In Silico Modeling Method for Computational Aquatic Toxicology of Endocrine Disruptors: A Software-Based Approach Using QSAR Toolbox

Summary

Quantitative structure-activity relationship (QSAR) modeling is a representative bioinformatics-assisted method in toxicological screening. This protocol demonstrates how to computationally assess the risks of endocrine disruptors (EDs) in aquatic environments. Utilizing the OECD QSAR Toolbox, the protocol implements an in silico assay for analyzing toxicity of EDs in fish.

Abstract

Computational analyses of toxicological processes enables high-throughput screening of chemical substances and prediction of their endpoints in biological systems. In particular, quantitative structure-activity relationship (QSAR) models have been increasingly applied to assess the environmental effects of a plethora of toxic materials. In recent years, some more highlighted types of toxicants are endocrine disruptors (EDs, which are chemicals that can interfere with any hormone-related metabolism). Because EDs may significantly affect animal development and reproduction, rapidly predicting the adverse effects of EDs using in silico techniques is required. This study presents an in silico method to generate prediction data on the effects of representative EDs in aquatic vertebrates, particularly fish species. The protocol describes an example utilizing the automated workflow of the QSAR Toolbox software developed by the Organization for Economic Co-operation and Development (OECD) to enable acute ecotoxicity predictions of EDs. As a result, the following are determined: (1) calculation of the numerical correlations between the concentration for 50% of lethality (LC₅₀) and octanol-water partition coefficient (K_ow), (2) output performances in which the LC₅₀ values determined in experiments are compared to those generated by computations, and (3) the dependence of estrogen receptor binding affinity on the relationship between K_ow and LC₅₀.

Introduction

New developments in informatics and computational technology have empowered the biological sciences with quantitative methodologies that offer high precision and reliability¹. In particular, algorithms used in molecular taxonomy and property classification have resulted in quantitative structure-activity relationship (QSAR) models². These models automatically correlate the chemical structures and biological activities of a given chemical database and implement rapid in silico screening of a wide range of chemical substrates according to their medicinal or toxicological actions³. QSAR tools can produce predictive toxicity profiles as a function of feature vectors of molecular descriptors (i.e., physicochemical parameters) of chemicals of interest to numerically create categorical endpoints⁴. Usually, each quantitative endpoint is displayed as a 2D scatterplot vs. changes in descriptor values. A QSAR model is then generated using (multiple) linear regression analyses. Once a dataset has been fully exploited to construct a QSAR model (called the training set), then the model is statistically validated by predicting the endpoints of a group of chemicals not included in the training set (called the test set). The model can then be used to predict the biological activities of untested compounds³.

Among many harmful chemicals, endocrine disruptors (EDs) have been highlighted as a group of toxicants that may interfere in numerous hormone-related metabolisms in mammals, amphibians, and fish^5,6. EDs are known to induce a variety of adverse effects, such as cancers and malformations, by blocking or altering normal hormonal pathways or activating abnormal hormone synthesis/degradation signals. As a consequence, these hormone-mimicking chemicals can perturb endocrine systems such that biological development and reproduction of wildlife animal populations are hampered. In particular, the ecotoxicological effects of EDs have been extensively investigated in aquatic vertebrates, which have nearly identical hormone receptor structures to those of mammals, including humans. Because all hormonal actions occur at low doses in vivo, predicting the potential toxicities of ED candidates using rapid in silico screening is critical to public and environmental health.

QSAR models based on the toxicology of EDs have been conducted utilizing both 2D and 3D descriptors (known as 2D and 3D QSAR, respectively), which reveal the ED ligand binding affinities of estrogen, androgen, and progesterone receptors⁷. Despite the high-precision advantages of 3D QSAR, in which conformational and electrostatic interactions are considered, 2D QSAR retains its own robustness in direct mathematical algorithms, rapid calculations, and extremely low computational loads. In addition, 2D-QSAR models are flexible for use in a wide range of applications while achieving relatively accurate prediction performance.

The OECD QSAR Toolbox is currently one of the most utilized computer software tools, providing freely available and pre-built QSAR models^8,9. Its profiler uses 2D descriptor databases. Since the release of the first version in 2008, the software has been applied in the fields of chemical and biological industries, public health, and environmental safety for full or partial analysis of the potential risks of natural and synthetic compounds, with special interests in carcinogenesis^10,11,12, mutagenicity^13,14,15, and developmental toxicity¹⁶. The application to aquatic toxicology has also been demonstrated, with focus on bioaccumulation and biotransformation¹⁷.

The QSAR Toolbox has been proven useful in predicting the short-term toxicity of a broad range of chemicals¹⁷, as well as the estrogen receptor (ER) binding affinities of EDs¹⁸. However, the acute ecotoxicities of EDs in aquatic vertebrates has not been analyzed using the QSAR Toolbox. In this study, a typical and facile protocol is presented to perform QSAR modeling on the acute adverse effects of EDs with a focus in fish species. The study shows that the QSAR Toolbox is a highly accessible software for calculating and predicting the lethality/mortality of aquatic vertebrates for some representative EDs. Statistical treatment methods for the derived in silico datasets are presented. Figure 1 shows the overall scheme for the general operation of the QSAR Toolbox. The workflow shown in Figure 2 provides straightforward instructions on how to operate the in silico assay to predict acute ecotoxicity of target substances such as endocrine disrupting chemicals.

