SApredictor: An Expert System for Screening Chemicals Against Structural Alerts

Abstract

The rapid and accurate evaluation of chemical toxicity is of great significance for estimation of chemical safety. In the past decades, a great number of excellent computational models have been developed for chemical toxicity prediction. But most machine learning models tend to be “black box”, which bring about poor interpretability. In the present study, we focused on the identification and collection of structural alerts (SAs) responsible for a series of important toxicity endpoints. Then, we carried out effective storage of these structural alerts and developed a web-server named SApredictor (www.sapredictor.cn) for screening chemicals against structural alerts. People can quickly estimate the toxicity of chemicals with SApredictor, and the specific key substructures which cause the chemical toxicity will be intuitively displayed to provide valuable information for the structural optimization by medicinal chemists.

Introduction

Nowadays, the development of chemical toxicology studies has provided us with extensive compound toxicity data. By analyzing and mining the existing toxicological experimental data, computational models can be established to predict the toxicity of chemical compounds around our lives. Compared with biological experimental methods, the computational methods were always green, fast, cheap, and accurate (Yang et al., 2018c). More importantly, toxicity can be predicted with computational models even before a chemical is synthesized or isolated. In the past decades, several expert systems, for example, Toxtree (Patlewicz et al., 2008) and OECD QSAR Toolbox (https://qsartoolbox.org/etc), and web-servers, for example, admetSAR (Yang et al., 2018a), ToxAlerts (Sushko et al., 2012), ADMETlab (Xiong et al., 2021), pkCSM (Pires et al., 2015), and vNN (Schyman et al., 2017) have also been proposed for in silico toxicity estimation.

The quantitative structure activity relationships (QSAR) method is one of the most widely used computational approaches for toxicity prediction, and many QSAR models are reported every year. However, these QSAR models based on machine learning methods tend to be “black box” models, which have limited the application in the prediction of various properties for regulatory agencies (Alves et al., 2016). In addition, the machine learning models need to face the problem of applicability domains (ADs) definition. Defining ADs are essential for regulatory acceptance of QSAR models, but there is less standard definition of AD for the global QSAR model nowadays, and many published QSAR models do not provide ADs (Wang et al., 2021).

Structural alert (SA) is another widely accepted tool for toxicity prediction in recent years, which can be defined as the key substructure which can cause specific toxicity. SA has been commonly used for assessment of many toxicity endpoints (Benigni et al., 2013; Li et al., 2017a; Limban et al., 2018; Kalgutkar, 2020; Cui et al., 2021; Huang et al., 2021; Shi et al., 2022) since Ashby and Tennant (1988) proposed the concept in 1985. The SAs can visually alert the toxicity of chemicals by displaying the key fragments responsible for drug toxicity because of the direct derivation from mechanistic knowledge. Therefore, SAs can provide valuable guidance and reference for structural optimization by medicinal chemists to reduce the risk (Yang et al., 2018c).

In the present study, we focused on screening chemicals against structural alerts, including 1) the identification of specific SAs responsible for the toxicity endpoints most concerned in drug studies based on a database with high quality toxicity data and the collection of reported SAs from research publications; and 2) the development of web-server for screening chemicals against structural alerts.

