, Salmonella Al-Otaibi, R., Kull, M., and Flach, P. (2014). “LaCova: a tree-based multi-label mutagenicity and extent of carcinogenicity as indicators of genotoxic classifier using label covariance as splitting criterion,” in 2014 13th International carcinogenesis among 222 chemicals tested in rodents by the U.S. Conference on Machine Learning and Applications, (Detroit, MI: Icmla), NCI/NTP. Mutat. Res. 204, 17–115. doi: 10.1016/0165-1218(88) 74–79. 90114-0 Frontiers in Chemistry | www.frontiersin.org 8 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity Avila, J. L., Gibaja, E. L., and Ventura, S. (2009). Multi-label Classification Demsar, J., Curk, T., Erjavec, A., Gorup, C., Hocevar, T., Milutinovic, M., et al. with gene expression programming. Hybrid Artif. Intell. Syst. 5572, 629–637. (2013). Orange: data mining toolbox in python. J. Mach. Learn. Res. 14, doi: 10.1007/978-3-642-02319-4_76 2349–2353. Available online at: https://orange.biolab.si/citation/ Barot, P., and Panchal, M. (2014). Review on various problem transformation Deng, L., Hinton, G., and Kingsbury, B. (2013). “New types of deep neural network methods for classifying multi-label data. Int. J. Data Min. Emerg. Technol. 4, learning for speech recognition and related applications: an overview,”in IEEE 45–52. doi: 10.5958/2249-3220.2014.00001.9 International Conference on Acoustics, Speech and Signal Processing (Vancouver, Bhatia, S., Schultz, T., Roberts, D., Shen, J., Kromidas, L., and Marie Api, A. BC), 8599–8603. (2015). Comparison of Cramer classification between Toxtree, the OECD Devillers, J., and Mombelli, E. (2010). Evaluation of the OECD QSAR QSAR Toolbox and expert judgment. Regul. Toxicol. Pharmacol. 71, 52–62. Application Toolbox and Toxtree for estimating the mutagenicity of doi: 10.1016/j.yrtph.2014.11.005 chemicals. Part 1. Aromatic amines. SAR QSAR Environ. Res. 21, 753-769. Bhhatarai, B., Wilson, D. M., Parks, A. K., Carney, E. W., and Spencer, P. J. doi: 10.1080/1062936X.2010.528959 (2016). Evaluation of TOPKAT, toxtree, and derek nexus in silico models Dik, S., Pennings, J. L., van Loveren, H., and Ezendam, J. (2015). Development for ocular irritation and development of a knowledge-based framework to of an in vitro test to identify respiratory sensitizers in bronchial epithelial improve the prediction of severe irritation. Chem. Res. Toxicol. 29, 810–822. cells using gene expression profiling. Toxicol. In Vitro 30(1 Pt B), 274-280. doi: 10.1021/acs.chemrestox.5b00531 doi: 10.1016/j.tiv.2015.10.010 Borgelt, C., and Berthold, M. R. (2002). “Mining molecular fragments: finding Diplaris, S., Tsoumakas, G., Mitkas, P. A., and Vlahavas, I. (2005). Protein relevant substructures of molecules,” in Data Mining, (2002). ICDM 2003. classification with multiple algorithms. Adv. Inform. Proc. 3746, 448–456. Proceedings 2002 IEEE International Conference (Maebashi: IEEE), 51–58. doi: 10.1007/11573036_42 Boutell, M. R., Luo, J. B., Shen, X. P., and Brown, C. M. (2004). Learning Doddareddy, M. R., Klaasse, E. C., Shagufta, Ijzerman, A. P., and Bender, A. multi-label scene classification. Pattern Recognit. 37, 1757–1771. (2010). Prospective validation of a comprehensive in silico hERG model and its doi: 10.1016/j.patcog.2004.03.009 applications to commercial compound and drug databases. Chem. Med. Chem. Breiman, L. (2001). Random forests. Mach. Learn. 45, 5–32. 5, 716–729. doi: 10.1002/cmdc.201000024 doi: 10.1023/A:1010933404324 Dong, J., Cao, D. S., Miao, H. Y., Liu, S., Deng, B. C., Yun, Y. H., et al. (2015). Capuzzi, S. J., Kim, I. S., Lam, W. I., Thornton, T. E., Muratov, E. N., Pozefsky, ChemDes: an integrated web-based platform for molecular descriptor and D., et al. (2017). Chembench: A publicly accessible, integrated cheminformatics fingerprint computation. J. Cheminform. 7:60. doi: 10.1186/s13321-015-0109-z portal. J. Chem. Inf. Model. 57, 105–108. doi: 10.1021/acs.jcim.6b00462 Dong, J., Wang, N.-N., Liu, K.-Y., Zhu, M.-F., Yun, Y.