[RASMB] NONLIN

[John Correia]

Good to see the revival of summer (winter downunder) fun - & worth
following up on - 

1st - stating guesses matter in NLLS is hardy a new reality & in fact
should be true of any nonlinear fitting - it just gets worse with global
fitting - using Nonlin to fit binding data will not be as difficult
until you go global, ie more than one data set under different
protein/ligand ratios for example.

2nd - the program is not the problem & new options don't make
exponential fitting any easier - exponential fitting of relaxation data
(the 1st published version of Nonlin was Johnson & Schuster for this
purpose)  is just as difficult, especially if the relaxation times are
not well separated - the advantage of ultra data is you know (hope?) the
sigma values are multiples of monomer MW (modified by vbar changes but
we'll skip that for now).

3rd - global is required to analyze ultra data and you can only trust
fits of global data - I often fit individual data sets to see the
convergence & parameter variance properties but I recently find
individual K's just as useful, but only once I have a "good" fit.  The
best & simplest example of this is monomers + disulfide crosslinked
dimers - you can fit a single channel to a monomer dimer & get a K2 vaue
& random residuals, but at different loading conc the K's values float
down with increasing loads.  Global fits fail and indiviual K2 fits show
the downward trend in MWs consistent with heterogeneity.

4th - this simply example is a reflection of the power of programs like
Omega or the original Roark & Yphantis program Biospin - different data
converted to MW by sliding slopes analysis do not superimpose when
plotted vs conc.  This means MW is not a single valued function of conc,
ie its not reversible - its gets to be "fun" when is doesn't superimpose
because of reversible reactions with heterogeneity or dead protein in
the background.   FACT:  if Omega function gives MW curves that
superimpose and can be fit to a reversible association model than the
same raw data can be fit with Nonlin.  Mathematically impossible
otherwise - someone can probably derive a proof for us.  Fact#2:  if you
change the model all parameters change including concentrations
(activities if you prefer) - its still the math..

5th - I started to describe the use of Biospin to pre-interpret the
goodness of a data set yesterday but decided since it only exists as a
DOS program & besides myself I don't know who is using Biopsin it
wouldn't help - furthermore I usually these days can decide the same
thing from failed attempts at global fitting with Nonlin so that is what
I do.  

Finally - But I agree with the suggestion Nonlin is hard to use if you
don't have a good data set & maybe therein lies the problem - data sets
you can play with if you're willing (this can't be a new idea? & I bet
data sets are on the RASMB already? or on the NAUF CD's already?) .  In
my XLA life I have produced enough good cases of data that I am
confident it works for monomers (usually boring but better than
aggregated monomers for the biology side) and 1-2-4-8 and 1-2-3 va 1-3
and tight dimers where K2 is less than 5 nM and thus fits best to a
single dimer species, etc etc etc - but my 1st attempt with the XLA was
an aggregated sample that had 1-2-4-8 in the background and I could fit
some sets & some channels but not all & K's floated all over the place &
I struggled for a month (yes I'm foolish & persistent) trying to make
sense & the solution was toss it all in the garbage & "make" my
collaborator redo the prep and check/verify activity and like magic good
data & a nice story.  (I wish that was my only bad data set & these days
I decide to trash the data a lot more quickly.)

Oh Ya:  I have never had a sample that can stand up to 4-5 speeds &
many of our samples require short column (which you have to fit to
single species, complex models with great caution) to get it over with
quickly & thus why we do so much quantitative velocity work instead &
especially before I waste time on an equilibrium run.  But if you are
doing 4-5 speeds & multiple waves I hope you are studying hetero-cases
because they need more data & its a waste of spin time to do monomers?
or simple monomer-dimer cases that way & more speeds and more scans will
never make the cross correlations get better - only better samples can
minimally do that - but who wants to bring up biology at a time like
this.......






-------------------------------------------------------------------
 Dr. John J. "Jack" Correia
 Department of Biochemistry
 University of Mississippi Medical Center
 2500 North State Street
 Jackson, MS  39216
 (601) 984-1522                                 
 fax (601) 984-1501                             
 email address: jcorreia at biochem.umsmed.edu     
 homepage location: http://biochemistry.umc.edu/correia.html  
 dept homepage location:    http://biochemistry.umc.edu
-------------------------------------------------------------------