The following questions / concerns were discussed:
- Review current tools and key (statistical) methodology, including assumptions about distributions of sensitivity, use of hierarchical models, interspecies correlations. Identify where there are important differences and what the implications of these could be.
The group considered a number of tools in turn, trying to evaluate the strengths and weaknesses of each. The tools considered were Web-ICE (a web-based interspecies correlation estimation tool), hSSD (a hierarchical SSD modelling tool), ETX 2.0 (a program for calculating estimates of the HC5 and potential fraction affected and confidence intervals for the estimates, based on a log-normal SSD model), MOSAIC_SSD (a web interface which calculates an estimate of the HCx for any x and provides a bootstrap confidence interval for the estimate) and the R software developed by Procter & Gamble and presented by Scott Belanger in this workshop (section 3.5). The group were aware of the existence of some other tools (e.g. BurrliOZ software used in Australia) but there was insufficient time to evaluate them.
The hierarchical SSD model (hSSD) was considered to be an experimental tool and still under development, whereas the other tools were available for use to construct SSDs and compute HCx estimates and confidence intervals.
The tools divided naturally into 2 groups:
Web-ICE and hSSD
Features in common: Both tools make use of taxonomic structure and have the potential to address data gaps by predicting toxicity for species which have not been tested. This was felt to be generally useful but especially in the context of reduced animal testing. It was noted that Web-ICE has a specific feature for prediction of the chemical sensitivity of endangered species based on simple regression models. Both tools were built using only data from acute tests and this was felt to be a significant limitation, especially in regulatory settings requiring chronic data. Both tools made no distinction between classes of chemical and it was felt that mode of toxic action is an important consideration which should be taken into account.
Contrasting features: The Web-ICE user can select or reject individual toxicity predictions whereas the hSSD user cannot. The hSSD programme accepts censored toxicity data whereas Web-ICE does not.
ETX, MOSAIC_SSD and R-SB
Features in common: None of these 3 tools has the ability to address data gaps although they can be used to indicate the influence of adding additional data points to the HC5. All of these tools would therefore require a toxicity dataset which is regarded as complete in terms of regulatory requirements for taxonomic representation, but the dataset might be supplemented with values predicted with/extrapolated from other models. However, all of these tools could be used with either acute or chronic data.
Contrasting features: Mosaic and R-SB both accept censored data whereas ETX does not. ETX is restricted to log-normal models whereas the others can fit other parametric distributions, in particular the log-logistic family. Notably, both MOSAIC_SSD and R-SB provide a feature to compute bootstrapped approximate confidence intervals for the HCx value.
In addition to this general division into 2 groups, some specific features were identified:
Some of the members in the syndicate group felt that the capacity to reduce data gaps would lead to reduced uncertainty attached to the resulting HC5 estimate while others felt that it was not clear that it would do so, as known uncertainty caused by a small sample size is replaced by uncertainty originating from modelled/extrapolated toxicity values, which may not be completely quantifiable. Members also considered that the method by which Web-ICE computes a confidence interval for the HC5 was not transparent and clear enough; some members questioned the validity of the method.
It was felt to be ecologically biased, in particular towards fish, in terms of the pairs of species for which a sufficient dataset was available to exploit the interspecies correlation. It was noted also that it is built mainly using species primarily from North America. Since the method is simply based on a simple linear regression model using toxicity endpoint data from the well-studied standard test species and those from a group of rarely tested species for various chemicals, there are foreseeable uncertainties due to the limited data, highly variable combination of chemicals and unknown mode of toxic actions of the chemicals being included in the regression.
In terms of validation of individual predictions, it was felt by some of our members that, although quality of prediction had been assessed by a cross-validation approach, this was still essentially a validation internal to the data used rather than a validation against data from a truly external source.
The hSSD model is highly parameterised; some of our members felt that this was beneficial, but others worried that it was over-parameterised.
The hSSD requires the user to specify an ecological scenario. This was felt to have some potential benefits, especially in terms of ability to test hypotheses about effect of the chemical on the site-specific community structure.
Up to now, the hSSD has not been validated against laboratory and/or field data.
The ETX tool has regulatory acceptance in Europe since it is easy to use and understand.
At the end of the detailed discussion on the pros and cons of the various tools, there was a short discussion on their implications. It was agreed that this remains an area in which progress is on-going and it is not clear which tool is the most useful nor is it clear whether a single approach should be used. It was suggested that the way in which any tool is used may be just as important as the choice of the tool.
2. As sensitivity to chemical stress seems to be related to taxonomic closeness, how could this be used in the construction and interpretation of SSDs?
The group devoted limited time to this question.
It was agreed that knowledge of mode of toxic action should be taken into account first. Resulting predictable differences between taxonomic groups are a key driver in ecological risk assessment.
Overall it was felt that the best choice of methodology and software tool will depend on the question to be answered. Consequently, it is very important for risk managers to specify clear protection goals and questions to be answered by the risk assessment. For example, one should be clear about purpose of the HC5; is it intended to protect 95% of species or 95% of ecosystems?
The question of the need to continue with testing of taxonomically diverse species was raised. A distinction was made between prospective and retrospective assessments and between biocides and other chemicals. For prospective assessments, it was suggested that for biocides, one may often know the mode of toxic action and therefore be able to target testing appropriately (e.g. insects and arthropods should be emphasised in SSDs for insecticides; plants and algae should be the focus of SSDs for herbicides) whereas this would probably not be so for an industrial chemical without clear a priori information on its mode of toxic action. For retrospective assessments, there appeared to be a greater need for taxonomically diverse data, especially for general biocides.
3. What are the research needs?
The group’s discussion was limited to a brief list of topics:
- Further validation for extrapolations in relevant models (i.e. hSSD and Web-ICE) and of consequences for HC5 uncertainty.
- Extending software tools to add the capacity to predict chronic toxicity.
- Validation of hSSD scenario-specific HC5s relative to the field and/or mesocosm studies.
- Uses of SSDs for purposes other than estimating the HC5 (e.g. using the entire SSD for probabilistic risk assessment and deriving other values (say HC50) for trigger management action).
- The role of SSDs in risk assessment for mixtures is still at an exploratory stage. It is likely to be dependent on the questions to be addressed and the rationale for using SSDs.
- Possible inclusion of microorganisms in SSDs to protect ecosystem functions was debated. For instance, when assessing the ecological risk of fungicides, we never consider including various fungal species in the test battery and incorporating their data into the SSD; if our management goal is to protect the ecosystem functions and services, we should try to protect the fungi as well. As such, microorganisms should be considered in the HCx derivation. Nonetheless, such a development is currently hindered by the lack of available approved testing procedures for different groups of microorganisms.