The Way Forward: (Q)SAR
Published: March 31, 2008
Early on in her career she joined Unilever’s Safety and Environmental Assurance Centre in the UK as a safety evaluation scientist before developing a keen interest in computational toxicology and moving on to a role which involved providing modeling and chemistry expertise to both internal and external projects. Ultimately she led the QSAR group.
She then joined the (Q)SAR group at the European Chemicals Bureau within the Joint Research Centre, Italy to gain a better appreciation of the regulatory application of (Q)SARs. There, she was involved in many of the activities related to the development of technical guidance for REACH, investigating the feasibility of using computational approaches in the development of chemical categories, developing and evaluating (Q)SAR models for human health as well as coordinating the technical development of software tools such as Toxtree (a tool which encodes the Cramer classification scheme amongst others) and Toxmatch (a tool for chemical similarity).
In February 2008, she made a return to Industry taking up a position within DuPont in the US as a computational modeller.
Dr. Grace Patlewicz (Tier)
DuPont Haskell Global Centers for Health and Environmental Sciences
PO Box 50
1090 Elkton Rd
Newark, DE 19714-0050
Predictive toxicology approaches include SARs, QSARs and expert systems which shall be collectively referred to as (Q)SAR models in this document. (Q)SAR models provide an efficient means of encoding and summarising existing knowledge for effective re-use. They can be used to screen and prioritise chemicals on the basis of toxicity, provide mechanistic insights to facilitate “safety by design” as well as help in the development of in vitro alternatives.
The basic underlying principle of a (Q)SAR model is that the properties of a chemical with respect to how it will interact with a defined system are inherent in its molecular structure. i.e. toxicity =ƒ (chemical). Hence attempts to develop (Q)SARs consist of looking for links between structure and biological activity.
QSARs may be categorised into 3 main types; statistical (empirically derived) QSARs which purport to make predictions for a wide range of chemicals, mechanistic QSARs which are developed on the basis of a common mechanism of action, and local QSARs which are typically restricted to structurally related homologues. Expert systems are designed to cover a wide range of chemicals.
One can consider QSARs to be collections of “mature” or “existing” knowledge that are packaged in a convenient form for subsequent and routine re-use. Hence instead of evaluating a chemical based on interpolating data from either one chemical or more (read-across), the QSAR approach means that information from many observations is effectively summarised in an algorithm which should enable a more robust evaluation to be made. The robustness of any prediction will obviously be greater if the underlying features connecting the chemicals together are mechanistically based.
Depending on the type of model, the QSAR can also provide useful insights to facilitate the direction of new research for in vitro methods. This depends on the endpoint under consideration and the degree to which its mechanism is already understood. For specific endpoints such as skin sensitisation, there has been considerable success in directing new research on alternatives based on the understanding derived from exploring relationships between chemicals and the sensitisation in vivo test result. Examples include works in references [1-3]. For more complex endpoints, such as carcinogenicity and developmental toxicity, success has been more limited (see later). At the very least, the principles underpinning QSAR approaches ensure that the selection of chemicals used to develop and validate a new in vitro method covers the widest possible chemistry space.
The Application of (Q)SARs
A common misconception surrounding (Q)SAR models is that they are only intended as non-animal replacements (a sort of “third class” alternative ) and as a result should be treated in much the same way as an in vitro test method. However (Q)SAR models are not test methods but sources of existing information, albeit packaged in summary form. Aside from providing mechanistic and scientific insights, their key value is in contributing to the information needs of a regulatory or safety assessment. Such assessments can be thought of as the outputs from so-called tiered testing strategies (Integrated Testing Strategy, ITS) where the (Q)SAR model(s) is just one of the inputs in much the same way as results from in vitro and in vivo studies are.
The prospects of using (Q)SAR models is very attractive in the current legislative environment (e.g. 7th Amendment to the Cosmetics Directive; Registration Evaluation, Authorisation of Chemicals REACH), given that they are relatively cheap and fast to apply. The challenge is to understand when a (Q)SAR model may sensibly be applied and the result relied upon. Instead of considering QSAR models as test methods, the issue is how a prediction result can be used in a risk/safety assessment as part of an ITS. This does require a shift in thinking; instead of a checkbox approach to risk assessment where a large amount of information is collected and evaluated, the risk assessment problem is first defined to identify what information is necessary and what confidence is required in the decision outcome. From there, the choice of (Q)SAR model is driven by whether it is fit for intended purpose. e.g. a (Q)SAR model may be entirely appropriate for screening but insufficiently robust to be relied upon in a risk management decision.
