3 Modelling Simple Toxicity Endpoints …
49
provide information and a rationale behind any exclusions along with the associated
alert.
There are some common limitations with statistical approaches which should be
considered during expert review where such a system is being used. These predictions
usually deal with association and do not assess causation. As a result, any positive
prediction made by the system should be checked with a view towards assessing the
likelihood that the feature identified in the query compound is the one likely to be
causing the positive result in the compounds used to make the prediction and that
activity of these compounds cannot be attributed to other features not present in the
predicted structure [24, 30, 33]. The in silico prediction should, therefore, provide
information about the toxicophore identified in the query structure as well as the
training set examples which also possess this feature in order to allow for analysis
of whether the feature attributed to toxicity is likely to be causative.
Many in silico prediction systems employ SARs based on structural alerts. It is
important to note that these alerts can sometimes be very general and, as a result,
specific substitution patterns representing steric or electronic factors which may
negate the hazard caused by the toxicophore may be missed [24]. For example,
structural alerts for mutagenicity based solely on broad toxicophores, such as those
proposed by Ashby and Tennant [1], will fail to take into consideration mitigating
factors that mean many compounds belonging to these structural classes do not show
mutagenic activity [38]. For these types of alerts, expert review should be carried out
to assess steric or electronic factors which may influence the activity of the training
set compounds, and the query compound. The scope of any alert activated should
be provided to allow the user to assess how broad the coverage of the alert is, along
with a rationale for the scope of the alert.
Some consideration should also be given to limitations of the software being used
to make the predictions. For example, in order that an accurate prediction be made for
a query compound, it is important that the structure is represented in the same way
as those used to build the model. Standardisation of the structures should, therefore,
be carried out in the same way in both cases to produce accurate predictions. If this
standardisation is not carried out automatically by the system, then it must be done
manually before making the prediction.
3.6 Modelling Complex Endpoints
The prediction of outcomes for complex endpoints, following oral exposure to
a chemical, presents new and different challenges. Additionally, the influence of
ADME factors requires a deeper consideration, which can be illustrated by several
questions, e.g.
1. Is the chemical stable in the gastrointestinal tract?
2. Will the chemical be absorbed?
3. Will first-pass metabolism detoxify the chemical?
4. Will the chemical be metabolised to a reactive species?
49
provide information and a rationale behind any exclusions along with the associated
alert.
There are some common limitations with statistical approaches which should be
considered during expert review where such a system is being used. These predictions
usually deal with association and do not assess causation. As a result, any positive
prediction made by the system should be checked with a view towards assessing the
likelihood that the feature identified in the query compound is the one likely to be
causing the positive result in the compounds used to make the prediction and that
activity of these compounds cannot be attributed to other features not present in the
predicted structure [24, 30, 33]. The in silico prediction should, therefore, provide
information about the toxicophore identified in the query structure as well as the
training set examples which also possess this feature in order to allow for analysis
of whether the feature attributed to toxicity is likely to be causative.
Many in silico prediction systems employ SARs based on structural alerts. It is
important to note that these alerts can sometimes be very general and, as a result,
specific substitution patterns representing steric or electronic factors which may
negate the hazard caused by the toxicophore may be missed [24]. For example,
structural alerts for mutagenicity based solely on broad toxicophores, such as those
proposed by Ashby and Tennant [1], will fail to take into consideration mitigating
factors that mean many compounds belonging to these structural classes do not show
mutagenic activity [38]. For these types of alerts, expert review should be carried out
to assess steric or electronic factors which may influence the activity of the training
set compounds, and the query compound. The scope of any alert activated should
be provided to allow the user to assess how broad the coverage of the alert is, along
with a rationale for the scope of the alert.
Some consideration should also be given to limitations of the software being used
to make the predictions. For example, in order that an accurate prediction be made for
a query compound, it is important that the structure is represented in the same way
as those used to build the model. Standardisation of the structures should, therefore,
be carried out in the same way in both cases to produce accurate predictions. If this
standardisation is not carried out automatically by the system, then it must be done
manually before making the prediction.
3.6 Modelling Complex Endpoints
The prediction of outcomes for complex endpoints, following oral exposure to
a chemical, presents new and different challenges. Additionally, the influence of
ADME factors requires a deeper consideration, which can be illustrated by several
questions, e.g.
1. Is the chemical stable in the gastrointestinal tract?
2. Will the chemical be absorbed?
3. Will first-pass metabolism detoxify the chemical?
4. Will the chemical be metabolised to a reactive species?