Author: Joop Hermens
Reviewers: Monika Nendza and Emiel Rorije
Date uploaded: 15th March 2024
Learning objectives:
You should be able to:
Keywords: quantitative structure-activity relationship (QSAR), Modes of Action (MOA) based classification schemes, octanol-water partition coefficient, excess toxicity
Introduction
The number of chemicals for which potential risks to the environment has to be estimated is enormous. Section 6.5 on ‘Regulatory Frameworks’ discusses the EU regulation on REACH (Registration, Evaluation, Authorisation and Restriction of Chemicals) and gives an indication of the number of chemicals that are registered under REACH. Because of the high number of chemicals, there is a strong need for predictive methods including Read-across from related chemicals, Weight of Evidence approaches, and calculations based on chemical structures (quantitative structure-activity relationships, QSARs).
This section discusses the following topics:
Prediction methodologies
A major in silico prediction methodology is based on quantitative structure-activity relationships (QSARs) (ECHA 2017a). A QSAR is a mathematical model that relates ecotoxicity data (the Y- variable) with one or a combination of structural descriptors and/or physical-chemical properties (the X-variable or variables) for a series of chemicals (see Figure 1).
Figure 1. The principle of a QSAR.
Note: LC50, EC50: concentrations with 50 % effect on survival (LCxx: Lethal Concentration xx%) or on sublethal parameters (ECxx: Effect Concentration xx%), NOEC: No-Observed Effect Concentrations regarding effects on growth or reproduction, or in general the most sensitive parameter. A QSAR is related to molecular events and, therefore, concentrations should always be based on molar units.
Most models are based on linear regression between Y and X. Different techniques can be used to develop a QSAR including a simple graphical presentation, linear regression equations between Y and X or multiple parameter equations based on more than one property (Y versus X1, X2, etc.). Also, multivariate techniques, such as Principal Component Analysis (PCA) and Partial Least Square Analysis (PLS), are applied. More information on these techniques can be found in section 3.4.3 ‘Quantitative structure-property relationships (QSPRs)’.
Multi-parameter linear regression takes the form of Y(i) = a1X1(i) + a2X2(i) + a3X3(i) + ... + b (1)
See Box 1 for more details.
Nowadays, Machine Learning techniques, like Support Vector Machines (SVM), Random Forest (RF) or neural networks, are also applied to establish a mathematical relationship between toxicological effect data and all kinds of chemical properties. The advantage is that it should allow to model non-linear relationships, but at the expense of interpretability of the model. Machine Learning techniques and QSAR models are outside the scope of this section.
Box 1: Statistics and validation of QSARs
Multiple-parameter linear regression
Multiple linear regression equations take the form of
Y(i) = a1X1(i) + a2X2(i) + a3X3(i) + … + b (1)
where Y(i) is the value of the dependent parameter of chemical i X1-X3(i) are values for the independent parameters (the chemical properties) of chemical i a1-a3 are regression coefficients and b is the intercept of the linear equation
Statistical quality of the model The overall quality of the equation is presented via the Pearson’s correlation coefficient (r) and the standard error of estimate (s.e.). The closer r is to 1.0, the better the fit of the relationship is. The square of r represents the percentage of information in the Y variable that is explained by the X-variable(s). The significance of the influence of a certain X parameter in the relationship is indicated by the confidence interval of the regression coefficient.
Validation of the model (Eriksson et al. 2003) The model is developed using a so-called “training set” that consists of a limited number of carefully selected chemicals. The validity of such a model should be tested by applying it to a "validation set" i.e., a set of compounds for which experimental data can be compared with the predictions, but which have not been used in the establishment of the (mathematical form of) the model. Another validation tool is cross-validation. In cross-validation, the data are divided in a number of groups and then a number of parallel models are developed from reduced data with one of the groups deleted. The predictions from the left-out chemicals are compared with actual data and the differences are used to calculate the so-called “cross-validated” r2 or Q2 from the correlation observed versus predicted of the left-out chemicals. In the so-called leave-one-out (LOO) approach, one chemical is left out and predicted from a model calculated from the remaining compounds. The LOO approach is often considered to yield a too optimistic value for the true model predictivity. Some modelling techniques apply a wide set (hundreds) of molecular descriptors (experimental and/or theoretical). This may lead to overfitted models and in these cases a good validation procedure is essential, as overfitting will automatically lead to poor external predictive performance (a low Q2).
