[NEUBIAS Academy@Home] Webinar “Deconstructing co-localisation workflows: A journey into the black boxes” + Questions & Answers

Dear all,

On September, the 29th, was given the NEUBIAS Academy webinar “Deconstructing co-localisation workflows: A journey into the black boxes”.

Speaker: @Fabrice_Cordelieres

Moderators: @aklemm, @RoccoDAntuono, @romainGuiet, @lagache, @MarionLouveaux

Video of the webinar: on the Youtube channel of NEUBIAS.

Slides: on Fabrice’s GitHub repository.

More than 400 persons attended the webinar. We are adding to this thread all the Questions & Answers collected during the webinar. Answers given by moderators are preceeded by @ references, additional information/discussion by @Fabrice_Cordelieres are given in italics.

Enjoy and post any missing question here!

Table of contents, Part 1

  • Checking data integrity
  • Pre-processing

  • Checking data integrity

    Q1: Can checks for bleethrough be done with control samples? sample only with one dye and sample with the other dye? Do you mean I should have one dye in one sample and then I should compare?

    Indeed ! It all depends on what you call “control samples”. By control, I mean the same type of sample as the one on which you are quantifying co-localization (sample 1), except you will have to mono-label them (samples 2). You may also want to have a non labelled sample (sample 3). First you’ll image sample 1, setting what seems to be appropriate imaging conditions. Then you’ll have to image samples 2 under same acquisition conditions. If everything goes right, you won’t see any noticeable cross-talk/bleethrough (bascically, no signal in other channels than in the one where you expect to see signal). In case you detect signal where it is not supposed to appear, you’ll have to tune your acquisition parameters or change the way you prepare the sample, for instance opting for fluorophores that are more spectrally distant. However, keep in mind that when shifting towards the far red part of the spectra, the resolution will be lowered…

    You may ask: “what about sample 3 ?”. Good question ! This sample will allow you to check that the signal is not corrupted by any endogenous fluorescence (autofluorescence).

    And finally: of course this type of control sample is not the only one that should be prepared. Always make samples by alternately omitting each of the primary antibodies, just to check that you don’t have a cross-reaction between your secondary antibodies…

    Q2: Can we simply trust the facility about the resolution of the microscope we plan to use or is it critical that we should define it before we start the assay?

    @lagache: Yes, trusting the facility is the easy way… Because assessing the resolution with nanometers beads is not easy

    Well, I agree you should trust your Imaging Facility… or at least ask them if they have performed this type of tests for the specific imaging configuration you are using. You may find a protocole on how to prepare reference samples and how to analyze them in the documentation of the MetroloJ plugin. In the specific context of colocalization, some protocols have also been described in P. Mascalchi P & F. P. Cordelières, “Which Elements to Build Co-localization Workflows? From Metrology to Analysis.” Methods Mol Biol. 2040:177-213, 2019..

    Q3: Is it okay to show this control images in the final manuscript?

    It might be hard to get the control images published, even as supplementary figure. You may however decide to make your datasets, together with the control datasets available through dedicated platforms such as Zenodo. Therefore, an easy way to “show” the control images might be to share a link to your original data in the paper


    Pre-processing

    Q4: How many deconvolutions algorithms are there and how do we choose among them?

    @romainGuiet: Here is a good place to start your journey about deconvolution : deconvolutionlab2

    They are MANY algorithms to restore 3D images. It would be recommanded to check if the one that performs best for you is conservative (i.e. the total intensity before and after is the same). Some parameters have to be set and we’ve described one way to set them (at least the number of iterations for iterative algorithms) here. Please keep in mind that most of the available tools are considering the unitary deformation (PSF: point spread function) to be the same all over the sample… which is a huge approximation. Image restoration is really usefull, but only an estimate of what your actual sample might be. Always have a critical view at its result and have a precise look at your images, and look for actefacts: a structure that is visible on the deconvolved image should at least be present as “seeds” on the raw image.

    Q5a: If spectra completely far apart we dont need unmixing, do we? For example Alexa 594 and 488

    Q5b: Can spectral unmixing help with the bleedthrough issue or would it create a problem?

