04-conclusion.Rmd

# Discussion

## Methods

Due to the pLLP’s relative simplicity (~100 cells) and excellent accessibility for advanced light-microscopes (~1 cell layer beneath the skin), it promises an _in toto_ understanding and complete model of factors influencing its development.

To create an accurate and robust model of a developmental process, it is necessary to have precise, meaningful and accurate data. Furthermore, to test a hypothesis thoroughly, it is beneficial to be able to analyze a single set of image data in different ways which gives the advantage to directly link records in the separate datasets to each other. _E.g._ with zebrafish you have a problem of exact stage matching. Even if at one point, when starting to live image, the embryos are stage-aligned, developmental speed may differ between embryos which results in a stage mis-alignment at a later point in time. Therefore you need biological and technical replicates to statistically test your hypothesis. This is true if you want to test for an effect in a single feature like the area of an organ. However, if you also want to know about other features that require a different kind of analysis you will have to record another set of image data, again with biological and technical replicates to say that at a given timepoint feature $\mathrm{a}$ and $\mathrm{b}$ relate to each other in a certain way. Doing so comes at cost of data consistency since (1) the samples are in fact different ones (2) they have to be prepared in a different way and (3) the measuring tool is a different one. 

One way to overcome this is standardization and automation which, if accomplished, would allow to generate large datasets of precise measurements. The idea here is to record a set of image data at a sufficient resolution and number of dimensions necessary to conduct the analyses, allowing to connect datapoints from a certain embryo and timepoint. Thereby different features don't need to be compared by their means, but directly, which reduces complexity and increases precision. 

The tools and methods developed for this work have proven easy to adopt even by novice users and very useful in extracting large scale quantitative information. 

### Standardized mounting

Although efforts have been made to standardize zebrafish embryo mounting and imaging, most protocols were designed for high throughput widefield screening [@Wittbrodt2014; @Yu2018; @Pulak2016; @Donoughe2018] rather than high-resolution confocal microscopy [@Herrgen2009; @A.2015; @Hirsinger2017]. In recent years a couple of studies tackled specifically this issue by developing standardized mounting methods using 3D printed stamps [@Campinho2018; @Donoughe2018; @Masselink2014; @Alessandri2017]. The solution provided here offers a much more efficient way of mounting zebrafish embryos for high resolution confocal microscopy.

Most methods concerned with specimen mounting are developed by microscopy specialized labs that are working with state-of-the-art microscopes. Usually, the most successful of these developments are those which find a collaboration partner that can provide a specimen that fits the microscope use case. Still, for most labs the microscope technology and mounting technique is not in reach since they are usually expensive and are targeted towards very specific scientific questions. The mounting method described in this thesis targets a gap that previously had not been filled. Today, the most useful and most spread microscope within the developmental biology community is the _spinning disc_ microscope. It allows to record relatively high-resolution time-lapse movies (depending on the size of the sample and number of channels) as well as 3-D imaging of multiple specimen at once [@Graf2005]. Sample preparation for spinning disc microscopy does not differ from preparing samples for standard brightfield imaging. Therefore, for many it is a door opener into advanced microscopy. 

For zebrafish, for many years the standard mounting technique was to mount zebrafish embryos on their side without assistance on a flat surface in a rather fast solidifying mounting medium. Since the embryos are not flat on their side they tend to fall over and have to be re-positioned frequently. Furthermore, the researcher needs to be quite skilled to know how to handle the embryo so it moves in a desired direction. Taken together, this limits the number of embryos that can be mounted simultaneously tremendously. Since one can harvest around a hundred eggs from a zebrafish female from a single crossing, for live imaging most of them could not be used or it would be extremely time consuming to mount and image all of them. Therefore, most of the eggs / larvae are not used and are discarded after some time. 

Since the method is easy to adopt and uses already existing microscopy resources, it increases the data output and time efficiency by a multitude without investment in new hardware and excessive training. I chose to mount in round dishes because (1) they are used classically and (2) the selection with the manufacturers is more suited for this application. However, round dishes require a manual alignment of the embryonic body axes with the microscope stage axes - which may compromise the use with pre-set positions if the dish is not placed accurately onto the microscope sample holder. By using rectangular dished together with a rectangular stamp it would be possible to minimize the angular degrees of freedom.

