Cronin:
Guidance, case studies, and workflows especially with respect to REACH.
What will the European Chemicals Agency (ECHA) accept? So far, it will
accept a valid prediction. In future it will need more data, especially
repeated dose data.
Matthews:
Most molecules have multiple off-target activity. We do not look at
these activities enough. We need to look again at QSAR: a compound and
its metabolites are a constellation.
Boyer:
There is a gap between some modelers and some experimentalists.
Knowledge about mechanism and basic facts would help in constructing
and judging models. Another problem is that models built on 50
compounds are applied to 1000. Reviews are transferred into
experimental systems.
Richard:
There is a problem with the question. What area of toxicology are we
predicting? What exactly are we modeling? Our models are limited by
regulatory requirements and they may or may not be relevant to humans.
What endpoints are we modeling? Each endpoint may need a different
approach.
Question:But
what about application in the pharmaceutical industry?
Richard:
Target to the most relevant endpoint for the drug.
Guengerich:
More mechanistic information must be built in, with relevance to human
toxicology.
Boyer:
One of the biggest gaps is cultural: the data and the data generators
are separate from the modelers. They sometimes do not understand the
term “multivariate data”. If they did they might
change the
data gathered. Iteration in relationships would help here. Clinicians
are less receptive. These guys have their own problems and we modelers
are giving them too complicated a message.
Benigni:
They appreciate simple tools such as Toxtree.
Greene:
Going back to the first question of the key gaps, we have a very
limited understanding of the biological processes behind toxicity. If
we understood them better we could model them better. The second
problem is access to data.
Richard:
One solution is to engage toxicologists to inject more biology into the
model at the level of a structured database and data mining. Chihae
Yang’s work matters here.
Greene:
We could look at how we describe things: people use different terms for
the same thing.
Richard:
ToxML tries to answer this. The International Life Sciences Institute
(ILSI) group will make the database available.
From
the floor:
We are good at prediction for rats and mice but we should be looking at
human-relevant toxicology. It is convenient to class things as drugs,
chemicals etc., but they are all xenobiotic. As Paracelsus said
“Everything is toxic”.
Richard:
We do not have the human data so we work on what we have, e.g., rodent
in vivo data.
Matthews:
There are two problems: the vocabulary (we use Medical Dictionary for
Regulatory Activities, MedDRA)35 and the denominator for exposure. In
the case of a rare occurrence in just two people it is hard to
establish a mechanism. You need to look at the majority of the
population, therefore most pharmaceutical companies use the whole
population as denominator to get the most significant results.
Question:
[Inaudible]
Matthews:
Gather a large database of adverse drug reactions (ADR) and add the
whole population exposure. This is a straightforward computer problem
but it has not been done.
Guengerich:
You can save blood samples in our hospital and track ADRs for any drug,
and then tie a hypothesis back to the DNA.
Matthews:
Tens of thousands of clinical trials are available at the Center for
Drugs. There is computer power there but not the other resources.
Question:
Has the time for in
silico come? AstraZeneca and Pfizer have put in
significant effort.
Matthews:
Many tools are designed to be used as tool boxes, for example the OECD
toolbox, and these tools are applied naively. If you developed your own
database it would be better. The whole area has taken off. It will
explode.
Question:
But in big pharma resources are challenged.
Greene:
There has been a massive expansion in the computational area recently.
Boyer:
This is true for AstraZeneca too but you have to make a reasoned case
for an activity.
Glen:
Large systems must be broken into smaller components so much research
is needed. We need to break QSAR down into understandable bits. How do
we calculate solubility? Why use octanol/water?
Richard:
We need a fundamental change in attitude. I know toxicologists who have
worked on one group of chemicals for 20 years. We need to pull all the
data together. It can cost $250,000 to do one Multigenerational
Developmental Toxicity study and $4 million to do a rodent bioassay
study. We can do lots with that sort of money. Even in companies you
can standardize data even if it is not shared. You need standardization
so you can look across data. Compare genomics. You can look at common
patterns of effects if you have standards, and at low cost too.
From
the floor:
It is not easy to detect small differences, e.g., small perturbances
that may cause a tumor in 30 years time. It is impossible to separate
these from the background.
Boyer:
We should be humble when we look at the magnitude of the task. My DNA
is 99% the same as that of a chimpanzee but the 1% difference results
in a large phenotypic change. We must recognize how large the problem
is and at the same time try to model the small details. We may not be
able to model ourselves.
Benigni:
Microarrays show that gene expression is not very specific: you can
reach the same point from different paths. We have big maps of pathways
on the wall but we need to know the kinetics. It is not simple to model
all this. In an interconnected network of biochemical pathways, the
presence of only one feedback loop makes modeling impossible. We have a
long way to go.
Guengerich:
People look for the things that change most, and the smaller changes
might be more important. We will understand the central points
eventually.
From
the floor: There will be unknowns but you can still get
useful predictions.
Question
from the floor:Will
it be easier with biologicals?
Boyer:
No. The problems are different but there is the same number of problems.