    Q5a, @romainGuiet: you should be safe but you still need your single stain controls to assess the bleedthrough

    Trying to get spectra appart is one thing that should be considered. But you also have to keep in mind that if you use longer wavelength dyes, you’ll end up with a lower resolution, increasing the chance to get co-localization: it’s all a matter of compromises.

    Q5b, @lagache: Yes it should help. but spectral unmixing is not that simple… Q5b, @romainGuiet: Can be considered as a “correction”, better to change dyes if possible

    When dealing with image analysis, it’s always best to be as close as possible from raw data. The first measure to take is trying to improve the sample preparation and image acquisition. Then, if you still can’t reach a clear separation of your dyes, you may turn to post-processing. Keep in mind that spectral unmixing is making some math on the images, based on references you’ll provide. One assumption you’ll make is that the cross-talk/bleedthroug is the same everywhere. However, depending on the fluorophore’s environment, this might not be totally true. As a consequence, you may “over” re-attribute signal to a channel, ending up with negative intensities on the other channel (oh gosh, we’ve just generated anti-light !!!).

    At least two ImageJ/Fiji plugins exist: LUMoS Spectral Unmixing (Learning Unsupervised Means of Spectra)and
    Spectral Unmixing Plugins

    Q6a: In a image processing software they always ask for a threshold. Can this step replace the background correction?

    Q6b: In different plugins such as coloc2 it is highly recomended to subtract background. Which method do you recommend? Rolling Ball radius?? Thank you!

    Q6c: For threshold, if I use costes I will not need control samples?

    Threshold and background correction are not the same ! When you perform a threshold, you only consider pixels that have an intensity within a certain range. If you then compute the minimum intensity of the thresholded pixels, the retrieved value will be equal to the minimum threshold you’ve set. When performing background correction, you subtract a unique value (ex: using the Process>Math>Subtract function from ImageJ/Fiji) or an image which is an estimate of the background (rolling ball algorithm, Process>Subtract background). In this case, the minimum possible intensity over the image will be zero. In other words, with a threshold, you select pixels and keep all the raw intensities, with background correction you set the baseline to zero.

    About the rolling ball correction, an estimate of the background is generated and used to correct the baseline. In case all the objects on the image are of same size and isotropic shape, you’ll have no problem finding a proper radius for the filter to work. For non isotropic shapes, a parameters exists to take this property into account. However, when dealing with objects of different sizes, in case the radius is chosen too small, you’ll start subtracting intensities within bigger structures.

    Weither you use threshold (any method) or other ways to correct the baseline, you should ALWAYS care about control samples !

    Q7: so AI-based denoising + colocalization is unadvisable? alternatively: segmentation with denoised -> colocalization with ROIs at raw images?

    Nothing is forbiden, as long as you caracterize precisely the impact of the processing methods you are using and make sure the conclusions you draw from them is not a side effect of the processing itself! If using innovative methods that have not yet been explored, the work to be done in terms of checks and controls might be heavier as compared to the use of “well known” methods. In any case, always document the full processing (the macro recorder from ImageJ/Fiji might help), make your workflow public (through GitHub,) make your datasets available (through Zeonodo and do not hesitate to reference both in BioImage Informatics Index.

    Q8a: How do you choose the plane that you want to colocalize? in many cases you have cells that are cut out of the stack in the middle. so automatic analysis cannot be done

    Q8b: When you are looking at a single slice in a confocal stack for co-localization, would you still recommend using the 3D especially if the slice is very thin say 0.42 um? Also, in FIJI, how do you see the 3D?

    Q8c: “Biology is not flat”, so what should we keep in mind when doing colocalization in 3D? (Clearly deconvolution is very useful in this case to correct for spherical abberations, and also sample drift + chromatic shift)

    Q8a, @romainGuiet: You should not choose! Analyse 3D first to check if you’re “quantifier” is sensitive to 2D.

    Q8b, @romainGuiet: From a 3D stack , you calculate your quantifier on the overall stack. Ideally you should check that your results is not sensitive to 2D.

    Q8a: I totally agree with @romainGuiet: in case you’ve acquired data in 3D, it probably means your problem can’t be summarized as 2D. So why would you want to extract only part of the information, running the risk to bias the data ?