-H., Zeng, W.-B., et al. Chawla, A., Repa, J. J., Evans, R. M., and Mangelsdorf, D. J. (2001). Nuclear (2017a). ChemBCPP: a freely available web server for calculating commonly receptors and lipid physiology: opening the X-files. Science 294, 1866–1870. used physicochemical properties. Chemometr. Intell. Lab. Syst. 171, 65–73. doi: 10.1126/science.294.5548.1866 doi: 10.1016/j.chemolab.2017.10.006 Chen, Y., Cheng, F., Sun, L., Li, W., Liu, G., and Tang, Y. (2014). Dong, J., Yao, Z. J., Zhu, M. F., Wang, N. N., Lu, B., Chen, A. F., et al. (2017b). Computational models to predict endocrine-disrupting chemical binding with ChemSAR: an online pipelining platform for molecular SAR modeling. J. androgen or oestrogen receptors. Ecotoxicol. Environ. Saf. 110, 280–287. Cheminform. 9:27. doi: 10.1186/s13321-017-0215-1 doi: 10.1016/j.ecoenv.2014.08.026 Du, H., Cai, Y., Yang, H., Zhang, H., Xue, Y., Liu, G., et al. (2017). In silico Cheng, F., Ikenaga, Y., Zhou, Y., Yu, Y., Li, W., Shen, J., et al. (2012a). In silico prediction of chemicals binding to aromatase with machine learning methods. assessment of chemical biodegradability. J. Chem. Inf. Model. 52, 655–669. Chem. Res. Toxicol. 30, 1209–1218. doi: 10.1021/acs.chemrestox.7b00037 doi: 10.1021/ci200622d Duvenaud, D., Maclaurin, D., Aguilera-Iparraguirre, J., Gómez-Bombarelli, R., Cheng, F., Li, W., Liu, G., and Tang, Y. (2013a). In silico ADMET prediction: Hirzel, T., and Aspuru-Guzik, A. N., et al. (2015). Convolutional Networks on recent advances, current challenges and future trends. Curr. Top. Med. Chem. Graphs for Learning Molecular Fingerprints. ArXiv e-prints [Online], (1509). 13, 1273–1289. doi: 10.2174/15680266113139990033 Available online at: http://adsabs.harvard.edu/abs/2015arXiv150909292D Cheng, F., Li, W., Wang, X., Zhou, Y., Wu, Z., Shen, J., et al. (2013b). Adverse drug (Accessed Sept 1, 2015). events: database construction and in silico prediction. J. Chem. Inf. Model. 53, Elith, J., Leathwick, J. R., and Hastie, T. (2008). A working 744–752. doi: 10.1021/ci4000079 guide to boosted regression trees. J. Anim. Ecol. 77, 802–813. Cheng, F., Li, W., Wu, Z., Wang, X., Zhang, C., Li, J., et al. (2013c). Prediction of doi: 10.1111/j.1365-2656.2008.01390.x polypharmacological profiles of drugs by the integration of chemical, side effect, Fan, D., Yang, H., Li, F., Sun, L., Di, P., Li, W., et al. (in press). In silico prediction and therapeutic space. J. Chem. Inf. Model. 53, 753–762. doi: 10.1021/ci400010x of chemical genotoxicity using machine learning methods and structural alerts. Cheng, F., Li, W., Zhou, Y., Shen, J., Wu, Z., Liu, G., et al. (2012b). Toxicol. Res. doi: 10.1039/C7TX00259A admetSAR: a comprehensive source and free tool for assessment of chemical Ferrari, T., Cattaneo, D., Gini, G., Golbamaki Bakhtyari, N., Manganaro, A., ADMET properties. J. Chem. Inf. Model. 52, 3099–3105. doi: 10.1021/ and Benfenati, E. (2013). Automatic knowledge extraction from chemical ci300367a structures: the case of mutagenicity prediction. SAR QSAR Environ. Res. 24, Cheng, F., Shen, J., Yu, Y., Li, W., Liu, G., Lee, P. W., et al. (2011a). In 631–649. doi: 10.1080/1062936X.2013.773376 silico prediction of Tetrahymena pyriformis toxicity for diverse industrial Floris, M., Raitano, G., Medda, R., and Benfenati, E. (2017). Fragment chemicals with substructure pattern recognition and machine learning prioritization on a large mutagenicity dataset. Mol. Inform. 36:1600133. methods. Chemosphere 82, 1636–1643. doi: 10.1016/j.chemosphere.2010.11.043 doi: 10.1002/minf.201600133 Cheng, F., Yu, Y., Zhou, Y., Shen, Z., Xiao, W., Liu, G., et al. (2011b). Insights into Fonger, G. C., Hakkinen, P., Jordan, S., and Publicker, S. (2014). The National molecular basis of cytochrome p450 inhibitory promiscuity of compounds. J. Library of Medicine’s (NLM) Hazardous Substances Data Bank (HSDB): Chem. Inf. Model. 