RE: MESF and correcting for isotype matched controls

From: James W. Jacobberger (jwj@po.cwru.edu)
Date: Mon Jan 10 2000 - 15:58:56 EST


I'd like to weigh in on this one because I disagree with most of what has
been said.  Correct me if I'm wrong.

Intuitively, I would expect the variance to increase when reporting
measurements of bound antibody or number of molecules precisely because you
subtract the background.  We do not usually get something-for-nothing; when
maKing two measurements to obtain one calculated value, presumably you have
two independent sources of error, so the calculated value should have more
error associated with it.

I did some quick modelling to see if my intuition might be correst.  By
randomly sampling from two normal distributions with equal CV's and then
subtracting the sampled value of distribution with the lesser mean from a
sampled value from the distribution with the greater mean, a large sample of
such calculated values was created and the mean, CV calculated of this
distribution.  The CV of the subtracted value distribution decreases and
approaches that of the two parent distributions as a function of the
distance between the distribution means.  The decrease fits an equation
something like CV = a * separation distance^b, therefore, when the two means
are close, the error in making the calculation is substantial, supporting
Calmun's observations and my intuition.

However, if you choose to not make the calculation, but instead report out
the number of bound antibodies or molecules of antigen as the fluorescence
of the positively stained sample, you are choosing precision over accuracy
(presumably the reason one would like things in molecules is to get closer
to the truth).  To me, it is more important to make more repeated measures,
get the CV down to something you can live with, and shoot for accuracy.  Why
would anyone bother to use beads that convert fluorescence to numbers of
bound antibodies when (say) half the measurement is not antibodies?  You
would be better off sticking to fluorescence in AU.

in table form: background equivalent to 10 molecules, and variable #of bound
antibodies,

background    10  10  10  10
molecules     10  20  30  40
total         20  30  40  50

if you report out that there are 20 molecules when there are really 10, you
are in error by 100%.  If you measure the background (and I think a
perfectly good way to do that is with an isotype control - if you can't use
a positively stained negative cell equal in size etc. to your positive
population) and subtract it, then you report the truth, 10, 20, 30, 40
molecules - of course subject to errors of making measurements.

If you are using the antibody beads solely to get better interlaboratory and
inter-machine consistency, and you do not subtract the background, then the
instrumental variation that you would like to eliminate is still there.  I
believe you would have accounted for staining variation, though.  In this
case, it would be misleading to report the answer in molecules or bound
antibody, and in my opinion, that practise should be avoided.

James W. Jacobberger, PhD
Associate Professor
Director Flow Cytometry
Case Western Reserve Univ
Cancer Research Center & Dept. Genetics
Ph. 216-368-4645

-----Original Message-----
From: Steve Perfetto [mailto:sperfetto@pasteur.hjf.org]
Sent: Friday, January 07, 2000 2:34 PM
To: cyto-inbox
Subject: Re: MESF and correcting for isotype matched controls



Calman,

I absolutely agree.  The same type of bias can be introduced in subtracting
out
instrument noise since this is essentially instrument specific it would
allow
for huge interlaboratory variation.  For this reason we always report the
raw
values without subtraction based on the Quantum Simply Cellular system.
However,  if a low cell control can be standardized to reference the low end
as
well as a high end cell control than these values should be valid for
subtraction.  Currently such a control does not exist.

Stephen P. Perfetto, MS.,MT. (ASCP)
Walter Reed Army Institute of Research
Department of Molecular Diagnostics and Pathogenesis
1600 East Gude Drive
Rockville, MD. 20850
____________________________________________________________________________
___
Subject: MESF and correcting for isotype matched controls
From:    Calman Prussin <CPRUSSIN@niaid.nih.gov> at Internet_Gateway
Date:    1/6/00  11:42 AM


We are quantitating cell associated antigens using standardized beads and
then generating an MESF from the standard curve generated from the beads.
The question has come up whether or not to "correct" the MESF values
obtained from the specific antibody by subtracting the MESF values obtained
for the  isotype matched controls.

