Back to the Top
I have a study looking at the effect of treatment with MS222, an anesthetic for fish, on disposition
of a test drug. Two sets of fish, each with 90 individuals, were dosed by gavage with the test
compound. Half of the fish were anesthetized with MS222 to facilitate gavage while half were not.
Ten fish each were killed at one of nine times post-dosing.
As there is only one time point for each subject, I thought it appropriate to use Monolix for a
population PK analysis, though there are no covariates. This analysis was performed on each
dataset, one with the anesthetic treatment and one w/o treatment. Both were best fit by a
one-compartment extravascular model with a Tlag included.
Had this been a data rich study with each animal having multiple time-concentration points I would
have analyzed each animal as an individual. I would then have tested each parameter for an effect
of treatment (anesthetic) by performing an F test followed by the appropriate t-test for the “with
anesthetic” versus “without anesthetic”. Though the Monolix file “indiv_parameters.txt” provides an
estimate of the PK parameters for each fish, my impression is that the individual estimates are not
suitable for the above hypothesis testing when derived in a population PK approach.
Question #1: Am I correct that it is inappropriate to perform a statistical test on “with
anesthetic” versus “without anesthetic” on the individual parameters for each animal as seen in
“indiv_parameters.txt w MS222.txt” and “indiv_parameters wo MS222.txt”. (Files are temporarily
uploaded to http://mymedia.msstate.edu/outputset.php?id=5848 if you wish to see them.)
Question #2: If it is inappropriate to use the approach in Question #1, how could I test for an
effect of treatment with the anesthetic? Would I combine the datasets and then perform the analysis
with treatment as a covariate?
Thanks,
Cory
Back to the Top
Dear Cory,
With naïve pooled data evaluation of robustness of the model and identifiability of the data for
proper model and parmameter estimations is a problem.
You can combine the datasets and then perform the analysis with treatment as a covariate. Then
compare the results.
The lag-time is significantly different in two groups (p<0.0001).
So are you sure that your model with lag-time is the best model? Try other models with combined
data.
Have you any mechanism suggestion for lag-time difference?
Dr Sima Sadrai
TUMS
Back to the Top
Dr. Sadrai,
Thank you for your reply, which brings up two very good points.
1) Is a Tlag appropriate?
Your question about the lag time is one I have considered. Physiologically I can hypothesize that
the handling involved in the gavage of the fish might have altered GI motility resulting in a lag
time. That is just a guess though. In terms of comparing the effect of the anesthetic on
absorption I would have preferred a model w/o a Tlag as then only ka describes the absorption rate
rather than having Tlag and ka together describing it. I mention this because in the original
analysis the Tlag is significant (as you pointed out), but the ka is not, thus confounding an answer
as to whether treatment effects absorption.
The AIC scores are better for a model including a Tlag but not by a lot:
pop_parameters w MS222.1 : Tlag AIC = -3602.01, noLag AIC = -3311.85
pop_parameters wo MS222.1 : Tlag AIC = -4412.36, noLag AIC = -4144.13
If you plot the predicted values however, the model with the Tlag is a much better fit. See file
"Plot_lag_versus no_lag.JPG" temporarily uploaded to
http://mymedia.msstate.edu/outputset.php?id=5848
If I were able to put limits on the parameters I might be able to get a good model w/o a Tlag, but
as far as I can tell neither Monolix nor WNL5 (Phoenix) allow you to fix a limit on a parameter.
(According to Pharsight, WNL5 evidently lets you set a starting upper and lower limit, but ignores
it subsequently in the fitting process.)
I will proceed as you suggested with creating a model with treatment as a covariate. If I use the
Phoenix model then I am told I can indeed set limits on parameters.
2) You commented that "The lag-time is significantly different in two groups (p<0.0001)." Does this
imply that it is permissible to use the individual pk estimates from Monolix in a statistical
comparison between groups? This was a question I have not really received an answer for from
anyone.
Thanks,
Cory
Back to the Top
Hi Cory,
Have you considered posting your project itself to the forum; www.pharsight.com/extranet,
Where you can get tips and discuss Phoenix products and models with other users;
Then we can examine more easily your reported issues with bounds.
Best regards,
Simon.
Simon.Davis.-a-.certara.com
Senior Scientific Consultant
[Simon, The old NONLIN would use the upper/lower limits to re-parameterize the whole problem and
proceed to the fitting without limits in the 'new' parameter space. [See equation 2.20 on page 9 of
A Users Manual for NONLIN and Associated Program, Metzler, Eflring and McEwen, Upjohn, 1974]. Sort
of like putting the problem in a big bowl. I would contrast this with Boomer which doesn't
re-parameterize like this but keeps the 'hard' limits like fitting in a box. Does the current
version do this conversion like the original? It might explain the confusion Cory has with the
current manual.
Cory, Regarding the tlag. If you don't want to use a tlag it may be possible to get better results
(i.e. similar to tlag) with two first order ka's, one after the other in series. I haven't looked at
the data this may not be appropriate. Can you do a deconvolution on the data to explore the
absorption process - db]
Back to the Top
Hi Simon,
Along the same theme, is it possible to evaluate a period or sequence effect for a nonlinear
function within Phoenix or NONMEM for a crossover type of study design? Not as a covariate, but on
the response. Similar to linear models where Response ~ f(Dose + Period + Sequence), can you
evaluate Response ~ f(PK + Period + Sequence) whereby the parameters of the exposure-response model
are adjusted for period and sequence effects?
-Max
Back to the Top
HI Max,
If I understand your question correctly; factors like a period effect in a x-over model can be
modeled using inter-occasion effects in NLME.
My colleague summarized using IOV for x-over as follows;
"the difference between categorical covariates and occasions is that covariates choose thetas, while
occasions choose etas.
So if you want to know if the there's a predictable difference between crossover period 1 and 2, I
would suppose that is a covariate effect.
If you simply want to allow an individual to have different etas (so different structural
parameters) between periods 1 and 2, without necessarily explaining why, then that's an
interoccasion effect."
If you want some tips with specific set/model then I would recommend posting your project/question
directly to the Phoenix forum; www.pharsight.com/extranet
Best regards,
Simon.
--
Simon.Davis.-a-.certara.com
Senior Scientific Consultant
Pharsight- A Certara™ Company
Back to the Top
Cory, I believe you've already had a response to this from our support group but I just wanted to
clarify this to the rest of the list;
In normal circumstances WNL Classic Modeling does not ignore limits; and gives a warning if it
converges near the limits. The only place I know of that limits are ignored is Phoenix model engine
when running the NLME QRPEM algorithm, and in summer release (1.4) we will be adding a warning
regarding that.
However you sent in a project that shows that WNL5 classic can ignore boundaries if ONLY the
boundaries are entered, without initial parameter estimates. In your example, you were trying to
restrict the values of the beta parameter between e.g. 0.08 and 0.1, yet the final parameters
reported beta values > 1.
The most straightforward solution is to enter an initial estimate somewhere within the boundary,
and then WNL5 classic will observe this boundary throughout the fitting process. Alternatively it is
very easy to freeze any individual PK parameters in a Phoenix model object and I believe that is the
most suitable solution for your specific problem.
Please let us know if this remains unclear.
Many thanks,
Simon.
--
Simon.Davis.aaa.certara.com
Senior Scientific Consultant
https://www.certara.com/training
|
Want to post a follow-up message on this topic?
If this link does not work with your browser send a follow-up message to PharmPK@lists.ucdenver.edu with "How to test effect of treatment on PK parameters" as the subject |
Copyright 1995-2014 David W. A. Bourne (david@boomer.org)