Biological relevance to the endpoint of interest
Here it is only possible to make some general remarks about biological relevance as this is obviously dependent on the toxicity of interest, the understanding of it and the types of (Q)SAR models that are available. For an endpoint as seemingly simple as skin irritation, there are still uncertainties about the mechanism of action and a distinct lack of models for their prediction. For other endpoints such as aquatic toxicity, there have been successes in developing QSARs based on modes of action. For more complex endpoints such as skin sensitisation, there is a clear hypothesis that has stood the test of time and which appears to be sufficiently robust to facilitate the development of mechanistic models which can predict both the potential and potency of skin sensitisation. The key issue is that in this case, there is sufficient understanding about the molecular initiating events governing skin sensitisation induction and moreover which events are the rate determining ones. This type of understanding paves the way for mechanistic models which are in turn of higher biological relevance.
Significantly more complex endpoints such as carcinogenicity and developmental toxicity, which are still poorly understood, can lead to models that have lesser biological relevance. However there have been some successes in relating electrophilic parameters and partitioning effects to small subsets of chemicals tested in a carcinogenicity assay [4-6].
We should also consider the way in which the QSAR community has evolved and the motivations that has led to some of these different types of models. To appreciate this more fully, a step back into history is required. In the very early days, QSARs were developed at a very local scale looking at homologous series and understanding or hypothesising the structural determining factors that would account for the trend in activity. Many of the early models by Hansch [7-8] have in fact used one or two parameters to relate small sets of chemicals to a range of different endpoints. With the explosion of computational approaches, notably from the pharmaceutical world, an impetus to develop more sophisticated models arose which moved away from the biological context but focused instead on models derived from using novel statistical approaches or selecting from large pools of descriptors. At least in the area of skin sensitisation, in recent years, there seems to have been a drive to develop models that apparently covered a wide range of chemicals but at the expense of limited mechanistic interpretability. Examples in this area include references [9-11].
Strengths and limitations of these approaches compared to current in vivo test methods
The strengths of (Q)SARs is that they can be routinely applied and do not require extensive resources in terms of experimental testing. Since the (Q)SARs themselves are summaries of the data taken from current methods, they will clearly suffer some of the same limitations as the test methods from which they were derived. As an example one could consider the LLNA skin sensitisation method from which the EC31 is derived. The EC3 in the LLNA can be effectively doubled or halved due to biological variability. Obviously a QSAR model predicting EC3 will also have this biological variability (though at least this is averaged for all chemicals in the training set) but in addition it will have some error from the modeling algorithm used as well as from the descriptors used, i.e. there will be some error in the descriptors as they are approximations and some error in the statistical algorithm used to relate the descriptors to the endpoint outcome. These types of errors are less pronounced if more mechanistic QSAR models are developed which rely on transparent descriptors and simpler modeling approaches.
Work still needed to be done for the development, validation and use of (Q)SARs
Firstly, a comprehensive catalogue of existing and available (Q)SARs is required for each of the different endpoints. Then these models need to be characterised to understand their scope and relevance for the context to which they will be applied. Characterising (Q)SARs with reference to the OECD Principles of (Q)SAR Validation is a convenient means of evaluating a (Q)SAR . These principles are specifically: 1) a defined endpoint 2) an unambiguous algorithm 3) a defined domain of applicability 4) appropriate measures of goodness-of-fit, robustness and predictivity 5) a mechanistic interpretation, if possible. Examples where this has already been carried out include [13-15].
A formal validation process is not thought to be a meaningful nor a useful concept for (Q)SARs since they are not test methods, but evaluating model validity with reference to the OECD Principles is thought to be helpful. Validation of a QSAR model to test its predictivity is captured in the OECD principles under Principle 4 – appropriate measures of goodness of fit, robustness and predictivity.
The characterisation of the (Q)SAR model is just one aspect to consider when using a (Q)SAR result; determining whether the model is appropriate (i.e. characterising the applicability domain) for a chemical of interest is just as if not more important. The characterisation of the domain is by no means trivial. Whilst research in systematic approaches has been on-going, more work in this area still needs to take place [16-18]. In addition, a consideration of the chemistry is also important; evaluating the predictions of close analogues with reference to their experimental data provides useful confirmation as to the reliability of the model prediction .
