Quantitative Imaging for

Colocalization Analysis
Spectroscopy,
not Photography
Daniel J. White
 Topics:

 Images
Images == “Information”
“Information”

(Digital
(Digital Images)
Images)


 Limitations
Limitations of
of microscopy
microscopy hardware
hardware


 Colocalization
Colocalization analysis
analysis

Coloc_2
Coloc_2
What is an Image anyway..?

1st: Go to Basics course slides

Spatial Digitisation
 Experimental Design - First Think...

Quantitative Experiments?
Am I trying to measure the
size/shape of some type of
object(s)

Am I trying to see
movement over time?

Am I trying to measure a
number, amount or
concentration?
Am I trying to measure the number
of some type of object?

Can I define how my objects

appear in images?
Segmentation

Image intensity - threshold

Size - threshold
Shape - circularity etc.
Am I trying to see something
move over time?
Can I define what movement is?
Linear - A to B?
Direction
Speed
Velocity

Rotation

Clustering
 Am I trying to measure an
amount or concentration?
Does that have a Biological
meaning?
Absolute or Relative?
Can I calibrate my image intensity
vs. something else / itself?
eg. Fluorescence signal vs.
Quantitative Assay or
Baseline, Control
Fluorescence response might
not be linear!
 Am I trying to measure an
“image parameter”?
Does that have a Biological
meaning?
Absolute or Relative?
Total / Mean / SD of signal
Background
Signal : Noise
Texture (smooth/spotty)
“Colocalization” between
“colours” / channels”
 Hardware issues
• Optical problems
• Spectral problems
• Intensity Digitisation
• Colour Channels
http://fiji.sc/Colocalization_-
_hardware_setup_and_image_acquisition
 Avoid Emission Bleed Through and
Crosstalk/Cross-excitation
Dye selection / Filter selection
Emission bleed through and/or excitation crosstalk...

Means you get: Overlapping emission - Quantitative? No!

Use multi tracking (Zeiss) / sequential (Olympus)

 Beware ! Crosstalk and Bleed Through
Alexa 488 Alexa 568

Wavelength (nm)

Cross talk (wrong excitation) Bleed through (wrong emission)

Watch Out - More Holes To Fall Into:
Correct objective lens / microscope setup for task
N.A / Resolution.
Apochromat for different colours (UV)
Calibrate Scanner / Check with multi-colour beads
 Check with multi-colour beads
Widefield: (Dvcore 1 micron tetraspek beads):

• Optimise Filter alignment / angle

• Lenses have residual aberrations, even expensive ones.
XZ slice
 Check with multi-colour beads
Confocal (Zeiss 510):

• Calibrate Scanner + Align pinholes (and collimator)

Measure error – then, correct for it!

XYslice
Watch Out - More Holes To Fall Into:
Required bit depth - 8 bit often enough for
LSCM imaging… and colocalization analysis.

More bits only for quantitative experiments where small

intensity differences are measured.

12 bit - bigger files than 8 bit.

(Olympus... 12 bit only. Zeiss 8,12. Leica 8,12,16.)

16 bit file is 2x bigger in RAM / on disk, than 8 bit !

CCD - many cases 12 bit might give better coloc info.

Watch Out - More Holes To Fall Into:
Laser power - don’t bleach area before imaging it.
Bleached sample
Lower signal : noise
Lost information
Set the HV and Offset quickly (Auto HV)
Live imaging, bleaching - big problem
Use low laser power (but more noise)
Colocalization / Correlation

The past:
“I see yellow - therefore there is colocalization”
but published images “look” over exposed.
No colocalization definition + No stats = No Science.

From Now On: 3D. Quantification. Correlation. Statistics.

Complementary methods: BioChemical, Optical (FRET, FLIM)
Colour Merge Images? Only for Art!
Channel Merge Images? What are they good for?
Apart from looking pretty... not much.
Scientific conclusions from the image below?
Colour blind people - see green and red the same!
• Use Magenta / Green or Yellow / Blue
 Colour Merge + Projection = Danger!