Materials and Methods

Data Collection and Preparation

The data for identification of structural alerts were collected from 1) the databases such as ChEMBL (Gaulton et al., 2011), ChemIDplus (Tomasulo, 2002), Comparative Toxicogenomics Database (CTD) (Davis et al., 2018), Carcinogenic Potency Database (CPDB) (Gold et al., 1984) and DrugBank (Wishart et al., 2017) and 2) peer-reviewed publications through manually filtering and processing. We focused on 22 toxicity endpoints which are of most concern in environmental toxicology and drug discovery, including acute oral toxicity (Li et al., 2014), chemical aquatic toxicity [Tetrahymena pyriformis (Cheng et al., 2011), Daphnia magna (Gajewicz-Skretna et al., 2021), and fathead minnow (Sun et al., 2015)], chemical-induced hematotoxicity (Hua et al., 2021), drug-induced neurotoxicity (Jiang et al., 2020), drug-induced autoimmune diseases (Wu et al., 2021), drug-induced ototoxicity (Huang et al., 2021), drug-induced rhabdomyolysis (Cui et al., 2019), endocrine disruption (Chen et al., 2014), eye irritation (Wang et al., 2017), hepatotoxicity (Li et al., 2018), hERG inhibition (Li et al., 2017c), honey bee toxicity (Li et al., 2017b), inhalation toxicity (Cui et al., 2021), mitochondrial toxicity (Nelms et al., 2015), mutagenicity (Yang et al., 2017), nephrotoxicity (Shi et al., 2022), non-genotoxic carcinogenicity (Benigni et al., 2013), reproductive and development toxicity (Fan et al., 2018; Jiang et al., 2019), skin sensitization (Di et al., 2019), and toxicity on avian species (Zhang et al., 2015). For each toxicity endpoint, we searched the literature separately and included the publications with the same definition of the toxicity endpoint and consistent toxic/non-toxic classification criteria.

The datasets were prepared in following steps to guarantee the quality and reliability of the data: 1) removing mixtures, inorganic, salts, and organic metallic compounds; 2) removing compounds without explicit description for toxicity properties or have inconsistent results in different experimental groups; 3) removing the fuzzy, uncertain, and obviously uncorrected data points; and 4) standardization and representing as canonical SMILES (O’Boyle, 2012).

Identification of Structural Alerts

The structural alerts were identified with two different methods, including SARpy (Ferrari et al., 2013) and fingerprints filter (Yang et al., 2020). Both the methods were based on frequency analysis, the general idea of which was to find some substructures presented more frequently in toxic compounds than in non-toxic ones (Yang et al., 2020). If a substructure presented far more frequently in toxic compounds than non-toxic compounds, the presence of such a substructure could alert to toxicity. Thus, this substructure should be regarded as a structural alert responsible for the specific toxicity. The flow of these two methods for identifying structural alerts was shown in Figure 1. SARpy is a python-based standalone software program for automated QSAR modeling. This program has been well-described in detail by Ferrari et al. (2013). Using SMILES-based algorithms, SARpy can cleave the compounds to obtain all possible fragments, and the potential structural alerts can be obtained by frequency analysis. In this study, rule sets were generated using standard settings; the substructures are composed of minimum two and maximum 18 atoms and occurring in a minimum of three substances. For the fingerprints filter method, the well-defined fingerprints of various lengths were utilized as the source of substructures. In the present study, the structural alerts were identified with a f-score and positive rate of each substructure from Klekota-Roth fingerprint (KRFP) calculated with a PaDEL-Descriptor (Yap, 2011), which contained 4,860 predefined structural fragments. The positive rate (PR) of a substructure is defined as Eq. 1:where N_{fragment_positive} is the number of toxic compounds containing the fragment, and N_fragment is the total number of compounds containing the fragment. For each specific endpoint, only the fragments presented in six or more compounds were maintained. The fragments with f-score ≥0.005 and positive rate ≥0.65 were identified as structural alerts.

FIGURE 1

The structural alerts were converted into SMARTS patterns (Hanson, 2016) and stored in the MySQL database. The SMARTS pattern is a language that allows users to specify substructures using the rules, which are straightforward extensions of SMILES. With SMARTS, flexible, and efficient substructure-search specifications can be made in terms that are meaningful to chemists, a compound can be matched against the alert in an automatic manner using one of the available chemical libraries (Sushko et al., 2012).

In addition, we also collected the structural alerts reported in the peer-reviewed publications. The collected structural alerts were also converted into SMARTS patterns, and the duplicates were removed.