Protocol

1. Equipment

1. Software: use OECD QSAR Toolbox 4.0 or newer (free download from <https://qsartoolbox.org/download/?) and data analysis software.
2. Computer: for the OECD QSAR Toolbox, use: (i) system type: 64 bit, Windows 7 or newer; (ii) processor: I5 at 2.4 GHz, or a faster processor or equivalent AMD CPU; (iii) installed memory (RAM): 6 GB; (iv) hard disk drive (HDD): 20 GB of free hard drive space (OECD QSAR Toolbox 4.3 Release Notes: <https://qsartoolbox.org/file/2019/02/Toolbox-4.3-Release-Notes-1.pdf>).

2. Procedure

1. OECD QSAR Toolbox
   NOTE: The QSAR Toolbox operates in six consecutive flow modules starting from Input and followed by Profiling, Data, Category Definition, Data Gap Filling, then Report, located at the top of the program interface.
   1. Explore the aforementioned six stages through six toolbar icons by left-clicking. First, look over the stages of Input, Data Gap Filling, and Report that are necessary to perform the automated workflow “Ecotoxicological endpoint” and to document its results.
   2. Take a short look over optional stages Profiling and Data. The Profiling stage provides an initial insight into the target substance’s (eco)toxicity potential and environmental fate characteristics. Optional Data stage enables searching for available experimental data related to the target substance.
2. Input
   1. Upon starting the QSAR Toolbox, the user begins at the Input toolbox stage by default. The QSAR Toolbox creates a working file named “Document 1” automatically, which is displayed in the stage option panel on the left of the program interface. Rename the file, if desired, by right-clicking the working file.
   2. Click on the CAS# button in the actions toolbar, enter the chemical abstract service (CAS) number of the target substance in the available text field, and click Search. The tool will then search for the target substance by CAS number.
   3. If required, choose other search options that are available in the action toolbar such as searching by substance name or simplified molecular-input line-entry system (SMILES) code. SMILES can be entered as 2D non-stereochemical or 3D stereochemical containing forms. Click Name or Structure, respectively. Use the Structure tool to draw the target substance.
   4. The search tool displays the search results through database records in a pop-up window. Choose the record reporting a “high” CAS-SMILES relation (CS Relation field) if multiple records are retrieved for the target substance by checking the box on the left of the record. Click OK.
      NOTE: Proceeding from this point is possible only if the retrieved record contains a SMILES code, as the SMILES code (2D non-stereochemical containing form) is the basis for computation.
   5. Batch mode: to perform the in silico assay for multiple target substances, write a simple substance list in a text editor in which each CAS number is listed in a single row (Supplementary Figure S3). Save the text file with an appropriate name and extension .txt on the computer.
   6. Batch mode: click Data. Then, go to Databases in the stage option panel on the left of the program interface. Make sure databases that are listed under Ecotoxicological Information are checked.
   7. Batch mode: click Input. Select Query from the actions toolbar. Accept the settings set in step 2.2.6 by clicking Yes in the dialog window.
   8. Batch mode: choose the CAS tab. Upload the substance list saved as text file through Load list from your computer.
   9. Batch mode: there are two Add buttons available; click the Add button on the bottom of the pop-up menu and then click Execute. The QSAR Toolbox will display a message on the number of substances that have been retrieved for the search.
      NOTE: Some substances of the loaded list may not be found by the search tool or that several entries may be available for one CAS number. It is not possible to delete substances from the retrieved set of substances.
3. Profiling
   NOTE: The following section is optional. If this is not required, skip to section 2.5.
   1. Click on the toolbox stage button Profiling. Go to Profiling methods in the stage option panel on the left of the program interface.
   2. Click Unselect All. Check all profilers listed under Predefined and those related to aquatic toxicity listed under Endpoint specific such as “Acute aquatic toxicity classification by Verhaar (Modified).”
   3. Finish the selection. Then click on the Apply button in the Actions toolbar.
      NOTE: The QSAR Toolbox provides recommendations on a set of profilers. These are highlighted in green (suitable) and orange (plausible) when choosing Options > Color by: > Endpoint selected in the data matrix in the upper left corner of Profiling methods. Left-click the data matrix field next to the endpoint of interest. Available endpoints are listed in the endpoint tree next to the stage option panel. The profiler Substance type will indicate whether the target substance is a “discrete chemical.” The information is displayed in the expanded endpoint tree “Profile”, “Predefined”, and “Substance type”. Only if the target substance is a discrete chemical can the automated workflow run successfully. “Acute aquatic toxicity classification by Verhaar (modified)” provides a first estimate of the acute aquatic toxicity mechanism of the target substance^19,20. The information is displayed in the expanded endpoint tree “Profile”, “Endpoint Specific”, and “Acute aquatic toxicity classification by Verhaar (modified)”. Five classes are available: (class 1) inert chemicals (baseline toxicity); (class 2) less inert chemicals; (class 3) reactive chemicals; (class 4) specifically acting chemicals; and (class 5) for chemicals not possible to classify.
   4. Right-click Parameter in the endpoint tree to run integrated 2D and 3D QSAR models available in the QSAR Toolbox, if desired. Click Calculate/extract all parameters for all chemicals in the pop-up menu.
   5. 2D and 3D QSAR models compiled in Parameter provide numeric values. Use “Profiling methods” for qualitative information (see step 2.3.1).
4. Data
   NOTE: This section is optional. If it is not required, skip to section 2.5.
   1. Click on the toolbox stage button Data. Then, click Gather from the Actions toolbar.
   2. Select All endpoints to gather all experimental data, then Choose to gather endpoint specific experimental data. As an example, if aquatic toxicity is the user’s focus, click Choose > Ecotoxicological Information > Aquatic toxicity > OK.
      NOTE: Choosing to gather experimental data for all endpoints may lead to extended processing time. The user can adapt the hierarchy of the endpoint tree to the specific purpose. This changes the manner in which data are displayed.
   3. If desired, right-click the endpoint of interest in the endpoint tree area. Choose Set tree hierarchy in the pop-up menu. Organize the endpoint tree in the preferred manner using the available terms and arrows and click OK.
   4. If desired, export the gathered data as an Excel file. Right-click on the endpoint of interest and choose Export Data matrix in the pop-up menu.
   5. A “Matrix export” wizard opens and enables adding other endpoints to the export list. Finish the selection, click Export, and save the file on the computer.
      NOTE: Exporting data from all databases is not possible. For example, data retrieved from the database “ECHA CHEM” cannot be saved.
5. Data gap filling
   1. Click on the toolbox stage button Data gap filling. Then, click Automated in the Actions toolbar.
   2. Select Ecotoxicological Endpoint > Fish, LC50 (lethal concentration, 50%) at 96 h for Pimephales promelas (mortality). Click OK. A “Workflow Controller” will appear, and processing will take up to several minutes, especially in batch mode.
      NOTE: The QSAR Toolbox automatically applies a defined set of profilers when searching for suitable substances with available experimental data for the prediction. The experimental data [e.g., effect concentrations 96 h LC₅₀ (P. promelas) or 96 h EC₅₀ (P. promelas, mortality)] are used to generate the prediction for the target substance by either linear approximation or nearest neighbor method. Note that the methods of linear approximation and nearest neighbor are referred to as trend analysis (labeled as “T”) and read-across (labeled as “R”), respectively.
   3. The user will receive a message if the prediction is performed successfully. Click OK and close the workflow controller indicating “Finished workflow” by clicking x in the upper right corner.
   4. Batch mode: upon starting the automated workflow, the user will be asked to specify the range of substances over which to execute the workflow. Accept the full range of substances selected by default in the dialog window by clicking OK.
   5. Batch mode: the user will not receive a message indicating whether a prediction was run successfully or unsuccessfully. Close the workflow controller indicating “Finished workflow” at the end of the batch processing by clicking x in the upper right corner.
6. Report
   1. If a prediction was successfully executed, click on the toolbox stage button Report.
      NOTE: No reports can be generated in batch mode.
   2. Scroll down and find the prediction value in matrix field located in a yellow highlighted row next to endpoint “96-h.” The predicted value is labelled with “T” or “R.” Activate this specific data matrix field by left-clicking it.
   3. Click Prediction in the Actions toolbar. Customize the report content and appearance in the pop-up wizard. Three types of reports are available: (i) prediction, (ii) category, and (iii) data matrix.
   4. The wizard allows the user to fill in the author’s name and contact details. Write a short summary, provide a detailed explanation of the mechanistic interpretation, or provide justification for the adequacy of the prediction.
   5. Include additional information related to the executed prediction, if desired. The extent of additional information depends on the user.
   6. Go through the wizard by clicking Next. Finally, click Create report and save the prediction and category reports as PDF files and the data matrix as an Excel spreadsheet on the computer.
   7. Find additional details on the functionalities of the QSAR Toolbox and automated workflows in the application manual for the OECD QSAR Toolbox v.4 (F1 help on the keyboard). Details on the algorithms and rationale behind the automated workflow are described by Dimitrov et al.⁸ and Yordanova et al.⁹.