Some modelling techniques apply a wide set (hundreds) of molecular descriptors (experimental and/or theoretical). This may lead to overfitted models and in these cases a good validation procedure is essential, as overfitting will automatically lead to poor external predictive performance (a low Q2).
OECD (2004) identified a number of principles for (Q)SAR validation. The principles state that “to facilitate the consideration of a (Q)SAR model for regulatory purposes, it should be associated with the following information: (i) a defined endpoint, (ii) an unambiguous algorithm, (iii) a defined domain of applicability, (iv) appropriate measures of goodness-of-fit, robustness and predictivity and (v) a mechanistic interpretation, if possible.” |
The Y-variable in a QSAR can for example be fish LC50 data (concentration killing 50 % of the fish) or NOEC (no-observed effect concentrations) for effects on the growth of Daphnia magna, after a specific exposure duration (e.g. LC50 to fish after 96 hours). The X-variable may include properties such as molecular weight, octanol water partition coefficient KOW, electronic and topological descriptors (e.g., quantum mechanics calculations), or descriptors related to the chemical structure such as the presence or absence or the number of different functional groups. Uptake and bioaccumulation of organic chemicals depend on their hydrophobicity and the octanol-water partition coefficient is a parameter that reflects differences in hydrophobicity. The effect of electronic or steric parameters is often related to the potency of chemicals to interact with the receptor or target, or more directly to reactivity (towards specific biological targets). More information on chemical properties is given in section 3.4.3 ‘Quantitative structure-property relationships (QSPR)’ and section 3.4.1 ‘Relevant chemical properties’.
Read-across is the appropriate data-gap filling method for “qualitative” endpoints like skin sensitisation or mutagenicity for which a limited number of results are possible (e.g. positive, negative, equivocal). Read-across is frequently applied in predicting human-health related endpoints. Furthermore read-across is recommended for “quantitative” endpoints (e.g., 96h-LC50 for fish) if only a low number of analogues with experimental results are identified. In that case it is simply assumed that the quantitative value of the endpoint for the substance of interest is identical to the value for the closest structural analogue for which experimental data is available. More information on read across can be found in ECHA (2017b).
Classification of chemicals based on chemical structure into modes of action (MOA) and QSAR equations
Information on mechanisms and mode of action is essential when developing predictive methods in integrated testing strategies (Vonk et al. 2009). “Mode of action” has a broader meaning than “mechanism of action”. Mode of action refers to changes at the cellular level while mechanism of action refers to the interaction of a chemical with a specific molecular target. In QSAR research the terminology is not always clearly defined and mode of action is used both in the broad sense (change at cellular level) as well as the narrow definition (interaction with a target). A QSAR should preferably be developed for a series of chemicals with a known and similar mechanism or mode of action (OECD 2004). Several schemes to classify chemicals according to their mode of action (MOA) are available. Well known MOA classification systems are those from Verhaar et al. (1992) and the US Environmental Protection Agency (US-EPA) (Russom et al. 1997). The latter classification scheme is based on a number of information sources, including results from fish physiological and behaviour studies, joint toxicity data and similarity in chemical structure. The EPA scheme includes a number of groups including: narcotics (or baseline toxicants), oxidative phosphorylation uncouplers, respiratory inhibitors, electrophiles/proelectrophiles, and Acetylcholinesterase (AChE) inhibitors. The Verhaar scheme is relatively simple and has identified four broad classes, including: Class 1, inert chemicals, Class 2, less inert chemicals, Class 3, reactive chemicals, and Class 4, specifically acting chemicals. Classes 1 and 2 are also known as non-polar and polar narcosis, respectively. Class 3 and 4 include chemicals with so-called “excess toxicity”, i.e. the chemicals are more toxic than base line toxicants (see Box 2 and Figure 4). Automated versions of the Verhaar classification system are available in the OECD QSAR Toolbox and in Toxtree (Enoch et al. 2008). Other classification systems apply more categories (Barron et al. 2015; Busch et al. 2016). More information about mechanisms and modes of action is given in section 4.2 ‘Toxicodynamics & Molecular Interactions’.
Expert systems can assign a MOA class to a chemical and predict toxicity of large data sets. Specific QSAR models may be available for a certain MOA (Figure 2), although one should realize that validated QSARs are available only for a limited number of MOAs (see also under ECOSAR). The Rule-based Expert systems are based on chemical-structure rules (using e.g. the presence of specific chemical substructures in a molecule) such as identified in Box 2 for a number of classes of chemicals and MOA.
Figure 2. The approach to select QSARs for predicting toxicity. The QSARs are MOA specific.
