Whither Predictive Toxicology??

advertisement
Full Speed into an
Alternative Future
Bob Chapin
Pfizer DART Group
Groton, CT
Background and Stage-Setting
• I spent 18 yrs at NIEHS doing male reproductive tox,
developing methods in primary cell culture, running a
lab and overseeing contract studies.
• Been at Pfizer for 9 yrs now, in a lab which develops
in vitro predictive assays and troubleshoots DART
issues.
• I am a very strong advocate of and believer in a
future for toxicology based increasingly on measuring
biochemical and genomic responses in cells in culture,
and in computational modeling of complex systems.
• My comments will come mostly from the pharma
perspective, we’ll spend less time on environmental
chemicals.
What are the motivations to get safety right?
• Chasing positives:
– The more specific and beneficial our meds can be, the more
people will be helped by them.
– Safer meds have an easier time finding a new indication.
– Safer and more specific pesticides (“plant protection
agents”) will provide more food for a hungry world with less
environmental damage
• Avoiding negatives
– Every adverse effect is undesired
– No one wants them; toll of human suffering, etc
Safety Assessment
• Current state: We give animals more and more of a
compound and watch what goes wrong. We then
examine the low end of the dose-response curve and
extrapolate that to humans based on experience and
some “safety factors”. It’s certainly not perfect, but
it’s worked pretty well (we think).
• “Alternative models” are cell culture or tissue culture
methods (or any other method) which reduce the
numbers of animals treated with toxicants.
Alternatives going mainstream:
• Future state: as laid out in the NAS book “Toxicity
Testing in the 21st Century”. Focused on cell culture
(with human cells) and progressively more
computational predictions.
• We’ll use the term “predictive model”. This can be
– a well-characterized cell culture system whose data are
massaged a certain way to give a prediction of toxicity
– A combination of different kinds of data from cell cultures
which, when combined using a certain statistical method,
yields an estimate of the in vivo activity of this compound
– A combination of different computer programs (each of which
predicts different things) which together predict the in vivo
toxicity of an exposure
Historically
Safety assessors have relied on animals because
– Assumptions of relevance based on conserved
evolutionary features
– Have integrated ADME (absorption, distribution,
metabolism, and excretion)
– Have integrated physiology that allows for
recovery and adaptation
– All target tissues are there in the relevant milieus
– The target tissues have all the appropriate cells
there in the right configuration
Animal-based systems
• May be all right, but the assumed predictivity may not be
all it’s cracked up to be.
• Harry Olson et al* did a survey and reported that there
was ≈ 43% predictivity of human response from rats alone.
Using all animal models, the predictivity was ≈ 70%. This
is depressingly low.
• Animals “get the basic physiology right” (control of heart rate
and blood pressure, steroid hormones, etc), but many toxic
responses differ between rodents and humans, which tells
me that there are subtle differences in biology that are
critically important, and we don’t yet know what they are.
*H Olson et
al. Concordance of the toxicity of pharmaceuticals in
humans and in animals. Regulatory Toxicology and Pharmacology 32:
56-67, 2000.
• And to build a tool (like a model for predicting toxicity), we
need to know what the critical parts are and how they
fit together. That is:
• If we commit to trying to model toxic responses using
in vitro and in silico tools, we will need to know all the
important component parts and how they interact and
feed-back and -forward.
– Absorption, distribution (protein binding, delivery to
peripheral tissues), metabolism (which of the P450’s is used
preferentially, whether a toxic metabolite will be formed,
whether this will interfere with basal body functions, etc)
and excretion (will it harm the kidneys? Will it deplete the
liver of critical glutathione?, etc, etc)
– What are the off-target targets? (the 800 lb gorilla)
– What will Cmpd X do to the foci that control heart rate,
blood pressure, breathing control, adrenal function, pacreatic
function, etc, etc.
• As it turns out, this is exceptionally difficult:
Current state (a very small sample of very focused efforts):
• Pfizer’s DART models seem
to have a glass ceiling at
about 70% predictivity (n >
80 compounds).
• ECVAM (DART): ≈ 80% for
the first 20 compounds,
then 2/13 for a second test
set (15% correct).
• A consortium of several
pharma is using Zebrafish
as a model for developmntal
tox, and is struggling to get
much past 50% predictive.
• A general safety prediction
group is struggling with the
false negative predictions
for truly-toxic compounds
True Negatives
Predicted
Negative
Predicted
Toxic
82
10
True Positives
(Toxic)
88
~8%
False
Positive
“clean”
113
“findings”
WTF??
• We probably don’t know the right things to measure
(but at least one of those previous examples has been
encyclopedic in their examination of endpoints)
• For those groups
reporting success,
Chapin’s Curve of Modeling Despair
they often are not
90 %
testing enough
chemicals:
• Or it could be
that the cultures Predictivity
don’t capture
enough (or the
right) biological
complexity
10
100
300
Number of compounds
Paracrine: the relationship when 1 cells secretes
something that affects its neighbor.
Schwartz and Holst
Another example of paracrine complexity
Ten Broek et al., J Cell Physiol.
224:7-16, 2010
Paracrine effects
• So are these key, or just some biological jawdroppers that we can marvel at and then ignore?
• So far we’ve been following the Einstein mantra
(“Make everything as simple as possible, but no
simpler”), and it’s often worked well enough, but
certainly not always.
• Perhaps the targeted addition of complexity to our
cultures will make the cells feel more at home, and
they’ll give better answers.
• We’ll also need to better intuit what the key
responses are (i.e., as scientists, we probably need to think
differently).
• Industry will embrace these when their predictivity
become dependable (“All models are wrong, but some
are useful.”) or  when they are required as part of a
regulatory submission package.
• Currently, predictive models (and their component
assays) are used for internal prioritization, but
because regulatory decision-makers have relatively
little experience with these models, in vitro data add
little to regulatory safety packages.
Role of stem cells
• Stem cells can (theoretically, and increasingly in real
life) turn into any cell type.
• They can also be used to model (or recapitulate) the
process of differentiation that an embryo normally
undergoes.
• Multiple cell types are (probably) important in a
toxicity response because of paracrinology. So any
model that gives rise to multiple related cell types
(normally found together in vivo) could be more useful
than a plate full of a single cell type (if we knew what
to measure).
Download