Stanley's ICA, Intensity Correlation Analysis; ICQ, Intensity Correlation Quotient

[EliseStanley]

Hi all,

This is my first post on the ImageJ forum.  I joined because I just discovered that there have been a number of queries on double-fluorescence staining analysis using the Intensity Correlation Analysis (ICA) method.  This method seems to have proved useful in many studies and has been cited in over 300 papers since its inception (including many high-profile reports).  Curiously, in all this time and despite its wide-ranging use its been very rare that I have been contacted to discuss or present the method itself.  Thus, on its 10th anniversary I thought it might be timely to come out of hiding!

ICA was first reported, discussed and evaluated in a paper in Journal of Neuroscience back in 2004:

[Li, Q.*, Lau, A.* Morris, T.J.*, Guo, L., Fordyce, C. and Stanley, E.F.#  A syntaxin 1, Gαo, and N-type calcium channel complex at a presynaptic nerve terminal: Analysis by quantitative immunocolocalization. J. Neurosci., 24 (2004) 4070-4081.  *equal contribution; #corresponding author.  

The authors were technicians (Q Li; L Guo) or trainees in my (A. Lau, T.J. Morris) or a colleagues lab (C.Fordyce).  These co-authors did an outstanding job generating the immunostaining and other data that form the basis of the paper that I applied the ICA method to.  However, as my coauthors would confirm, the method itself, its rationalization and test and its discussion were entirely my work.  

As conceived ICA was effective - but laborious since all the computations had to be done serially on a calculator.  Because of this I asked Tony Morris, then the manger of the Wright Cellular Imaging facility at our facility (Toronto Western Research Institute) and an ImageJ expert (I am sure many of you know of his contribiutions) if he could automate the analysis.  The result was the ImageJ plugin which was first introduced in a paper that we coauthored:
[Khanna, R., *Li, Q., Li Sun, Collins, T.J. and Stanley, E.F.#  N type Ca channels and RIM scaffold protein covary at the presynaptic transmitter release face but are components of independent protein complexes.  Neurosci., 140 (2006) 1201-8. *Equal contribution.]

Thus, while I conceived of the method I think we are all indebted to Tony for making it generally available.  He also cleverly and independently included other costaining indices in the results outcome for the convenience of the user.

A couple of notes:  
First, ICA/ICQ analysis does NOT test for colocalization (despite the reference in the title of the paper).  The best way to illustrate this is to use a test area that is maximally stained by both fluorophores.  The resulting ICQ value is 0.  What
ICA/ICQ test for is covariance - if when one stain goes up the other varies in synchrony.  In a sense, ICA tests for protein complexes in situ (see the Khanna et all paper referred to above).  

Second, I do not consider it valid to use Pearsons for analysis of immunostaining covariance. For a discussion please see the original paper.  The reason that for immunostaining the opposite of covariance is NOT inverse variance (as it is for Pearson's) but RANDOM staining (a negative correlation is almost unheard of in biology - one example might be two types of subunits that contribute to a four stave ion channel).  Indeed, this was one of the main motivations for the invention of ICA.

Third, a tool is only as good as its wielder.  Please be careful about the selection of areas to apply ICA analysis to.  The area should be one within which it is POSSIBLE for the proteins to be random - such as cytoplasm of a cell.  tireless work not only this plug-in but all of the amazing ImageJ platform would not be possible.

I am of course very pleased with the general acceptance of this method.  I supported its release on ImageJ (rather than selling rights) to make sure it would be generally available to the scientific community.  It would be nice to see the method also on commercial platforms - but that is of course outside my control.

Finally, I wish to give my Big Thanks to Wayne Rasband - my ex-colleague at NINDS - without who's
tireless dedication the ImageJ platform would obviously not exist.

Elise F. Stanley PhD
Senior Scientist
Toronto Western Research Institute




 
 
 

[EliseStanley]

I see I told a lie!!  I searched the forum and found that yes, indeed, I have been here long ago to answer an ICA question but under a forgotten user name.  Sorry for the double identity...
ee

[Jeremy Adler]

I have some doubts about the ICQ method for measuring correlation.