51, 2482–2495. doi: 10.1021/ci200317s background, recent enhancements and future plans. Toxicology 325, 209–216. Colborn, T. (1995). Environmental estrogens: health implications for humans and doi: 10.1016/j.tox.2014.09.003 wildlife. Environ. Health Perspect. 103 (Suppl. 7), 135–136. Fowler, S., and Schnall, J. G. (2014). TOXNET: information on Cortes, C., and Vapnik, V. (1995). Support-Vector Networks. Mach. Learn. 20, toxicology and environmental health. Am. J. Nurs. 114, 61–63. 273–297. doi: 10.1007/BF00994018 doi: 10.1097/01.NAJ.0000443783.75162.79 Cortes-Ciriano, I. (2016). Bioalerts: a python library for the derivation of Frank, E., Hall, M., Trigg, L., Holmes, G., and Witten, I. H. (2004). structural alerts from bioactivity and toxicity data sets. J. Cheminform. 8:13. Data mining in bioinformatics using Weka. Bioinformatics 20, 2479–2481. doi: 10.1186/s13321-016-0125-7 doi: 10.1093/bioinformatics/bth261 Cover, T., and Hart, P. (1967). Nearest neighbor pattern classification. IEEE Trans. Gaulton, A., Hersey, A., Nowotka, M., Bento, A. P., Chambers, J., Mendez, D., Inform. Theory 13, 21–27. doi: 10.1109/TIT.1967.1053964 et al. (2017). The ChEMBL database in 2017. Nucleic Acids Res. 45, D945–D954. Davis, A. P., Grondin, C. J., Johnson, R. J., Sciaky, D., King, B. L., McMorran, R., doi: 10.1093/nar/gkw1074 et al. (2017). The comparative toxicogenomics database: update 2017. Nucleic Gibaja, E., and Ventura, S. (2015). A tutorial on multilabel learning. Acm Comput. Acids Res. 45, D972–D978. doi: 10.1093/nar/gkw838 Surveys 47, 1–38. doi: 10.1145/2716262 Frontiers in Chemistry | www.frontiersin.org 9 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity Gilson, M. K., Liu, T., Baitaluk, M., Nicola, G., Hwang, L., and Chong, J. (2016). Kazius, J., McGuire, R., and Bursi, R. (2005). Derivation and validation BindingDB in 2015: a public database for medicinal chemistry, computational of toxicophores for mutagenicity prediction. J. Med. Chem. 48, 312–320. chemistry and systems pharmacology. Nucleic Acids Res. 44, D1045–D1053. doi: 10.1021/jm040835a doi: 10.1093/nar/gkv1072 Kazius, J., Nijssen, S., Kok, J., Bäck, T., and Ijzerman, A. P. (2006). Substructure Gintant, G., Sager, P. T., and Stockbridge, N. (2016). Evolution of strategies to mining using elaborate chemical representation. J. Chem. Inf. Model. 46, improve preclinical cardiac safety testing. Nat. Rev. Drug Discov. 15, 457–471. 597–605. doi: 10.1021/ci0503715 doi: 10.1038/nrd.2015.34 Klekota, J., and Roth, F. P. (2008). Chemical substructures that enrich for biological Giri, S., and Bader, A. (2015). A low-cost, high-quality new drug discovery process activity. Bioinformatics 24, 2518–2525. doi: 10.1093/bioinformatics/btn479 using patient-derived induced pluripotent stem cells. Drug Discov. Today 20, Kolle, S. N., van Ravenzwaay, B., and Landsiedel, R. (2017). Regulatory accepted 37–49. doi: 10.1016/j.drudis.2014.10.011 but out of domain: in vitro skin irritation tests for agrochemical formulations. Godbole, S., and Sarawagi, S. (2004). Discriminative methods for multi- Regul. Toxicol. Pharmacol. 89, 125–130. doi: 10.1016/j.yrtph.2017.07.016 labeled classification. Adv. Knowl. Discov. Data Min. Proc. 3056, 22–30. Kuhn, M., Campillos, M., Letunic, I., Jensen, L. J., and Bork, P. (2010). A side doi: 10.1007/978-3-540-24775-3_5 effect resource to capture phenotypic effects of drugs. Mol. Syst. Biol. 6:343. Goh, G. B., Hodas, N. O., and Vishnu, A. (2017). Deep learning for computational doi: 10.1038/msb.2009.98 chemistry. J. Comput. Chem. 38, 1291–1307. doi: 10.1002/jcc.24764 Landrum, G. (2017). RDKit. Available online at: http://www.rdkit.org. Golbamaki, A., and Benfenati, E. (2016). In silico methods for LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436–444. carcinogenicity assessment. Methods Mol. Biol. 1425, 107–119. doi: 10.1038/nature14539 doi: 10.1007/978-1-4939-3609-0_6 Lei, T., Chen, F., Liu, H., Sun, H., Kang, Y., Li, D., et al. (2017). ADMET Evaluation Golbamaki, A., Benfenati, E., Golbamaki, N., Manganaro, A., Merdivan, in drug discovery. part 17: development of quantitative and qualitative E., Roncaglioni, A., et al. (2016). New clues on carcinogenicity-related prediction models for chemical-induced respiratory toxicity. Mol. Pharm. 