50 replies to this topic

[Elliot Offman]

Well I will prove your hypothesis, and you have helped me understand the coding process immensely over this weekend.

I've already run the model the other way with the infusion into the central compartment and it does look "ok". I wanted to see if I could do better.

 

regardless, with these TMDD models, are you familiar with the most practical approach for calculating total body clearance? I've got CL from the Vc and also clearance via the TMDD (kon * C * free receptor) and I'm not sure how to convert this over to a total body clearance.

any ideas would be appreciated.

Elliot

[serge guzy]

I am afraid I am not the right person to give you an advise on that one. I am more a Pharmacometrician dealing with modeling and simulation and my clearance is more a model based clearance than a physiologically based clearance. I guess you can look at all the elimination terms in terms of clearance and just sum them up but be careful with the units (do not add apple with oranges).

Best

Serge

[serge guzy]

The addition of the lymph compartment if assuming zero order from it to central and bolus into lymph will not add anything other than numerical difficulty compare to the zero order infusion into central. It is mathematically identical. It was indeed a nice exercise with a lot of thinking.

note that I am giving through Pharsight a 2 days workshop at ACOP as well as at PGAE where at Page I will concentrate on categorical response Population analysis.

best

Serge

[Elliot Offman]

I got it to run but as you predicted it was overparameterized. I wouldn't converge. Lesson here is to go with the simplest model.

[Simon Davis]

Elliot,

  If your browser has renamed it to a Zip file then you can simply rename the extension back to PHXPROJ

 

  Simon.

[Elliot Offman]

Hi again,

I have two questions about the XF vs. F.

 

(1) can you please explain how one would convert an XF theta value to a relative proportion?

(2) In some of the replies, I've seen the expression expressed:

 

(a) (XF/(1+XF)

( (exp(XF/(1+expXF)

 

which is correct and why?

 

thank you,

Elliot

[serge guzy]

First one is if xf is assumed to be normal. Second one if xf is assumed to be lognormal.

Best

Serge

[Elliot Offman]

Thanks. And to convert it to a proportion? what is the relationsip? You previously mentioned that

F=0.5 is the same as XF=0. But say XF =2, what would that be in terms of a prorpotion?

 

Also I verified that the input into the central compartment via zero order is a better model than the lymph input as you had noted it would be, however it wont run as a QRPEM method even with 5 iterations, 3000 samples and MAP Assist selectted.

 

I wonder if you wouldn't mind confirming this code is correct. This is definitely better fit than absorption with a single ka or 2 ka values (AIC and fit plots are superior with the model using FO, but as you've stated before a bad model can have good fit. I'd just like to confirm my model is not the reason the QRPEM is not working.

Thanks

Elliot

:

 

test(){

deriv(A1 = - (Cl * C) + (Aa * Ka)- (Cl2 * (C - C2))-(kon*C*Rf)+(koff*DR))

deriv(DR=(C*kon*Rf)-(koff+kint)*DR)

deriv(Aa = - (Aa * Ka))

deriv(A2 = (Cl2 * (C - C2)))

Rf=Rmax-DR

 

    urinecpt(A0 = (Cl * C))

    C = A1 / V

    dosepoint(A1, tlag = (Tlag2), bioavail = 1-((exp(XF))/(1+exp(XF))), duration = (dur), idosevar = A1Dose, infdosevar = A1InfDose, infratevar = A1InfRate)

    dosepoint(Aa, tlag = (Tlag), bioavail = exp(XF)/(1+exp(XF)), idosevar = AaDose, infdosevar = AaInfDose, infratevar = AaInfRate)