3. Application

1. If using the predicted effect concentration (i.e., 96-h LC₅₀ of P. promelas) in the environmental risk assessment, use the lower limit of the 95% confidence interval. Find the data on the first page of the saved prediction report (PDF) in “Prediction summary”, “Predicted value: <mean> (from <lower_limit> to <upper_limit>).”
   NOTE: The notes given here are based on results of the comparison between predicted and experimental data for a set of target substances reported in this study. Selecting the lower end of the 95% confidence range will increase the likelihood that the predicted effect concentration will not underestimate the real toxicity of the substance (see the representative results). The predicted effective concentration of the lower limit of the 95% confidence interval will therefore present a safer basis for risk assessment.

Results

The example described in this study was implemented for quantitative analysis and prediction of acute toxicities of selected EDs in fish. When the predicted data points were plotted versus experimental data points as a log-log scale, a positive correlation between both was found for all fish and a representative species, namely, Pimephales promelas (fathead minnow; Figure 3). In both cases, the slope of the linear regression appeared to be comparable (predicted LC₅₀/experimental LC₅₀ = 0.611 and 0.602 for all fish and P. promelas, respectively). Because of the limited amount of experimental data, the number of available values from experimental observation was usually smaller than that from computational prediction. Application of the tolerance factor as 5-fold for the computational capability²¹ resulted in 94% (34/36) and 96% (26/27) of the protective prediction for all fish and P. promelas, respectively. Based on this prediction, 3',5,7-trihydroxy-4',6-dimethoxyisoflavone and 1,4-benzenediol appeared to exhibit calculated LC₅₀ values greater than the tolerance limit.