14, substructures derived from mining two large datasets of chemical compounds. 2407-2421. doi: 10.1021/acs.molpharmaceut.7b00317 J. Environ. Sci. Health C Environ. Carcinog. Ecotoxicol. Rev. 34, 97–113. Lei, T., Li, Y., Song, Y., Li, D., Sun, H., and Hou, T. (2016). ADMET evaluation doi: 10.1080/10590501.2016.1166879 in drug discovery: 15. Accurate prediction of rat oral acute toxicity using Gold, L. S., Manley, N. B., Slone, T. H., Rohrbach, L., and Garfinkel, G. B. relevance vector machine and consensus modeling. J. Cheminformat. 8:6. (2005). Supplement to the Carcinogenic Potency Database (CPDB): results doi: 10.1186/S13321-016-0117-7. of animal bioassays published in the general literature through 1997 and by Lepailleur, A., Poezevara, G., and Bureau, R. (2013). Automated detection of the National Toxicology Program in 1997-1998. Toxicol. Sci. 85, 747–808. structural alerts (chemical fragments) in (eco)toxicology. Comput. Struct. doi: 10.1093/toxsci/kfi161 Biotechnol. J. 5:e201302013. doi: 10.5936/csbj.201302013 Grün, F., and Blumberg, B. (2007). Perturbed nuclear receptor signaling by Li, F., Fan, D., Wang, H., Yang, H., Li, W., Tang, Y., et al. (2017). In silico prediction environmental obesogens as emerging factors in the obesity crisis. Rev. Endocr. of pesticide aquatic toxicity with chemical category approaches. Toxicol. Res. 6, Metab. Disord. 8, 161–171. doi: 10.1007/s11154-007-9049-x 831–842. doi: 10.1039/C7TX00144D Halling-Sørensen, B., Nors Nielsen, S., Lanzky, P. F., Ingerslev, F., Holten Li, X., Chen, L., Cheng, F., Wu, Z., Bian, H., Xu, C., et al. (2014). In silico prediction Lützhøft, H. C., and Jørgensen, S. E. (1998). Occurrence, fate and effects of chemical acute oral toxicity using multi-classification methods. J. Chem. Inf. of pharmaceutical substances in the environment–a review. Chemosphere 36, Model. 54, 1061–1069. doi: 10.1021/ci5000467 357–393. doi: 10.1016/S0045-6535(97)00354-8 Li, X., Du, Z., Wang, J., Wu, Z. R., Li, W. H., Liu, G. X., et al. (2015). In silico Hansen, K., Mika, S., Schroeter, T., Sutter, A., ter Laak, A., Steger-Hartmann, estimation of chemical carcinogenicity with binary and ternary classification T., et al. (2009). Benchmark data set for in Silico prediction of ames methods. Mol. Inform. 34, 228–235. doi: 10.1002/minf.201400127 mutagenicity. J. Chem. Inf. Model. 49, 2077–2081. doi: 10.1021/ci90 Matsuno, O. (2012). Drug-induced interstitial lung disease: mechanisms and best 0161g diagnostic approaches. Respir. Res. 13:39. doi: 10.1186/1465-9921-13-39 He, J., Peng, T., Yang, X., and Liu, H. (2017). Development of QSAR Maunz, A., Gütlein, M., Rautenberg, M., Vorgrimmler, D., Gebele, D., and Helma, models for predicting the binding affinity of endocrine disrupting chemicals C. (2013). lazar: a modular predictive toxicology framework. Front. Pharmacol. to eight fish estrogen receptor. Ecotoxicol. Environ. Saf. 148, 211–219. 4:38. doi: 10.3389/fphar.2013.00038 doi: 10.1016/j.ecoenv.2017.10.023 Mayr, A., Klambauer, G., Unterthiner, T., and Hochreiter, S. (2016). Hewitt, M., and Przybylak, K. (2016). In silico models for hepatotoxicity. Methods DeepTox: toxicity prediction using deep learning. Front. Environ. Sci. Mol. Biol. 1425, 201–236. doi: 10.1007/978-1-4939-3609-0_11 3:80. doi: 10.3389/fenvs.2015.00080 Hsieh, J. H., Sedykh, A., Huang, R., Xia, M., and Tice, R. R. (2015). A Mazzatorta, P., Cronin, M. T. D., and Benfenati, E. (2006). A QSAR study of data analysis pipeline accounting for artifacts in Tox21 quantitative avian oral toxicity using support vector machines and genetic algorithms. QSAR high-throughput screening assays. J. Biomol. Screen. 