1 of 157

■

A localization Tale...

Romain Guiet

55 of 157

DiAna

An “easy” concept of Nearest Neighbor

Works on non-blob objects

Randomization is only center to center (so far)

Edge effects

Limitations

■

A localization Tale...

Romain Guiet

56 of 157

DiAna

exercise-1

■

A localization Tale...

Romain Guiet

57 of 157

Distances analysis

Distances

Object

Blobs to Blobs

Ripley’s K function

Blobs/Else to Blobs/Else

DiAna

Lagache et al 2013

■

A localization Tale...

Romain Guiet

58 of 157

Co-Localization Analysis

Object Based

Object

Features

Distances

Intensities

■

A localization Tale...

Romain Guiet

59 of 157

Co-Localization Analysis

Pixels Based

Image Coefficient(s)

Image

■

A localization Tale...

Romain Guiet

60 of 157

Are Tubulin and MYH9 co-localizing?

MYH9 (ABs-647)

Tubulin (ABs-555)

R.GUIET, EPFL

■

A localization Tale...

Romain Guiet

61 of 157

Are Tubulin and MYH9 co-localizing?

MYH9 (ABs-647)

Tubulin (ABs-555)

We can’t segment each protein

Global Analysis

R.GUIET, EPFL

■

A localization Tale...

Romain Guiet

62 of 157

Pearson Correlation Coefficient

R.GUIET, EPFL

Source-C0

Principle

Pearson 1895

Pearson and Lee 1903

■

A localization Tale...

Romain Guiet

63 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Source-C0

■

A localization Tale...

Romain Guiet

64 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Cytofluorogram

Pixel intensities Tubulin channel

Pixel intensities MYH9 channel

■

A localization Tale...

Romain Guiet

65 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Cytofluorogram

Pixel intensities Tubulin channel

Pixel intensities MYH9 channel

Max Pixel Count

■

A localization Tale...

Romain Guiet

66 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Cytofluorogram

Pixel intensities Tubulin channel

Pixel intensities MYH9 channel

Max Pixel Count

■

A localization Tale...

Romain Guiet

67 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

MYH9 (ABs-647)

Tubulin (ABs-555)

R.GUIET, EPFL

■

A localization Tale...

Romain Guiet

68 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

MYH9 (ABs-647)

Tubulin (ABs-555)

This is not enough!

We have a single value

R.GUIET, EPFL

■

A localization Tale...

Romain Guiet

69 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Actin (Phalloidin-488)

MYH9 (ABs-647)

■

A localization Tale...

Romain Guiet

70 of 157

Pearson Correlation Coefficient

Limitations

■

A localization Tale...

Romain Guiet

71 of 157

Pearson Correlation Coefficient

Limitations

■

A localization Tale...

Romain Guiet

72 of 157

Pearson Correlation Coefficient

Limitations

■

A localization Tale...

Romain Guiet

73 of 157

■

A localization Tale...

Romain Guiet

74 of 157

■

A localization Tale...

Romain Guiet

75 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Actin (Phalloidin-488)

MYH9 (ABs-647)

Acceptable Range

■

A localization Tale...

Romain Guiet

76 of 157

Pearson Correlation Coefficient

Are Tubulin and MYH9 co-localized?

R.GUIET, EPFL

MYH9 (ABs-647)

Tubulin (ABs-555)

Actin (Phalloidin-488)

MYH9 (ABs-647)

Acceptable Range

■

A localization Tale...

Romain Guiet

77 of 157

Pearson Correlation Coefficient

It is a Global Analysis:

1.Correlation of pixels intensities in both channels
Correlated (1)
Random (0)
Anti-correlated (-1)
2.Meaningful if signals have similar areas
Otherwise more difficult to interpret
3. Better to compare
Control conditions (WT vs KO, ...)
Another protein staining

■

A localization Tale...

Romain Guiet

78 of 157

Pearson Correlation Coefficient

It is a Global Analysis:

1.Correlation of pixels intensities in both channels
Correlated (1)
Random (0)
Anti-correlated (-1)
2.Meaningful if signals have similar areas
Otherwise more difficult to interpret
3. Better to compare
Control conditions (WT vs KO, ...)
Another protein staining
4. Costes’ randomization

■

A localization Tale...

