Mitochondrial event localiser (MEL) to quantitativelydescribe fission, fusion and depolarisation in the three-dimensional space

Mitochondrial fission and fusion play an important role not only in maintaining mitochondrial homeostasis but also in preserving overall cellular viability. However, quantitative analysis based on the three-dimensional localisation of these highly dynamic mitochondrial events in the cellular context has not yet been accomplished. Moreover, it remains largely uncertain where in the mitochondrial network depolarisation is most likely to occur. We present the mitochondrial event localiser (MEL), a method that allows high-throughput, automated and deterministic localisation and quantification of mitochondrial fission, fusion and depolarisation events in large three-dimensional microscopy time-lapse sequences. In addition, MEL calculates the number of mitochondrial structures as well as their combined and average volume for each image frame in the time-lapse sequence. The mitochondrial event locations can subsequently be visualised by superposition over the fluorescence micrograph z-stack. We apply MEL to both control samples as well as to cells before and after treatment with hydrogen peroxide (H2O2). An average of 9.3/7.2/2.3 fusion/fission/depolarisation events per cell were observed respectively for every 10 sec in the control cells. With peroxide treatment, the rate initially shifted toward fusion with and average of 15/6/3 events per cell, before returning to a new equilibrium not far from that of the control cells, with an average of 6.2/6.4/3.4 events per cell. These MEL results indicate that both pre-treatment and control cells maintain a fission/fusion equilibrium, and that depolarisation is higher in the post-treatment cells. When individually validating mitochondrial events detected with MEL, for a representative cell for the control and treated samples, the true-positive events were 47%/49%/14% respectively for fusion/fission/depolarisation events. We conclude that MEL is a viable method of quantitative mitochondrial event analysis.

The manuscript entitled "Mitochondrial event localiser (MEL) ..." by Theart et al. describes an automated analysis tool for 3D image stacks of mitochondrial fusion, fission and depolarisation. The manuscript is very well written and pleasent to read. The scientific topic of this manuscript is highly important as -to the best of my knowledge -an objective and automated analysis tool does not exist today; yet it is highly desirable, because the mitochondrial dynamics can only be accurately quantified in this way. The quantification of these processes is highly important, because it provides a firm basis for models of mitochondrial dynamics.
We thank the reviewer for the kind words and also acknowledging the importance of the problem that we set out to solve in this paper. We are also not aware of an automated analysis tool that can achieve this. The software implementation of our method has been included with the revision for use by other researchers.

Reviewer's comment 2.
I have a number of serious issues that need to be addressed before the manuscript may be considered for publication: 1) In line 166-169, the authors list the parameters that enter the image analysis. I suggest that the authors give here the range in which they have tested these parameters for each of them. Absolute numbers are required in order to understand statements that are otherwise meaningless, because one does not know what they compare to; e.g. "overlapping volumes that account for less than 1% of the volume ... are eliminated" (caption of Fig. 2) -what makes 1% the magic number?
"with a small percentage" (line 228) -what is small? what is large? how to decide this?
We appreciate the reviewer mentioning the lack of specificity of some assumptions made with MEL. We have now corrected these aspects, focussing on absolute numbers, and clearly indicating the statements as follows: • Line 166-169: The three tunable parameters in question are the two different standard deviations used for the 2D and 3D Gaussian blurs and the volume of structures considered as noise. The values of these parameters were chosen empirically by considering their effect on the sample data that we analysed in this paper, and the ranges we provide might not be appropriate to all conditions and acquisition parameters. Regardless, we have adjusted the paragraph in question by adding the following text: "The volume should be determined empirically by considering the average size of the binarised noise without erroneously removing true mitochondrial structures. From our testing, the value of σ2D in the normalisation step should take values between 0.5 and 1.5. σ3D in the 3D Gaussian filter step should take values between 0.2 and 0.5 (0.02-0.04 µm in x-y and 0.1-0.25 µm in z) due to the limited resolution in the z-dimension. These values are, however, subject to different image acquisition parameters." • Line 112: We added the following sentence as explanation: "The required time interval for accurate results is dependent on the motility of the mitochondria, but in our experimentation time intervals of 10-30 seconds produced good results. Shorter time intervals such as 5 seconds may also be used." • What is appropriate?: We updated the section in which this comment was made as follows: "Since it has been observed that some noise remnants were also being binarised, we removed any structure containing less than 40 voxels in the upscaled images. We determined this value empirically by considering the average size of the binarised noise structures. In concrete terms, this is equivalent to a circle with a diameter of 5 pixels in the original image, or in our case a circular structure with a physical diameter of about 0.6 µm." • Fig 2 Caption and Line 228: We thank the reviewer for allowing us to reconsider our method design. In the process of refining the MEL method for the revised submission we realised that there was no need for the small percentage to remove 'coincidental' events due to slight overlaps between structures since the overlap percentages that are used at a later stage already eliminated these events. As this in no longer part of the method we have removed the corresponding sentences from the revised manuscript.