20, 887–897. Comb. Sci. 25, 616–628. doi: 10.1002/qsar.200530189 doi: 10.1177/1087057115581317 Mekenyan, O., Patlewicz, G., Kuseva, C., Popova, I., Mehmed, A., Kotov, S., et al. Huang, R., Xia, M., Sakamuru, S., Zhao, J., Shahane, S. A., Attene-Ramos, (2014). A mechanistic approach to modeling respiratory sensitization. Chem. M., et al. (2016). Modelling the Tox21 10 K chemical profiles for in vivo Res. Toxicol. 27, 219–239. doi: 10.1021/tx400345b toxicity prediction and mechanism characterization. Nat. Commun. 7:10425. Mombelli, E., and Devillers, J. (2010). Evaluation of the OECD (Q)SAR doi: 10.1038/ncomms10425 Application Toolbox and Toxtree for predicting and profiling the Huh, D., Hamilton, G. A., and Ingber, D. E. (2011). From 3D cell culture carcinogenic potential of chemicals. SAR QSAR Environ. Res. 21, 731–752. to organs-on-chips. Trends Cell Biol. 21, 745–754. doi: 10.1016/j.tcb.2011. doi: 10.1080/1062936X.2010.528598 09.005 Mulliner, D., Schmidt, F., Stolte, M., Spirkl, H. P., Czich, A., and Amberg, Huh, D., Matthews, B. D., Mammoto, A., Montoya-Zavala, M., Hsin, H. Y., and A. (2016). Computational models for human and animal hepatotoxicity Ingber, D. E. (2010). Reconstituting organ-level lung functions on a chip. with a global application scope. Chem. Res. Toxicol. 29, 757–767. Science 328, 1662–1668. doi: 10.1126/science.1188302 doi: 10.1021/acs.chemrestox.5b00465 Jack, J., Wambaugh, J., and Shah, I. (2013). Systems toxicology from genes O’Boyle, N. M., Banck, M., James, C. A., Morley, C., Vandermeersch, T., and to organs. Methods Mol. Biol. 930, 375–397. doi: 10.1007/978-1-62703- Hutchison, G. R. (2011). Open Babel: an open chemical toolbox. J. Cheminform. 059-5_17 3:33. doi: 10.1186/1758-2946-3-33 Jarvis, J., Seed, M. J., Stocks, S. J., and Agius, R. M. (2015). A refined QSAR model Patlewicz, G., Jeliazkova, N., Safford, R. J., Worth, A. P., and Aleksiev, B. for prediction of chemical asthma hazard. Occup. Med. (Lond). 65, 659–666. (2008). An evaluation of the implementation of the cramer classification doi: 10.1093/occmed/kqv105 scheme in the toxtree software. SAR QSAR Environ. Res. 19, 495–524. Judson, R., Richard, A., Dix, D., Houck, K., Elloumi, F., Martin, M., et al. doi: 10.1080/10629360802083871 (2008). ACToR–Aggregated computational toxicology resource. Toxicol. Appl. Paul, S. M., Mytelka, D. S., Dunwiddie, C. T., Persinger, C. C., Munos, B. Pharmacol. 233, 7–13. doi: 10.1016/j.taap.2007.12.037 H., Lindborg, S. R., et al. (2010). How to improve R&D productivity: the Frontiers in Chemistry | www.frontiersin.org 10 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity pharmaceutical industry’s grand challenge. Nat. Rev. Drug Discov. 9, 203–214. Sushko, I., Salmina, E., Potemkin, V. A., Poda, G., and Tetko, I. V. (2012). doi: 10.1038/nrd3078 ToxAlerts: a web server of structural alerts for toxfic chemicals and compounds Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Weiss, R., Dubourg, V., with potential adverse reactions. J. Chem. Inf. Model. 52, 2310–2316. et al. (2011). Scikit-learn: machine learning in python. J. Mach. Learn. Res. doi: 10.1021/ci300245q 12, 2825–2830. Available online at: http://scikit-learn.org/stable/about.html# Tanabe, K., Kurita, T., Nishida, K., Lucić, B., Amić, D., and Suzuki, T. (2013). citing-scikit-learn Improvement of carcinogenicity prediction performances based on sensitivity PhRMA (2015). 2015 Biopharmaceutical Research Industry Profle. Washington, analysis in variable selection of SVM models. SAR QSAR Environ. Res. 24, DC: Pharmaceutical Research and Manufacturers of America. 565–580. doi: 10.1080/1062936X.2012.762425 Pizzo, F., Gadaleta, D., Lombardo, A., Nicolotti, O., and Benfenati, E. (2015). Tatonetti, N. P., Ye, P. P., Daneshjou, R., and Altman, R. B. (2012). Data- Identification of structural alerts for liver and kidney toxicity using repeated driven prediction of drug effects and interactions. Sci. Transl. Med. 4:125ra131. dose toxicity data. Chem. Cent. J. 9:62. doi: 10.1186/s13065-015-0139-7 doi: 10.1126/scitranslmed.3003377 Raymond, J. W., Rogers, T. N., Shonnard, D. R., and Kline, A. A. (2001). A review Tetko, I. V., Gasteiger, J., Todeschini, R., Mauri, A., Livingstone, D., Ertl, P., et al. of structure-based biodegradation estimation methods. J. Hazard. Mater. 