Another aspect is in characterising the overall chemical space of interest, e.g. a regulatory inventory such as EINECS for REACH and the extent to which existing (Q)SARs map onto this chemical space. This is a massive undertaking that needs to occur in order to determine what is known but also to identify the data gaps and hence research needs (i.e. such as what new experimental testing might need to be performed).
Challenges and barriers
One major challenge in the development of new (Q)SARs and the evaluation of existing ones lies with the lack of toxicity data and, to an extent, the lack of this data being structured in a form that makes it readily usable for modeling purposes. Formats–protocols of test guidelines–are not always adequately described and terms of reference may differ from one method to another. Harmonising how toxicity data is reported, to facilitate comparison between available data sources and databases, remains a critical need. Another problem is that often such databases, if they exist, are rarely structurally annotated i.e. names and toxicity data are reported but not the chemical structural identifiers. Efforts have been initiated to address these issues but more is needed. Notable successes include the work of the Leadscope LIST consortium  and the DSStox effort at the US-EPA .
Another factor is the fragmented way in which (Q)SAR models often are developed, i.e., in a vacuum without consideration of their potential and practical use. There is no real strategic framework for the development of models. Hence the same experimental data frequently is re-used, resulting in limited novel insights and at worse duplication of effort. It is also unclear what synergies could be realized if (Q)SAR model development cut across different industry sectors – there is still a huge disconnect between the (Q)SAR developers within chemical industry and those within the pharmaceutical sector. Some expert systems such as Derek for Windows, Pipeline Pilot or Leadscope have user groups which span across industry sectors which itself has led to substantive improvements in software due to the shared expertise. More collaborative efforts like this are needed.
Another issue is that whilst we have a wealth of different descriptors for different chemical features that can be readily calculated, there still is no good parameter for reactivity that effectively encodes electrophilicity information and does so in a form that facilitates routine use. Global parameters which serve as crude surrogates for electrophilicity do exist such as LUMO energy but they are of limited practical value. In the area of sensitisation, a robust parameter for reactivity would facilitate the development of new mechanistic QSAR models (so-called QMMs) [19, 22]. At the present time, experimental rate constants are being derived to help in the development of new QSAR models for skin sensitisation as well as other endpoints where reactivity is a key factor, but resources to realise a comprehensive database of reactivity measurements is still very much required .
It is worth noting too that whilst there are likely to be gaps regarding the coverage of the available (Q)SARs on the chemical domain of interest, the availability of experimental in vivo toxicity data to fill those gaps is probably close to being exhausted notwithstanding that data which is subject to CBI (confidential business information) and hence inaccessible. Surrogate information sources such as endpoint to endpoint correlations should be a focus to identify possible links between endpoints and as means of exploiting available information as far as possible.
The future role of this technology
The immediate needs are clear – the cataloguing and characterisation of available models, Information Technology (IT) implementation, raising awareness and building capability with relative stakeholders – are all critical requirements to ensure the successful uptake of (Q)SARs for at least the foreseeable future. In the next 10 and 15 years it becomes more difficult to speculate. Suffice to say there are plenty of challenges to overcome within the next 5 years before judging whether the application of (Q)SARs is perceived to be a worthwhile strategy or not.
1 The EC3 is the effective concentration for stimulation of a 3-fold increase in lymph node cell proliferation
©2007 Grace Patlewicz
- Roberts, D.W., Aptula, A.O. & Patlewicz, G. (2007). Electrophilic Chemistry Related to Skin Sensitization. Reaction Mechanistic Applicability Domain Classification for a Published Data Set of 106 Chemicals Tested in the Mouse Local Lymph Node. Chem. Res. Toxicol.20(1), 44-60.
- Dimitrov, S., Low, L., Patlewicz, G., Kern, P.S., Dimitrova, G.D., Comber, M.H., Phillips, R.D., Niemela, J., Bailey, P.T. & Mekenyan, O.G. (2005) Skin sensitization: Modelling based on skin metabolism simulation. Int. J. Toxicol.24(4), 189-204.
- Roberts, D.W., Aptula, A.O. & Patlewicz, G. (2006). Mechanistic Applicability Domains for Non-Animal Based Prediction of Toxicological Endpoints. QSAR Analysis of the Schiff Base Applicability Domain for Skin Sensitization. Chem. Res. Toxicol.19(9), 1228-1233.
- Franke, R., Gruska, A., Giuliani, A. & Benigni. R. (2001) Prediction of rodent carcinogenicity of aromatic amines: a quantitative structure-activity relationships model. Carcinogenesis.22(9), 1561-1571.