Never make colour merge / overlay images from projections

of 3D / z stacks... why not?

Lose 3D info - are the objects overlapping in 3D, or is one in

front of the other one, in the z-stack.

False overlaps!!! Easy to make false interpretation

colour merged projection 3D

What does “Colocalisation” mean anyway…?

That depends who you ask…

… and what BIOLOGY you are thinking about

+ =
 Colocalisation/Correlation?
Think about the biology!

What is the biological/biochemical question?

Are you looking for Co-Compartmentalisation?
Are you looking for exclusion / anti correlation?
Are you looking for interacting molecules?
Then you must also do biochemsitry
(Immuno Co-precip, Fluo Correlation Spectroscopy)
FRET / FLIM might be very informative
Colocalisation / Correlation / Concurrence?
“Colocalisation” covers two qualitatively different conditions:

1) that objects have both

fluorophores present
(Object Based Coloc)
Segmentation needed.
Biology?

2) there is some relationship

between the intensities of
the fluorophores in a pixel.
(Pixel Intensity Based Coloc)
Interaction - BioChemistry?
Colocalization / Correlation / Concurrence?
2 fluorophores are there in a pixel
Binary information

Is it Random?
Is it Real?

Little or no biological meaning?

…unless you are confident about
how to segment objects out from
the background.
 Definition of Terms

“Concurrence” = “co-presence” “there is red and green”

“Colocalisation” = Relationship between channel intensities
Eg. “Red is only found with Green”

Special case - “Correlation”

Intensity Correlation over Space
 Define what is
Colocalization / Correlation?

2 objects overlap
Colocalisation is #1 Binary information
No intensity information

Concurrence?
Image Segmentation!

Biological Meaning?
Colocalisation is #2

Some objects appear to

overlap
with another object
Binary information
No intensity information

Colocalisation?

Biological Meaning?
Colocalisation is: #3

Intensity profiles overlap

Image “Correlation”
pixel Biological Meaning?
intensity Co-compartmentalisation?
Physical interaction?
0
X
Colocalisation/Correlation -Think about:

Are your “objects” smaller than optical resolution?

Vesicles? Small Organelles?
Check channel overlap with sub resolution beads!
Are your objects large?
Large single homogenous blobs?
Large reticular networks / membranes
Resolution required?
Complementary “correlation” methods
Fluorescence correlation spectroscopy (FCS in live cells)
Flow Cytometry? Multiple markers in a cell. Good stats.
 Colour Merge Images = Bad
… so what should I do instead?
“Colocalisation Analysis”
Statistical Significance of Colocalisation
Single image - random / insignificant.
Statistical P value (significance), Manders coefficients, and
Scatter Plot. (ImageJ, BioImageXD, Huygens and others)

But remember…
Don’t merge projections of stacks
(you lose 3D info, false coloc)
Don’t believe your eyes, they lie.
Machines don’t make mistakes…
 Colocalization
Analysis

vs.

How can I measure the amount of colocalisation or rather

“correlation” between these two images?

BioImageXD, ImageJ and others have methods to do that!

 QuickTimeﾪ and a

Colocalization decompressor
are needed to see this picture.

Analysis

Scatter plot
2D histogram
Publish it?

Coloc stats:
Pearsons r
M1, M2,
Costes P-val,

Automatic
thresholding
Coloc Stats - Costes et al. 2004 Biophysical J. vol 86 p3993
 Pearson’s Image Correlation Coefficient
(Manders et al., 1993)

Don’t panic - it’s not that complicated!

Correlation between images, r ranges from -1 to +1

+1 means full correlation (images are the same)
0 means no correlation (random)
-1 means full anti correlation (no red where there is green)
 Pearson’s Image Correlation Coefficient

In English…per pixel and summed for the whole image:

r = +1 r = -1 r=0 r>0
Pearson’s Image Correlation Coefficient is…

Insensitive to diff. intensity of the 2 images. Why?