84, (2005). Virtual computational chemistry laboratory–design and description. J. 189–215. doi: 10.1016/S0304-3894(01)00207-2 Comput. Aided Mol. Des. 19, 453–463. doi: 10.1007/s10822-005-8694-y Read, J., Pfahringer, B., and Holmes, G. (2008). “Multi-label classification using Toropov, A. A., and Benfenati, E. (2006). QSAR models of quail dietary toxicity ensembles of pruned sets,” ICDM 2008: Eighth IEEE International Conference based on the graph of atomic orbitals. Bioorg. Med. Chem. Lett. 16, 1941–1943. on Data Mining, Proceedings (Pisa), 995-1000. doi: 10.1016/j.bmcl.2005.12.085 Read, J., Pfahringer, B., Holmes, G., and Frank, E. (2011). Classifier Tsoumakas, G., Katakis, I., and Taniar, D. (2007). Multi-label chains for multi-label classification. Mach. Learn. 85, 333–359. classification: an overview. Int. J. Data Warehousing Min. 3, 1–13. doi: 10.1007/s10994-011-5256-5 doi: 10.4018/jdwm.2007070101 Sauer, J. M., Hartung, T., Leist, M., Knudsen, T. B., Hoeng, J., and Hayes, A. W. Tsoumakas, G., Katakis, I., and Vlahavas, I. (2008). “Effective and efficient (2015). Systems toxicology: the future of risk assessment. Int. J. Toxicol. 34, multilabel classification in domains with large number of labels,” in Ecml/pkdd 346–348. doi: 10.1177/1091581815576551 Workshop on Mining Multidimensional Data (Ho Chi Minh City). Sauvant, M. P., Pepin, D., and Piccinni, E. (1999). Tetrahymena pyriformis: Tsoumakas, G., Katakis, I., and Vlahavas, I. (2010). “Mining Multi-label Data,” a tool for toxicological studies. A review. Chemosphere 38, 1631–1669. in Data Mining and Knowledge Discovery Handbook, eds O. Maimon and L. doi: 10.1016/S0045-6535(98)00381-6 Rokach. (Boston, MA: Springer), 667–685. Schmidt, U., Struck, S., Gruening, B., Hossbach, J., Jaeger, I. S., Parol, R., et al. Tsoumakas, G., and Vlahavas, I. (2007). “Random k-labelsets: an ensemble method (2009). SuperToxic: a comprehensive database of toxic compounds. Nucleic for multilabel classification,” in Proceedings of Machine Learning. ECML 2007 Acids Res. 37, D295-D299. doi: 10.1093/nar/gkn850 (Warsaw). Schuster, D., Laggner, C., and Langer, T. (2005). Why drugs fail - A study Verheyen, G. R., Braeken, E., Van Deun, K., and Van Miert, S. (2017). Evaluation on side effects in new chemical entities. Curr. Pharm. Des. 11, 3545–3559. of existing (Q)SAR models for skin and eye irritation and corrosion to use for doi: 10.2174/138161205774414510 REACH registration. Toxicol. Lett. 265, 47–52. doi: 10.1016/j.toxlet.2016.11.007 Segall, M. D., and Barber, C. (2014). Addressing toxicity risk when designing and Verma, R. P., and Matthews, E. J. (2015a). Estimation of the chemical-induced eye selecting compounds in early drug discovery. Drug Discov. Today 19, 688–693. injury using a weight-of-evidence (WoE) battery of 21 artificial neural network doi: 10.1016/j.drudis.2014.01.006 (ANN) c-QSAR models (QSAR-21): part I: irritation potential. Regul. Toxicol. Shah, I., Liu, J., Judson, R. S., Thomas, R. S., and Patlewicz, G. (2016). Pharmacol. 71, 318–330. doi: 10.1016/j.yrtph.2014.11.011 Systematically evaluating read-across prediction and performance Verma, R. P., and Matthews, E. J. (2015b). Estimation of the chemical-induced eye using a local validity approach characterized by chemical structure injury using a Weight-of-Evidence (WoE) battery of 21 artificial neural network and bioactivity information. Regul. Toxicol. Pharmacol. 79, 12–24. (ANN) c-QSAR models (QSAR-21): part II: corrosion potential. Regul. Toxicol. doi: 10.1016/j.yrtph.2016.05.008 Pharmacol. 71, 331–336. doi: 10.1016/j.yrtph.2014.12.004 Shen, J., Cheng, F., Xu, Y., Li, W., and Tang, Y. (2010). Estimation of ADME Wang, J. R., Feng, J., Sun, X., Chen, S. S., and Chen, B. (2014). Simplified properties with substructure pattern recognition. J. Chem. Inf. Model 50, Constraints Rank-SVM for Multi-label Classification. Pattern Recogn. 