Applicability Domain Definition

As emphasized by OECD principles, a well-defined applicability domain (AD) was preferred to make models more precise and robust (Yang et al., 2020). From the viewpoint of predictive performance, AD can also be helpful for improving the accuracy of SAs. Numerical relationships between chemical descriptors and toxicity values from training set are the basis of many applicability domain definition techniques, especially for QSAR models. However, AD for SAs have not been defined using these methods since the alerts are always a combination of structural information, toxic or non-toxic testing outcomes, and expert knowledge, which are used to directly link substructures with potential activity (Ellison et al., 2011). To date, there has been no single generally accepted algorithm for determining the AD on SAs. Since no chemical descriptors were used for SA model building, structural similarity could be the most appropriate measure to define the AD. Structural similarity is a popular AD definition method based on the concept that if a query chemical can be defined as similar to the chemicals in the training data, then it can be considered within the AD (Kühne et al., 2006; Ellison et al., 2011). In the present study, the similarity matrix was calculated employing the Tanimoto coefficient (Tc) (Godden et al., 2000; Bajusz et al., 2015) based on the KRFP fingerprint. The Tc is defined as Tc = Nab/(Na+ Nb− Nab), with Na being the number of bits set on in molecule a, Nb is the number of bits set on in molecule b, and Nab is the number of bits set on common to both molecules (Godden et al., 2000). The cutoff similarity value was defined as 0.5, thus if a query compound had a similarity value of ≥0.5 to at least one compound in the training set, it would be considered to be within the AD.

Toxicity Prediction With Structural Alerts

The structural alerts were assessed with the specific dataset of each endpoint. The compounds were input as SMILES and queried for matching the specific structural alerts with RDKit (Lovrić et al., 2019). If a compound contains one or more structural alerts, it would be predicted to have the specific toxicity. The evaluation was based on the counts of true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Several statistical parameters were also calculated, including the total accuracy (Q), sensitivity (SE), specificity (SP), and positive predictive value (PPV). These parameters are calculated with Eqs. 2–5:

Web-Server Implementation

The prediction system was developed employing the Python web framework of Django. The system was deployed on an elastic compute service from Huawei Cloud running an Ubuntu Linux system. The web access was enabled via the Nginx web-server and the interactions between Django and proxy server were supported by mod_wsgi v3.3. A user-friendly web interface was provided for computational prediction using a cascading style sheet (CSS) and Python script.

Results and Discussion

Compound Libraries and Sets of Structural Alerts

In total, more than 35,716 annotated measurements of about 27,500 unique compounds were collected, including thousands of FDA-approved and experimental drugs, pesticides, environmental agents, and industrial chemicals. As shown in Table 1, these chemicals were checked and divided into 22 subsets, according to different toxicity endpoints.

Through the identification and literature retrieval, a total of 1,834 structural alerts were identified and collected for the aforementioned 22 toxicity endpoints, as shown in Table 1. ToxAlerts and Toxtree are two popular tools for the estimation of potential adverse reactions of chemicals. ToxAlerts is a web-server of structural alerts, which collected SAs defined by experts or detected by computational tools. The latest ToxAlerts (accessed on 8 April 2022) contains 814 structural alerts for 13 toxicity endpoints, as shown in Supplementary Table S1. Toxtree is another user-friendly open-source application, which is able to estimate toxic hazard by applying a decision tree approach. In the latest version (Toxtree 3.1.0), in addition to the three Cramer Decision Trees (Cramer Rules, Revised Cramer Decision Tree, and Cramer Rules, with Extensions), it contains 499 structural alerts for 13 toxicity endpoints, as shown in Supplementary Table S2. To our knowledge, this may be the largest structural alert database with specific toxicity endpoints until now. In addition, several toxicity endpoints which get a lot of concerns (hepatotoxicity, nephrotoxicity, reproductive and development toxicity, hERG inhibition, hematotoxicity, mitochondrial toxicity, etc.) were included in SApredictor while not in ToxAlerts or Toxtree.

Performance of Toxicity Prediction With Structural Alerts

The performances of the structural alerts on toxicity prediction are shown in Table 2. The results suggested that for most endpoints, the structural alerts can well distinguish toxic compounds from non-toxic ones. For different toxicity endpoints, the disparity was observed in the performance. This can be attributed to that the complexity of the mechanisms of action (MOAs) of different toxicity endpoints vary greatly, and the sizes of the data are also different, which lead to differences in the representativeness of SAs and the ability to distinguish between toxic and non-toxic compounds.