    Q8b: If your acquisition in 2D, there is unfortunately nothing you could do… except go back to the microscope and re-acquire the data. This is of course true only if your sample/region of interest is not encompassed into your 2D acquisition.

    Q8c: True: deconvolution can help. About things to keep in mind… well, it all depends on the method you are using. For Pearson and Manders you analyze pixel wise, so there is not that much you have to do. When dealing with object-based, when summarizing the objects as single spots, the disparity of resolution should be taken into account. Let’s imagine two centers in the same plane or one on top of the other, both having same (x, y) coordinates, the reference resolution is the xy resolution or the z resolution respectively. This means that if the distance between the two spots is below the optical resolution they do co-localize (or, more precisely, knowing the current resolution, you can’t exclude that they do co-localize). Now imagine the two spots close one from the other, but one slightly off-centered as compared to the other: what reference resolution would you take ? The xy resolution ? The z resolution ? In fact, you would have to take into account both resolutions, in a weighted fashion, depending on the orientation between the two dots. This is what the JACoP plugin does (see the ImageJ conference 2008 paper for an extensive explanation).

5 Likes

Table of contents, Part 2

  • Choosing a reporter/metric
  • Comparing/interpreting
  • Assembling a workflow

  • Choosing a reporter/metric

    Q9: Sorry I’m pretty new to all this; what is it that you call metric please?

    @romainGuiet:The function to measure the colocalization between objects, for instance Pearson or Manders coefficient

    A metric is a system for measuring something (source: Cambridge dictionnary)

    Q10a: might be a philosphical question: How do you convince the user that “this looks orange, so it is colocalised” is not enough?

    Q10b: I always tell my student not to trust “orange” as a sign of colocalization since it implies that both markers have the same intensity. Instead I look at the shape of the “object”. Unfortunately all the softwares quantify co-loc by taking largely in account the intensity. Are there any software correcting for this?

    So true ! When giving practicals about co-localization, I like to give the same dataset to all participants, ask them to process the images before overlaying the channels. When asked to look at their neighbours’ screen, they will all realize that processing the histogram for simple presentation will lead to as many results as participants…

    I would tune a bit the second part of the question. It is true all software will work on intensities: this is what we get out of the microscope. Intensities mostly account for fluorescence, but may as well reflect other parameters such as polarisation, fluorescence life-time etc. Quantifiers/indicators are not all based directly on intensities: some will extract geometrical features that will be used to assess co-localization (ex: object-based methods, tesselation, spatial distribution etc).

    Q11: Can you please comment on this: plotting intensity line graphs to describe colocalization?

    This has been the first published method, largely inspired by flow cytometry. We use the cytofluorogram, plotting for each pixel a dot hich coordinates are intensities from channel 1 and 2. When a linear relationship between the quantities of material between the 2 channels is expected, the dots’ cloud adopts the shape of a line. If you get a line, you may have some kind of association between the two. if you don’t, it all depends on which shape you get. No shape means either no correlation of intensities or a hidden co-localization. Try using a region of interest to see if this is everywhere the case or if you find some spots where you do have a correlation. This is how this “intensity line graph” could help and is a good starting point. In case you expect co-localization but can’t find intensities correlation, it might mean there is no favored stoechiometry of association between your two proteins of interest. You will then have to explore overlap-based or geometry-based methods.

    Q12: Manders coefficient: what do you recommend on choosing a roi (yes/no) and how to subtract background and what to use as control? Manders seems highly sensitive to the choices.

    Unfortunately, I won’t be able to answer by picking one of the options you are proposing: my answer will be “it all depends on the type of structure/phenomenon you are working on”. As explained during the talk, you should really be careful when using Manders’ coefficient in the way you interpret the data as the same percentage of overlap may correspond to different biological realities. I would recommand presenting additionnal data such as the distribution of the object’s volume, the number of structures etc. This would then enforce the proofs that eveything else, and in particular the geometry of the “positive” sites is not changing. As for the regions of interest, they will impact on the actual value of the coefficient, especially if cherry picking… Having additionnal proof that the “picking” has been made wisely and is a proper sampling is once more what I would recommand.

