[RASMB] NONLIN

Borries Demeler demeler at bioc09.v19.uthscsa.edu
Thu May 9 06:19:00 PDT 2002


> 
> Having said that: (i) The use of NONLIN would definitely benefit from a decent manual.
> (ii) I have not compared NONLIN to other NONLIN-like programs, which may, for whatever
> reasons (e.g., the nonlinear regression algorithm used) prove more tractable.
> 
> Time now for the authors of those programs to sales-pitch their product.
> 

Well, here goes my shameless pitch:

UltraScan (www.ultrascan.uthscsa.edu) has a global NLS equilibrium fitter

here are some of the differences compared to Nonlin:

* unlimited number of scans can be included
* automatic (and sensible!) initialization of fitting parameters by use
of iterative linesearch algorithm
* automatic screening of data for inclusion into the fit based on 
quality criteria
* absorbance scans only
* extinction profiles can be loaded (from a global wavelength fitter)
to correct extinction coefficients for different lambdas when scans from
different lambdas are included
* no need to convert between molar and absorbance units, program will
report the answer in 3 ways: association constant - molar units, log
(association constants) - absorbance units, dissociation constant
- molar units
* integrated vbar corrections
* integrated buffer corrections on a per scan basis is possible
* 4 different nonlin algorithms available + one linear algorithm
(Levenberg/Marquardt, modified
Gauss/Newton, Quasi Newton with BFGS and linesearch algorithm, hybrid
method and general linear least squares)
* predefined and user defined models
* data can be edited on the fly during a fitting session
* html reports
* concentration histogram
* association species plot
* MW vs r^2 plot
* ln(C) vs r^2 plot
* MW vs C plot
* single/all residuals plot
* single/all overlay plot (Nonlin probably has something similar, but 
it has been a while since I used it)
* Monte Carlo analysis (Linux Beowulf cluster module available)
* a very detailed manual and tutorials (online help)
* Unix only.
 
The version on the web is over a year old now and we are planning to
release a new version in the fall which will include the following 
additional features:

* centerpiece calibration database
* rotor stretching calibration considered
* interference equilibrium
* SQL relational database support for data (raw and analyzed)
* improved initialization routine
* Monte Carlo for velocity/approach to equilibrium finite element fits
* an automatic differentiation module for finite element fits
* many bug fixes
* for the first time we will support Windows (I wished people wouldn't 
ask for it) and Mac OSX natively. In addition all Linux hardware platforms
(Mac/PPC, Intel, Sparc, Alpha, SGI and anything that will run *BSD Unix)
will be supported, as well as the commercial Unix's (Solaris, OSF, Irix)
* new simulation modules for equilibrium experiments to accurately 
estimate equilibrium time and equilibrium concentration distributions

UltraScan also supports many velocity and simulation methods and includes
many utilities for AUC research. UltraScan is free and can be downloaded
from www.ultrascan.uthscsa.edu

If you think that you need a pre-release/preview of the new version 
(Intel Linux only), please contact me off-line.

about the fitting business:
I agree with everything that has been said on this list, but want
to point out one important additional aspect:

Most of the problems with fitting complicated models such as self-association
and multi-component systems can be traced back to inadequate data. Either
poor quality data, incorrect model, or not enough data. Assuming you have
the correct model and have decent quality data (without which nonlinear
fitting makes little or no sense), the quantity and spectrum of the data
will help to alleviate the problems with cross-correlations, parameters
drifting into infinity, failed fits and so forth. A good initialization
is absolutely essential. I have found that disturbing the "best fit"
by perturbing the parameters often allows you to find a better fit, meaning
that the fit was stuck in a local minimum. A good nonlin algorithm can help
here. Use data from a wide range of concentrations to get *enough signal* from
the dissociated and associated species. I doesn't do you any good to fit
for a monomer when your association is so tight, and your concentration is
so high that you can't see one! The same is true for the associated species.
Use several different speeds to get many descriptions of the same data
with different equilibrium profiles. The more data you include in the fit
the better the fit will get. I usually fit 6 different concentrations from
2 different wavelengths (280 and 230), and 4-5 speeds. That's 24-30 scans
from a big concentration spectrum. It not only improves the fitting 
characteristics but also your statistics. Running a Monte Carlo analysis
of your fit will tell you how reliable your fit really is and I think
that is also very important to do. In my opinion fitting 5-6 scans is
sufficient only for the most simple models like single ideal species.

Enuf said...happy fitting!

-Borries



More information about the RASMB mailing list