It was worth pointing out that compared with QSAR models, the prediction accuracy of structural alerts did not have any advantage in most cases. However, different from the QSAR’s black box model, the structural alerts can visually display the fragments that lead to specific toxicity of compounds, which is conducive to the targeted optimization of toxic structures and the study of toxic mechanisms (Yang et al., 2018c; Shi et al., 2022).

To ensure usefulness of the prediction system, it will be updated regularly with additional structural alerts based on available data, whether identified by ourselves or reported by peer-reviewed publications. If high quality datasets with new endpoints are reported, new structural alerts will be identified and implemented in our database.

Web Interface and Usage of the Structural Alerts

Based on distributed storage architectures, a piece of software for the estimation of chemical toxicity with structural alert was developed. The software provides a user-friendly interface viawww.sapredictor.cn. A screenshot of the web-server is shown in Figure 2. Users can submit compound structures in two different ways: 1) enter the SMILES of small compounds in the dialog box; 2) click the “Select File” button to upload the structure file of compounds with SMILES format. After entering the verification code, users can click the “Predict” button to complete the task submission.

FIGURE 2

After the matching of the structural alerts, it will be redirected to the results page, as shown in Figure 3. On the left is the 2D structure of the query compound and on the right is the toxicity endpoints and corresponding predicted result. Where “Yes” indicates that the query structure contains one or more structural alerts of the specific toxic property, that is, the compound has the potential of the specific toxicity, while “No” indicates the query compound does not have the potential of the specific toxicity. For the toxicity endpoint with the result of “Yes,” click the name of the toxicity endpoint and a drop-down list will appear listing the ID of structural alerts. When clicking the ID, the fragments of the compound will be highlighted in red, and the SMARTS of the alert will also be available on the page. The researchers can view the specific substructure that causes the toxicity of the compound.

FIGURE 3

Conclusion and Perspectives

In summary, we have described here a web-server, named SApredictor, for screening chemicals against structural alerts viawww.sapredictor.cn. In SApredictor, 1,834 structural alerts for 22 different toxicity endpoints were extracted from more than 35,716 toxicity annotated data points or collected from peer-reviewed publications. Users can quickly estimate the toxicity of compounds and visually display the fragments, which contribute to their toxicity. The web-server will never retain any information submitted to it because of the confidentiality of users’ projects. We hope that the software should facilitate the process of drug discovery and development by enabling the rapid and rational screening, design, evaluation, and prioritization of drug candidates.

It is worth pointing out that use of structural alerts alone may suffer from false positives, such as skin sensitization and toxicity on avian species in the present study. The structural alerts were always identified by statistics-based methods or knowledge of toxic mechanisms, which would make them be overtly common and lead to many non-toxic structures being estimated as toxic. On the other hand, it is debatable whether compounds which do not contain any SA can be classified as non-toxic. Toxicity prediction based on SA is based on the existing knowledge. The compounds with SAs are always toxic, but whether those without SA are non-toxic needs more toxicity data support. Yang et al. (2018b) proposed a concept of non-toxic substructures, whose appearance will reduce the probability of a compound becoming toxic (Yang et al., 2018b). In Wang et al. (2012) work, modulating factors that suppressed the toxic effects of SAs were extracted and practice on carcinogens (Wang et al., 2012). Non-toxic substructures and modulating factors could be beneficial supplements to SAs. In addition to optimizing the identification method of structural alerts, defining the applicability domain for the structural alerts in a reasonable strategy may be helpful to improve the predictive performance and eliminate the worries. We will continue to work in both directions to improve the predictive ability of structural alerts and make them more useful.

References

• 1

  AlvesV. M.MuratovE. N.CapuzziS. J.PolitiR.LowY.BragaR. C.et al (2016). Alarms about Structural Alerts. Green Chem.18 (16), 4348–4360. 10.1039/C6GC01492E

  • CrossRef
  • Google Scholar

• 2

  AshbyJ.TennantR. W. (1988). Chemical Structure, Salmonella Mutagenicity and Extent of Carcinogenicity as Indicators of Genotoxic Carcinogenesis Among 222 Chemicals Tested in Rodents by the U.S. NCI/NTP. Mutat. Research/Genetic Toxicol.204 (1), 17–115. 10.1016/0165-1218(88)90114-0