Reviewer's comment 3.
2) Line 305-306: I do not understand why the experiments and controls were recorded with different time steps. One would expect that every condition is recorded with identical settings and it would be the best if time steps would be chosen as small as possible. This would not only increase accuracy of the analysis, but would also allow to systematically study the impact of increasing time steps: for experiment and control the analysis could be repeated with larger time steps (leaving frames in between out) -but still having the exact time stepping in experiment and control. Otherwise, I do not see how to uniquely attribute changes in the analysis performance to the difference in the experiment versus control, because such differences may as well stem from technical issues like different time stepping in the imaging. While having such differences should be avoided in the experimental design, the reasoning of the authors for their data set with incommensurate time steps is not at all clear.

Authors' response 3.
We wish to thank the reviewer for this important feedback. These differences were due to distinct acquisition runs that were performed during the development of MEL, and we wished to showcase the application of MEL under various conditions.
We have now acquired substantial additional data sets, repeated imaging acquisitions under control and treatment conditions, keeping a consistent time interval of 10 seconds for both control and treated samples. Moreover, we have used the same cells pre-and post-treatment, which was technically not trivial, using a 100x oil immersion objective. Hence, this now enables a direct comparison of the different mitochondrial events under the conditions assessed. In addition to the control cells, the pre-and post-treated cells were also compared which allowed the assessment of dynamic changes of the mitochondrial events, as opposed to merely endpoints. In order to induce a mitochondrial response, we have now utilized H2O2 1 , to achieve a generic and more commonly implemented cellular stress response. Care was taken to implement similar and controlled acquisition parameters that favour dynamic range, minimize bleaching and allow rapid acquisition in z with high spatial and temporal resolution.