483(Pt 1034–1041. doi: 10.1021/ci100104j I), 229–236. doi: 10.1007/978-3-662-45646-0_23 Shen, Q., Jiang, J. H., Jiao, C. X., Lin, W. Q., Shen, G. L., and Yu, R. Q. Wang, Q., Li, X., Yang, H., Cai, Y., Wang, Y., Wang, Z., et al. (2017). In silico (2004). Hybridized particle swarm algorithm for adaptive structure training of prediction of serious eye irritation or corrosion potential of chemicals. RSC multilayer feed-forward neural network: QSAR studies of bioactivity of organic Adv. 7, 6697–6703. doi: 10.1039/C6RA25267B compounds. J. Comput. Chem. 25, 1726–1735. doi: 10.1002/jcc.20094 Wang, S., Li, Y., Wang, J., Chen, L., Zhang, L., Yu, H., et al. (2012). ADMET Singh, K. P., Gupta, S., and Rai, P. (2013). Predicting carcinogenicity of diverse evaluation in drug discovery. 12. Development of binary classification models chemicals using probabilistic neural network modeling approaches. Toxicol. for prediction of hERG potassium channel blockage. Mol. Pharm. 9, 996-1010. Appl. Pharmacol. 272, 465–475. doi: 10.1016/j.taap.2013.06.029 doi: 10.1021/mp300023x Sokolova, M., and Szpakowicz, S. (2010). In Handbook of Research on Machine Wang, S., Sun, H., Liu, H., Li, D., Li, Y., and Hou, T. (2016). ADMET evaluation Learning Applications and Trends: Algorithms, Methods, and Techniques. in drug discovery. 16. Predicting hERG Blockers by combining multiple Hershey, PA: IGI Global. pharmacophores and machine learning approaches. Mol. Pharm. 13, 2855– Sonnet, P., Guillon, J., Enguehard, C., Dallemagne, P., Bureau, R., Rault S. 2866. doi: 10.1021/acs.molpharmaceut.6b00471 Auvray P., et al. (1998). Design and synthesis of a new type of non Wang, Y., Xiao, J., Suzek, T. O., Zhang, J., Wang, J., and Bryant, S. H. steroidal human aromatase inhibitors. Bioorg. Med. Chem. Lett. 8, 1041–1044. (2009). PubChem: a public information system for analyzing bioactivities doi: 10.1016/S0960-894X(98)00157-7 of small molecules. Nucleic Acids Res. 37(Suppl. 2), W623–W633. Steinbeck, C., Han, Y. Q., Kuhn, S., Horlacher, O., Luttmann, E., and Willighagen, doi: 10.1093/nar/gkp456 E. (2003). The Chemistry Development Kit (CDK): an open-source Java Wilhelmus, K. R. (2001). The Draize eye test. Surv. Ophthalmol. 45, 493–515. library for chemo- and bioinformatics. J. Chem. Inf. Comput. Sci. 43, 493–500. doi: 10.1016/S0039-6257(01)00211-9 doi: 10.1021/ci025584y Williams-DeVane, C. R., Wolf, M. A., and Richard, A. M. (2009). DSSTox Sun, L., Yang, H., Li, J., Wang, T., Li, W., Liu, G., et al. (2017). In silico chemical-index files for exposure-related experiments in ArrayExpress and prediction of compounds binding to human plasma proteins by QSAR models. Gene Expression Omnibus: enabling toxico-chemogenomics data linkages. ChemMedChem. doi: 10.1002/cmdc.201700582. [Epub ahead of print]. Bioinformatics 25, 692–694. doi: 10.1093/bioinformatics/btp042 Sun, L., Zhang, C., Chen, Y. J., Li, X., Zhuang, S. L., Li, W. H., et al. (2015). In silico Wishart, D., Arndt, D., Pon, A., Sajed, T., Guo, A. C., Djoumbou, Y., et al. prediction of chemical aquatic toxicity with chemical category approaches and (2015). T3DB: the toxic exposome database. Nucleic Acids Res. 43, D928-D934. substructural alerts. Toxicol. Res. 4, 452–463. doi: 10.1039/C4TX00174E doi: 10.1093/nar/gku1004 Frontiers in Chemistry | www.frontiersin.org 11 February 2018 | Volume 6 | Article 30 Yang et al. In Silico Prediction of Toxicity Xu, C., Cheng, F., Chen, L., Du, Z., Li, W., Liu, G., et al. (2012). In silico Zhang, H., Cao, Z., Li, M., Li, Y., and Peng, C. (2016). Novel naive Bayes prediction of chemical Ames mutagenicity. J. Chem. Inf. Model. 52, 2840–2847. classification models for predicting the carcinogenicity of chemicals. Food doi: 10.1021/ci300400a Chem. Toxicol. 