The Verhaar classification scheme is developed based on acute fish toxicity data. A major class of chemicals are compounds with a non-specific mode of action, also called narcosis type chemicals or baseline toxicity. This class 1 in this classification scheme includes aromatic and aliphatic (chloro)hydrocarbons, alcohols, ethers and ketones. In ecotoxicology, baseline (or narcosis-level) toxicity denotes the minimal effects caused by unspecific non-covalent interactions of xenobiotics with membrane components, i.e. membrane perturbations (Nendza et al. 2017). This MOA is non-specific and each organic chemical has this MOA as a base-line or minimum effect (see section 4.2). The effect (mortality or a sublethal effect) will occur at a constant concentration in the cell membrane and the internal lethal concentration (ILC) is around 50 mmol/kg lipid and is independent of the octanol-water partition coefficient (KOW). Box 2 gives an overview of the Verhaar classification scheme and also includes chemical structures within each class and short descriptions of the mode of action.
Box 2: Examples of chemicals in each of the classes (Verhaar class 1 to class 4) | |
---|---|
Class 1 chemicals: inert chemicals
MOA: non-polar narcosis Non-specific mechanism. Effect is related to presence of a chemical in cell membranes. Effect will occur at a constant concentration in a cell membrane. |
Class 2 chemicals: less inert chemicals
MOA: polar narcosis Similar to class 1, with hydrogen bonding in addition to thermodynamic partitioning.
|
Class 3 chemicals: reactive chemicals (electrophiles)
MOA: related to reactivity Electrophiles may react with a nucleophile. Nucleophilic groups are for example NH2, OH, SH groups and are present in amino acids (and proteins) and DNA bases. Exposure to these chemicals may lead for example to mutagenicity or carcinogenicity (DNA damage), protein damage or skin irritation.
|
Class 4 chemicals: specific acting chemicals
MOA: specific mechanism Several chemicals have a specific MOA. Insecticides such as lindane and DDT specifically interact with the nervous system. Organophosphates are neurotoxicants that interact with the enzyme Acetylcholine-esterase.
|
LC50 data of class 1 chemicals show a strong inverse relationship with the hydrophobicity (Kow). This decrease of LC50 with KOW is logical because the LC50 is inversely related to the bioconcentration (BCF) and BCF increases with KOW (see equation 2 and Figure 3).
(2)
Figure 3. Relation between (i) concentration with 50 % mortality (log LC50), (ii) bioconcentration factor (log BCF) and (iii) internal lethal concentration (ILC) and the octanol-water partition coefficient (log KOW). Also see Figure 2 in section 4.1.7.
Figure 4A shows the relationship between guppy log LC50 and log KOW for 50 compounds that act via narcosis (class 1 chemicals). The line in Figure 4A represents the so-called minimum or base-line toxicity. Figure 4B additionally shows LC50 data for the other classes (class 2, 3 and 4). LC50s of class 2 compounds (polar narcosis) are significantly lower on the log KOW scale. The distinction into non-polar and polar narcosis was introduced by Schultz and Veith (Schultz et al. 1986). The LC50 values of reactive and specifically acting chemicals (Classes 3 and 4, respectively) are mostly below base-line toxicity (see Figure 4B).
Figure 4. Correlation between log LC50 data and the octanol-water partition coefficients (log KOW) for class 1 (Figure 4A, top) and classes 2, 3 and 4 chemicals (Figure 4B, bottom). Data are from Verhaar et al. (1992).
Several QSARs are published for class 1 chemicals for different species including fish, crustaceans and algae and effects on survival (LC50), growth (EC50) or no-observed effect concentrations (NOEC). Some examples are presented in Table 1. The equations have the following format:
log 14-d LC50 (mol/L) = -0.869 log Kow - 1.19, n=50, r2=0.969, Q2=0.957, s.e.=0.31 (3)
The intercept in the equations gives information about the sensitivity of the test. The intercept in equation 7 (-2.30) is 1.11 lower than the intercept of equation 5 (1.19). The difference of 1.11 is on a logarithmic scale and the slopes of the equations are similar (-0.869 versus -0.898). This means that the test on sublethal effects (NOEC) is a factor 13 (101.11) more sensitive than the LC50 test and this is in agreement that a factor of 10 is the standard assessment factor for extrapolating from LC50 to NOEC.