    Indeed, picking the right threshold is tricky. In case you are looking for co-localization over a certain structure/subcellular compartment, why not have a third channel which would be a structural maker for this compartment ? You would then compare the percentage of the signal or of a marker’s volume involved in co-localization on this compartment, ensuring a segmentation independent of the signal you are trying to quantify.

    As for controls, once more it all depends on the biological problematics. Are you performing time-lapse experiments or comparing different experimental conditions ? Depending on which situation you are describing, first time point or a non treated samples might be used as control.


    Comparing/interpreting

    Q13a: So for the data collection if there are multiple cells having PCC and MCc data, would you average them?

    Q13b: How do you really report this pearson’s correlation coefficient. Most times people only report a value and not the curve (is it enough to show only values?).Another question is about M1 and M2. How do you plot these on the graph?

    Processing the data extracted from co-localization analysis might be tricky and instead of developping here how to proceed, which might be different from one situation to another, I would recommand reading this excellent paper by McDonald & Dunn: McDonald JH, Dunn KW. Statistical tests for measures of colocalization in biological microscopy. J Microsc. 2013;252(3):295-302. doi:10.1111/jmi.12093

    I would definitely suggest not relying on a single value, which may not mean much, but have a set of values for each experimental situation. Presenting the data as bar graph is also not the best way to visualize the variability within your data. I would therefore suggest two reading: first this article by Weissgerber et al., second the excellent tutorial @MarionLouveaux gave as part of NeuBIAS Training school 15.

    Q14: Can we do Manders and PCC analysis in a single cell level instead of field of view?

    @romainGuiet: Yes! it would be recommended if you see heterogeneity of phenotypes

    Yes ! Imagine your phenomenon depends on the cell cycle status or varies depending on where the cell is (inside an islet or at the border): you wouldn’t want to miss this information ! Additionally, this would allow you to exclude cropped cells located at the borders of your image. In short: everything is possible, you choose your scale and document your process.

    Q15: In the case of very small particles with the center of the particles not colocalizing. How can you say that they are not colocalyzing when there is a part of the particles that do colocalize? if they are very small they could be closer than 40nm, then they should be colocalyzing.

    @lagache: When the size of particles & microscope resolution is below the typical distance between colocalized particles, you need to use object-based methods (see also the webinar next week: https://neubiasacademy.org/)

    Indeed, it might seem confusing. Let’s imagine two dots, each one having a size below the optical resolution: they are supposed to appear on the image as simple pixels… Unfortunately, due to the way a microscope is imaging objects, you’ll end up with a disc (sort of) in 2D known as the Airy disc. In case the two particles are close one from the other, you might see part of the discs overlapping, while the centres are not. This is a bit confusing, but shows that depending on the metric you choose, you might end up with different conclusions… For small particles, looking for overlap might not be appropriate: a particles that appears on few pixels, let’s say 4, will give you only a sparse number of possible results: 0%, 25%, 50%, 75% or 100% of overlap (0, 1, 2, 3 or 4 pixels). Is it really the metric you want to use in this case ? We would need a bit more sensitivity and using object-based methods is surely the way to go.


    Assembling a workflow

    Q16a: Can we calculate the co-localisation or co-occurance of more than 2 channels?

    Q16b: can we calculate the co-localisation or co-occurance of more than 2 channels?

    Q16c: Are there any tools to study colocalization of three differently labelled objects? One of our users would like to do it. And how to present the data?

    Q16a, @romainGuiet: I guess you could but then it might be tricky to understand ! Usually it’s done 2 by 2.

    Q16b, @romainGuiet: there is not implemented method for directly analyzing 3 colors all together; but you can combine the information you obtained about all the 2-by-2 colocalization (see for example Lagache et al. Nature comm. 2018).

    For correlation analysis between more than two channels, I would definitely go two-by-two. Having a proper visual representation for the data might be a bit tricky and I might want to explore solutions outside of the ImageJ/Fiji ecosystem. I would try importing the data (maybe images) into R and use dedicated packages to create proper diagrams. You may get inspiration from what exists in high-throughput data analysis: this link gives a nice overview of how to visualise such data.

