MW distribution test/calculations for Dextran 70 (USP)

[Alfred88]

Dear Chromatographers:

We have got a new project: to perform the MW distribution test for Dextran 70 per USP (using SEC).

Commercial off the shelf (COTS) software packages available can do the calculations if there are narrow standards with known MW; or if the viscosity info is available for standards (Mark-Houwink approx.).
However, COTS packages cannot do the calculations for the top 10% or bottom 10% of the MW distribution, as required by USP.

USP states that an appropriate method to calculate the constants b1 thru b5 is the Gauss-Newton (modified by Hartley), or Nilsson-Nilsson method. I have got the article by Nilssons, and found that the authors referred to the Hartley modification of the Gauss-Newton method.
There were no sample calculation data provided in the article, nor chromatogram, and the number of iterations done was not listed. Info on the Gauss-Newton can be found in Wikipedia:
http://en.wikipedia.org/wiki/Gauss-Newton


We have contacted tech dept. of USP, and asked referral for a contract lab capable of doing this test, but USP declined to provide referral.

Do you know a contract lab that can do this test? Or anyone can function as a consultant? Or can someone provide a reference to a better/latest reference/article?

I realize the potential problem when I did some background reading. The initial selected values in a Gauss-Newton method may lead to non-convergent series. Also, it may be impossible to achieve the "best-fit" curve (Gauss-Newton method was designed to achieve a "good-enough-fit").

This project is rather urgent for us, because we need to certify a supplier. Thanks all for your input.

Alfred.

[ivanvins]

You can try our Clarity software with the GPC/SEC extension module. The MW calculations include a ranges calculations as for the USP requirements. The calibration options support besides a narrow standards also the multiple broad linear or broad integral standard as well as broad on narrow calibration. The software is available also in Offline version to process chromatograms imported as AIA or ASCII from other CDS.

Please contact me for further details or download a demo version or manuals from our web pages

Ivan Vins
DataApex
www.dataapex.com

[Alfred88]

Dear chromatographers:
I realized that the calculation for MW distribution per USP is not an exciting topic. In fact, I could not find a lab in the US to contract out this test.
I have completed the data acquisitions for several Dextran 70 USP samples per USP. I have not successful in the calculation yet, but I have learned a few new things.

So far I have tried Clarity demo software (DataApex.com), but it seems that the software cannot do the fitting per USP, of the form W= A+exp(B+C*T+ D*T^2+E*T^3). The interface looks very good though.
I have installed the Scilab package (http://www.scilab.org), and wrote *.sce scripts to run the minimization (optimization). My compiled programs always crashed, because the series were not converging.

My next step may be to try to solve this problem using Matlab (similar to Scilab), or a demo version of WinNonLin (I just found out about this powerful fitting package, which was approved by FDA to do fitting in PD/PK-Pharmacodynamics/Pharmacokinetics).
I would like to give up, but I love the challenge.
Thanks all for reading.

Alfred

[aceto_81]

Maybe you should first consider R (www.r-project.org), before purchasing some programs.

With the optim function, you can optimise through minimisation.
I'm not familiar with the Dextran USP monograph, but if you can give me a clear view on how things should be done, I can provide you an R script which will do the job.

If you want you can mail me at: aceto81_at_gmail.com


Ace

[Alfred88]

Dear aceto_81:
Thanks a lot for your goodwill. I appreciate greatly your offer.

I have installed the R (version 2.7.1 for Windows) package last week, and have used it as "evaluation" to compare against Scilab. I believe that it can run faster than Scilab, or Gaussian.

After several weeks of failures (and sleepless nights), I have a moment of triumph. I had to try so many combinations of "initial conditions." The series (fitted curve) finally converged, and the results from R software are below:

> out
$par
[1] -277.97420 16.69112 -17.42709 20.64162 -13.70928

$value
[1] 5.322548e-05

$counts
function gradient
501 NA

$convergence
[1] 1

$message
NULL

$hessian
[,1] [,2] [,3] [,4] [,5]
[1,] 6.188192e-05 0.02911048 0.02904458 0.02902762 0.02904272
[2,] 2.911048e-02 23.40917645 19.37226960 17.34406006 16.18140716
[3,] 2.904458e-02 19.37226960 17.28037756 16.11816014 15.40423942
[4,] 2.902762e-02 17.34406006 16.11816014 15.38727230 14.91744263
[5,] 2.904272e-02 16.18140716 15.40423942 14.91744263 14.59524239

Note that the residual sum square is very small (5.322548e-05).

Thanks for reading and sharing your failures and successes. Thank you for the developers of the R package.

Alfred

[aceto_81]

Hi Alfred,

seems like you found the optim function
The optim function use Nelder-Mead as default, instead of Gauss-Newton.
I don't think this is a problem for you, as the USP states "a suitable method".
But in case you want to try Gauss-Newton, you can use the nls function.
(tip: with your optim$par as starting values). But I must confess that I used nls with mixed success. Optim has never let me down....

Also it seems like you reached the maximum iterations (500 for Nelder-Mead), so maybe it's possible to find a "better" fit when setting maxit higher, or use the found parameters as starting values for a next optim.

If you have more questions, don't have sleepless nights, just ask.

Ace