Romain Guiet

79 of 157

Co-Localization Analysis

Pixels Based

Image Coefficient(s)

Image

■

A localization Tale...

Romain Guiet

80 of 157

Is Protein X in Golgi or Mitochondria?

Golgi

Protein X

Mitochondria

Protein X

■

A localization Tale...

Romain Guiet

81 of 157

Is Protein X in Golgi or Mitochondria?

Manders’ coefficient

Golgi

Protein X

Mitochondria

Protein X

■

A localization Tale...

Romain Guiet

82 of 157

■

A localization Tale...

Romain Guiet

83 of 157

■

A localization Tale...

Romain Guiet

84 of 157

■

A localization Tale...

Romain Guiet

85 of 157

Manders’ coefficients

Golgi

Protein X

Manders’ Coef. Protein X

Manders’ Coef. Golgi

■

A localization Tale...

Romain Guiet

86 of 157

Manders’ coefficients

Golgi

Protein X

0.0

Manders’ Coef. Protein X

Manders’ Coef. Golgi

■

A localization Tale...

Romain Guiet

87 of 157

Manders’ coefficients

Golgi

Protein X

Manders’ Coef. Protein X

Manders’ Coef. Golgi

■

A localization Tale...

Romain Guiet

88 of 157

Manders’ coefficients

Golgi

Protein X

Manders’ Coef. Protein X

Manders’ Coef. Golgi

■

A localization Tale...

Romain Guiet

89 of 157

Manders’ Coef. Results

Golgi

Protein X

Mitochondria

Manders’ Coef. Protein X

Protein X

Manders’ Coef. Golgi/Mito.

■

A localization Tale...

Romain Guiet

90 of 157

Manders’ Coef. Limitations

On the whole image Manders’ coef. are always close to 1

We need to define Thresholds !

Golgi

Protein X

Mitochondria

Protein X

Limitations

■

A localization Tale...

Romain Guiet

91 of 157

Manders’ Coefficients

It is a Global Analysis:

1.Contribution of signal to the colocalization
2.Requires to define threshold values
you need mono staining controls
3. Better to compare
Control conditions (WT vs KO, ...)
Another protein staining

Mitochondria

Golgi

Protein X

■

A localization Tale...

Romain Guiet

92 of 157

Co-Localization Analysis

Conclusion

Segmentable �Objects?

Only Blobs Objects?

Ripley’s K function

Nearest Neighbor

Similar Areas?

Pearson Correlation Coefficient

Manders’ coefficients

YES

Objects: Spatial Analysis

Image:�Global Analysis

■

A localization Tale...

Romain Guiet

93 of 157

Co-localization Analysis

A Case Study

■

A localization Tale...

Romain Guiet

94 of 157

The Pilot Experiment

Confocal Acquisition:

2D Dataset
Sequential
Pixel Size : 76nm
Pinhole matched to 1Airy Unit (A.U.) for each channel
Very little noise in the image
No / very little saturation
Negative staining controls acquired the same way !

■

A localization Tale...

Romain Guiet

95 of 157

The Pilot Experiment

Confocal Acquisition:

2D Dataset

we would like to compare �2D to 3D results

Sequential

optimal to reduce crosstalk/bleedthrough

Pixel Size : 76nm

Match the resolution limit sampling

Pinhole matched to 1Airy Unit (A.U.) for each channel

better to Match Optical Thickness.

Very little noise in the image
No / very little saturation
Negative staining controls acquired the same way !

■

A localization Tale...

Romain Guiet

96 of 157

The Pilot Experiment

Co-localization analysis:

Objects of Interest are sub-resolution�=> Global Analysis

Signals have very different Areas

Pearson alone will not be enough

Manders’ coef.

Need threshold

■

A localization Tale...

Romain Guiet

97 of 157

Defining Thresholds

Manually

Automatically

Costes auto threshold
False Positive hypothesis

■

A localization Tale...

Romain Guiet

98 of 157

Defining Thresholds

Manual Selection

Threshold = 4

Too much Background

Threshold value is too low

■

A localization Tale...

Romain Guiet

99 of 157

Defining Thresholds

Manual Selection

Threshold = 60

Too Few pixels

Threshold value is too high

■

A localization Tale...

Romain Guiet

100 of 157

Defining Thresholds

Manual Selection

100

Threshold = 23

THE RIGHT VALUE!

