New Horizons in Toxicity Prediction.
Lhasa Limited Symposium Event in Collaboration with
the University of Cambridge - February 2009
A Report by Wendy A. Warr, wendy@warr.com, http://www.warr.com Knowledge-based
approaches for toxicity prediction
Nigel Greene, Pfizer
Early hazard identification in the pharmaceutical industry is very
important because development costs increase exponentially over time
and stage. Adverse safety effects may be due to primary pharmacology
(e.g., phosphodiesterase-4, PDE-4, inhibitors are linked to emesis and
vasculitis), secondary pharmacology (e.g., D-1 activity is linked to
tremors), chemical structure (e.g., clozapine causes agranulocytosis
and forms reactive metabolites), or physicochemical properties (e.g.,
lipophilic basic compounds have a risk of causing phospholipidosis,
hepatotoxicity, and QT interval prolongation].
Approaches to toxicity prediction based on machine learning are fast to
find relationships and can deal with complex data and relationships,
but they are dependent on high quality data and can be difficult to
interpret. Knowledge-based approaches can cope with
“fuzzy”
data sets and the results are easy to interpret, but they are slow to
develop and they may not identify complex relationships. It is always
important to consider exposure. Greene illustrated this point with a
data set where area under the curve (AUC) and maximum concentration
(Cmax) vary over seven orders of magnitude for compounds that are
administered at the same dose. [In bioequivalence studies, Cmax is
shown to reflect not only the rate but also the extent of absorption.
Cmax is highly correlated with AUC contrasting blood concentration with
time. Therefore, use of the Cmax/AUC ratio is recommended for assessing
the equivalence of absorption rates.]
In the hit identification to lead optimization phases of drug
discovery, knowledge-based approaches can quickly identify a
toxicophore and a potential mechanism. In addition any new information
can be added quickly to the knowledge base. In later phases,
knowledge-based approaches can help with risk assessment for synthetic
intermediates, low-level impurities, metabolites, degredants and
excipients and the information may form part of a regulatory submission.
Legacy and public data are fairly readily available for in vivo
outcomes that can be used for building in silico models.
It is easy to
correlate the predictions of these models to clinical outcomes but
often it is not clear if there is an in vitro assay to
use for
confirming the prediction. It is difficult and costly to validate these
systems and often model performance may be limited by the complexity of
the biological systems. Model development is also hindered by a lack of
exposure information in preclinical species.
in
silico models to predict the results of in vitro assays can
be used
to prioritize screening and thus reduce assay capacity requirements,
and give a clear next step for exploring a potential safety issue. It
is relatively easy to validate these models, but they require a
training set of compounds, and often the in vitro assay
correlates
poorly with in vivo
outcomes. Cell based systems reduce complexity of
the system to some extent because fewer mechanisms are involved.
in
vitro assays have higher throughput and are cheaper to run
than in
vivo ones. They reduce the time taken to make a decision
and enable SAR
and comparison of series. They implement the “three
R’s”: reducing, refining and replacing animals. The
relatively simple readouts are easy to understand. Thus many common in
vitro assays (e.g., hERG patch clamp assay, Ames test) are
in use,
although they frequently have poor correlation to in vivo toxicity and
may not be broadly applicable. A high false positive rate would
eliminate too many compounds that might be useful, but the assays may
help resolve mechanisms of toxicity were an in vivo issue
identified.
Derek for Windows is an example of a knowledge-based approach. It is
continually improving, users can store their own knowledge using the
editor, and it is not an isolated system since it can interact with
other software, such as physicochemical property predictors, through an
adapter. Greene gave an example of the development of an alert. From
156 compounds in a database, more than 71 with a 2-aminopyrimidine
substructure were positive in an in
vitro micronucleus assay. SAR
revealed several distinct subclasses. Mechanisms of action were
studied. Greene showed some colored matrices of ranges of values for
inhibition, showing that one structural class consisted of
non-selective kinase inhibitors while another contained selective
inhibitors of an unrevealed enzyme.
Greene has also developed hepatotoxicity SARs. About 50 structural
classes known to cause liver injury in humans were identified and
implemented in Derek for Windows. Performance was evaluated against
about 600 compounds compiled from internal and external sources. The
validation set contained both idiosyncratic and dose-dependent
hepatotoxicants. Derek for Windows predictivity for positives was 45%
and for negatives was 76.3%. The results were rather better if weak and
animal-only results were ignored. Sensitivity was reduced due to animal
hepatotoxicants not identified correctly and specificity suffered due
to compounds that have fewer than 10 case reports of liver injury. The
development is addressing these issues.
Predictions based on physicochemical properties are also being
developed. The Ploeman model22 involving CLogP
and pKa has been
implemented in Derek using the adapter to computational models. A
statistically significant correlation between CLogP and Topological
Polar Surface Area (TPSA) and increased incidence of findings in in
vivo toxicology studies is also being considered.
There has been much research looking at using batteries of in vitro
assays to predict an in
vivo outcome, for example in the
EPA’s
ToxCast, CEREP Bioprint profiling, and work by Roche and Pfizer on
kinase selectivity as a surrogate for in vitro
micronucleus. Some
success stories have been reported but the patterns may not be broadly
applicable.
