3 Modelling Simple Toxicity Endpoints …
39
Additionally, data for training mutagenicity and skin sensitisation models comes
from standardised assays such as the Ames test and murine local lymph node assay
(LLNA) which have OECD test guidelines and thus high levels of reproducibility
[7].
As a result, there has been increasing regulatory acceptance of models for simple
endpoints. Mutagenicity predictions are now accepted in lieu of data from an Ames
test for certain pharmaceutical impurities or plant protection product metabolites [8,
9]. In silico skin sensitisation predictions, especially when used as part of a defined
approach (DA), can now be used in place of traditional in vivo LLNA or Guinea
pig maximisation tests (GPMT) [10–12]. Predictions from such models are also
straightforward to validate, e.g. by running an in vitro assay to confirm or refute the
in silico result. This data can then be supplied to the developers to incrementally
improve model performance [13].
It is now straightforward (and accepted) to produce qualitative models for simple
toxicity endpoints. However, the relative success of these tools has unveiled further challenges relating to interpreting and applying the results of models. These
include: how to make negative predictions, moving from qualitative to quantitative
predictions, the necessity for expert review of predictions and how to model complex
endpoints where these methods are not suitable.
3.3 Making Negative Predictions
The first challenge to be addressed is whether the lack of toxicity can be predicted
using structural alerts. These define chemical fragments that are causative of an
adverse outcome. When no alerts are found in a compound, is it reasonable to assume
that the lack of a positive prediction is enough evidence to make an explicit negative
prediction? In this context, two questions need further consideration:
1. Do the existing structural alerts in the appropriate chemical space cover the
known mechanisms of toxicity well?
2. Is the adverse outcome driven by a single MIE?
If the answer to both questions is yes, then it can be expected that an absence of
structural alerts will be indicative of a lack of toxicity. This has been demonstrated
for some endpoints including bacterial in vitro mutagenicity and in vivo skin sensitisation [14]. Both endpoints have been studied for numerous years, resulting in many
structural alerts, and both are largely dependent on a reactivity-driven MIE. However,
methodologies are also required to assess the reliability of individual predictions, to
identify (and justify) whether these can be treated with a higher or lower level of
confidence.
Two differing methodologies for assessing the reliability of alert-based negative
predictions have previously been investigated for bacterial mutagenicity: an expert
knowledge-based approach considered whether a non-alerting chemical had been
purposely excluded from the scope of a structural alert, and a data-driven approach
39
Additionally, data for training mutagenicity and skin sensitisation models comes
from standardised assays such as the Ames test and murine local lymph node assay
(LLNA) which have OECD test guidelines and thus high levels of reproducibility
[7].
As a result, there has been increasing regulatory acceptance of models for simple
endpoints. Mutagenicity predictions are now accepted in lieu of data from an Ames
test for certain pharmaceutical impurities or plant protection product metabolites [8,
9]. In silico skin sensitisation predictions, especially when used as part of a defined
approach (DA), can now be used in place of traditional in vivo LLNA or Guinea
pig maximisation tests (GPMT) [10–12]. Predictions from such models are also
straightforward to validate, e.g. by running an in vitro assay to confirm or refute the
in silico result. This data can then be supplied to the developers to incrementally
improve model performance [13].
It is now straightforward (and accepted) to produce qualitative models for simple
toxicity endpoints. However, the relative success of these tools has unveiled further challenges relating to interpreting and applying the results of models. These
include: how to make negative predictions, moving from qualitative to quantitative
predictions, the necessity for expert review of predictions and how to model complex
endpoints where these methods are not suitable.
3.3 Making Negative Predictions
The first challenge to be addressed is whether the lack of toxicity can be predicted
using structural alerts. These define chemical fragments that are causative of an
adverse outcome. When no alerts are found in a compound, is it reasonable to assume
that the lack of a positive prediction is enough evidence to make an explicit negative
prediction? In this context, two questions need further consideration:
1. Do the existing structural alerts in the appropriate chemical space cover the
known mechanisms of toxicity well?
2. Is the adverse outcome driven by a single MIE?
If the answer to both questions is yes, then it can be expected that an absence of
structural alerts will be indicative of a lack of toxicity. This has been demonstrated
for some endpoints including bacterial in vitro mutagenicity and in vivo skin sensitisation [14]. Both endpoints have been studied for numerous years, resulting in many
structural alerts, and both are largely dependent on a reactivity-driven MIE. However,
methodologies are also required to assess the reliability of individual predictions, to
identify (and justify) whether these can be treated with a higher or lower level of
confidence.
Two differing methodologies for assessing the reliability of alert-based negative
predictions have previously been investigated for bacterial mutagenicity: an expert
knowledge-based approach considered whether a non-alerting chemical had been
purposely excluded from the scope of a structural alert, and a data-driven approach