■

A localization Tale...

Romain Guiet

101 of 157

Defining Thresholds

Manual Selection

You get the result that you want

you decided of the value

Fast

true if you do it badly

101

Pros

Cons

■

A localization Tale...

Romain Guiet

102 of 157

Defining Thresholds

Manual Selection

You get the result that you want

you decided of the value

Fast

true if you do it badly

It’s difficult to explain how you decided of the value

gut feeling = no method

Not reproducible

ask some else to do the same = different results

Slow

doing it seriously will take a lot of time

102

Pros

Cons

■

A localization Tale...

Romain Guiet

103 of 157

Defining Thresholds

Manual Selection

You get the result that you want
you decide of the value

Fast

if you do it badly it’s fast

Difficult to explain how you decide of the value

no method
ask some else = different results

Not reproducible

Slow

doing it seriously will take a lot of time

103

Pros

Cons

DON’T �DO �THAT!

■

A localization Tale...

Romain Guiet

104 of 157

Defining Thresholds

Manually

Automatically

Costes auto threshold
False positive assomption

104

■

A localization Tale...

Romain Guiet

105 of 157

Defining Thresholds

Automatic Selection

make use of your experimental controls (mono-stained samples, … ) to define threshold value with a clear method

assume you want to find a correlation of your signals, and define thresholds values where you lose the correlation

Costes et al 2004

105

Based on the controls

Based on the results

Relies on Biology

Relies on Statistics

■

A localization Tale...

Romain Guiet

106 of 157

Defining Thresholds

Automatic Selection

106

RAW

Biologist Threshold

■

A localization Tale...

Romain Guiet

107 of 157

Defining Thresholds

Automatic Selection

107

RAW

Biologist Threshold

Costes’ AutoThreshold

■

A localization Tale...

Romain Guiet

108 of 157

Defining Thresholds

Automatic Selection

make use of your experimental controls (mono-stained samples, … ) to define threshold value with a clear method

assume you want to find a correlation of your signals, to define thresholds value where you lose the correlation

Costes et al 2004

108

Based on controls

Based on results

Relies on Biology

Relies on Statistics

■

A localization Tale...

Romain Guiet

109 of 157

Defining Thresholds

Automatic Selection

Based on controls

109

Protein X

Golgi

DAPI

■

A localization Tale...

Romain Guiet

110 of 157

Defining Thresholds

Automatic Selection

Based on controls

110

Protein X

Golgi

DAPI

Protein X

....

DAPI

■

A localization Tale...

Romain Guiet

111 of 157

Defining Thresholds

Automatic Selection

Based on controls

111

Protein X

Golgi

DAPI

Same Acquisition Settings

BUT �No staining in this channel

Protein X

....

DAPI

■

A localization Tale...

Romain Guiet

112 of 157

Defining Thresholds

Automatic Selection

Based on controls

112

Protein X

....

DAPI

Same Acquisition Settings

BUT �No staining in this channel

Histogram

■

A localization Tale...

Romain Guiet

113 of 157

Defining Thresholds

Automatic Selection

Based on controls

113

We can now define a “Rule” (method) on the control channel

Accept ZERO False Positive pixels (FP-0): Max= 23 => Threshold=24
Accept a Tolerable amount of False Positive pixels (FP-Tol)

you need to define what is tolerable...

Histogram

■

A localization Tale...

Romain Guiet

114 of 157

Defining Thresholds

Automatic Selection

Based on controls

114

We can now define a “Rule” (method) on the control channel

Accept ZERO False Positive pixels (FP-0): Max= 23 => Threshold=24
Accept a Tolerable amount of False Positive pixels (FP-Tol)

you need to define what is tolerable...

Histogram

■

A localization Tale...

Romain Guiet

115 of 157

Defining Thresholds

Automatic Selection

Based on controls

115

We can now define a “Rule” (method) on the control channel

Accept ZERO False Positive pixels (FP-0)
Accept a Tolerable amount of False Positive pixels (FP-Tol)

you need to define what is tolerable...

Histogram

■

A localization Tale...