The problem with correcting the values is that it has little effect on the
samples with high density of Ag, but has a large effect on those with low
expression. As such, if we run replicate samples   on the low expressing
samples we find a larger amount of  variation in the corrected MESFs. My
bias is not to use it, as it seems to introduce more "noise" into the
system. Your thoughts?

Thanks!

> _______________________
> Calman Prussin
> Laboratory of Allergic Diseases
> NIAID/ National Institutes of Health
>
Calman,
I, personally, agree entirely with your logic.  We wrestled with the same
issue when we were quantifying TCR's on stimulated T cells (which has now
fallen out of popularity).  I always argued against normalizing against
isotype quantitative values (for the reason you stated), and I used to use
a silly analogy in the attempt to make this logic inescapable:  If you and
I were sitting in the stands at Veterans Stadium at night when the lights
were off and you had your flashlight on, then I accidently turned mine on,
the increase in the intensity of light would be two-fold.  In this case,
you and I are the background, and my turning on my flashlight represents a
(two-fold) fluctuation in background due to noise.  Now, if you again had
only your flashlight on and we were joined by 998 fans, each with their
flashlights on, and then I turned my flashlight on, does the intensity of
light which fills the stadium go up two-fold?  Of course not.  It goes up
by 1/1000th of the intensity before I turned my light on.  But the former
is what one would be implying by normalizing samples with high
fluoresence/MESF to the fluoresence/MESF of background, and thus one
introduces the potential for tremendous artifact.

Now, as you imply, this analogy/argument begins to break down when the
antigen density decreases (say to 3- to 5-fold above background.  My
opinion is that is that if you can show that the variance in the MESF of
the specific stain from sample-to-sample is significantly less than the
variance of the isotype control values, then it would be most appropriate
to NOT normalize against isotype.

These rationalizations make me feel better, I hope they help you also.

AW




Andrew D. Wells, Ph.D.
University of Pennsylvania
Department of Medicine
728 Clinical Research Building
415 Curie Boulevard
Philadelphia, PA  19104
(215) 573-1840 (office)
(215) 898-1951 (lab)
(215) 573-2880 (FAX)
adwells@mail.med.upenn.edu

Calman,

I've spoken out against using isotype "matched" controls many times,
and won't go into that again, except to say, don't use them for
antigen density measurements!  Instead, use a parallel aliquot of
cells that are stained with everything EXCEPT for the color on which
you are doing the measurement (and leave that one unstained).  It is
very important to include the other antibodies, not just isotypes for
them--to fully correct for problems with compensation, spectral
shifts, etc.

Eventually, you should subtract the background value.  But this
assumes that you are only interested in a single value, and single
values are not good for representing distributions (unless the
distribution happens to be normal, or, more precisely, log-normal).
Our approach was to calculate a series of values, including not only
the median (which is far better than the mean when you are using
log-scale), but also the 10th, 25th, 75th, and 90th percentiles of
fluorescence.  This allowed us to build up a picture of the
distribution of fluorescences.  Correcting these values by
background, however, is not simple.  i.e., from the 10th percentile,
do you subtract the median of the unstained?  the 10th percentile of
the unstained?  Once you start considering questions like this, you
realize that subtracting the MESF of the control from the MESF of the
sample has significant pitfalls of its own!

Our approach now is to use the calibration platform in FlowJo--this
platform allows you to rescale any fluorescence channel based on a
standard fluorescence measurement (either by fitting to a bead set or
just by entering the appropriate conversion value).  At this point,
all graphs & statistics are shown in terms of absolute molecules
rather than relative fluorescence.  Thereby you can easily calculate
the above statistics (or show histograms etc) in units which are
meaningful for your final desired answer, and you can then decide
what type of analysis is appropriate.

So, while the simplistic answer to your question is "Yes, subtract
background values", there is an enormous complexity underlying the
process which really requires considerable more thought.

mr



This archive was generated by hypermail 2b29 : Sat Mar 10 2001 - 19:31:03 EST