To enable safety assessment at the highest reliability, further computational analysis was performed by plotting the predicted lower limit of the 95% confidence interval of LC₅₀ (instead of the mean values used in Figure 3) versus the experimentally derived values (Figure 4). In this evaluation with an elevated safety threshold, 92% (33/36) of the total tested endocrine disrupting compounds were shown to fall into the protective range when compared to the experimentally derived values except for: 3',5,7-trihydroxy-4',6-dimethoxyisoflavone; 1,4-benzenediol; and 4-hexylphenol.

Based on assessments of the entire species available from the database, values for the predicted and experimental 96-h log₁₀LC₅₀ exhibited linearity with the log₁₀K_OW values in the domain between -1 and 7, indicating a hyperbolic correlation between LC₅₀ and K_OW. An overall trend existed whereby the LC₅₀ decreased for higher K_OW values of EDs for the data obtained from both computational predictions and experiments, suggesting increasing acute toxicity in fish species for EDs with higher hydrophobicity (Supplementary Figure S1).

By the rule-based ER profiler embedded in the OECD QSAR Toolbox, the ER binding affinities of the EDs were categorized as non-binding as well as weak, moderate, strong, and very strong binders, in order of increasing binding affinity¹⁸. Accordingly, the statistical distribution of log₁₀K_ow could be displayed as a qualitative classification of ER binding affinity (Supplementary Figure S2). Overall, the changes in K_ow distribution ranges and their mean levels appeared to not have a defined tendency. Similarly, the distributions of predicted and experimental LC₅₀ were shown as the extent of ER binding affinity (Figure 5). In this case, mean levels of predicted LC₅₀ for ER binders were higher than those of non-binders. By contrast, for the experimental LC₅₀, the mean levels of non- and weak binders were higher than those of stronger ER binders.

Figure 1: Basic scheme of the general workflow of the OECD QSAR Toolbox.
Please click here to view a larger version of this figure.

Figure 2: Workflow.
Shown is the workflow conceptualizing the modules and sequences applied to predict the acute toxicities of endocrine disruptors (EDs) in fish using the OECD QSAR Toolbox. Please click here to view a larger version of this figure.

Figure 3: Predicted vs. experimental 96-h LC₅₀ of EDs in Table 1 for all fish (blue diamonds, n = 36) and a selected species P. promelas (cyan diamonds, n = 27).
For the predicted LC₅₀, the average (“AVE”) values are displayed. The dashed lines represent linear regressions for the two groups: for all fish (light blue), predicted LC₅₀^AVE = 0.611 x (experimental LC₅₀) + 0.277 (adjusted r² = 0.408); and for P. promelas (light cyan), predicted LC₅₀^AVE = 0.602 x (experimental LC₅₀) + 0.385 (adjusted r² = 0.441). The solid diagonal line shows unity in which the predicted and experimental values are equal²¹. The dotted gray line shows the 5-fold tolerance limit of the computational capability¹⁹. Outliers: 3',5,7-trihydroxy-4',6-dimethoxyisoflavone (*) and 1,4-benzenediol (**). Please click here to view a larger version of this figure.

Figure 4: Predicted (lower limit of 95% confidence interval, “low-95%”) vs. experimental 96-h LC₅₀ of EDs in Table 1 for all fish (n = 36).
The dashed line represents the linear regression: predicted LC₅₀^low-95% = 0.470 x (experimental LC₅₀) - 0.312, where adjusted r² = 0.193. The solid diagonal line indicates unity where the predicted and experimental values are equal to each other¹⁹. Outliers: 3',5,7-trihydroxy-4',6-dimethoxyisoflavone (*), 1,4-benzenediol (**), and 4-hexylphenol (***). Please click here to view a larger version of this figure.

Figure 5: Distributions of predicted (solid boxes, n = 8–20 for each category) and experimental (dashed boxes; n = 3–16 for each category) 96-h LC₅₀ depending on ER binding affinity of EDs in Table 1 for all fish.
A box plot represents: (A) mean (small square with a bold horizontal bar), (B) 1^st and 3^rd quartiles (lower and upper– ends of the box, respectively), (C) median (horizontal segment inside the box), (D) 5^th and 95^th percentile (lower and upper error bars, respectively), (E) 1^st and 99^th percentile (lower and upper x, respectively), and (F) minimum and maximum (lower and upper –, respectively). Please click here to view a larger version of this figure.