Romain Guiet

116 of 157

Defining Thresholds

Automatic Selection

Based on controls

116

False Positive Tolerance (%)	Pixel Number	Threshold value			Average	Median	MAX
False Positive Tolerance (%)	Pixel Number	Control Image 1	Image 2	Image 3	Average	Median	MAX
0.1	1049	9	10	10	10	10	11
0.01	105	11	13	13	12	13	14
0.001	10	13	17	16	15	16	18
0	0	18	22	18	19	18	23

Tolerable amount of False Positive

Image : 1024 x 1024 > 1000 000 pixels

■

A localization Tale...

Romain Guiet

117 of 157

Defining Thresholds

Automatic Selection

Based on controls

117

False Positive Tolerance (%)	Pixel Number	Threshold value			Average	Median	MAX
False Positive Tolerance (%)	Pixel Number	Control Image 1	Image 2	Image 3	Average	Median	MAX
0.1	1049	9	10	10	10	10	11
0.01	105	11	13	13	12	13	14
0.001	10	13	17	16	15	16	18
0	0	18	22	18	19	18	23

Tolerable amount of False Positive

Threshold = 18

Image : 1024 x 1024 > 1000 000 pixels

Threshold = 12

■

A localization Tale...

Romain Guiet

118 of 157

Defining Thresholds

Automatic Selection

Based on controls

118

Tolerable amount of False Positive

False Positive Tolerance (%)	Pixel Number	Threshold value			Average	Median	MAX
False Positive Tolerance (%)	Pixel Number	Control Image 1	Image 2	Image 3	Average	Median	MAX
0.1	1049	9	10	10	10	10	11
0.01	105	11	13	13	12	13	14
0.001	10	13	17	16	15	16	18
0	0	18	22	18	19	18	23

Image : 1024 x 1024 > 1000 000 pixels

■

A localization Tale...

Romain Guiet

119 of 157

Defining Thresholds

Automatic Selection

Based on controls

119

Tolerable amount of False Positive

False Positive Tolerance (%)	Pixel Number	Threshold value			Average	Median	MAX
False Positive Tolerance (%)	Pixel Number	Control Image 1	Image 2	Image 3	Average	Median	MAX
0.1	1049	9	10	10	10	10	11
0.01	105	11	13	13	12	13	14
0.001	10	13	17	16	15	16	18
0	0	18	22	18	19	18	23

Image : 1024 x 1024 > 1000 000 pixels

Thresholds for each channel were defined on the corresponding negative staining control, set with a false positive pixels tolerance of 0.001% (0 or 0.1% or 0.01%), using the average value (or median, or max) from 3 independent control images.

■

A localization Tale...

Romain Guiet

120 of 157

Exercise - 3

15-20min

BIOP JACoP Tutorial

https://go.epfl.ch/2020-biop-jacop-tuto

120

■

A localization Tale...

Romain Guiet

121 of 157

Results

121

Raw�FPTol0.001%

Raw �FP0

■

A localization Tale...

Romain Guiet

122 of 157

Defining Thresholds

Automatic Selection

Based on controls

Explainable

You have a method

Reproducible

Run a script to get the threshold value

Fast

Just run a script

You need controls

but who does an experiment without control?

122

Pros

Cons

■

A localization Tale...

Romain Guiet

123 of 157

The Pilot Experiment

Confocal Acquisition:

2D Dataset

we would like to compare �2D to 3D results

Sequential

optimal to reduce crosstalk/bleedthrough

Pixel Size : 76nm

Match the resolution limit sampling

Pinhole matched to 1Airy Unit (A.U.) for each channel

better to Match Optical Thickness.

Very little noise in the image
No / very little saturation
Negative staining controls acquired the same way !

123

■

A localization Tale...

Romain Guiet

124 of 157

Some Practical Aspects

Sampling

Deconvolution

2D versus 3D analysis

124

■

A localization Tale...

Romain Guiet

125 of 157

Sampling

125

Raw

Downsample by 2

Pixel size = 75 nm

Pixel size = 150 nm

■

A localization Tale...

Romain Guiet

126 of 157

Sampling

126

�FPTol0.001%

� FPTol0.001%

In that case we can acquire twice less pixels (in x AND in y) without changing the results

Raw

Downsample by 2

Pixel size = 75 nm

Pixel size = 150 nm

■

A localization Tale...