Insensitive to intensity offset.
If red is 1/2 as bright as green…
Still can get r = 1
… so Pearsons r is is robust for biological imaging…
 Manders' Coefficients
Biologically meaningful
coloc coefficients:

Proportion of each dye

colocalised with the other
(Manders et al., 1993)

Ri,coloc = colocalized red signal

Ri,total = total red signal

Great! … but how do I know which pixels are

colocalized and which are not…?
“Thresholding” and “% colocalisation”
The calculated
“% colocalisation”
depends on what
thresholds you set.

… so how should
one set them?

..until you get the

result you want?

No science here!
 Automatic Thresholding?
How should I set the thresholds of the 2 channels?
Manually? No! Subjective user bias, not reproducible...
Need a robust reproducible method!
Find thresholds where Pearson correlation below thresholds <= 0

Auto Threshold - Costes et al. 2004 Biophysical J. vol 86 p3993

 2D Histograms / Scatterplots
Display 2 colour channel image data in 2D:
colour merge / overlay or 2D histogram?
2D histogram: Ch1 - y axis (left), Ch2 - x axis (bottom)
Colour mapped to number of pixels with that R and G value (right)
A

D
 2D Histograms / Scatterplots
See correlation qualitatively - better than colour merge
See problems from imaging:

Saturated Saturated Wrong offset

Wrong offset
Noisy No correlation? Bleed through
 Automatic Thresholding?
Does it work in a biological experiment? Yes!
Time course of Rev-CRM1 dissociation, nucleolus to nucleus
The dissociation rate constant kd =1.25 ± 0.31 x 10 -3 s-1

auto threshold - Costes et al. 2004 Biophysical J. vol 86 p3993

 One more thing…
Statistical significance!
Are coloc results better than random chance?
A busy image might give high correlation and Manders
Lots of signal = larger chance of random signal overlap.

17 / 40 pixels
overlap !!!
vs.
Is that significant
or just random?

Statistical confidence P - Costes et al. 2004 Biophysical J. vol 86 p3993

 Costes' Method - Randomisation…
Measure Pearson’s correlation for:
Randomised 1st channel image data (PSF sized chunks)
Repeat 100 times
How many randomised have <= correlation than real image.
If > 95% of randomised are worse, then we believe Manders.

P = 0.5 = 50% (no)

P = 0.95 = 95% (yes)
P = 1 = 100% (YES!)
vs.
confidence

Statistical confidence P - Costes et al. 2004 Biophysical J. vol 86 p3993

 Colocalization example: virus entry to caveolae

32% of virus colocalized

10 min P.I. Costes P-value 0.00

0% chance it’s real

39% of virus colocalized

20 min P.I. Costes P-value 1.00

100% chance it’s real

Without significance test, we wrongly assume virus is colocalised

with caveolae at 10 min P.I.
It is not! Only at 20 min is there signficant correlation.
 Examples:
No Correlation?
Pearson r 0.024
M1 0.0354
M2 0.0471

Why high
Thresholds?
 Noisy Saturated Images
Good Correlation?
Pearson r 0.747
M1 0.7291
M2 0.7420

Thresholds
Include
noise?

Badly
Saturated!
 Bad detector settings
Good Correlation?
Pearson r 0.68
M1 0.77
M2 0.63

Offset wrong
+ Saturated

Thresholds
Handle it?
No?
Bleed Through!
DAPI into GFP
 Bad detector settings
Good Correlation? Bleed through?
Bad detector settings…
…gives wrong results!!!
 Software for Colocalization

ImageJ - Colocalization plugins

●
Coloc_2, JACoP, older plugins.

BioImageXD (Coloc Task - Pixel Intensity and Object based methods)

Huygens (RBNCC)

Imaris (Coloc module)

Matlab (J-Y. Tinevez, MPI-CBG / Pasteur)

 Thanks to: MPI-CBG LMF and IPF
Fiji, Heino, Pahajoki,
Kankaanpää, Marjomäki
Uuksalainen, Paavolainen,
TEKES, Tom Kazimiers

Thanks for listening