These QSAR equations for class 1 are relatively simple and include the octanol-water partition coefficient KOW) as the only parameter. QSAR models for reactive chemicals and specific acting compounds are far more complex because the intrinsic toxicity (reactivity and potency to interact with the target) and also biotransformation to active metabolites will affect the toxicity and effect concentration.
An example of a QSAR for reactive chemicals is presented in Box 3. This example also shows how a QSAR is derived.
The ‘excess toxicity’ value (Te), also called toxic ratio (TR), presents an easy way to interpret tox data. Excess toxicity (Te) is calculated as the ratio of the estimated LC50 value for base-line toxicity (using the Kow regression) and the experimental LC50 value (equation 4).
(4)
Table 1. QSARs for class 1 chemicals.
Species |
Endpoint |
QSAR |
Eqn. # |
FISH |
|
|
|
Poecilia reticulata |
log 14-d LC50 |
-0.869 log Kow - 1.19 |
5 |
|
(mol/L) |
n=50 r2=0.969 Q2=0.957 s.e.=0.31 |
|
Pimephales promelas |
log 96-h LC50 |
-0.846 log Kow - 1.39 |
6 |
|
(mol/L) |
n=58 r2=0.937 Q2=0.932 s.e.=0.36 |
|
Branchidanio rerio |
log 28-d NOEC |
-0.898 log Kow - 2.30 |
7 |
|
(mol/L) |
n=27 r2=0.917 Q2=0.906 s.e.=0.33 |
|
CRUSTACEANS |
|
|
|
Daphnia magna |
log 48-h LC50 |
-0.941 log Kow - 1.32 |
8 |
|
(mol/L) |
n=49 r2=0.948 Q2=0.944 s.e.=0.34 |
|
Daphnia magna |
log 16-d NOEC |
-1.047 log Kow - 1.85 |
9 |
|
(mol/L) |
n=10 r2=0.968 Q2=0.954 s.e.=0.39 |
|
ALGAE |
|
|
|
Chlorella vulgaris |
log 3-h EC50 |
-0.954 log Kow - 0.34 |
10 |
|
(mol/L) |
n=34 r2=0.916 Q2=0.905 s.e.=0.32 |
|
n is the number of compounds, r2 is the correlation coefficient, Q2 is the cross-validated r2 and s.e. is the standard error of estimate
LC50: concentrations with 50 % effect on survival
NOEC: no-observed effect concentrations for sublethal effects (growth, reproduction)
EC50: concentrations with 50 % effect on growth
The equations are taken from EC_project (1995)
Box 3: Example of a QSAR
Data set: acute toxicity (LC50) of reactive chemicals
Chemicals: 15 reactive chemicals including a, b-unsaturated carboxylates Y: log LC50 to Pimephales promelas (in mol/L) X1: log kGSH reaction rate to glutathione (in (mol/L)-1 min-1) X2: log KOW octanol-water partition coefficient
Te: excess toxicity in comparison with calculated base-line toxicity (calculated with equation 4). Log LC50 base-line (mmol/L) = 0.846 log KOW – 1.39 (see equation 6 in Table 1)
Dataset
QSAR based on two parameters: log 96-h LC50 = -0.67 ± 0.09 log kGSH - 0.31 ± 0.11 log KOW - 3.33 ± 0.21 r2 = 0.82, s.e. = 0.47
The relatively low standard deviation in the regression coefficients show that both parameters are significant. The LC50 decreases with increasing KOW – related to effect of hydrophobicity on accumulation The LC50 decreases with increasing reactivity – more reactive chemicals are more toxic. |
The discussed examples are QSARs with one or only a few X variables. Other QSPR approaches use large numbers of parameters derived from chemical graphs. The CODESSA software for example, generates molecular (494) and fragment (944) descriptors, classified as (i) constitutional, (ii) topological, (iii) geometrical, (iv) charge related, and (v) quantum chemical (Katritzky et al. 2009). Some models are based on structural fragments in a molecule. Fish toxicity data were analysed with this approach and up to 941 descriptors were calculated for each chemical in the data sets studied (Katritzky et al. 2001). Most of the data are the same as the ones presented in Figure 3. Two to five parameter correlations were calculated for the four Verhaar classes. The correlations for class 4 toxins were less satisfactory, most likely because the QSAR included different mechanisms into one model. This approach applies a wide set (hundreds) of molecular descriptors and this may lead to overfitted models. In such a case, validation of the model is essential (Eriksson et al. 2003).
