sequential PK/PD

From: Sriram.Krishnaswami@aventis.com Subject: [NMusers] sequential PK/PD Date: Wed, 1 May 2002 16:10:44 -0500 Dear all, can someone shed light on whether or not retaining the Observed Concentrations (PK) in the data file during sequential fitting of PK/PD data as given below would affect the results in anyway? For example, Step 1: Fit PK alone $PK CL=Theta(1)*exp(eta(1)) V=Theta(2)*exp(eta(2)) Step 2: Take the individual predicted PK parameters from the table output of step 1 and feed them into the data file for PD fitting $PK CL=CLI where CLI is the individual predicted CL V=VI where V is the individual predicted V.. $ERROR Conc=F*(1+err(1)) Eff=Emax*F/(EC50+F)+err(2) Y=Indicator * Conc + (1-indicator) * Eff thanks Sriram ------------------------------------------------ Sriram Krishnaswami, Ph.D. Global Biopharmaceutics Aventis Pharma, NJ

From: Liping Zhang. [mailto:lpz@itsa.ucsf.edu] Sent: Wednesday, May 01, 2002 7:17 PM To: Krishnaswami, Sriram PH/US Cc: nmusers@globomaxnm.com Subject: Re: [NMusers] sequential PK/PD Dear Sriram, can someone shed light on whether or not retaining the Observed Concentrations (PK) in the data file during sequential fitting of PK/PD data as given below would affect the results in anyway? to me, the way to analysis PK/PD data you asked is different with the one you described below. To make matter clear, lets call the one that retains the observed PK data (and population PK parameter estiatmes) for PD fitting Method 1, and the one that uses individual PK parameter estiamtes for PD fitting Method 2. From my experience, if you are using FO, Method 2 is better. If you are using FOCE, Method 1 is better (however it takes longer time than Method 1). Hope this helps. Best, Liping

From: Sriram.Krishnaswami@aventis.com Subject: RE: [NMusers] sequential PK/PD Date: Wed, 1 May 2002 19:37:52 -0500 Liping, thanks for the clarification. However, my question was slightly different. I was only intending to use individual PK parameter estimates for PD fitting and while doing so, I was interested to know if retaining the observed concentrations in the data file would make any difference to the results at all? In other words, when you use individual predicted CL and V from the PK fit from step 1 to fit PD data in step 2, why do you still need PK DV to be present in the data file? or dont we? I am not sure if and how the objective function and other diagnostics etc would be affected by not having the PK DV. hence the question. thanks Sriram

From: "Gobburu, Jogarao V" Subject: RE: [NMusers] sequential PK/PD Date: Thu, 2 May 2002 09:55:20 -0400 Dear Sriram, Unusually not many readers have responded to this question. I would also like to know the answer. I can only report what I heard and this is not first hand experience. In general, sequential PKPD modeling is more prevelant, although there are cases when one might want (or have) to use simultaneous PKPD modeling. I remember a presentation by Mats Karlsson et al, who compared these 2 approaches (1. post-hoc estimates only used to drive PD versus 2. post-hoc estimates + raw PK data driving PD). If my memory serves me right Mats obtained different results for the 2 approaches. I think they found that method#2 was better. The reasons are not obvious to me. But may be Mats can direct us to the reference and/or help us to understand better about these approaches. On the other hand, have you tried both approaches on your specific dataset? If so, what have you found? Regards, Joga Gobburu Team Leader, Pharmacometrics, CDER, FDA.

From: Mats Karlsson Subject: Re: [NMusers] sequential PK/PD Date: Thu, 02 May 2002 16:16:13 +0200 Dear Joga and Sriram, Actually the three approaches that Janet Wade and I compared were: 1) Posthoc PK parameters only to drive the PD 2) Simultaneous PK and PD analysis 3) Simultaneous PK and PD analysis but where the *population* PK parameters were fixed to values found in an analysis of the PK data alone. We found that 2 and 3 behaved equally well under the (limited) circumstances we tried whereas 1 behaved slightly worse. The rational for 3 can be made as follows: We may to fix some PK parameters a) to speed up computation, or b) to assure that misspecification of the PD model doesn't translate into misspecification of the PK parameters. The choice of fixing on the population parameters can be justified because these are often determined with good precision (as opposed to the individual posthoc parameters, which in addition are biased with sparse data). Surprisingly to many (incl. us in the beginning), the PK data will influence the population PD parameters for method 3, even if the FO method is used. The reason is that all data are included in the calculation of the objective function value. The work was presented at PAGE, Saintes, 1999 Best regards, Mats -- Mats Karlsson, PhD Professor of Pharmacometrics Div. of Pharmacokinetics and Drug Therapy Dept. of Pharmaceutical Biosciences Faculty of Pharmacy Uppsala University Box 591 SE-751 24 Uppsala Sweden phone +46 18 471 4105 fax +46 18 471 4003 mats.karlsson@farmbio.uu.se