No.	CAS Registry Number		Substance Name		SMILES Formula (2D non-stereochemical form)						Log Kow	AVE predicted 96-h LC50 (mg/L)	LOWER 95% CI predicted 96-h LC50 (mg/L)	Profiler - Estrogen Receptor Binding
1	50-28-2		17-β Estradiol		CC12CCC3C(CCc4cc(O)ccc34)C1CCC2O						4.01	3.62	1.42	Very strong binder, OH group
2	57-63-6		17-α Ethinyl- estradiol		CC12CCC3C(CCc4cc(O)ccc34)C1CCC2(O)C#C						3.67	3.00	1.18	Strong binder, OH group
3	80-05-7		2,2-bis(4-hydroxyphe-nyl)propane (Bisphenol A)		CC(C)(c1ccc(O)cc1)c1ccc(O)cc1						3.32	4.68	1.80	Very strong binder, OH group
4	80-46-6		4-tert-Pentylphenol		CCC(C)(C)c1ccc(O)cc1						3.91	2.27	0.87	Weak binder, OH group
5	140-66-9		4-tert-Octylphenol		CC(C)(C)CC(C)(C)c1ccc(O)cc1						5.28	0.38	0.14	Strong binder, OH group
6	446-72-0		Genistein [3',5,7-trihydroxy-4',6-dimethoxyisoflavone]		Oc1ccc(cc1)C1=COc2cc(O)cc(O)c2C1=O						2.84	32.00	10.03	Very strong binder, OH group
7	10161-33-8		17β-Trenbolone		CC12C=CC3C(CCC4=CC(=O)CCC=34)C1CCC2O						2.65	124.72	19.75	Strong binder, OH group
8	67747-09-5		Prochloraz (DMI fungicide)		CCCN(CCOc1c(Cl)cc(Cl)cc1Cl)C(=O)n1ccnc1						4.1	5.19	1.74	Non binder, without OH or NH2 group
9	84852-15-3		4-Nonylphenol		CC(C)CCCCCCc1ccc(O)cc1						5.92	0.21	0.07	Strong binder, OH group
10	69-72-7		salicylic acid		OC(=O)c1ccccc1O						2.26	24.07	9.31	Weak binder, OH group
11	80-09-1		4,4’-dihydroxydiphenyl sulphone (Bisfenol S)		Oc1ccc(cc1)S(=O)(=O)c1ccc(O)cc1						1.65	48.67	10.67	Very strong binder, OH group
12	84-74-2		phthalic acid, dibutyl ester		CCCCOC(=O)c1ccccc1C(=O)OCCCC						4.5	0.76	0.06	Non binder, without OH or NH2 group
13	92-88-6		4,4′-dihydroxybiphenyl		Oc1ccc(cc1)-c1ccc(O)cc1						2.8	12.05	4.20	Moderate binder, OH grooup
14	94-13-3		4-hydroxybenzoic acid, propyl ester		CCCOC(=O)c1ccc(O)cc1						3.04	10.32	3.86	Moderate binder, OH grooup
15	98-54-4		4-tert-butylphenol		CC(C)(C)c1ccc(O)cc1						3.31	4.36	1.68	Weak binder, OH group
16	97-23-4		2,2′-dihydroxy-–5,5′-dichlorodiphenyl-methane		Oc1ccc(Cl)cc1Cc1cc(Cl)ccc1O						4.26	0.48	0.10	Very strong binder, OH group
17	97-53-0		eugenol		COc1cc(CC=C)ccc1O						2.27	14.70	5.60	Weak binder, OH group
18	99-76-3		4-hydroxybenzoic acid, methyl ester		COC(=O)c1ccc(O)cc1						1.96	38.20	14.01	Weak binder, OH group
19	103-90-2		N-(4-hydroxyphenyl) acetamide		CC(=O)Nc1ccc(O)cc1						0.46	338.97	43.39	Weak binder, OH group
20	106-44-5		p-cresol		Cc1ccc(O)cc1						1.94	20.47	7.14	Weak binder, OH group
21	108-39-4		m-cresol		Cc1cccc(O)c1						1.96	23.45	9.17	Weak binder, OH group
22	108-45-2		1,3-phenylenediamine		Nc1cccc(N)c1						-0.33	34.60	0.00	Weak binder, NH2 group
23	108-46-3		1,3-dihydroxybenzene		Oc1cccc(O)c1						0.8	123.03	27.06	Weak binder, OH group
24	108-91-8		cyclohexylamine		NC1CCCCC1						1.49	28.08	1.40	Weak binder, NH2 group
25	119-36-8		salicylic acid, methyl ester		COC(=O)c1ccccc1O						2.55	16.16	5.68	Weak binder, OH group
26	120-47-8		4-hydroxybenzoic acid, ethyl ester		CCOC(=O)c1ccc(O)cc1						2.47	19.93	7.40	Weak binder, OH group
27	120-80-9		1,2-dihydroxybenzene		Oc1ccccc1O						0.88	11.14	0.01	Weak binder, OH group
28	123-31-9		1,4-dihydroxybenzene [1,4-benzenediol]		Oc1ccc(O)cc1						0.59	90.75	33.19	Weak binder, OH group
29	131-53-3		2,2′-dihydroxy-4-methoxybenzophenone		COc1ccc(C(=O)c2ccccc2O)c(O)c1						3.82	3.97	1.46	Very strong binder, OH group
30	131-56-6		2,4-dihydroxybenzophenone		Oc1ccc(c(O)c1)C(=O)c1ccccc1						2.96	12.04	4.73	Strong binder, OH group
31	131-57-7		2-hydroxy-4-methoxybenzophenone		COc1ccc(C(=O)c2ccccc2)c(O)c1						3.79	5.96	2.27	Strong binder, OH group
32	599-64-4		4-cumylphenol		CC(C)(c1ccccc1)c1ccc(O)cc1						4.12	2.15	0.84	Strong binder, OH group
33	2855-13-2		1-amino-3-aminomethyl-3,5,5-trimethyl-cyclohexane		CC1(C)CC(N)CC(C)(CN)C1						1.9	30.65	1.53	Moderate binder, NH2 group
34	6864-37-5		3,3′-dimethyl-4,4′-diaminodicyclohexylmethane		CC1CC(CCC1N)CC1CCC(N)C(C)C1						4.1	1.07	0.05	Strong binder, NH2 group
35	25013-16-5		tert-butyl-4-hydroxyanisole		COc1ccc(O)c(c1)C(C)(C)C						3.5	4.85	1.85	Moderate binder, OH grooup
36	147315-50-2		2-(4,6-diphenyl-1,3,5-triazin-2-yl)-5-(hexyloxy)phenol		CCCCCCOc1ccc(c(O)c1)-c1nc(nc(n1)-c1ccccc1)-c1ccccc1						6.24	0.17	0.06	Strong binder, OH group