### Automated image analysis

Humans depend on vision to orient themselves in their environment. Even though, from a technical perspective, eyesight is not at all a perfect system (blind spot, foveal centralis, farsightedness, shortsightedness, ...), humans usually don't perceive these shortcomings. This is because vision is mostly just a heavily processed and augmented version of the faulty and blurry data coming from the eye, processed by the brain. The brain on the other hand is a very advanced pattern recognition system [@Mattson2014].

For humans, analyzing image data is a tedious and tiring process. First, one has to be trained to detect certain features in images, then one has to recognize those features in many different images (from different angles, different sizes, ...), sometimes going into the thousands. Furthermore, even though humans can orient themselves in a 3-D space, for most people recognizing image features in 3-D is a spacial challenge. Whether analyzing images in 2-D or 3-D, it demands a high degree of attention which, in order to ensure high quality data, in turn requires the analyst to have regular breaks. Computers are made to process data. Therefore, to overcome the shortcomings of a human image analyst it is instead possible to train a computer program to specifically detect image features. 

#### 2D Analysis

The anallyzr2D and anallyzr2DT detect rather simple image features - the deposited Neuromasts, the pLLP and the nuclei within each structure. While this could, even for large datasets, still be done by a human analyst, one also needs to take into account human _confirmation bias_^[Assuming the selection of images analysed is randomized and the analyst doesn't know what he is looking at, the analyst could still _think_ to be looking at a certain mutant or wildype and therefore draw a line about a shape more or less generously. The algorithm on the other hand, depending on how its programmed, usually is not biased. Therefore, it generates more neutral and trustworthy measurements.] during an analysis session. Since the algorithm is not subject to fatigue, it is also less error prone.

In addition to the analytical improvement, the anallzr2DT processes and prepares the images for further downstream processing in a way pLLP timelapses could not be analyzed before. In a timelapse of the LL, the pLLP migrates from _anterior_ to _posterior_, while the embryo still grows. Since this means a shift of the position of the pLLP from timepoint to timepoint, it is hard to detect and analyze cellular dynamics within the pLLP. During processing of the anallzr2DT the migratory dimension as well as sideward movements of the pLLP are eliminated, making it possible to investigate single cell behaviour during migration more thoroughly. Furthermore, while CC count and CC position had been analysed before [@Matsuda2010b; @Durdu2014a], using this technique allowed us to conduct the most thoroughly and most accurate analysis conducted on cell cluster deposition so far.

#### 3D Analysis

The anallyzr3D is able to segment and analyze cells within the pLLP in 3D and measure a cells apical area at a certain distance from the _apex_. Since the fluorescence signal in some areas of the cell can be very faint (e.g. in basal regions) or very high (e.g. in rosette centers, where many cell apices come together) it is almost impossible for a human analyst to track and detect the cells boundaries correctly. While manual 3D cell segmentation in the pLLP has been done before [@Harding2013], here I provided a method that allows for segmentation of all the cells of the pLLP and large datasets. Furthermore, for certain measurements and when comparing positions in 3D, a normalized orientation of cells and between specimen is necessary, which is taken care of by the standardized mounting method.

For computational image analysis, the researcher has different options in methodology where each have their pros and cons. (1) For segmentation, traditionally one would use thresholding or a _watershed_ algorithm. For the latter, the image is treated like a topological map, assuming higher gray values represent something topologically high and lower gray values something topologically low, the watershed algorithm then 'fills' the lower regions stating from the _peak_ of each maximum until it comes in touch with another filled region [@Vincent1991]. (2) Alternatively, one can also _train_ a computer to detect cell boundaries. With this technique, an artificial _neural network_ (part of Machine- and _Deep - Learning_) is trained to detect image features on a _training_ dataset where the features (cell membranes, nuclei, ...) were already marked by a human analyst. When the network has reached a certain precision through learning, it is asked to find the same features on a _test_ dataset [@Ueda1993] to validate the training results. For the segmentation results, for the moment, neither option has a clear advantage over the other.

For my analyses, I wanted to make sure they were usable and reproducible by other people in the field. While there are people working on making Machine Learning approaches like Deep Learning more approachable to non-technical people (like most biologists are), they still require a very good understanding and knowledge of the subject and its respective implementation (typically programming languages like _Python_ or _R_). Meanwhile, many biologists are familiar with the image analysis software _ImageJ_ respectively _FIJI_, which has a rich catalog of actively developed and maintained libraries or _update sites_. ImageJ again has the ability to record, parameterize and execute macro code, where it is possible to not only incorporate the core features, but also many of the update site plugins. This way, it is possible to compose a macro program that automates even challenging image analysis pipelines and to easily distribute it to a wide audience - instantly able to reproduce the analyses^[provided the necessary computing power]. In addition, this method is much more transparent in terms of how the features are acutally extracted and may be _forked_^[to open a parallel branch of development] to develop it further or adjust it to the users own needs.