Mathematically it is basically a binary version of the Pearson correlation coefficient, but with a different scale (-.5 to +.5) instead of the conventional -1 to  +1, simply doubling the ICQ scale to align the scales has been suggested.

The ICQ is binary in that the only considerations are whether each intensity is above or below its mean and then whether the two fluorophores are both above or both below their mean. In correlation I can see a case for ranking the intensity values, which turns the Pearson into the Spearman rank coefficient, but I cannot see the argument for totally disregarding the actual intensities - we work hard to get them and they are meaningful.

It’s finally down to the weighting given to different combinations of intensities. In the Pearson and Spearman correlations pairs of intensities that are strongly high or low are given substantial weight while the ICQ just regards them as matched.  So the ICQ gives equal weight to (i) a pair of intensities just above or below the mean, (ii) to a high intensity combined with a just above the mean intensity and (iii) to two high intensities – which seems to be a poor way of calculating correlation.  Note that marginal changes to an intensity around the mean can suddenly remove  the whole weight– and that with normal distributions the data is clustered around the mean, so a little noise can easily move the combination of intensities into a different category and affect the ICQ but not the Pearson, or at least to a much smaller extent.

The ICA is a way of replotting the classic scattergram, on pure Pearson correlation lines – using the numerator of the Pearson.  So what the ICA shows and the ICQ measures, remember that ICQ binarises the intensities that make the ICA work, are disconnected.

So I suggest sticking with the Pearson and rank Spearman for correlation analysis, they have a long history. If you really want to go binary, try Kendall’s tau.

-----Original Message-----
From: ImageJ Interest Group [mailto:[hidden email]] On Behalf Of EliseStanley
Sent: den 11 mars 2014 21:02
To: [hidden email]
Subject: Stanley's ICA, Intensity Correlation Analysis; ICQ, Intensity Correlation Quotient

Hi all,

This is my first post on the ImageJ forum.  I joined because I just discovered that there have been a number of queries on double-fluorescence staining analysis using the Intensity Correlation Analysis (ICA) method.
This method seems to have proved useful in many studies and has been cited in over 300 papers since its inception (including many high-profile reports).  Curiously, in all this time and despite its wide-ranging use its been very rare that I have been contacted to discuss or present the method itself.  Thus, on its 10th anniversary I thought it might be timely to come out of hiding!

ICA was first reported, discussed and evaluated in a paper in Journal of Neuroscience back in 2004:

[Li, Q.*, Lau, A.* Morris, T.J.*, Guo, L., Fordyce, C. and Stanley, E.F.#  A syntaxin 1, Gαo, and N-type calcium channel complex at a presynaptic nerve
terminal: Analysis by quantitative immunocolocalization. J. Neurosci., 24
(2004) 4070-4081.  *equal contribution; #corresponding author.  

The authors were technicians (Q Li; L Guo) or trainees in my (A. Lau, T.J.
Morris) or a colleagues lab (C.Fordyce).  These co-authors did an outstanding job generating the immunostaining and other data that form the basis of the paper that I applied the ICA method to.  However, as my coauthors would confirm, the method itself, its rationalization and test and
its discussion were entirely my work.  

As conceived ICA was effective - but laborious since all the computations had to be done serially on a calculator.  Because of this I asked Tony Morris, then the manger of the Wright Cellular Imaging facility at our facility (Toronto Western Research Institute) and an ImageJ expert (I am sure many of you know of his contribiutions) if he could automate the analysis.  The result was the ImageJ plugin which was first introduced in a paper that we coauthored:
[Khanna, R., *Li, Q., Li Sun, Collins, T.J. and Stanley, E.F.#  N type Ca channels and RIM scaffold protein covary at the presynaptic transmitter release face but are components of independent protein complexes.
Neurosci., 140 (2006) 1201-8. *Equal contribution.]

Thus, while I conceived of the method I think we are all indebted to Tony for making it generally available.  He also cleverly and independently included other costaining indices in the results outcome for the convenience of the user.