    For overlap analysis, I would go for expending what I’ve been presenting when dealing with masks on two channels, combining the 3, 4 or 5 channels using a logical AND: this would give the triple or more expressors. Extracting the Manders coefficient from there would be the same as for 2 images (number of common positive pixels between all channels/total number of positive pixels for the channel). Once more, I would rather present the data in a table or a graph, like the ones described here.

    Q17: Can you assign objects to “groups” based on their degree of colocalization? E.g. Object A is colocalized to Objects 1 and 2 but a higher volume of A colocalizes to Object 1 so, I want to assign such object to Group 1. If the colocalization were higher respect to Object 2, I would have assign it to group 2. And so one and so forth… Thanks!

    This is possible. You would have to segment the objects, compute a co-localization metric for each and assign the extracted value to each object. You will find a simple example workflow in the NeuBIAS textbook, “Bioimage Data Analysis Workflows”, here. Once the output image has been created, you may use partitionning algorithms to extract several populations out of your objects.

    Q18: How should we perform temporal colocalization? In other words, colocalization over time.

    @lagache: No easy answer! the only paper(s) I know about the subject is the use of boolean models for ~ static objects.

    It all depends on what you call “co-localization over time” and if you need to track moving objects. In case you need a general overview of the phenomenon, you might want to repeat regular analysis over each individual timepoint, using regular tools. Using an ImageJ/Fiji macro to do that might not be that complicated: we have a NeuBIAS Academy@Home about macro programming here by @aklemm. In case you need to track objects first, have a look at this video about TrackMate here by @tinevez.

    Q19a: Are you using some software tools? if yes,which one?

    @romainGuiet: JACoP (ImageJ/Fiji), coloc2 (ImageJ/Fiji), Colocalisation Studio (Icy), SODA suite (Icy).

    @from_the_audience: CellProfiler is able to calculate the colocalization metrics that Fabrice described, using MeasureColocalization module. I wasn’t able to use this module yet, but I do know that CellProfiler is automated and can analyse a lot of images using the same metrics!

    @from_the_audience: I have been using Colocalization colourmap for 3D images

    Here is a table extracted from this paper referencing most of the open-source-based solutions:

    Reporter type

    Package/Plugin [software]

    Analyzed entity

    Coordinates as input

    Reference

    Correlation-based indicators

    Coloc2 ImageJ/Fiji

    Whole image or ROI

    No

    [1]

    Correlation-based indicators

    JACoP ImageJ/Fiji

    Whole image

    No

    [2], [3], [4]

    Correlation-based indicators

    Colocalization Studio Icy(d)

    Whole image or ROI

    No

    [5]

    Correlation-based indicators

    MeasureCorrelation CellProfiler

    Whole image or ROI Individual objects

    No

    [6]

    Intensities’ overlap-based quantifiers

    Coloc2 ImageJ/Fiji

    Whole image or ROI

    No

    [1]

    Intensities’ overlap-based quantifiers

    JACoP ImageJ/Fiji

    Whole image

    No

    [2], [3], [4]

    Intensities’ overlap-based quantifiers

    Colocalization Studio Icy(d)

    Whole image or ROI

    No

    [5]

    Intensities’ overlap-based quantifiers

    MeasureCorrelation CellProfiler

    Whole image or ROI Individual objects

    No

    [6]

    Pixels’/voxels’ overlap-based quantifiers

    DiAna ImageJ/Fiji(c, d)

    Individual objects

    No

    [7]

    Pixels’/voxels’ overlap-based quantifiers

    JACoP ImageJ/Fiji

    Whole image

    No

    [2], [3], [4]

    Pixels’/voxels’ overlap-based quantifiers

    Squassh/MosaicSuite ImageJ/Fiji(a, c, d)

    Whole image or ROI Individual objects

    No

    [8]

    Pixels’/voxels’ overlap-based quantifiers

    GcoPS Icy(c, d)

    Whole image or ROI

    No

    [9]

    Pixels’/voxels’ overlap-based quantifiers

    CalculateImageOverlap CellProfiler

    Whole image or ROI Individual objects

    No

    [6]

    Centre/objects overlap-based quantifiers

    JACoP ImageJ/Fiji

    Individual objects

    No

    [2], [3], [4]