Romain Guiet

127 of 157

Sampling

127

�FPTol0.001%

� FPTol0.001%

Raw

Downsample by 2

Pixel size = 75 nm

Pixel size = 150 nm

3D at a lower cost

In that case we can acquire twice less pixels (in x AND in y) without changing the results

■

A localization Tale...

Romain Guiet

128 of 157

The next Experiment

Confocal Acquisition:

3D Dataset
Sequential
Pixel Size : 76nm -> 140/150 nm
Pinhole matched to 1Airy Unit (A.U.) for each channel
Very little noise in the image
No / very little saturation
Negative staining controls acquired the same way !

128

■

A localization Tale...

Romain Guiet

129 of 157

Some Practical Aspects

Sampling

Deconvolution

2D versus 3D analysis

129

■

A localization Tale...

Romain Guiet

130 of 157

Deconvolution

130

Decrease the noise

Increase resolution �(mostly in z)

raw 3D

deconvolved 3D

■

A localization Tale...

Romain Guiet

131 of 157

Defining Thresholds

Automatic Selection

Based on controls

131

RAW

Deconvolved

■

A localization Tale...

Romain Guiet

132 of 157

Defining Thresholds

Automatic Selection

Based on controls

132

RAW

Deconvolved

■

A localization Tale...

Romain Guiet

133 of 157

Defining Thresholds

Automatic Selection

Based on controls

133

RAW

Deconvolved

0.000%	0	9	12	11	11	11	12
0.001%	29	7	9	7	8	7	9
0.010%	288	6	8	5	6	6	8
0.100%	2884	4	6	4	5	4	6

False Positive Tolerance	Pixel Number	Threshold value			Average	Median	MAX
False Positive Tolerance	Pixel Number	Control Image 1	Image 2	Image 3	Average	Median	MAX
0	0	21	24	21	22	21	24
0.001%	29	12	16	12	13	12	16
0.010%	288	9	13	10	11	10	13
0.100%	2884	7	10	7	8	7	10

■

A localization Tale...

Romain Guiet

134 of 157

Raw vs Deconvolved

134

RAW-FP0

Deconvolved-FP0

■

A localization Tale...

Romain Guiet

135 of 157

Raw vs Deconvolved

135

RAW-FPTol

Deconvolved-FPTol

■

A localization Tale...

Romain Guiet

136 of 157

Defining Thresholds

Automatic Selection

Based on controls

Very sensitive to noise

usable with deconvolved datasets

Accept a bit of noise

you can explain it anyway

136

False Positive Zéro

False Positive Tolerance

■

A localization Tale...

Romain Guiet

137 of 157

Some Practical Aspects

Sampling

Deconvolution

2D versus 3D analysis

137

■

A localization Tale...

Romain Guiet

138 of 157

Exercise - 5

5-10min

BIOP JACoP Tutorial

https://go.epfl.ch/2020-biop-jacop-tuto-3D

Use NEW 3D-dataset
Use the script to define thresholds
Compare results obtained if the 3D-stacks are analyzed:

as individual 2D slices
as a whole

138

■

A localization Tale...

Romain Guiet

139 of 157

Compare individual 2D to whole 3D results

139

whole stack = 1 value

1 slice = 1 value

In that case 3D stacks are not required

Could acquire more 2D

■

A localization Tale...

Romain Guiet

140 of 157

The next Experiment

Confocal Acquisition:

2D Dataset

acquire more fields , more cells

Sequential
Pixel Size : 140nm
Pinhole matched to get the same optical thickness
Very little noise in the image
No / very little saturation
Negative staining controls acquired the same way !

140

■

A localization Tale...

Romain Guiet

141 of 157

Co-Localization Analysis

Pixels Based

141

Image Coefficient(s)

Image

■

A localization Tale...

Romain Guiet

142 of 157

Costes’ randomization

142

Principle

Golgi

Protein X

Merge

■

A localization Tale...

Romain Guiet

143 of 157

Costes’ randomization

143

Randomize channel B

Original channel A

Measure PCC

Principle

Golgi

Protein X

Merge

■

A localization Tale...

Romain Guiet

144 of 157

Costes’ randomization

144

...

Randomize channel B

Original channel A

Measure PCC

...

Principle

Golgi

Protein X

Merge

■

A localization Tale...

Romain Guiet

145 of 157

Costes’ randomization

145

Image PCC

Random Images PCCs

Principle

Golgi

Protein X

Merge

Random