In addition to visualization, large multidimensional datasets offer the possibility for advanced computational methods such as Machine- and Deep Learning. For example one could label cells of a dataset as either leading, trailing, rosette, lateral, ... etc. to train an Machine learning model on. The model could then be used on unlabeled data to assign the previous learned labels to cells that fit the right parameters, a strategy that that has recently been published by J.Hartmann _et al._ [@Hartmann2020]. Yet another application example would be to use the Ground Truth image data generated for my studies to train a CNN that potentially would be more robust to data of different quality / resolution. For the future it would be interesting to apply this method on timeplapse movies too.

### Rosette Detection

While the anallzr3D actually detects and counts pLLP rosette centers using traditional image processing techniques, it does not quantify the rosettes maturity respectively to which degree it actually resembles a wild-type rosette. An inherent feature of object detection _via_ neural networks is that they tell us how safe it is for an object detected to acually be what it was trained to detect. In other words, neural networks will also detect objects which do not specifically resemble what they know from the training data - but with a lower detection _score_. While the interpretation of neural network results should be taken cautiously [@Ghorbani2019], for my analyses I interpreted the detection score as the _rosettiness_, or how much over all the pLLP is rosettized. 

For classification tasks one can use _discriminative_ or _generative_ models. The main difference here is that (1) generative models will model the positives (the distribution or label to detect) as well as the negatives (the part of the data that is not labeled). When asking a generative model to classify data it will look for a boundary where one model becomes more plausible than the other. Hence they are probabilistic. (2) Discriminative models on the other hand focus on the boundary that separates the labeled from the unlabeled data. Hence they are not probabilistic.

The model used for rosette detection is a discriminative one. During the course of training the model is optimized to reach a _softmax_^[An important metric in object detection is the _softmax_-score (also "detection score", the final result of all weights of the NN). It is a metric that tells about the security of the network how safe it is in its prediction.] of 1 which later allows to make a statement if the pattern is more similar to a pattern of one class seen in the training data than to patterns of the other class. The gradation however is not necessarily linear, which makes a statement like 'the mutant rosette is 50$\%$ wildtype' impossible. However, even for generative models we have the problem of chosing the right set of training data that covers all the aspects and expressions of possibly occuring real patterns. In addition there is the problem of modelling multi-modal distributions. Many questions that cannot be answered conclusively, but do not prevent us from using these models in the hope that they will do something useful and keep us busy for many years to come.

## Shroom3

### Lateral Line

Till the end of migration, on average two additional CCs are deposited in _shroom3_$^{-/-}$ embryos. Proliferation is not increased in the pLLP - but in CCs once they are deposited (section \@ref(res-prolpLLP)) reaching at ~ 2 dpf wild-type levels of cells _per_ area. Therefore the increase in CC count is unrelated to the amount of proliferation in the pLLP. While I did not quantify the size of CCs directly after deposition, it seems likely that the observation of an increase in proliferation after CCs are deposited is due to a compensatory effect. Interestingly, there are reports showing an _inexhaustable_ hair cell generation [@Pinto-Teixeira2015] which seems to be triggered by an interaction between WNT and Notch Pathways [@Romero-Carvajal2015; @Head2013; @Wada2013]. The cell count and area per CC show an average reduction of 6$\%$ respectively and no difference in density (section \@ref(res-llmorph)).

```{conjecture, echo = TRUE}
The increase in deposited cell clusters is independent of proliferation
```

### Rosette Formation

One of the most important and most interesting results of my work is the correlation between rosettiness and CC count as it clearly shows the interdependence between a high median rosettiness and low number of deposited CCs and _vice versa_ - therefore highlighting the role of Shroom3 as a stabilizing factor during NM formation. 

```{conjecture, echo = TRUE}
The increase in deposited cell clusters is due to a reduction in rosettiness.
```

