Posted by
Daniel James White on
URL: http://confocal-microscopy-list.275.s1.nabble.com/colocalization-analysis-tp786850p838395.html
Search the CONFOCAL archive at
http://listserv.acsu.buffalo.edu/cgi-bin/wa?S1=confocalDear Judy and Julio,
Some more thoughts to add to that...
>
>
>
> There have been a number of papers on colocalization. I just found a
> recent one in Current Protocols in Cell Biology that gives a working
> protocol so to say for doing colocalization analysis with the ImageJ
> colocalization plugins:
>
> Zinchuck and Zinchuk, Quantitative Colocalization Analysis of
> Confocal Fluorescence Microscopy Images. CPCB 4.19.1-4.19.16, June
> 2008.
I would be a little weary here.
The methods they show in this paper have no objective way of setting
the thresholds?
this means you can easily fiddle the results until you get what you
want.
the method of Costes provides a good (but not totally fool proof)
method for objectively and reproducibly setting the thresholds of the
2 channels
using pearsons correlation coefficient to find where there is
and is not correlation.
Alos, if you look at the 2D histograms / scatter plots in the
figures of this paper, most of the data is very clearly intensity
saturated,
meaning most of the higher intensities are crashed onto the highest
value
of the detector (usually 225 for 8 bit scale in confocal).
The gain / laser power was too high,
and the highest intensity information (usually the most interesting!)
has bee LOST!
This makes the all the calculation turn into nonsense
as the saturated numbers arent as high as tyhey should be
compared to the in range data.
This is not good.
>
>
> If my concern were to evaluate the presence of three proteins (or
> other markers) inside a cell, my approach would be to select suitable
> thresholds for each signal, generate a mask for each channel with the
> selected thresholds, and then combine the three masks through Boolean
> or Image Arithmetic operations. This would generate a channel where
> all three markers are present (above set threshold).
This is a good approach, but again there must be a repeatable,
non subjective, math/stats sensible method to set these thresholds,
else you can maniplulate then until you get what you want,
instead of what actually might be there.
>
> My main issue with the standard correlation approach is that it is
> pixel-based, as opposed to object-based, and that the results are
> highly dependent on how the images were collected (sampling) and
> processed, how much noise and background they have, how well
> registered the different channels are (always true, but even more so
> in this case, especially if looking at colocalization of very small
> objects) and which regions and thresholds are used for analysis.
These are very important points.
Noisy confocal images can still be measured for colocalisation,
but they will give lower Manders coefficients and Peasrons
than is really true, as these are both very sensitive to noise.
THere is a way around that, be measuring the Pearsons of 2 consecutive
images of the
same channel of the same object, then using that to back correct the
between channel Pearsons coefficient.
... but Pearsons coefficient is sensitive to different relative
intensities of the 2 channels,
so its a poor measure of colo in many situations.
Manders are not sensitive to this.
> Used
> properly, these methods are quite powerful though, and will give very
> useful information such as colocalization coefficients, % of A
> colocalized, % of B colocalized, etc.... but I have seen so many
> results that didn't make sense just because someone was using a
> poorly sampled image, or a noisy image (such as a confocal image
> under limiting signal conditions), that I am very careful and tend to
> prefer a more visual approach where objects are identified based on
> thresholds, and then analyzed. Anyhow, the paper above discusses
> some of the issues to take into consideration when performing the
> pixel-based type of colocalization analysis.
Here is where I disagree fundamentally.
The way one should approach a problem is to define a hypothesis,
then measure something, and test the hypothesis using appropriate stats.
This data analysis must be blind and objective.
As soon as a researcher, who has a strong idea of what the result
needs to be so they can get their next grant,
is allowed to subjectively manipulate parameters of the analysis
(moving sliders until you see what you wanted to see)
the whole scientific process fall apart.
The brain is a visually biased computer,
and your eyes seldom tell you the truth,
especially when looking at a 2 colour merge image.
We are evolved to find bright coloured fruit
with high contrast against the green of leaves.
We are very bad at asessing quantitatively a whole grey scale range,
as our brains only really pick out very contrasty objects,
and we miss much information.
I think colour merge images are a very bad idea,
as they are misleading for the reader.
Much better to get the reader to understand
what they see in a 2D hisotgram (its just like a FACS data presentation,
which everyone has no trouble understanding... at 2D dot plot if you
like)
Here you get a much better visualisation of the corellation between 2
images
which is not intensity dependent (in the actual visualisation)
For single channel images, I much preffer to use
at the very least grey scalem but much better
a rainbow colour look up table, where each intensity iod a different
bright colour.
Our eyes and brain are much better at interpreting that than greyscale
where the lower intensitites just look like black, and the higher
intensities
all look nearly white, and you cant tell the difference. With
different colours it is much clearer.
No to colour merge images!!!
cheers
Dan
Dr. Daniel James White BSc. (Hons.) PhD
Senior Microscopist / Image Processing and Analysis
Light Microscopy Facility
Max Planck Institute of Molecular Cell Biology and Genetics
Pfotenhauerstrasse 108
01307 DRESDEN
Germany
New Mobile Number!!!
+49 (0)15114966933 (German Mobile)
+49 (0)351 210 2627 (Work phone at MPI-CBG)
+49 (0)351 210 1078 (Fax MPI-CBG LMF)
http://www.bioimagexd.nethttp://www.chalkie.org.uk[hidden email]
(
[hidden email] )