Back to the Top
Hi,
I was interested in the accuracy of the currently available methods of
predicting human PK parameters.
a) What is the range of accuracy of "purely" in silico methods
b) What is accuracy range of in vitro methods
c) How well does animal PK correlate with human PK.
Any pointers will be highly appreciated.
Thanks
Kas
Back to the Top
The following message was posted to: PharmPK
Dear Kas,
My name is Juergen Bulitta and I am currently working on a QSPKR
(quantitative structure pharmacokinetics relationships) project. You
already
mentioned three methods how to predict human PK data. I cannot comment
on
in-vitro methods and on allometric scaling, rather than on pure
in-silico
methods. However, to my understanding, the primary problem in using
experimental data is the high inherent degree of experimental error for
example in solubility values of compounds - besides the time and costs
those
experiments require.
If you refer to in silico methods there are three sections that are
necessarily involved:
1) PK-datasets
2) Molecular descriptors
3) any kind of statistical model linking 1) and 2)
1) It is not easy to gather a consistent PK dataset from literature,
since
sources of variability are manifold. I guess, it really depends on what
you
are looking for. Speaking in general, it might be useful to be rather
picky
with the choice of PK-parameters.
2) There is also a huge variety of molecular descriptors. And to the
best of
my knowledge different authors rarely use the same descriptors. There
are
also other computational details that possibly need to be considered in
constructing a set of descriptors.
3) The statistical model links the molecular descriptor dataset with the
PK-parameters. Most often, either artificial neural networks or multiple
linear regression models (usually with a principal components analysis
or
partial least squares) are applied.
It might be helpful, if you can provide some more details on what kind
of
PK-prediction you are looking for. Some helpful references with further
citations are possibly:
Turner JV, Maddalena DJ, Cutler DJ, Agatonovic-Kustrin S.
Multiple pharmacokinetic parameter prediction for a series of
cephalosporins.
J Pharm Sci. 2003 Mar;92(3):552-9.
Mager DE, Jusko WJ.
Quantitative structure-pharmacokinetic/pharmacodynamic relationships of
corticosteroids in man.
J Pharm Sci. 2002 Nov;91(11):2441-51.
Van der Graaf PH, Nilsson J, Van Schaick EA, Danhof M.
Multivariate quantitative structure-pharmacokinetic relationships
(QSPKR)
analysis of adenosine A1 receptor agonists in rat.
J Pharm Sci. 1999 Mar;88(3):306-12.
Some helpful keywords in PubMed would be:
"quantitative structure pharmacokinetics"
"QSPKR"
"in silico pharmacokinetics"
"ADME"
"ADMET"
In general, the results of the published models in literature are
promising.
However, this is still an open field of research and the QSPKRs
published
are still sparse. The same applies to which set of molecular descriptors
should be used.
I hope that these initial comments were helpful for you. If you specify
your
problem in more detail, I will try to give you a more specific answer.
Best regards
Juergen
Back to the Top
The following message was posted to: PharmPK
Dear Juergen,
For some elaboration on my original question...
We have been working on predicting PK parameters measured in humans -
specifically values of absorption and elimination rates, volume of
distribution, protein binding and bioavailability. The values have been
taken from the FDA database. Our methods of prediction are based on AI
(machine learning). If we plot the predicted values against the actual
values, we are able to get R sq. values greater than 80% and for some
parameters > 90%.
I wanted to estimate how well this compares with existing methods,
whether
in silico or otherwise.
Thank you for your reply and the list of references. I will follow up on
those to see if I can glean further insights what has been done and what
accuracies have been achieved.
Kas
Back to the Top
What FDA database are you talking about?
Is this already summarized or you have to go drug by drug?
Do you have an electronic version of the database that you could share?
Thanks very much for your help,
Dario
dario_doller.-a-.yahoo.com
Back to the Top
The following message was posted to: PharmPK
Dear Kas,
Most likely, you have already considered this point many times
before. If I understand correctly, you have a rather large
database with say 30 or more compounds. I believe you should
invest a reasonable amount of time not only to correlate PK
with any kinds of descirptors, rather than you should also
spend much time in validation of your system.
Personally I do not favor r^2 values, since these are pure
numerical values and are often misleading. Focusing on
q^2 for cross validation, you might try to validate your
system not only be leave-one-out rather than by leave-20%-out
or whatever. It may also be advisable to make a "2nd stage"
validation. You wrote, you are actually predicting also
rate constants. However, in drug development it might be
more important to predict plasma concentrations than model
parameters. Therefore, looking onto the real-life usefulness
by predicting actual concentrations might be a nice add-on
in validation. Hope this helps.
All the best
Juergen
Back to the Top
Hi all,
I am also currently using in silico methods for PK predictions.
However, given the wide range of values available for a single PK
parameter in the literature, which will be an appropriate value for in
silico use? For example, the protein binding of a drug gathered from
various literatures or even from a single literature might range from
50 to 80%. Which protein binding value for that drug should we use for
training? Do we use the average which is 65%? Assuming we use 65% for
training. If the result of the prediction is 78%, do we consider it as
a correct prediction or should we calculate any errors (e.g. square
errors) for it?
The reason for asking these questions is to gather any general
consensus for the evaluation of in silico prediction methods.
Thanks
Chun Wei
PharmPK Discussion List Archive Index page
Copyright 1995-2010 David W. A. Bourne (david@boomer.org)