Model estimates based on 1 sample per subject

From: tgordi@buffalo.edu Subject:[NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 2:48 PM Dear all, I have been asked by a friend to estimate the parameters of a simple binding study. Different infusion concentrations were administered to rats and 1 single sample was collected at 30 seconds after the start of he infusion (rats were sacrificed at this time point). There are several "dose" levels in the study. The data was first analysed using Winnonlin and Km, Bmax (maximum binding), and base line were estimated. The estimates look reasonable and in line with previous studies. I tried the same model (simple Hill equation) in NONMEM, running the model first in FOCE and then FO mode. The parameter estimates are very off from the expected range, not even close to the more reasonable Winnonlin estimates. I started with different initial values and the model has problems in getting to the same final estimates. Could the nature of the data, i.e. single sample per subject, explain the poor performance of the run? Is there any remedy for this? regards, Toufigh Gordi

From: GUZY@xoma.com Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 4:27 PM I am surprised that you could evaluate 2 parameters with one single data point using Winnonlin (does not do a population fit). Using Population modeling, I think that the same time point for each rat (30 minutes) can easily make the problem non identifiable. If it is identifiable, you need anyway to fix the intraindividual variance to some value. Serge Guzy President POP-PHARM Head of Preclinical Statistics and Pharmacometrics, Xoma

From: tgordi@buffalo.edu Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 4:37 PM Dear Serge, I am not trying to fit a kinetic model. This is data from a cell binding study. If one plots the incoming infusion (different concentrations) vs. the cell binding data from each rat, a saturable system is observed, where a Hill function can be applied. I don't see the problem of estimating Vmax and Km from such data. There is no time involved in this model. This observation vs. effect. I didn't do the Winnonlin analysis myself. Toufigh

From: lgibiansky@emmes.com Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 5:03 PM Toufigh, I think you will need to provide more info concerning your model (e.g., NONMEM control stream and data sample). Otherwise, it is difficult to discuss the problem without knowing what exactly you are trying to do. In general, NONMEM can fit nonlinear model without difficulties. If you have one sample per animal, you cannot fit population model, you will need to approach this as an "average" pooled data fit. In this case you should not face problems if your data support the model (i.e., describe the full profile). Leonid

From: GUZY@xoma.com Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 5:37 PM Dear Toufigh If I understand, you plot effect versus concentration and fit the data using Winnonlin which gave you reasonable values for Vm and Km. When you used NONMEM, my understanding is that you took into account that each observation came from a different rat and you are trying to retrieve both the fixed effect (Mean values) as well as random effects (variability across the population of rats). Since you have only one data point per rat (am I right?) you must fix the intra-individual variance parameters to a specific value. Did you do that? Serge

From: GUZY@xoma.com Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 5:55 PM If you have a good idea about the assay errors, you can assume fixed intra-individual variance parameters (let say for example 15% CV) and then perform a population fit. I did it many times with our MCPEM (Monte-Carlo Parametric Expectation Maximization) algorithm which optimize the same objective function as NONMEM (FOCE with interaction). If you are interested I could get the data and try our software to retrieve both fixed and random effects. Serge

From: tgordi@buffalo.edu Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/19/2003 6:55 PM I apologize for not having provided the control stream. It is now given at the end of this mail. As you might guess, the V1 and CL values are arbitrary. Beside this model, I also tried one where DADT(1) was set equal to -A(1)*CL/V1. That model gives almost identical result as the one presented here. Thank you! Toufigh $PROB CELL $INPUT ID AMT TIME DV REG CMT $DATA ind.csv IGNORE=# $SUBROUTINE ADVAN6 TRANS1 TOL=5 $MODEL NCOMP=1 COMP=CENT $PK V1=1 CL=0.00001 S1=V1 IMAX=THETA(1)*EXP(ETA(1)) KM=THETA(2) KNS =THETA(3) $DES DADT(1)=0 $ERROR CONC=F LAM=(IMAX*CONC/(KM+CONC))+KNS*CONC Y=LAM*(1+EPS(1)) $THETA (0, 500) ;1 IMAX $THETA (0, 10000) ;2 KM $THETA (0, 0.001) ;3 KNS $OMEGA 0.1 ;1 IIV IMAX $SIGMA 0.1 ;1 PROP ERROR $ESTIMATION NOABORT POSTHOC MAXEVAL=9999 PRINT=3 METHOD=1 MSFO=MSF1 $COVARIANCE $TABLE ID TIME Y NOPRINT ONEHEADER FILE=SDTAB1 $TABLE ID TIME IMAX KM KNS ETA1 NOPRINT ONEHEADER FILE=PATAB1 $TABLE ID REG NOPRINT ONEHEADER FILE=CATAB1

From: GibianskyE@guilfordpharm.com Subject: RE: [NMusers] Model estimates based on 1 sample per subject Date: 8/20/2003 9:26 AM Toufigh, as Leonid and Serge pointed out, with one point per rat you can't fit a population model (i.e have both ETA and EPS in the control stream). Also, to fit an algebraic function you do not need to have a differential equation and a ficticious compartment. You can use $PRED instead (then you do not need $SUBROUTINE, $MODEL, $PK and $ERROR). Katya _______________________________________________________