- Benigni, R., Passerini, L. (2002) Carcinogenicity of the aromatic amines: from structure-activity relationships to mechanisms of action and risk assessment. Mutat. Res.511(3), 191-206.
- Benigni, R., Netzeva, T.I., Benfenati, E., Bossa, C., Franke, R., Helma, C., Hulzebos, E., Marchant, C., Richard, A., Woo, Y.T., Yang, C. (2007). The Expanding Role of Predictive Toxicology: An Update on the (Q)SAR Models for Mutagens and Carcinogens. J. Environ. Sci. Health C Environ. Carcinogen. Ecotoxicol. Rev.25(1), 53-97.
- Hansch, C., Maloney, P.P., Fujita, T. & Muir, R.M. (1962) Correlation of Biological Activity of Phenoxyacetic Acids with Hammett Substituent Constants and Partition Coefficients. Nature.194, 178-180.
- Hansch, C., Muir, R.M., Fujita, T., Maloney, P.P., Geiger, F., Streich, M. (1963) The Correlation of Biological Activity of Plant Growth Regulators and Chloromycetin Derivatives with Hammett Constants and Partition Coefficients. J. Am. Chem. Soc.85, 2817-2824.
- Fedorowicz, A., Zheng, L., Singh, H. & Demchuk, E. (2004) Structure-Activity Models for Contact Sensitisation. Int. J. Mol. Sci.5, 56-66.
- Li, Y., Tseng, Y.J., Pan, D., Liu, J., Kern, P.S., Gerberick, G.F. & Hopfinger, A.J. (2007). 4D-Fingerprint Categorical QSAR Models for skin sensitisation based on the classification of Local Lymph Node assay measures. Chem. Res. Toxicol.20(1), 114-128.
- Miller, M.D., Yourtee, D.M., Glaros, A.G., Chappelow, C.C., Eick, J.D. & Holder, A.J. (2005) Quantum Mechanical Structure-Activity Relationship Analyses for Skin Sensitisation. J. Chem. Inf. Modeling.45, 924-929.
- European Chemicals Bureau
- Liu, H., Papa, E., Gramatica, P. (2006). QSAR prediction of estrogen activity for a large set of diverse chemicals under the guidance of OECD principles. Chem. Res. Toxicol.19(11), 1540-1548.
- Pavan, M., Netzeva, T.I. & Worth, A.P. (2006). Validation of a QSAR model for acute toxicity. SAR QSAR Environ. Res., 17(2): 147-171.
- Saliner, A.G., Netzeva, T.I., Worth, A.P. (2006). Prediction of estrogenicity: validation of a classification model. SAR QSAR Environ. Res.17(2), 195-223.
- Netzeva, T.I., Worth, A.P. Aldenberg, T., Benigni, R., Cronin, M.T.D., Gramatica, P., Jaworska, J.S., Kahn, S., Klopman, G., Marchant, C.A., Myatt, G., Nikolova-Jeliazkova, N., Patlewicz, G.Y., Perkins, R., Roberts, D.W., Schultz, T.W., Stanton, D.T., van de Sandt, J.J.M., Tong, W., Veith, G. & Yang, C. (2005) Current Status of Methods for Defining the Applicability Domain of (Quantitative) Structure-Activity Relationships. The Report and Recommendations of ECVAM Workshop 52. Altern. Lab. Anim.33, 155-173.
- Nikolova-Jeliazkova, N. & Jaworska, J. (2005) An approach to determining applicability domains for QSAR group contribution models: an analysis of SRC KOWWIN. Altern. Lab. Anim.33(5), 461-470.
- Dimitrov, S., Dimitrova, G., Pavlov, T., Dimitrova, N., Patlewicz, G., Niemela, J. & Mekenyan, O. (2005) A stepwise approach for defining the applicability domain of SAR and QSAR models. J. Chem. Inf. Modeling45(4), 839-849.
<li>Roberts, D.W., Aptula, A.O., Patlewicz, G. & Pease, C. (2007). Chemical Reactivity Indices and Mechanism-based read across for non-animal based assessment of skin sensitization potential. J. Appl. Toxicol.in press.
- Leadscope, Inc.
- Distributed Structure-Searchable Toxicity (DSSTox) Database Network
- Patlewicz, G., Aptula, A.O., Roberts, D.W. & Uriarte, E. (2007). A mini review of available (Q)SARs and Expert systems for skin sensitisation. QSAR & Comb. Sci. in press.