As explained in section \@ref(intro-types), cells in the leading region have a more mesenchymal (more loose and more migrating) while cells in the trailing region have a more epithelial character. The cells in the trailing region therefore are more adhesive to their substrate, the basement membrane, while at the same time they are _pulled_ by the more mesenchymal leading cells towards the direction of migration. This results in traction forces in the trailing cells, becoming stronger at times of acceleration and towards the more trailing part, which potentially destabilizes the migrating cluster. To resist these forces, the cells within the cluster need to adhere strongly, which is made sure by Shroom3. Interestingly this also means that there are no external factors regulating the pattern of LL development, but that it is completely controlled by internal factors.

```{conjecture, echo = TRUE}
The pattern of deposited Neuromasts is controlled by internal factors of the posterior Lateral Line Primordium.
```

While the CC count can be derived from images of the LL at end of migration, the rosettiness at this timepoint can not be measured anymore since the pLLP disperses into three terminal NMs. Likewise, the rosettiness can be measured in images during migration - but the final CC count is only revealed at end of migration. One possible solution to this is to first take single timepoint images of the pLLP during migration, leave the embryos embedded in agarose while incubating them till end of migration, and then image them again at end of migration to derive the CC count. While this way one can image more emrbyos, one only obtains a single measurement for a single timepoint - making the statistics much poorer. This is mostly due to the measurement of rosettiness. As explained, the result of the _weight_ from the rosette detector does not exactly depict a linear relationship between 0 - no rosette and 1 - perfect rosette. Therefore the variance in measurements may be quite high. Taking however the median of a series of rosettiness measurements from the same pLLP reduces this error. Additionally this kind of procedure produces a number of other useful measurements like speed and acceleration, pLLP area and roundness that may be subject to investigate possible interdependencies.

As figure \@ref(fig:rdtreg)B-B' shows, the method allows to measure velocity and acceleration quite accurately (as indicated by the standard error). Especially for that time scale, there are no reports where researchers were able to successfully derive these measurements - even though it was long known that phases of acceleration and deceleration existed [@Ghysen2007a]. One of our hypothesis was that the higher number of CCs may be related to phases of acceleration, since here cells within CCs would get stretched more which in absence of Shroom3 may lead to a more easy and therefore more frequent deposition of CCs. Even though I have shown a there is a relationship between velocity / acceleration and rosettiness I think the approach is not flawless. As shown in figure \@ref(fig:rdtreg)B' acceleration has a wave-like pattern of acceleration and deceleration through time. Therefore, to correlate the measures acceleration and rosettiness precisely it would be necessary to compare acceleration and rosettiness at a certain timepoint and to replicate this wave-like pattern in the correlation coefficient. However, the dataset recorded does not allow for such an analysis. To calculate a correlation one needs at least two data-points that are precisely stage-matched for each timepoint. For a convincing correlation analysis it is necessary to have much more than two datapoints. In general, the more, the better. However, even in wild-type embryos the cluster deposition cycle is not exactly the same throughout time and between embryos. To make things easier it might be sufficient to stage-match just a single deposition cycle, but even here it would be necessary to record a multiple of embryos I recorded for this analysis to find similar deposition cycle patterns.

### Apical Constriction

The main learning from the quantification of the apical index is that it is far from easy not only to accurately quantify a single cell 3D morphometric feature, but also to interpret it.

#### Apical Index

The main findings from this study are that the difference in A.I. between the cells of _shroom3_$^{+/+}$ and _shroom3_$^{-/-}$ pLLPs is more pronounced along the major than the minor. For recap, the A.I. is lateral height over the respective fit ellipsoid axis of a cells cross-section at a relative distance from the apex (section \@ref(ACI-param)). If the lateral height was 5 $\mu$m and major / minor were 2 / 1 $\mu$m, then the A.I. major was 2.5 and the A.I. minor 5. The inversion of values represents the reality, where the minor axis is more constricted than the major axis. The axes however do not reflect a specific body axis (anterior - posterior or dorsal - ventral), but essentially the longer (major) and the shorter (minor) axis of the ellipsoid. If the reduction in apical constriction in _shroom3_$^{-/-}$ embryos would occur isotropically we would expect a similar increase in axis length (major and minor) as compared to _shroom3_$^{+/+}$ controls. At 32 hpf we see about a 10$\%$ reduction in both the major and the minor, while at the later stages of 36 and 40 hpf we see a stronger reduction in the major only (section \@ref(res-ai)). A main difference between the earliest stage and the later ones is the speed of migration, which reaches its maximum at around 36 - 40 hpf (section \@ref(res-rosreg)). Therefore one possible explanation could be that at lower speed of migration apical constriction is reduced less, while at higher speeds the cells are stretched more strongly along the axis parallel to the direction of migration. Interestingly, this explanation would also concur with a study by Kozlovskaja-Gumbrienė _at al._ [@Kozlovskaja-Gumbriene2017], who report Shroom3 independent rosette formation at early stages and active Notch signaling. 

```{conjecture, echo = TRUE}
The reduction in apical constriction is anisotropic. The anisotropicity may be related to speed of migration.
```