From: Leonid Gibiansky Subject: RE: [NMusers] sequential PK/PD Date: Thu, 02 May 2002 11:01:56 -0400 Sriram, I think concentration DV should not influence your PK/PD model if you use F in Emax model expression. However, the objective function will differ with and without those observations simply because you have err1 term that is not fixed, but chosen by NONMEM based on the concentration DV. With you control stream, objective function is the combination of the PK fit OF and PD fit OF with the PD model is not connected with the PK model in any way. Effectively, you are fitting two models in step 2: PK/PD model, and error model for the PK part. These two models are not connected. Since PK model does no change when you change parameters of the PD model, the results (parameters of the PD model) should be identical with and without PK observations, although run time should be different. Have you observed this in your experiments ? There may be different opinions and situations how to choose between simultaneous and sequential PK/PD modeling. But if you prefer sequential modeling as described in your message, I would suggest to exclude all the concentration DV and modify step 2: >$ERROR >Y=Emax*F/(EC50+F)+err(1) (Y here is PD observation) Leonid

From: Nelamangala V Nagaraja Subject: Re: [NMusers] sequential PK/PD Date: Thu, 02 May 2002 11:15:56 -0400 Hello all, I think the results from 1. fixing population parameters (THETA, ETA, and SIGMA) and doing a simultaneous PK-PD fitting and 2. fixing the individual THETAs (feeding through the dataset) and doing PD fitting will be different. What will be the criteria for the comparison? OBJ values may not be suitable. In Method 1, although we have fixed the POPULATION parameters, is there a chance that errors due to PD model misspecification will leak into PK model because the program is generating the individual PK parameters based on the Population values. This problem may not be there in Method 2. Any suggestions please. Thanks Raj

From: Mats Karlsson Subject: Re: [NMusers] sequential PK/PD Date: Thu, 02 May 2002 17:43:41 +0200 Dear Raj, I assume that method 1, corresponds to method 3 in my notation and your method 2 corresponds to my method 1 (just to keep things clear :-)). If so, I would agree that OFV is not a suitable method of comparison. If I took the trouble of doing things more than one way, I would look for consistency in PD parameter estimates between methods. If I wouldn't get that, I would be worried and look into it more closely. I would assume that there is a explanation, probably rather case specific. Best regards, Mats -- Mats Karlsson, PhD Professor of Pharmacometrics Div. of Pharmacokinetics and Drug Therapy Dept. of Pharmaceutical Biosciences Faculty of Pharmacy Uppsala University Box 591 SE-751 24 Uppsala Sweden phone +46 18 471 4105 fax +46 18 471 4003 mats.karlsson@farmbio.uu.se

From: "Lewis B. Sheiner" Subject: Re: [NMusers] sequential PK/PD Date: Thu, 02 May 2002 10:44:05 -0700 I had hoped that Liping would clarify this more than she did. As Mats discussed, there are a number of ways to do sequential vs simultaneous fitting. Liping and I have been studying these, following up on Mats' and Janet Wade's work. The short answer is that Mat's method #3, whereby one fits the PK data alone, then fits ALL the data to the PKPD model with all pop PK parameters fixed to values obtained in the previous PK fit, generally saves computation time relative to the simultaneous method (Mats' #2), and produces results that are just as good. Mats' method #1, whereby one the fits PK data alone, generates Post-hoc PK parameters, and uses these for the second step PD model fit using only the PD data in the second step, saves more time, and is almost as good. There is no point to including the PK data in the last step of here since they will not influence the PD fit at all. LBS. -- _/ _/ _/_/ _/_/_/ _/_/_/ Lewis B Sheiner, MD (lewis@c255.ucsf.edu) _/ _/ _/ _/_ _/_/ Professor: Lab. Med., Bioph. Sci., Med. _/ _/ _/ _/ _/ Box 0626, UCSF, SF, CA, 94143-0626 _/_/ _/_/ _/_/_/ _/ 415-476-1965 (v), 415-476-2796 (fax)

From: Liping Zhang. Subject: Re: [NMusers] sequential PK/PD Date: Wed, 8 May 2002 10:50:52 -0400 Hi, Raj, > Hello all, > I think the results from > 1. fixing population parameters (THETA, ETA, and SIGMA) and doing a simultaneous > PK-PD fitting and > 2. fixing the individual THETAs (feeding through the dataset) and doing PD fitting > will be different. What will be the criteria for the comparison? OBJ values may not > be suitable. Mats anwered your first question. > In Method 1, although we have fixed the POPULATION parameters, is there a chance > that errors due to PD model misspecification will leak into PK model because the > program is generating the individual PK parameters based on the Population values. > This problem may not be there in Method 2. I guess your interest is in getting PD models and parameter estimates since you are doing PK/PD analysis. By doing a simulation study, we have found that when the PD model is misspecified, your method 1 and method 2 behavior very similarly measured in respect to their prediciton performance, since the PD fits are still dominated by the PD model. If in method 1 there are errors in generating individual PK parameter due to PD model misspecification, the influnece in PD fits must be very minimal, if any. -- _/ _/ _/_/ _/_/_/ _/_/_/ Liping Zhang (liping@c255.ucsf.edu) _/ _/ _/ _/_ _/_/ Graduate Student in Dr. Lewis Sheiner's group _/ _/ _/ _/ _/ Biological and Medical Informatics Program _/_/ _/_/ _/_/_/ _/ Box 0626, UCSF, SF, CA, 94143-0626 415-502-1989 (v), 415-476-2796 (fax)