37	88-68-6		2-aminobenzamide		NC(=O)c1ccccc1N						0.35	694.00	84.30	Weak binder, NH2 group
38	611-99-4		4,4′-dihydroxybenzophenone		Oc1ccc(cc1)C(=O)c1ccc(O)cc1						2.19	37.74	14.67	Very strong binder, OH group
39	27955-94-8		1,1,1-tris(4-hydroxyphenol)ethane		CC(c1ccc(O)cc1)(c1ccc(O)cc1)c1ccc(O)cc1						4.38	2.09	0.82	Very strong binder, OH group
40	87-18-3		salicylic acid, 4-tert-butylphenyl ester		CC(C)(C)c1ccc(OC(=O)c2ccccc2O)cc1						5.73	0.24	0.09	Strong binder, OH group
41	47465-97-4		3,3-bis(3-methyl-4-hydroxyphenyl)2-indolinone		Cc1cc(ccc1O)C1(C(=O)Nc2ccccc12)c1ccc(O)c(C)c1						4.48	2.07	0.77	Very strong binder, OH group
42	99-96-7		p-hydroxybenzoic acid		OC(=O)c1ccc(O)cc1						1.58	8.54	0.00	Weak binder, OH group
43	80-07-9		1-Chloro-4-(4- chlorophenyl)sulfonylbenz		Clc1ccc(cc1)S(=O)(=O)c1ccc(Cl)cc1						3.9	3.92	0.85	Non binder, without OH or NH2 group
44	84-65-1		9,10-Anthraquinone		O=C1c2ccccc2C(=O)c2ccccc12						3.39	7.00	3.54	Non binder, without OH or NH2 group
45	85-44-9		2-benzofuran-1,3-dione		O=C1OC(=O)c2ccccc12						1.6	2.69	0.00	Non binder, without OH or NH2 group
46	92-84-2		10H-Phenothiazine		N1c2ccccc2Sc2ccccc12						4.15	1.07	0.08	Non binder, without OH or NH2 group
47	2855-13-2		1-amino-3-aminomethyl-3,5,5-trimethyl-cyclohexane		CC1(C)CC(N)CC(C)(CN)C1						1.9	30.65	1.53	Moderate binder, NH2 group
48	50-27-1		Estriol		CC12CCC3C(CCc4cc(O)ccc34)C1CC(O)C2O						2.45	21.21	8.29	Very strong binder, OH group
49	50-50-0		beta-Estradiol-3-benzoate		CC12CCC3C(CCc4cc(OC(=O)c5ccccc5)ccc34)C1CCC2O						5.47	0.36	0.02	Strong binder, OH group
50	53-16-7		Estrone		CC12CCC3C(CCc4cc(O)ccc34)C1CCC2=O						3.13	7.78	3.06	Strong binder, OH group
51	92-52-4		Biphenyl		c1ccc(cc1)-c1ccccc1						4.01	4.10	0.47	Non binder, without OH or NH2 group
52	92-69-3		p-Phenylphenol		Oc1ccc(cc1)-c1ccccc1						3.2	5.99	1.82	Moderate binder, OH grooup
53	96-29-7		2-Butanone oxime		CCC(C)=NO						0.63	32.67	2.49	Non binder, non cyclic structure
54	121-75-5		Malathon		CCOC(=O)CC(SP(=S)(OC)OC)C(=O)OCC						2.36	37.73	3.33	Non binder, non cyclic structure
55	123-07-9		4-Ethylphenol		CCc1ccc(O)cc1						2.58	13.63	4.65	Weak binder, OH group
56	645-56-7		4-n-Propylpehnol		CCCc1ccc(O)cc1						3.2	7.32	2.55	Weak binder, OH group
57	1638-22-8		p-Butyl phenol		CCCCc1ccc(O)cc1						3.65	4.09	1.39	Weak binder, OH group
58	1912-24-9		Atrazine		CCNc1nc(Cl)nc(NC(C)C)n1						2.61	30.87	4.63	Non binder, without OH or NH2 group
59	40596-69-8		Methoprene		COC(C)(C)CCCC(C)CC=CC(C)=CC(=O)OC(C)C						5.5	0.08	0.00	Non binder, non cyclic structure
60	1987-50-4		4-Heptylphenol		CCCCCCCc1ccc(O)cc1						5.01	0.66	0.22	Moderate binder, OH grooup
61	92-86-4		p,p'-Dibromobiphenyl		Brc1ccc(cc1)-c1ccc(Br)cc1						5.72	0.11	0.02	Non binder, without OH or NH2 group
62	480-41-1		Naringenin		Oc1ccc(cc1)C1CC(=O)c2c(O)cc(O)cc2O1						2.52	27.84	10.87	Very strong binder, OH group
63	486-66-8		Daidzein		Oc1ccc(cc1)C1=COc2cc(O)ccc2C1=O						2.55	36.47	11.71	Very strong binder, OH group
64	491-70-3		Luteolin		Oc1cc(O)c2C(=O)C=C(Oc2c1)c1ccc(O)c(O)c1						2.53	43.75	14.28	Very strong binder, OH group
65	491-80-5		Biochanin A		COc1ccc(cc1)C1=COc2cc(O)cc(O)c2C1=O						3.41	15.87	3.70	Strong binder, OH group
66	520-18-3		Kaempferol		Oc1ccc(cc1)C1Oc2cc(O)cc(O)c2C(=O)C=1O						1.96	70.98	8.05	Very strong binder, OH group
67	2051-60-7		2-Chlorobiphenyl (PCB 1)		Clc1ccccc1-c1ccccc1						4.53	0.77	0.16	Non binder, without OH or NH2 group
68	2051-61-8		3-Chlorobiphenyl (PCB 2)		Clc1cccc(c1)-c1ccccc1						4.58	0.77	0.16	Non binder, without OH or NH2 group
69	2051-62-9		4-Chloro-1,1'-biphenyl		Clc1ccc(cc1)-c1ccccc1						4.61	0.77	0.16	Non binder, without OH or NH2 group
70	2446-69-7		p-n-Hexylphenol [4-hexylphenol]		CCCCCCc1ccc(O)cc1						4.52	1.22	0.42	Moderate binder, OH grooup
71	14938-35-3		4-n-Amylphenol		CCCCCc1ccc(O)cc1						4.06	2.44	0.89	Weak binder, OH group
72	17924-92-4		Zearalenone		CC1CCCC(=O)CCCC=Cc2cc(O)cc(O)c2C(=O)O1						3.58	7.22	2.66	Strong binder, OH group
73	1743-60-8		beta-Estradiol 3-benzoate 17-nbutyrate		CC(=O)OC1CCC2C3CCc4cc(O)ccc4C3CCC12C						4.95	0.91	0.35	Strong binder, OH group
74	479-13-0		Coumestrol		Oc1ccc2c(OC(=O)c3c-2oc2cc(O)ccc32)c1						1.57	52.16	11.44	Very strong binder, OH group