    Centre/objects overlap-based quantifiers

    ExpandOrShrinkObjects, using “Shrink objects to a point option" CellProfiler(a, b, c)

    Individual objects

    No

    [6]

    Distance-based quantifiers

    DiAna ImageJ/Fiji(c, d)

    Individual objects

    No

    [7]

    Distance-based quantifiers

    JACoP ImageJ/Fiji

    Individual objects

    No

    [2], [3], [4]

    Distance-based quantifiers

    ThunderSTORM ImageJ/Fiji(a, b, c)

    Individual objects

    Yes

    [10]

    Distance-based quantifiers

    Colocalizer Icy

    Individual objects

    Yes

    [11]

    Distance-based quantifiers

    Colocalization Studio Icy(d)

    Individual objects

    Yes

    [5]

    Distance-based quantifiers

    SODA Icy

    Individual objects

    Yes

    [12]

    Distance-based quantifiers

    MeasureObjectNeighbor CellProfiler

    Individual objects

    Yes

    [6]

    (a): includes preprocessing; (b): includes image corrections; (c): includes detection options; (d): includes tools for comparison/statistical tests.

    References

    1. J. Schindelin, J. Eglinger, L. Guizzetti, M. Hiner, and J.-Y. Tinevez, “Coloc2.” 2018.
    2. S. Bolte and F. P. Cordelières, “A guided tour into subcellular colocalization analysis in light microscopy.” Journal of microscopy, vol. 224, no. 3, pp. 213–32, Dec. 2006.
    3. F. P. Cordelières and S. Bolte, “JACoP v2.0 : improving the user experience with co-localization studies,” in ImageJ user & developer conference, 2008, pp. 174–181.
    4. F. P. Cordelières and S. Bolte, “JaCoP, just another co-localization plugin v2.” 2018.
    5. T. Lagache, N. Sauvonnet, L. Danglot, and J. C. Olivo-Marin, “Statistical analysis of molecule colocalization in bioimaging,” Cytometry Part A, vol. 87, no. 6, pp. 568–579, 2015.
    6. A. E. Carpenter et al., “CellProfiler: image analysis software for identifying and quantifying cell phenotypes,” Genome Biology, vol. 7, no. 10, p. R100, 2006.
    7. J. F. Gilles, M. Dos Santos, T. Boudier, S. Bolte, and N. Heck, “DiAna, an ImageJ tool for object-based 3D co-localization and distance analysis,” Methods, vol. 115, pp. 55–64, 2017.
    8. A. Rizk et al., “Segmentation and quantification of subcellular structures in fluorescence microscopy images using Squassh,” Nature Protocols, vol. 9, no. 3, pp. 586–596, 2014.
    9. F. Lavancier, et al., “Testing independence between two random sets for the analysis of colocalization in bioimaging,” Biometrics, 76(1):36-46, 2020.
    10. M. Ovesný, P. Křížek, J. Borkovec, Z. Švindrych, and G. M. Hagen, “ThunderSTORM: A comprehensive ImageJ plug-in for PALM and STORM data analysis and super-resolution imaging,” Bioinformatics, vol. 30, no. 16, pp. 2389–2390, 2014.
    11. F. De Chaumont, “Colocalizer.” 2018.
    12. [12] T. Lagache et al., “Mapping molecular assemblies with fluorescence microscopy and object-based spatial statistics,” Nature Communications, vol. 9, no. 1, pp. 102–108, 2018.

    Q19b: For the coloc software section : can you comment also on Commercial software - eg Imaris coloc module

    I don’t have a specific comment to make about commercial software: use the one tool that you feel is the most appropriate for your specific question. But please do not use it as a black box: get insights about what precise processings are performed !!! In case you are working on a consequent dataset, you may also want to check first that the commercial software you are using is providing a batch mode: you wouldn’t want to make it all manually.

    8 Likes

    Updated link to the YouTube video with a proper image/sound synchro: https://youtu.be/P2JvFe0hB_M

    5 Likes

    Thx @Fabrice_Cordelieres. Just watched the entire talk and was awesome!!
    I am just wondering whether the metrics that you mentioned can be used as an evaluation of image registration. What’s your opinion on that?