#### Cell orientation

The second finding is the difference in cell radial orientation. A rosette can be divided in four quadrants of 0 - 90$^\circ$, where 0$^\circ$ represents the anterior - posterior axis (section \@ref(res-rosreg)A'). The measurement taken represents the angle between the ellipse major axis and the horizon, which is oriented along the pLLPs horizontal midline. Interestingly, the results show an increase in cell count in cells oriented along the horizontal midline in _shroom3_$^{-/-}$ embryos which suggests a shift of organ morphology from a radial to a more wedge like shape (section \@ref(sum-ai)). Since a wedge is also anisotropically apically contricted, this is also in line with the first results. In the range of 30 - 45$^\circ$ we see the opposite ratio, where now the count of cells in _shroom3_$^{-/-}$ embryos is reduced, suggesting that those cells are under more stress to apically bind to the rosette center. Again, this result is in favor of the theory that Shroom3 acts as a stabilizer against cell stretching and maintenance of radial cell arrangements under extreme conditions.

```{conjecture, echo = TRUE}
Organ morphology shifts from a radial, to a more wedge-like shape.
```

### Hair Cell specification

In the pLLP, Notch signaling is important for selection and specification of hair-cells (section \@ref(intro-notch)). For our observation this could mean that (1) either expression of _atoh1a_ is up-regulated due to a compensation mechanism to ensure CC deposition or (2) that the cellular rearrangements lead to a _bias_ in Notch signaling. A phenomenon that in fact has been reported [@Shaya2017a]. For the latter, the proposed model would be that in each micro-rosette the HC is in contact with less cells due to a more wedge-like organ morphology, but the amount of Notch ligand stays the same. Therefore, lateral inhibition and feedback is stronger, leading to an increase in expression of _atoh1a_ and ultimately to premature CC deposition.

```{conjecture, echo = TRUE}
Expression of atoh1a appears pre-mature and more frequent
```

Even though the dataset for hair-cell specification provides a solid base for hypothesis testing, the count of Atoh1a expressing cells shows only a relatively low significance difference. This could be attributed to an imprecision in measurements. However, the same dataset also shows an earlier activity of the _atoh1a_ promotor \@ref(fig:hctl)B-C, which is independent of cell count and makes the results more trustworthy. Furthermore, the control experiment of rescuing the _shroom3_$^{-/-}$ phenotype of more deposited CCs _via_ an _atoh1a_ knockdown shows that neither in _shroom3_$^{-/-}$ nor in _shroom3_$^{+/+}$ embryos the first NM is deposited anymore revealing a possibly novel role for Atoh1a that has not been reported before. Arguing - in theory - for an Atoh1a dependent mechanism of CC deposition. Furthermore, even though an MO for p53 was co-injected, MO injection has been shown to potentially have many more unspecific effects due to the high concentrations usually injected [@Schulte-Merker2014c]. For this reason double mutants ( _shroom3_$^{-/-}$; _atoh1a_-\-) were generated - which at this timepoints are not analyzed yet.

### Model

Based on my results the phenotype of the _shroom3_$^{-/-}$ pLLP is defined by a greater amount of rosettes that are smaller and are less pronounced. In this model Shroom3 stabilizes the cell clusters as more cells are integrated and more traction forces are built up. Without this stabilizing force the cell clusters could not aggregate above a certain threshold, which results in 'micro-rosettes'. In those micro-rosettes the cells are more wedge-like shaped and are less densely packed, which results in pre-mature hair cell specification and deposition. The results are summarized and graphically modeled in figure \@ref(fig:summodel).

(ref:summodel) Shroom3 dependent rosette formation. First morphogen FGF binds to FGFR1, leading to expression of _shroom3_, interaction with Rock and actin network contraction through phosphorylated NMII. Without Shroom3 there's only a partial contraction of the action network which leads to smaller and more rosettes. Altered cellular morphology then leads to a premature expression of _atoh1a_ and hair-cell specification.

```{r summodel, out.width = '95%', fig.pos = 'H', fig.cap = "(ref:summodel)", fig.scap = "Shroom3 dependent rosette formation"}
knitr::include_graphics("figures/summary/CurrentModel_new-01.png")
```