97, 141–149. doi: 10.1016/j.fct.2016.09.005 Xu, Y., Dai, Z., Chen, F., Gao, S., Pei, J., and Lai, L. (2015). Deep Zhang, L., Ai, H., Chen, W., Yin, Z., Hu, H., Zhu, J., et al. (2017). CarcinoPred- learning for drug-induced liver injury. J. Chem. Inf. Model. 55, 2085–2093. EL: novel models for predicting the carcinogenicity of chemicals using doi: 10.1021/acs.jcim.5b00238 molecular fingerprints and ensemble learning methods. Sci. Rep. 7:2118. Xu, Y., Pei, J., and Lai, L. (2017). Deep learning based regression and multiclass doi: 10.1038/s41598-017-02365-0 models for acute oral toxicity prediction with automatic chemical feature Zhang, M. L., and Zhou, Z. H. (2006). Multilabel neural networks with applications extraction. J. Chem. Inf. Model. 57, 2672-2685. doi: 10.1021/acs.jcim.7b00244 to functional genomics and text categorization. IEEE Trans. Knowl. Data Eng. Xue, Y., Li, Z. R., Yap, C. W., Sun, L. Z., Chen, X., and Chen, Y. Z. (2004). Effect of 18, 1338–1351. doi: 10.1109/TKDE.2006.162 molecular descriptor feature selection in support vector machine classification Zhang, M. L., and Zhou, Z. H. (2007). ML-KNN: a lazy learning of pharmacokinetic and toxicological properties of chemical agents. J. Chem. approach to multi-label leaming. Pattern Recognit. 40, 2038–2048. Inf. Comput. Sci. 44, 1630–1638. doi: 10.1021/ci049869h doi: 10.1016/j.patcog.2006.12.019 Yang, H., Li, J., Wu, Z., Li, W., Liu, G., and Tang, Y. (2017a). Evaluation of Zhang, M. L., and Zhou, Z. H. (2014). A review on multi-label learning algorithms. different methods for identification of structural alerts using chemical ames IEEE Trans. Knowl. Data Eng. 26, 1819–1837. doi: 10.1109/TKDE.2013.39 mutagenicity data set as a benchmark. Chem. Res. Toxicol. 30, 1355–1364. Zhu, H., Martin, T. M., Ye, L., Sedykh, A., Young, D. M., and Tropsha, A. (2009). doi: 10.1021/acs.chemrestox.7b00083 Quantitative structure-activity relationship modeling of rat acute toxicity by Yang, H., Li, X., Cai, Y., Wang, Q., Li, W., Liu, G., et al. (2017b). In silico oral exposure. Chem. Res. Toxicol. 22, 1913–1921. doi: 10.1021/tx900189p prediction of chemical subcellular localization via multi-classification methods. Zhu, X., and Kruhlak, N. L. (2014). Construction and analysis of a human Medchemcomm. 8, 1225-1234 doi: 10.1039/C7MD00074J hepatotoxicity database suitable for QSAR modeling using post-market safety Yap, C. W. (2011). PaDEL-descriptor: an open source software to calculate data. Toxicology 321, 62–72. doi: 10.1016/j.tox.2014.03.009 molecular descriptors and fingerprints. J. Comput. Chem. 32, 1466–1474. doi: 10.1002/jcc.21707 Conflict of Interest Statement: The authors declare that the research was Zhang, C., Cheng, F., Li, W., Liu, G., Lee, P. W., and Tang, Y. (2016a). In silico conducted in the absence of any commercial or financial relationships that could prediction of drug induced liver toxicity using substructure pattern recognition be construed as a potential conflict of interest. method. Mol. Inform. 35, 136–144. doi: 10.1002/minf.201500055 Zhang, C., Cheng, F., Sun, L., Zhuang, S., Li, W., Liu, G., et al. Copyright © 2018 Yang, Sun, Li, Liu and Tang. This is an open-access article (2015). In silico prediction of chemical toxicity on avian species distributed under the terms of the Creative Commons Attribution License (CC using chemical category approaches. Chemosphere 122, 280–287. BY). The use, distribution or reproduction in other forums is permitted, provided doi: 10.1016/j.chemosphere.2014.12.001 the original author(s) and the copyright owner are credited and that the original Zhang, C., Zhou, Y., Gu, S., Wu, Z., Wu, W., Liu, C., et al. (2016b). In publication in this journal is cited, in accordance with accepted academic practice. silico prediction of hERG potassium channel blockage by chemical category No use, distribution or reproduction is permitted which does not comply with these approaches. Toxicol. Res. 5, 570–582. doi: 10.1039/C5TX00294J terms. Frontiers in Chemistry | www.frontiersin.org 12 February 2018 | Volume 6 | Article 30