A couple of notes:  
First, ICA/ICQ analysis does NOT test for colocalization (despite the reference in the title of the paper).  The best way to illustrate this is to use a test area that is maximally stained by both fluorophores.  The resulting ICQ value is 0.  What ICA/ICQ test for is covariance - if when one stain goes up the other varies in synchrony.  In a sense, ICA tests for protein complexes in situ (see the
Khanna et all paper referred to above).  

Second, I do not consider it valid to use Pearsons for analysis of immunostaining covariance. For a discussion please see the original paper.
The reason that for immunostaining the opposite of covariance is NOT inverse variance (as it is for Pearson's) but RANDOM staining (a negative correlation is almost unheard of in biology - one example might be two types of subunits that contribute to a four stave ion channel).  Indeed, this was one of the main motivations for the invention of ICA.

Third, a tool is only as good as its wielder.  Please be careful about the selection of areas to apply ICA analysis to.  The area should be one within which it is POSSIBLE for the proteins to be random - such as cytoplasm of a cell.  tireless work not only this plug-in but all of the amazing ImageJ platform would not be possible.

I am of course very pleased with the general acceptance of this method.  I supported its release on ImageJ (rather than selling rights) to make sure it would be generally available to the scientific community.  It would be nice to see the method also on commercial platforms - but that is of course outside my control.

Finally, I wish to give my Big Thanks to Wayne Rasband - my ex-colleague at NINDS - without who's tireless dedication the ImageJ platform would obviously not exist.

Elise F. Stanley PhD
Senior Scientist
Toronto Western Research Institute




 
 
 



--
View this message in context: http://imagej.1557.x6.nabble.com/Stanley-s-ICA-Intensity-Correlation-Analysis-ICQ-Intensity-Correlation-Quotient-tp5006871.html
Sent from the ImageJ mailing list archive at Nabble.com.

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

[EliseStanley]

Hi Jeremy,

Everyone is looking for a single value to declare whether two proteins associate. The sad truth is that this is almost always not possible because even proteins proteins that associate reliably - such as myosin and actin, say, may do not do so in some regions of the cell - such as during synthesis and transport.  Worse, some proteins are generally in different locations and only associate transiently - at the point where they are of most interest.  An example would be a protein on a secretory granule that is involved with exocytosis.

Thus, the obsession with single-value association units is fraught with problems from the start.

Intensity correlation analysis was invented to permit the investigator to detect where proteins vary in synchrony in the cell - and for that the scalar, not the discrete (yes or no staining) value is used.  The result is a plot within which it is possible to detect quantitative correlations of immunostaining and it is also possible (though most analysis methods do not yet take advantage of this) to go back to the original image and identify where in the cell the two proteins associate and where they do not.  

When I invented that method I realized that elegant as the ICA method is, statistical comparisons between treatments required a way to get a single-value estimate of whether the two proteins associate.  That's the origin of the ICQ - it really throws away 99% of the (spatial) information in the ICA analysis for the sake of a simple statistically testable single value output.  And yes, this is not perfect!  I suppose we could have a subroutine that allowed you to calculate the ICQ values in select regions of the cell - I think that would be amazing.  I had hoped that a commercial software vendor would have picked up on what I think is an astonishing potential of ICA analysis in two or three dimensions.  Unfortunately, that has not happened.  

I hope you have looked at the original paper in which most of these issues are discussed.  But ICA analysis and the ICQ value are unrelated to Pearsons - and I maintain that the latter is invalid because of how it handles random staining vs negative-correlated staining.  Further, it is not possible to pull out covariance within subregions of your sample.

ee

[dscho]

Hi Jeremy,

On Sun, 16 Mar 2014, Jeremy Adler wrote:

> So I suggest sticking with the Pearson and rank Spearman for correlation
> analysis, they have a long history. If you really want to go binary, try
> Kendall’s tau.

For your interest: I am happy to report that Coloc 2 recently learnt to
calculate Kendall's Tau.

Ciao,
Johannes

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html