  • CrossRef
  • Google Scholar

• 3

  BajuszD.RáczA.HébergerK. (2015). Why Is Tanimoto Index an Appropriate Choice for Fingerprint-Based Similarity Calculations?J. Cheminform7 (1), 20. 10.1186/s13321-015-0069-3

  • CrossRef
  • Google Scholar

• 4

  BenigniR.BossaC.TcheremenskaiaO. (2013). Nongenotoxic Carcinogenicity of Chemicals: Mechanisms of Action and Early Recognition through a New Set of Structural Alerts. Chem. Rev.113 (5), 2940–2957. 10.1021/cr300206t

  • CrossRef
  • Google Scholar

• 5

  ChenY.ChengF.SunL.LiW.LiuG.TangY. (2014). Computational Models to Predict Endocrine-Disrupting Chemical Binding with Androgen or Oestrogen Receptors. Ecotoxicol. Environ. Saf.110, 280–287. 10.1016/j.ecoenv.2014.08.026

  • CrossRef
  • Google Scholar

• 6

  ChengF.ShenJ.YuY.LiW.LiuG.LeeP. W.et al (2011). In Silico prediction of Tetrahymena Pyriformis Toxicity for Diverse Industrial Chemicals with Substructure Pattern Recognition and Machine Learning Methods. Chemosphere82 (11), 1636–1643. 10.1016/j.chemosphere.2010.11.043

  • CrossRef
  • Google Scholar

• 7

  CuiX.LiuJ.ZhangJ.WuQ.LiX. (2019). In Silico prediction of Drug‐induced Rhabdomyolysis with Machine‐learning Models and Structural Alerts. J. Appl. Toxicol.39 (8), 1224–1232. 10.1002/jat.3808

  • CrossRef
  • Google Scholar

• 8

  CuiX.YangR.LiS.LiuJ.WuQ.LiX. (2021). Modeling and Insights into Molecular Basis of Low Molecular Weight Respiratory Sensitizers. Mol. Divers25 (2), 847–859. 10.1007/s11030-020-10069-3

  • CrossRef
  • Google Scholar

• 9

  DavisA. P.GrondinC. J.JohnsonR. J.SciakyD.McMorranR.WiegersJ.et al (2018). The Comparative Toxicogenomics Database: Update 2019. Nucleic Acids Res.47 (D1), D948–D954. 10.1093/nar/gky868

  • CrossRef
  • Google Scholar

• 10

  DiP.YinY.JiangC.CaiY.LiW.TangY.et al (2019). Prediction of the Skin Sensitising Potential and Potency of Compounds via Mechanism-Based Binary and Ternary Classification Models. Toxicol. Vitro59, 204–214. 10.1016/j.tiv.2019.01.004

  • CrossRef
  • Google Scholar

• 11

  EllisonC. M.SherhodR.CroninM. T. D.EnochS. J.MaddenJ. C.JudsonP. N. (2011). Assessment of Methods to Define the Applicability Domain of Structural Alert Models. J. Chem. Inf. Model.51 (5), 975–985. 10.1021/ci1000967

  • CrossRef
  • Google Scholar

• 12

  FanD.YangH.LiF.SunL.DiP.LiW.et al (2018). In Silico prediction of Chemical Genotoxicity Using Machine Learning Methods and Structural Alerts. Toxicol. Res.7 (2), 211–220. 10.1039/C7TX00259A

  • CrossRef
  • Google Scholar

• 13

  FerrariT.CattaneoD.GiniG.Golbamaki BakhtyariN.ManganaroA.BenfenatiE. (2013). Automatic Knowledge Extraction from Chemical Structures: the Case of Mutagenicity Prediction. SAR QSAR Environ. Res.24 (5), 365–383. 10.1080/1062936X.2013.773376

  • CrossRef
  • Google Scholar

• 14

  Gajewicz-SkretnaA.FuruhamaA.YamamotoH.SuzukiN. (2021). Generating Accurate In Silico Predictions of Acute Aquatic Toxicity for a Range of Organic Chemicals: Towards Similarity-Based Machine Learning Methods. Chemosphere280, 130681. 10.1016/j.chemosphere.2021.130681

