48
R. Williams et al.
6. Generality of any toxicophore identified
7. Limitations of the software being used.
If a model is being generated based on data coming from an in vitro or in vivo
assay linked to an ultimate toxicity endpoint, the limitations of the assay should
be considered either as part of the model building process or, more likely, during
expert review [32, 34, 35]. For example, there are a minority of compound classes
for which the (in vitro) Ames mutagenicity assay may not reflect the true DNA
reactivity and consequently the hazard may be over- or underestimated. Models
produced from these data will not, therefore, reflect the true (in vivo) carcinogenicity
hazard caused by the compounds. Amberg et al. have highlighted the fact that the
Ames test may over- or underestimate the mutagenicity of compounds containing
the acid halide group depending on the solvent chosen to carry out the test [32, 36].
Water would hydrolyse the acid halide deactivating it to a carboxylic acid, whereas
dimethyl sulphoxide (DMSO) would react directly with the acid halide producing
halodimethyl sulphides which then act as the DNA-reactive species. To get accurate
results for this compound class, the compounds should be tested in an inert solvent
such as acetonitrile.
The in silico prediction should be transparent enough that the user can interrogate
the data on which the prediction is based. It is important that the data supporting
each prediction can be probed to the lowest captured level, to allow for adequate
assessment of the weighting of this evidence. Any inadequacies in the testing protocol
of the data being used to make the prediction should be noted during expert review
[33]. This is particularly important for data points which are pivotal for the prediction
being made, and as a consequence the in silico prediction system should identify those
key compounds used to make the prediction and their relative weighting.
Many in silico prediction systems account for the ability of the model to make
accurate predictions based on the similarity of the predicted structure to the training
set chemicals, by employing an applicability domain and/or confidence metrics. In
some cases, these machine-generated boundaries may not be as good as a human
in assessing the relevance of the data used, and the adequacy of the extrapolation
should always be assessed by the expert as part of the review process [33]. Again,
this requires that the data supporting the derivation of the model and the training set
compounds should be provided, with the prediction, for assessment.
Where a potential toxicophore has been identified and a negative prediction has
been made, it is important to determine whether the toxicophore has been assessed
adequately by the model and that the factors negating it are appropriate [30]. If the
toxicophore has been considered and the negating factors are acceptable, then the
prediction can be upheld. However, if the toxicophore was not considered or the
negating factors are not relevant, then the prediction may have to be overturned. As
a consequence, the in silico prediction should provide some information relating to
the chemical space covered and when a potential toxicophore may not be adequately
covered (e.g. the misclassified and unclassified features in Derek Nexus mutagenicity
and skin sensitisation predictions, and the overturned hypotheses in Sarah Nexus
[37]). Where structural alerts are being used to make a prediction, it is also useful to
R. Williams et al.
6. Generality of any toxicophore identified
7. Limitations of the software being used.
If a model is being generated based on data coming from an in vitro or in vivo
assay linked to an ultimate toxicity endpoint, the limitations of the assay should
be considered either as part of the model building process or, more likely, during
expert review [32, 34, 35]. For example, there are a minority of compound classes
for which the (in vitro) Ames mutagenicity assay may not reflect the true DNA
reactivity and consequently the hazard may be over- or underestimated. Models
produced from these data will not, therefore, reflect the true (in vivo) carcinogenicity
hazard caused by the compounds. Amberg et al. have highlighted the fact that the
Ames test may over- or underestimate the mutagenicity of compounds containing
the acid halide group depending on the solvent chosen to carry out the test [32, 36].
Water would hydrolyse the acid halide deactivating it to a carboxylic acid, whereas
dimethyl sulphoxide (DMSO) would react directly with the acid halide producing
halodimethyl sulphides which then act as the DNA-reactive species. To get accurate
results for this compound class, the compounds should be tested in an inert solvent
such as acetonitrile.
The in silico prediction should be transparent enough that the user can interrogate
the data on which the prediction is based. It is important that the data supporting
each prediction can be probed to the lowest captured level, to allow for adequate
assessment of the weighting of this evidence. Any inadequacies in the testing protocol
of the data being used to make the prediction should be noted during expert review
[33]. This is particularly important for data points which are pivotal for the prediction
being made, and as a consequence the in silico prediction system should identify those
key compounds used to make the prediction and their relative weighting.
Many in silico prediction systems account for the ability of the model to make
accurate predictions based on the similarity of the predicted structure to the training
set chemicals, by employing an applicability domain and/or confidence metrics. In
some cases, these machine-generated boundaries may not be as good as a human
in assessing the relevance of the data used, and the adequacy of the extrapolation
should always be assessed by the expert as part of the review process [33]. Again,
this requires that the data supporting the derivation of the model and the training set
compounds should be provided, with the prediction, for assessment.
Where a potential toxicophore has been identified and a negative prediction has
been made, it is important to determine whether the toxicophore has been assessed
adequately by the model and that the factors negating it are appropriate [30]. If the
toxicophore has been considered and the negating factors are acceptable, then the
prediction can be upheld. However, if the toxicophore was not considered or the
negating factors are not relevant, then the prediction may have to be overturned. As
a consequence, the in silico prediction should provide some information relating to
the chemical space covered and when a potential toxicophore may not be adequately
covered (e.g. the misclassified and unclassified features in Derek Nexus mutagenicity
and skin sensitisation predictions, and the overturned hypotheses in Sarah Nexus
[37]). Where structural alerts are being used to make a prediction, it is also useful to