Table 1: List of evaluated endocrine disrupting chemicals. Average mean (AVE) and lower 95% confidence interval (CI) effective concentrations (95-h LC₅₀, Pimephales promelas) as well as Estrogen Receptor Binding were predicted with the QSAR Toolbox version 4.3 Automated Workflow. Log₁₀K_ow was retrieved via QSAR Toolbox version 4.3 from KOWWIN v1.68, 2000, U.S. Environmental Protection Agency. Experimental log₁₀K_ow values were preferred over predicted values. The target substance list was compiled from previously reported lists of EDs^22,23,24.

Supplementary Information. Please click here to download this file.

Discussion

The versatility of the OECD QSAR Toolbox as analytic software for ecotoxicology is shown here with specific interest in the adverse effects of endocrine disrupting chemicals on aquatic vertebrates. In addition, a simple and standard protocol was demonstrated for predicting acute toxicity (96-h LC₅₀) of 74 representative EDs (Table 1) for fish species. This was achieved by applying category building, data gap filling, and ER profiling modules embedded in the QSAR Toolbox (Figure 1, Figure 2).

The linear correlation between log₁₀LC₅₀ and log₁₀K_OW with a negative slope (as shown in Supplementary Figure S1) has long been known as a standard quantitative relationship in QSAR analyses²⁵, where higher toxicity is shown the more hydrophobic a given chemical is. As can be seen from a simple calculation, the general mathematical relation that includes Equation S1 and Equation S2 (Supplementary Information) is a converted expression from the following power function²⁶:

From the plot of (Equation 2), characterizing an intermediate range of K_OW²⁶ may be possible by adjusting the parameters a and b, where a certain variation in hydrophobicity (or hydrophilicity) does not significantly change the endpoint of acute toxicity.

Comparative analyses between the computational predictions and experimental observations on the LC₅₀, as shown in Figure 3 and Figure 4, have been typically reported in studies of QSAR for various aquatic toxicants, including technical nonionic surfactants²⁷, triazole fungicides²⁸, and pesticide metabolites²¹. This type of retrospective validation provides information on how far a given QSAR tool can reach in terms of comparative performance to experimental results. In this study of acute toxicity in fish, the QSAR Toolbox was proven to provide protective predictions for over 90% of tested EDs in all fish and in a single species, Pimephales promelas.