  • CrossRef
  • Google Scholar

• 15

  GaultonA.BellisL. J.BentoA. P.ChambersJ.DaviesM.HerseyA.et al (2011). ChEMBL: a Large-Scale Bioactivity Database for Drug Discovery. Nucleic Acids Res.40 (D1), D1100–D1107. 10.1093/nar/gkr777

  • CrossRef
  • Google Scholar

• 16

  GoddenJ. W.XueL.BajorathJ. (2000). Combinatorial Preferences Affect Molecular Similarity/diversity Calculations Using Binary Fingerprints and Tanimoto Coefficients. J. Chem. Inf. Comput. Sci.40 (1), 163–166. 10.1021/ci990316u

  • CrossRef
  • Google Scholar

• 17

  GoldL. S.SawyerC. B.MagawR.BackmanG. M.De VecianaM.LevinsonR.et al (1984). A Carcinogenic Potency Database of the Standardized Results of Animal Bioassays. Environ. Health Perspect.58, 9–319. 10.1289/ehp.84589

  • CrossRef
  • Google Scholar

• 18

  HansonR. M. (2016). Jmol SMILES and Jmol SMARTS: Specifications and Applications. J. Cheminform8 (1), 50. 10.1186/s13321-016-0160-4

  • CrossRef
  • Google Scholar

• 19

  HuaY.ShiY.CuiX.LiX. (2021). In Silico prediction of Chemical-Induced Hematotoxicity with Machine Learning and Deep Learning Methods. Mol. Divers25 (3), 1585–1596. 10.1007/s11030-021-10255-x

  • CrossRef
  • Google Scholar

• 20

  HuangX.TangF.HuaY.LiX. (2021). In Silico prediction of Drug‐induced Ototoxicity Using Machine Learning and Deep Learning Methods. Chem. Biol. Drug Des.98 (2), 248–257. 10.1111/cbdd.13894

  • CrossRef
  • Google Scholar

• 21

  JiangC.YangH.DiP.LiW.TangY.LiuG. (2019). In Silico prediction of Chemical Reproductive Toxicity Using Machine Learning. J. Appl. Toxicol.39 (6), 844–854. 10.1002/jat.3772

  • CrossRef
  • Google Scholar

• 22

  JiangC.ZhaoP.LiW.TangY.LiuG. (2020). In Silico prediction of Chemical Neurotoxicity Using Machine Learning. Toxicol. Res.9 (3), 164–172. 10.1093/toxres/tfaa016

  • CrossRef
  • Google Scholar

• 23

  KalgutkarA. S. (2020). Designing Around Structural Alerts in Drug Discovery. J. Med. Chem.63 (12), 6276–6302. 10.1021/acs.jmedchem.9b00917

  • CrossRef
  • Google Scholar

• 24

  KühneR.EbertR.-U.SchüürmannG. (2006). Model Selection Based on Structural Similarity−Method Description and Application to Water Solubility Prediction. J. Chem. Inf. Model.46 (2), 636–641. 10.1021/ci0503762

  • CrossRef
  • Google Scholar

• 25

  LiX.ChenL.ChengF.WuZ.BianH.XuC.et al (2014). In Silico prediction of Chemical Acute Oral Toxicity Using Multi-Classification Methods. J. Chem. Inf. Model.54 (4), 1061–1069. 10.1021/ci5000467

  • CrossRef
  • Google Scholar

• 26

  LiX.ChenY.SongX.ZhangY.LiH.ZhaoY. (2018). The Development and Application of In Silico Models for Drug Induced Liver Injury. RSC Adv.8 (15), 8101–8111. 10.1039/C7RA12957B

  • CrossRef
  • Google Scholar

• 27

  LiX.ZhangY.ChenH.LiH.ZhaoY. (2017a). In Silico prediction of Chronic Toxicity with Chemical Category Approaches. RSC Adv.7 (66), 41330–41338. 10.1039/C7RA08415C