Reviewer's comment 4.
3) In line 348ff, the visualization in virtual reality is emphasized. I do not understand this point, may be; however, what I would have liked to see is a number of videos (e.g. with changing perspectives) that gives some idea how the dynamics looks in three spatial dimensions. Why did the author not opt for this possibility?
We thank the reviewer for the comment. The MEL method is independent of the use of virtual reality, and in the initial submission we resorted to using a previously developed virtual reality tool 2 that enables precise three-dimensional region of interest selections.
As to not distract from the focus of the manuscript, we have now removed the reference to virtual reality and instead performed the 3D ROI selections and visualisation with Fiji (ImageJ).
Reviewer's comment 5. 4) Related to the last issue, in general, I am missing a quantitative validation of the analysis tool. The Results section and the Discussion section contain a lot of statements that are not quantitatively supported by a rigorous validation. No comparison is made; neither to manually annotated data, nor by using another analysis tool (which may not have been developed for mitochondria, but for other data, e.g. for cell migration and interaction). Without any quantitative validation of this analysis tool, I cannot see how the value of the presented tool can be objectively judged. Without such an analysis, I think, this work can not be published. I would like to know, how many false positives and false negatives you have in detecting and not detecting events of fusion/fission/depolarisation; please, provide the numbers of your validation in terms of standard performance measures.
We thank the reviewer for this important aspect. We have now performed an extensive analysis on the large, newly acquired data sets, so as to better indicate quantitative validation of the analysis tool. As far as has been possible, we have assessed the number of false-positive events.
The type of analysis that is performed by MEL is, however, challenging to perform accurately 'by eye', especially on complex networked mitochondrial structures in three-dimensional space. Although we were not able to locate a tool that achieves something similar enough to be used as a comparison with MEL, we have developed a tool, which has been submitted as part of the revision, that life science experts used to verify the mitochondrial events detected by MEL individually. This allowed us to determine the number of false positives that were detected by MEL and it is hoped to address the reviewer's quantitative validation concern.
Due to the incredible abundance of such events for multiple cells over a time-lapse sequence we have limited this validation to a data set of a representative cell from the control, pre-and post-treated groups. The quantitative validation results are summarised in Table 1 and Fig 6. In addition, to enhance clarity and user implementation, we have now indicated and described common scenarios in which false-positive events are detected (Fig 5). Our analysis confirms that MEL operates favourably, especially in the context of detecting changes in mitochondrial events over time and comparing different treatment groups with each other. In addition, we have now included a supplementary section titled "General considerations for MEL" which outlines technical considerations that may assist the user in generating raw data that are least prone to inaccurate analysis, for example, due to poor thresholding. This may serve as a transparent point of departure for further implementation by the imaging community.
Moreover, we have added the comment of validating the events to line 105 in the method: "Since some of the detected events could be false positives as a consequence of the thresholding step joining two mitochondrial structures together that are close to each other but not in fact fused, MEL also allows for each event to be subsequently validated and removed from the visualisation and event counts." Reviewer's comment 6. 5) Once issue 4) has been done, the authors can remove statements like "it was often observed" (line 408), "From the analysis it seems that MEL is robust..." (line 426), "the false detection of these events is reduced" (line 432), "the best 3D quantification" (line 442), "it produces the most consistent results" (line 446) and replace them by quantiatively concrete statements. Please, check the whole manuscript for many of these kinds of statements, which all need to be removed and replaced by quantitative statements with clear reference as to how properties like "robust" and "best" are measured.
Authors' response 6. We appreciate the reviewer's comment. We have edited the comments in the manuscript as follows: • Line 408: This statement has been removed and the text now reads: "MEL can also aid in the detection of mitochondrial structures that alternate between fission and fusion events as shown in Fig 11." • Line 426: This line has been removed since the statement is no longer relevant to the new data that we showcased. • Line 432: We have changed the manuscript to read: "However, since MEL considers the entire threedimensional structure of mitochondria, it can distinguish mitochondrial structures passing each other from true fusion between these structures. For this to be possible, the microscope that is used should have sufficient resolving power and the z-stack should contain enough micrographs. We found using a 100x oil immersion objective with the microscope and acquiring image stacks with 4-6 micrographs, with a 0.5 µm step width, produced favourable results." • Line 442: This statement was in reference to other work that was cited. We have, however, changed the text to read: "Such a strategy was used, for example, for an automated image analysis algorithm that determines the optimal filtering parameters to produce an optimised 3D quantification of mitochondrial morphology." • Line 446: This statement now reads: "It is for this reason that we first normalise the images before applying hysteresis thresholding to ensure it produces more consistent results." • We have also corrected similar statements throughout the manuscript.

Reviewer's comment 7.
6) Table 2 contains a list of all tunable parameters, which need to be given in realistic units; the symbol \sigma should not be used repeatedly for different quantitites.

Authors' response 7.
We thank the reviewer for the comment and for picking up that we erroneously used \sigma to represent different quantities. We have now distinguished between the two-dimensional Gaussian filter used during normalisation by using \sigma_{2D} and the three-dimensional Gaussian filter as \sigma_{3D}.
In the revised manuscript we have also included all the values in realistic units throughout the paper, as well as in S1 All these models (and may be others as well) will benefit from a tool like you have developed here, underlining the importance of it once it has been appropriately validated.

