3 Modelling Simple Toxicity Endpoints …
51
presented [43–45]. Despite the relativity high sensitivity of these models for the
training sets, their performance for independent test sets was often low.
QSAR models using histopathology data have been published for liver and nephrotoxicity [46]. However, the performance of these models (with initially high sensitivity and specificity against external test sets) dropped when subsequently validated
with new data [46]. This and earlier examples indicate that currently global QSAR
models for complex endpoints have limited coverage of relevant chemical space.
As knowledge of different mechanisms leading to organ toxicity increases, it is
increasingly being presented in the form of AOPs [47]. There are several in vitro
assays that measure MIEs or KEs for some endpoints (e.g. BSEP inhibition, mitochondrial toxicity, etc.). The output of these assays in the form of activity towards a
specific target can be used as indicators of whether events on an AOP are likely or not.
A battery of the assays describing a selection of the key mechanisms in combination
with daily dose (and covalent binding to proteins) has been shown to be useful in
identification of potential hepatotoxic liabilities of compounds [48]. However, this
approach somewhat lacks comprehensive coverage for an AOP, as it is limited to a
selection of mechanisms. As the array of mechanistically derived tests develops, it
should lead to a wider coverage of MIEs and KEs for each AOP.
Several publications describe approaches for modelling various MIE models,
where the physicochemical properties of chemicals [49] or structural information
in combination with multitarget bioactivity [50] are used. The bioactivity data for
the modelling is obtained from in vitro assays, that can be used to derive SARs or
for machine-learning algorithms. Different approaches for using this data have been
adopted: whilst in some cases all available data was utilised, in other cases only a subset of assays that were assumed to be relevant for the adverse outcome were taken
into consideration. For example, from the available in vitro high-throughput data
generated in the ToxCast project, data from only nine MIE endpoints (peroxisome
proliferator-activated receptor (PPAR) alpha, PPAR beta, PPAR gamma, constitutive
androstane receptor, pregnane X receptor, aryl hydrocarbon receptor, liver X receptor, nuclear factor (erythroid-derived 2)-like 2, farnesoid X receptor) was used to
develop random forest models for MIEs relevant to hepatic steatosis [51]. Similarly,
when creating a liver cholestasis model, the authors utilised existing knowledge and
selected data for the inhibition of hepatic transporters (BSEP, breast cancer resistance protein, P-glycoprotein, organic anion transporting polypeptide(OATP)1B1
and OATP1B3) previously shown to disrupt the bile flow [52]. In other cases, when
data from a high number of in vitro assays was applied, the predictive capacity of
almost all models (regardless of the algorithm) was improved when combined with
structural information about the chemical itself [53]. This was exemplified for the
hepatotoxicity endpoint [54], with a similar outcome for other organ toxicities [55].
Other types of data, such as those from gene expression studies, are also becoming
more accessible and have been used for predicting organ toxicity in a systems biology
approach [56], or to generate machine-learning models [57].
Utilising in vitro assays data for building (Q)SAR models enables the prediction of
possible liabilities for a new chemical; these global models for individual MIE/KEs or
apical endpoints are likely to ‘catch’ potential intrinsic toxicities of compounds and
51
presented [43–45]. Despite the relativity high sensitivity of these models for the
training sets, their performance for independent test sets was often low.
QSAR models using histopathology data have been published for liver and nephrotoxicity [46]. However, the performance of these models (with initially high sensitivity and specificity against external test sets) dropped when subsequently validated
with new data [46]. This and earlier examples indicate that currently global QSAR
models for complex endpoints have limited coverage of relevant chemical space.
As knowledge of different mechanisms leading to organ toxicity increases, it is
increasingly being presented in the form of AOPs [47]. There are several in vitro
assays that measure MIEs or KEs for some endpoints (e.g. BSEP inhibition, mitochondrial toxicity, etc.). The output of these assays in the form of activity towards a
specific target can be used as indicators of whether events on an AOP are likely or not.
A battery of the assays describing a selection of the key mechanisms in combination
with daily dose (and covalent binding to proteins) has been shown to be useful in
identification of potential hepatotoxic liabilities of compounds [48]. However, this
approach somewhat lacks comprehensive coverage for an AOP, as it is limited to a
selection of mechanisms. As the array of mechanistically derived tests develops, it
should lead to a wider coverage of MIEs and KEs for each AOP.
Several publications describe approaches for modelling various MIE models,
where the physicochemical properties of chemicals [49] or structural information
in combination with multitarget bioactivity [50] are used. The bioactivity data for
the modelling is obtained from in vitro assays, that can be used to derive SARs or
for machine-learning algorithms. Different approaches for using this data have been
adopted: whilst in some cases all available data was utilised, in other cases only a subset of assays that were assumed to be relevant for the adverse outcome were taken
into consideration. For example, from the available in vitro high-throughput data
generated in the ToxCast project, data from only nine MIE endpoints (peroxisome
proliferator-activated receptor (PPAR) alpha, PPAR beta, PPAR gamma, constitutive
androstane receptor, pregnane X receptor, aryl hydrocarbon receptor, liver X receptor, nuclear factor (erythroid-derived 2)-like 2, farnesoid X receptor) was used to
develop random forest models for MIEs relevant to hepatic steatosis [51]. Similarly,
when creating a liver cholestasis model, the authors utilised existing knowledge and
selected data for the inhibition of hepatic transporters (BSEP, breast cancer resistance protein, P-glycoprotein, organic anion transporting polypeptide(OATP)1B1
and OATP1B3) previously shown to disrupt the bile flow [52]. In other cases, when
data from a high number of in vitro assays was applied, the predictive capacity of
almost all models (regardless of the algorithm) was improved when combined with
structural information about the chemical itself [53]. This was exemplified for the
hepatotoxicity endpoint [54], with a similar outcome for other organ toxicities [55].
Other types of data, such as those from gene expression studies, are also becoming
more accessible and have been used for predicting organ toxicity in a systems biology
approach [56], or to generate machine-learning models [57].
Utilising in vitro assays data for building (Q)SAR models enables the prediction of
possible liabilities for a new chemical; these global models for individual MIE/KEs or
apical endpoints are likely to ‘catch’ potential intrinsic toxicities of compounds and