  • CrossRef
  • Google Scholar

• 28

  LiX.ZhangY.ChenH.LiH.ZhaoY. (2017b). Insights into the Molecular Basis of the Acute Contact Toxicity of Diverse Organic Chemicals in the Honey Bee. J. Chem. Inf. Model.57 (12), 2948–2957. 10.1021/acs.jcim.7b00476

  • CrossRef
  • Google Scholar

• 29

  LiX.ZhangY.LiH.ZhaoY. (2017c). Modeling of the hERG K+ Channel Blockage Using Online Chemical Database and Modeling Environment (OCHEM). Mol. Inf.36 (12), 1700074. 10.1002/minf.201700074

  • CrossRef
  • Google Scholar

• 30

  LimbanC.NuţăD. C.ChiriţăC.NegreșS.ArseneA. L.GoumenouM.et al (2018). The Use of Structural Alerts to Avoid the Toxicity of Pharmaceuticals. Toxicol. Rep.5, 943–953. 10.1016/j.toxrep.2018.08.017

  • CrossRef
  • Google Scholar

• 31

  LovrićM.MoleroJ. M.KernR. (2019). PySpark and RDKit: Moving towards Big Data in Cheminformatics. Mol. Inf.38 (6), 1800082. 10.1002/minf.201800082

  • CrossRef
  • Google Scholar

• 32

  NelmsM. D.MellorC. L.CroninM. T. D.MaddenJ. C.EnochS. J. (2015). Development of an In Silico Profiler for Mitochondrial Toxicity. Chem. Res. Toxicol.28 (10), 1891–1902. 10.1021/acs.chemrestox.5b00275

  • CrossRef
  • Google Scholar

• 33

  O’BoyleN. M. (2012). Towards a Universal Smiles Representation - a Standard Method to Generate Canonical Smiles Based on the InChI. J. Cheminform4 (1), 22. 10.1186/1758-2946-4-22

  • CrossRef
  • Google Scholar

• 34

  PatlewiczG.JeliazkovaN.SaffordR. J.WorthA. P.AleksievB. (2008). An Evaluation of the Implementation of the Cramer Classification Scheme in the Toxtree Software. SAR QSAR Environ. Res.19 (5-6), 495–524. 10.1080/10629360802083871

  • CrossRef
  • Google Scholar

• 35

  PiresD. E. V.BlundellT. L.AscherD. B. (2015). pkCSM: Predicting Small-Molecule Pharmacokinetic and Toxicity Properties Using Graph-Based Signatures. J. Med. Chem.58 (9), 4066–4072. 10.1021/acs.jmedchem.5b00104

  • CrossRef
  • Google Scholar

• 36

  SchymanP.LiuR.DesaiV.WallqvistA. (2017). vNN Web Server for ADMET Predictions. Front. Pharmacol.8, 889. 10.3389/fphar.2017.00889

  • CrossRef
  • Google Scholar

• 37

  ShiY.HuaY.WangB.ZhangR.LiX. (2022). In Silico Prediction and Insights into the Structural Basis of Drug Induced Nephrotoxicity. Front. Pharmacol.12, 793332. 10.3389/fphar.2021.793332

  • CrossRef
  • Google Scholar

• 38

  SunL.ZhangC.ChenY.LiX.ZhuangS.LiW.et al (2015). In Silico prediction of Chemical Aquatic Toxicity with Chemical Category Approaches and Substructural Alerts. Toxicol. Res.4 (2), 452–463. 10.1039/C4TX00174E

  • CrossRef
  • Google Scholar

• 39

  SushkoI.SalminaE.PotemkinV. A.PodaG.TetkoI. V. (2012). ToxAlerts: a Web Server of Structural Alerts for Toxic Chemicals and Compounds with Potential Adverse Reactions. J. Chem. Inf. Model.52 (8), 2310–2316. 10.1021/ci300245q

  • CrossRef
  • Google Scholar

• 40

  TomasuloP. (2002). ChemIDplus-super Source for Chemical and Drug Information. Med. Ref. Serv. Q.21 (1), 53–59. 10.1300/J115v21n01_04

  • CrossRef
  • Google Scholar

• 41

  WangQ.LiX.YangH.CaiY.WangY.WangZ.et al (2017). In Silico prediction of Serious Eye Irritation or Corrosion Potential of Chemicals. RSC Adv.7 (11), 6697–6703. 10.1039/C6RA25267B