Expert systems
Several expert systems are developed that apply QSAR and other in silico methods to predict ecotoxicity profiles and fill data gaps. The following two are briefly discussed: the ECOSAR program from the US-EPA (Environmental Protection Agency) model and the QSAR toolbox from the OECD (Organisation for Economic Cooperation and Development).
ECOSAR
The Ecological Structure Activity Relationships (ECOSAR) Class Program is a computerized predictive system that estimates aquatic toxicity. The program has been developed by the US-EPA. As mentioned on their website: “The program estimates a chemical's acute (short-term) toxicity and chronic (long-term or delayed) toxicity to aquatic organisms, such as fish, aquatic invertebrates, and aquatic plants, by using computerized Structure Activity Relationships (SARs)".
Key characteristics of the program include:
ECOSAR software is available for free and is posted below as a downloadable software program without licensing requirements. Information on use and set-up is provided in the ECOSAR Operation Manual v2.0 and ECOSAR Methodology Document v2.0.
OECD QSAR Toolbox
The OECD Toolbox is a software application intended for filling data gaps in (eco)toxicity. The toolbox includes the following features:
Data gaps can be filled via classical read-across or trend analysis using data from analogues or via the application of QSAR models.
The OECD QSAR Toolbox is a very big and powerful system that requires expertise and experience to use it. The OECD QSAR Toolbox can be downloaded at https://www.oecd.org/chemicalsafety/risk-assessment/oecd-qsar-toolbox.htm. Guidance documents and training materials are also available there, as well as a link to the video tutorials on ECHA’s YouTube channel.
When using the OECD QSAR toolbox to identify suitable analogues for a Read-Across approach to estimate substance properties, it is very important to not only look at structural similarity of the chemical structures, but also to take into account any information (from experimental data, or from estimation models – the so called ‘profiles’ in the OECD QSAR Toolbox. The example in Box 4 on the importance of assigning the correct MOA underlines this.
Box 4: Chemical domain: small change in structure - large consequences for toxicity. The importance of assigning the correct MOA
To illustrate the limitations of the read-across approach, as well as underlining the importance of being able to correctly assign the ‘real’ MOA to a chemical structure we can look at two very close structural analogues:
Both substances have the same three functional groups; aromatic 6-ring, nitro-substituent and chloro-substituents, the same substitution pattern on the ring (1,2,4-positions). The only structural difference between them is the number of substituents, as one nitro-substituent is replaced with another chloro-substituent. When calculating Chemical Similarity coefficients between the two substances (often used as a start to determine the ‘best’ structural analogues for Read Across purposes) these two substances will be considered 100% similar by the majority of existing Chemical Similarity coefficients, as these often only compare the presence/absence of functional groups, and not the number.
Looking at the chemical structures and the examples given for the Verhaar classification scheme (Box 2) one could easily come to the conclusion that these two substances both belong to the Class 2: less inert, or polar narcosis type chemicals.
Applying the class 2 polar narcosis QSAR for Pimephales promelas 96 hr- LC50, as reported in EC_project (1995):
LC50 (mol/L) log LC50 = -0.73 log KOW – 2.16 n = 86, r2 =0.90, Q2 = 0.90, s.e. = 0.33
yield estimates of the LC50 of 36.5 mg/L for the dinitro-compound and 8 mg/L for the dichloro-compound (see Table below). When looking at experimentally determined acute (96hr) fish toxicity data for these two compounds, the estimate for the dichloro-compound is quite close to reality (96hr LC50 for Oryzias latipes of 4.7 mg/L), even though we do not have data for the exact same species Pimephales promelas). But the estimate for the dinitro-compound is largely underestimating the toxicity as the experimental 96hr LC50 for Oryzias latipes is as low as 0.16 mg/L, a factor of 230 times lower than estimated by the polar-narcosis type QSAR.
The explanation is in the MOA assignment, as 1,2-dichloro-4-nitrobenzene has indeed a polar narcosis type MOA, but the 1-chloro-2,4-dinitrobenzene is actually an alkylating substance (unspecific reactive, Class 3 MOA) as the electronic interactions of the 2,4-dinitro substitution make the 1-chloro substituent highly reactive towards nucleophilic groups (e.g. DNA, or proteins). This reactivity leads to an increased toxicity.