Further identifying the three outlier chemicals in Figure 3 and Figure 4, which showed higher predicted LC₅₀ on average and at a minimum, respectively, is required. First, the 3',5,7-trihydroxy-4',6-dimethoxyisoflavone is a type of flavonoid (more specifically, an isoflavone), which is considered to be generally safe and used in herbal pharmaceuticals; however, it still has estrogen-related concerns²⁹ and may cause acute toxicity probably through oxidative phosphorylation uncoupling³⁰. Next, the 1,4-benzenediol, called hydroquinone, is a phenolic compound that can trigger a non-specific and cytotoxic immune response in fish³¹. Finally, the 4-hexylphenol has been known to exhibit sufficient positive estrogenic activity to be classified as an ED³². It has been well-studied that the main reason of the acute toxicity of hydroquinone is the reduction-oxidation (redox) cycling. The hydroquinone is oxidized to benzoquinone and reduced back to semi-quinone or hydroquinone repeatedly, with depleting cofactors and generating reactive oxygen species³³. The other two chemicals may require deeper investigations to reveal their mechanisms of action in acute ecotoxicity using molecular docking approaches such as that used by Panche et al.³⁴, which cannot be covered by the QSAR Toolbox.

EDs interfere with the endocrine system mainly through physicochemical interactions with steroid receptors such as estrogen and androgen receptors, which are of considerable interest in QSAR modeling studies³⁵. Considering this, the QSAR Toolbox is robust in terms of facile and rapid classification of ER binding affinities for a set of chemicals based only on the 2D descriptors of molecular structures. When this ER profiler system was applied to our list of EDs, no clear correlation was found between ER binding affinity and hydrophobicity (Supplementary Figure S2). This result may be explained by the fact that the formation of a steroid-receptor complex is not a direct consequence of a hydrophobic bonding contribution but should be accompanied by a conformational change in the active-site receptor structure³⁶. The receptor binding can be also due to hydrogen-bonding and π-stacking.

Additionally, the position of each chemical group on the molecule may affect the receptor binding, even if the hydrophobicity and number of hydrogen-bond acceptors-donors remain the same. Second, the ER profiler produced contrary trends between predicted and experimental LC₅₀ mean levels with increasing ER binding affinity (Figure 5). This may be because the lethality of parents in an acute toxicity test are not due to ER binding but rather to narcosis in most cases, or to redox cycling in the case of hydroquinone. For example, more extensive analysis, including the chronic toxicity, is required for a larger set of EDs to define predictive limitations of the current version of the QSAR Toolbox.

This preliminary research may also have public health implications because steroids (androgens, estrogens, progestines, and corticoids) and their receptors exhibit similar or even identical macromolecular structures across vertebrates⁵. These types of analogous endocrine signaling systems may operate using a common mechanism in key events of EDs⁵. Nevertheless, additional and complementary methodologies are required to illuminate this vast and complex aspect [for example, by performing computational modeling of absorption, distribution, metabolism, and excretion (ADME), and/or adverse outcome pathway (AOP)]³⁸. Furthermore, because most of the scientific and public concerns raised about the adverse effects of EDs are related to their chronic toxicities, improving the databases and algorithms in the QSAR Toolbox and producing reliable long-term ecotoxicology predictions for EDs are both necessary.

This paper demonstrates the application of QSAR Toolbox to compare ecotoxicological LC₅₀ values for fish with log₁₀K_ow values of EDs. Throughout the protocol, it results in weak relationships between the two parameters, as it has been revealed by previous studies (e.g., Kim et al.³⁹) that log₁₀K_ow is not a good direct predictor of aquatic LC₅₀. In spite of this limitation, this protocol provides a general review or “vignette” to describe how to use the dashboard for a given purpose, since it is a valid application to use the QSAR Toolbox for investigating correlations between LC₅₀ (or ER binding affinity) and log₁₀K_ow, or as a tool for rapid acute ecotoxicity screening. Nevertheless, it should be noted that (1) illuminating the link between estrogen receptor binding and chronic toxicity, rather than acute toxicity (lethality), is more relevant so that clearer correlations may be found, and (2) the androgen receptor, together with that of estrogen, also plays a critical role in reproductive toxicity. Therefore, it is required for the future version of the QSAR Toolbox to improve the prediction functions in light of those two points.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Adobe Acrobat Reader DC	Adobe Systems Software Ireland Limited	NA	Required to view prediction and category report
Computer	System: Microsoft Corporation	NA	Recommended system properties: (i) system type: 64 bit, Microsoft Windows 7 or newer, (ii) processor: I5 at 2.4 GHz or faster processor or equivalent AMD CPU, (iii) Installed memory (RAM): 6 GB of RAM, (iv) Hard Disk Drive (HDD): 20 GB free hard drive space
Microsoft Editor	Microsoft Corporation	NA	Required to upload a substance list of CAS numbers (batch mode) to the OECD QSAR Toolbox as .txt file (text file)
Microsoft Excel 2016	Microsoft Corporation	NA	Required to export data from OECD QSAR Toolbox as .cvs, .xls or .xlsx files
OECD QSAR Toolbox version 4.0 or newer	Organisation for Economic Co-operation and Development	NA	Required to run OECD QSAR Toolbox Automated Workflows; free download: https://qsartoolbox.org/download/
OriginPro 9	OriginLab Corporation	NA	Optional program for data analysis; similar tools possible