  • CrossRef
  • Google Scholar

• 42

  WangY.LuJ.WangF.ShenQ.ZhengM.LuoX.et al (2012). Estimation of Carcinogenicity Using Molecular Fragments Tree. J. Chem. Inf. Model.52 (8), 1994–2003. 10.1021/ci300266p

  • CrossRef
  • Google Scholar

• 43

  WangZ.ChenJ.HongH. (2021). Developing QSAR Models with Defined Applicability Domains on PPARγ Binding Affinity Using Large Data Sets and Machine Learning Algorithms. Environ. Sci. Technol.55 (10), 6857–6866. 10.1021/acs.est.0c07040

  • CrossRef
  • Google Scholar

• 44

  WishartD. S.FeunangY. D.GuoA. C.LoE. J.MarcuA.GrantJ. R.et al (2017). DrugBank 5.0: a Major Update to the DrugBank Database for 2018. Nucleic Acids Res.46 (D1), D1074–D1082. 10.1093/nar/gkx1037

  • CrossRef
  • Google Scholar

• 45

  WuY.ZhuJ.FuP.TongW.HongH.ChenM. (2021). Machine Learning for Predicting Risk of Drug-Induced Autoimmune Diseases by Structural Alerts and Daily Dose. Ijerph18 (13), 7139. 10.3390/ijerph18137139

  • CrossRef
  • Google Scholar

• 46

  XiongG.WuZ.YiJ.FuL.YangZ.HsiehC.et al (2021). ADMETlab 2.0: an Integrated Online Platform for Accurate and Comprehensive Predictions of ADMET Properties. Nucleic Acids Res.49 (W1), W5–W14. 10.1093/nar/gkab255

  • CrossRef
  • Google Scholar

• 47

  YangH.LiJ.WuZ.LiW.LiuG.TangY. (2017). Evaluation of Different Methods for Identification of Structural Alerts Using Chemical Ames Mutagenicity Data Set as a Benchmark. Chem. Res. Toxicol.30 (6), 1355–1364. 10.1021/acs.chemrestox.7b00083

  • CrossRef
  • Google Scholar

• 48

  YangH.LouC.LiW.LiuG.TangY. (2020). Computational Approaches to Identify Structural Alerts and Their Applications in Environmental Toxicology and Drug Discovery. Chem. Res. Toxicol.33 (6), 1312–1322. 10.1021/acs.chemrestox.0c00006

  • CrossRef
  • Google Scholar

• 49

  YangH.LouC.SunL.LiJ.CaiY.WangZ.et al (2018a). admetSAR 2.0: Web-Service for Prediction and Optimization of Chemical ADMET Properties. Bioinformatics35 (6), 1067–1069. 10.1093/bioinformatics/bty707

  • CrossRef
  • Google Scholar

• 50

  YangH.SunL.LiW.LiuG.TangY. (2018b). Identification of Nontoxic Substructures: a New Strategy to Avoid Potential Toxicity Risk. Toxicol. Sci.165 (2), 396–407. 10.1093/toxsci/kfy146

  • CrossRef
  • Google Scholar

• 51

  YangH.SunL.LiW.LiuG.TangY. (2018c). In Silico prediction of Chemical Toxicity for Drug Design Using Machine Learning Methods and Structural Alerts. Front. Chem.6, 30. 10.3389/fchem.2018.00030

  • CrossRef
  • Google Scholar

• 52

  YapC. W. (2011). PaDEL-descriptor: An Open Source Software to Calculate Molecular Descriptors and Fingerprints. J. Comput. Chem.32 (7), 1466–1474. 10.1002/jcc.21707

  • CrossRef
  • Google Scholar

• 53

  ZhangC.ChengF.SunL.ZhuangS.LiW.LiuG.et al (2015). In Silico prediction of Chemical Toxicity on Avian Species Using Chemical Category Approaches. Chemosphere122, 280–287. 10.1016/j.chemosphere.2014.12.001

  • CrossRef
  • Google Scholar