It should be noted that software implementations of MOA classification schemes, like the Verhaar classification scheme in the ToxTree software, or as implemented in the OECD QSAR Toolbox, do identify both nitro-benzene substances as Class 3, unspecified reactive MOA. The OASIS MOA classification scheme, and also the ECOSAR classification scheme do distinguish between mono-nitrobenzenes as inert and di-nitrobenzenes as (potentially) reactive substances and (correctly) assign different MOA to these two substances. ECOSAR subsequently has a separate polynitrobenzene grouping, with its own log KOW based linear regression QSAR for fish toxicity. In the summary below the ECOSAR estimates for 96hr LC50 for fish in general are also given for comparison. The polynitrobenzene model still underestimates the toxicity of the alkylating agent 1-chloro-2,4-dinitrobenzene by a factor of 25.
|
References
Barron, M.G., Lilavois, C.R., Martin, T.M. (2015). MOAtox: A comprehensive mode of action and acute aquatic toxicity database for predictive model development. Aquatic Toxicology 161, 102-107.
Busch, W., Schmidt, S., Kuhne, R., Schulze, T., Krauss, M., Altenburger, R. (2016). Micropollutants in European rivers: A mode of action survey to support the development of effect-based tools for water monitoring. Environmental Toxicology and Chemistry 35, 1887-1899.
EC_project (1995). Overview of structure-activity relationships for environmental endpoints. Report prepared within the framework of the project "QSAR for Prediction of Fate and Effects of Chemicals in the Environment", an international project of the Environmental Technologies RTD Programme (DG XII/D-1) of the European Commission under contract number EV5V-CT92-0211. Research Institute of Toxicology, Utrecht University, Utrecht, The Netherlands.
ECHA (2017a). Non-animal approaches: Current status of regulatory applicability under the REACH, CLP and Biocidal Products regulations. European Chemicals Agency, Helsinki, Finland.
ECHA (2017b). Read-Across Assessment Framework (RAAF). https://echa.europa.eu/documents/10162/13628/raaf_en.pdf. European Chemicals Agency, Helsinki, Finland.
Enoch, S.J., Hewitt, M., Cronin, M.T.D., Azam, S., Madden, J.C. (2008). Classification of chemicals according to mechanism of aquatic toxicity: An evaluation of the implementation of the Verhaar scheme in Toxtree. Chemosphere 73, 243-248.
Eriksson, L., Jaworska, J., Worth, A.P., Cronin, M.T.D., McDowell, R.M., Gramatica, P. (2003). Methods for reliability and uncertainty assessment and for applicability evaluations of classification- and regression-based QSARs. Environmental Health Perspectives 111, 1361-1375.
Katritzky, A.R., Slavov, S., Radzvilovits, M., Stoyanova-Slavova, I., Karelson, M. (2009). Computational chemistry approaches for understanding how structure determines properties. Zeitschrift Fur Naturforschung Section B-a Journal of Chemical Sciences 64, 773-777.
Katritzky, A.R., Tatham, D.B., Maran, U. (2001). Theoretical descriptors for the correlation of aquatic toxicity of environmental pollutants by quantitative structure-toxicity relationships. Journal of Chemical Information and Computer Sciences 41, 1162-1176.
Nendza, M., Müller, M., Wenzel, A. (2017). Classification of baseline toxicants for QSAR predictions to replace fish acute toxicity studies. Environmental Science: Processes Impacts 19, 429-437.
OECD (2004). The report from the expert group on (quantitative) structure-activity relationships [(q)sars] on the principles for the validation of (Q)SARs, OECD series on testing and assessment, number 49. Organisation for Economic Cooperation and Development, Paris, France.
Russom,C.L., Bradbury, S.P., Broderius, S.J., Hammermeister, D.E., Drummond, R.A. (1997). Predicting modes of toxic action from chemical structure: Acute toxicity in the fathead minnow (Pimephales promelas). Environmental Toxicology and Chemistry 16, 948-967.
Schultz, T.W., Holcombe, G.W., Phipps, G.L. (1986). Relationships of quantitative structure-activity to comparative toxicity of selected phenols in the Pimephales-promelas and Tetrahymena-pyriformis test systems. Ecotoxicology and Environmental Safety 12, 146-153.
Verhaar, H.J.M., van Leeuwen, C.J., Hermens, J.L.M. (1992). Classifying environmental pollutants. 1: Structure-activity relationships for prediction of aquatic toxicity. Chemosphere 25, 471-491.
Vonk, J.A., Benigni, R., Hewitt, M., Nendza, M., Segner, H., van de Meent D., et al. (2009). The Use of Mechanisms and Modes of Toxic Action in Integrated Testing Strategies: The Report and Recommendations of a Workshop held as part of the European Union OSIRIS Integrated Project. Atla-Alternatives to Laboratory Animals 37, 557-571.