Authors' response 8.
We wish to thank the reviewer for this comment and the recommendations of referring to existing tools.
We have now included these in the manuscript (line 24), so as to contextualize the strength and potential importance of MEL better.
"Furthermore, major interest exists in quantitatively describing mitochondrial dynamics to better discern the role of mitochondrial dysfunction in pathology and pathogenesis of major human diseases and to unravel whether changes in fission and fusion are adaptive or maladaptive. This includes the role of mitochondrial function in ageing (Ref 1 and 2), in context of mitochondrial quality control and mitophagy (Ref 3) as well as in the context of mitochondrial directionality and network distribution (Ref 4). However, a tool that assists in the rapid characterization, localization and quantification of mitochondrial fission, fusion and depolarisation events in the subcellular, three-dimensional space is currently not available." The purpose of MEL has been clarified in the introduction in Line 77: "[MEL] provides insight into the spatio-temporal change of mitochondrial events in a time-lapse sequence and can also be used to compare mitochondrial event dynamics under different treatment conditions" Reviewer's comment 9.
8) I somehow miss the information where the code can be downloaded and where the data can be downloaded. Did I overlook this?
We thank the reviewer for requesting the code. This was an oversight on our part, and we have accordingly submitted it as part of the revised manuscript.
Overall, I think this is an interesting paper that with more rigorous analysis and experimental replicates with statistical analysis has the chance to contribute to the field of mitochondria morphometry.
My most concerning critique is that the authors present the data of a single biological experimental replicate per condition. How can the community be assured of the repeatability of your method? This must be addressed prior to publication. Please describe and discuss the variance between accuracy (including false positive and false negative events as discussed in the critique below), results, and biological implications.
We wish to thank the reviewer for this important comment. We have now attended to this aspect and have carefully acquired large additional data sets, so as to allow better and direct comparison and statistical analysis.
It is challenging to precisely evaluate the accuracy of MEL, since to our knowledge no tool exists currently which achieves something similar. To address the reviewer's comment, we have developed a tool which was used by life science experts to verify the mitochondrial events detected by MEL individually. In this way we were able to determine the number of false positives that were detected by MEL under different conditions. The result of this is summarised in Table 1  A supplementary section titled "General considerations for MEL" have also been added, in which we outline some technical considerations to generate raw data that are least prone to inaccurate analysis, for example, due to poor thresholding. This may serve as a transparent point of departure for further implementation by the imaging community.
We have also added the comment of validating the events to line 105 in the method: "Since some of the detected events could be false positives as a consequence of the thresholding step joining two mitochondrial structures together that are close to each other but not in fact fused, MEL also allows for each event to be subsequently validated and removed from the visualisation and event counts."

Reviewer's comment 2.
How do you distinguish between events that are truly depolarization and loss of florescence signal from photobleaching? One way to describe this would be an overall signal to background fluorescence intensity from frame to frame, but discussion of the lower limit of detection of the algorithm is important given the assumption that a disappearing label is a depolarization event. An alternative approach could be the use of a covalently or otherwise permanently bound fluorophore (e.g MitoTracker Deep Red or fluorescent-tagged membrane bound protein).
The reviewer raises an important point. For our analysis that was performed and is presented in the revised manuscript, we have now acquired data in a highly controlled environment, using a very rapid scanning speed, very low (down to 0.2) laser power and have optimized dynamic range and photonmultipliertube gain, so as to avoid photobleaching. We have, for that purpose, implemented a time-lapse sequence that is resolved and extensive enough, but does not result in bleaching of signal. There was therefore no requirement to compensate for the detection of depolarisation in such conditions.
In the revised manuscript we now make use of hysteresis thresholding which can distinguish better between background and dim mitochondrial structures. The lower limit of detection is therefore determined by the threshold values that are used. We have added the following explanation in Line 139: "Hysteresis thresholding is very effective in removing background voxels that have intensities similar to the mitochondrial object voxels. It uses two threshold values. Voxels that have an intensity above the high threshold are considered to belong to the object and voxels that have an intensity below the low threshold are considered to belong to the background. Voxels that have an intensity between the low and high thresholds are considered to belong to the object if they are connected to other object voxels. We automatically calculated the low threshold at the edge of the histogram valley of the background voxels intensities, and the high threshold at the halfway point between the low threshold and the maximum intensity." We agree, a permanently bound fluorphore could equally be used for this analysis. Although we also frequently use MitoTrackers, for the purpose of this study, we chose TMRE, due to its wide use and application in assessing mitochondrial dysfunction, and due its exceptionally good signal-to-noise ratio. However, MEL would equally operate on image data acquired using other fluorescent probes. We have now included this in the manuscript, to enhance applicability and user range.

Reviewer's comment 3.
I would argue that the tunable parameter of minimum voxel volume or the otsu thresholding sensitivity should be retuned. In both the regions of interest image series displayed in figures 4 and 5, multiple incorrect categorizations of events can be observed by looking at the entire time lapse. For instance, in figure 5, a more dimly lit small spherical shaped mitochondrion is identified by the algorithm to undergo depolarization 3 times but by the naked eye it's appearance changes very little and yet it exists across the entire image series. Similarly, the fission and fusion events the authors describe as kiss and run do not convincingly separate and join. If this is representative of the algorithm's performance, then the reader must conclude that the false positive event rate is far too high to be acceptable.