    C2 = A2 / V2

    error(CEps = 0.15157)

    observe(CObs = C * (1 + CEps))

    stparm(V = tvV * exp(nV))

    stparm(Cl = tvCl * exp(nCl))

    stparm(Ka = tvKa * exp(nKa))

    stparm(V2 = tvV2 * exp(nV2))

    stparm(Cl2 = tvCl2 * exp(nCl2))

    stparm(XF = tvXF + nXF)

    stparm(Tlag = tvTlag * exp(nTlag))

    stparm(dur = tvdur * exp(ndur))

    stparm(Tlag2 = tvTlag2 * exp(nTlag2))

    stparm(kon = tvkon * exp(nkon))

    stparm(koff = tvkoff * exp(nkoff))

    stparm(kint = tvkint * exp(nkint))

    stparm(Rmax = tvRmax * exp(nRmax))

 

 

    fixef(tvV = c(, 5716.97, ))

    fixef(tvCl = c(, 4005.89, ))

    fixef(tvKa = c(, 0.0145181, ))

    fixef(tvV2 = c(, 60784.1, ))

    fixef(tvCl2 = c(, 24298.5, ))

    fixef(tvXF = c(, 2.67071, ))

    fixef(tvTlag = c(, 2.67563, ))

    fixef(tvdur = c(, 17.436, ))

    fixef(tvTlag2 = c(, 0.237751, ))

    fixef(tvkon = c(0, 2.00808, ))

    fixef(tvkoff = c(, 0.100393, ))

    fixef(tvkint = c(, 0.699986, ))

    fixef(tvRmax = c(, 5.01539, ))

 

 

    ranef(diag(nV, nCl, nKa, nV2, nCl2, nXF, nTlag, ndur, nTlag2, nkon, nkoff, nkint, nRmax) = c(1.7356216, 0.23072074, 0.1149578, 0.70325565, 0.47666378, 0.0039289403, 3.8734189, 8.3010688E-05, 5.0150691, 1.0005434, 1.0005223, 1.0003276, 0.99989155))

}

[serge guzy]

Dear Elliot

I think we used the relationship exp(XF)/(1+exp(XF))

Therefore XF=9 means F=exp(0)/(1+exp(0))=0.5

 

XF=2, then F=exp(2)/(1+exp(2))=0.88

 

You model is way too much overparametrized. I suggest you to simplify it.

Best

Serge

[Elliot Offman]

Thanks for the advice.

Structurally, this is the best model when run in FO.I removed pretty much all the etas for fixed effect parameters other than V, CL and ka, but it doesn't run under QRPEM and under ELS it stalls after about 40-50 iterations (Which I understand is common with overparameterized models).

 

However the fit just looks awful with other model so its a catch 22 and seems to go against good modeling practice when your best fitting model and your lowerst AIC won't run using the algorithms in the application.

 

Any additional advice on how you might simplify such a model?

[serge guzy]

Dear Elliott

I have indeed an advice for you. I looked again at all your data and in the log-domain. What you can see is:

1: only one dose per patient and the same dose

2: In the log_domain you can not see any sign of non linearity. Non linearity is the response to TMDD based processes that you have but it does not mean that you can see it and this for many reasons. The time range may be not enough to see the non linearity or just the number of molecules going through this non linear type of process may be too small compare to the other portion that does not suffer from saturable based processes. The net result is that only a linear based process model will fit better the data even though the reality is that these non linear processes exist. Therefore, you can start with non compartmental analysis, then you take the average V and Cl as initial estimates for your first model. That model will assume linearity, duration as a random variable for the infusion rate and Ka for the first absorption process. We can see 2 phases and therefore 2 compartment model will be the choice.

Take as in itial estimates V2=V and Cl2 =Cl as you have a good estimates of CL and V.

Start with duration of about 20 based on the observed data.

 

Fit with QRPEM, 3000 iterations, 15 iteration, MAP check box in the advanced option and no standard errors.

 

 

I did it and the project is attached. What we learn is that the processes modeled and the real processes are sometimes disconnected because of statistical based optimization issues and also because the observation can hide a part of the processes.

Let me know what you think.

Look at the last model (edit as textual). [file name=elliot_simplest_model.phxproj size=1769408]http://www.pharsight.com/extranet/media/kunena/attachments/legacy/files/elliot_simplest_model.phxproj[/file]

Attached Files

•  elliot_simplest_model.phxproj   1.69MB   0 downloads