50
R. Williams et al.
5. Will the chemical (or any metabolite) generate a molecular initiating event?
6. Will the chemical show high levels of protein plasma binding?
7. Will the chemical be ‘quickly’ cleared from the body?
Each of these processes will have a unique structure–activity relationship
(SAR—which may be increasingly granular, e.g. interactions with specific transporters), which is masked by the apical endpoint data. For example, if a chemical is classified as ‘not hepatotoxic’ based on experimental data—is this because
the chemical does not ‘activate’ an MIE, or because it is not absorbed or quickly
excreted? From this, it becomes obvious that summary data from apical endpoints is
an oversimplified summary of the biological reality. Additionally, factors that promote activity in some of the component models could curtail activity in another. For
example, the addition of a methyl group to an arene-containing chemical creates a
potential ‘metabolic hook’ which can lead to the generation of reactive metabolites,
whereas the same methyl group attached close to a reactive moiety can deactivate
this by preventing access to a biological nucleophile, such as DNA, through steric
blocking. Moreover, to model all these processes to the fine detail required to make
accurate predictions requires much data. If in silico models of mutagenicity are now
based on greater than 10,000 chemicals, how much data would be required to create
broadly applicable models of more complex apical endpoints to a similar level of
precision? This is without considering the increasing heterogeneity within the data
when moving away from reproducible standardised assays.
Hence, to accurately model multifactorial endpoints such as hepatotoxicity, ideally the multiple SARs that relate to the multiple processes leading to the high-level
outcomes to activity or inactivity need to be accounted for. This would enable persuasive arguments to be made; e.g., a chemical is predicted to be hepatotoxic (due to
cholestasis) because it blocks the bile salt export pump (BSEP) channel as well as
being straightforward to validate (e.g. in this case by running a BSEP assay).
Some of the complication inherent in complex global models can be rectified by
using local models. However, local models are designed to only predict the activity of
a narrow series of congeneric chemicals and lack general applicability. A knowledge
base containing a series of structural alerts is, in effect, a collection of local models
which obviates the problem of applicability. Structural alerts can be designed to be
activated by a given chemical class operating through a single MIE, taking account of
heterogeneity in the training data [39]. Thus, one method which has been employed
to model in vivo endpoints is through extrapolation of alerts from a related in vitro
endpoint that share a common MIE. This has been demonstrated with some success
for the chromosome damage endpoint [40], though the resulting alerts still carry the
complication of additional ADME factors.
With respect to creating SARs for complex endpoints, there are several models
available which use available in vivo data to predict potential liabilities for new
chemicals. In many of these cases, the toxicophores were identified either statistically
[41] or by an expert [42]. In other cases, automated clustering followed by human
evaluation where the relevance of statistically determined toxicophores was assessed
by an expert and/or by searching public literature for mechanistic evidence has been
R. Williams et al.
5. Will the chemical (or any metabolite) generate a molecular initiating event?
6. Will the chemical show high levels of protein plasma binding?
7. Will the chemical be ‘quickly’ cleared from the body?
Each of these processes will have a unique structure–activity relationship
(SAR—which may be increasingly granular, e.g. interactions with specific transporters), which is masked by the apical endpoint data. For example, if a chemical is classified as ‘not hepatotoxic’ based on experimental data—is this because
the chemical does not ‘activate’ an MIE, or because it is not absorbed or quickly
excreted? From this, it becomes obvious that summary data from apical endpoints is
an oversimplified summary of the biological reality. Additionally, factors that promote activity in some of the component models could curtail activity in another. For
example, the addition of a methyl group to an arene-containing chemical creates a
potential ‘metabolic hook’ which can lead to the generation of reactive metabolites,
whereas the same methyl group attached close to a reactive moiety can deactivate
this by preventing access to a biological nucleophile, such as DNA, through steric
blocking. Moreover, to model all these processes to the fine detail required to make
accurate predictions requires much data. If in silico models of mutagenicity are now
based on greater than 10,000 chemicals, how much data would be required to create
broadly applicable models of more complex apical endpoints to a similar level of
precision? This is without considering the increasing heterogeneity within the data
when moving away from reproducible standardised assays.
Hence, to accurately model multifactorial endpoints such as hepatotoxicity, ideally the multiple SARs that relate to the multiple processes leading to the high-level
outcomes to activity or inactivity need to be accounted for. This would enable persuasive arguments to be made; e.g., a chemical is predicted to be hepatotoxic (due to
cholestasis) because it blocks the bile salt export pump (BSEP) channel as well as
being straightforward to validate (e.g. in this case by running a BSEP assay).
Some of the complication inherent in complex global models can be rectified by
using local models. However, local models are designed to only predict the activity of
a narrow series of congeneric chemicals and lack general applicability. A knowledge
base containing a series of structural alerts is, in effect, a collection of local models
which obviates the problem of applicability. Structural alerts can be designed to be
activated by a given chemical class operating through a single MIE, taking account of
heterogeneity in the training data [39]. Thus, one method which has been employed
to model in vivo endpoints is through extrapolation of alerts from a related in vitro
endpoint that share a common MIE. This has been demonstrated with some success
for the chromosome damage endpoint [40], though the resulting alerts still carry the
complication of additional ADME factors.
With respect to creating SARs for complex endpoints, there are several models
available which use available in vivo data to predict potential liabilities for new
chemicals. In many of these cases, the toxicophores were identified either statistically
[41] or by an expert [42]. In other cases, automated clustering followed by human
evaluation where the relevance of statistically determined toxicophores was assessed
by an expert and/or by searching public literature for mechanistic evidence has been