Neural crest cells bulldoze through the microenvironment using Aquaporin 1 to stabilize filopodia

Cell migration is essential during embryogenesis in order to gastrulate, elongate the vertebrate axis and distribute cells into the periphery to contribute to organ development. Despite the importance of cell migration to human development and disease, it is still unclear what mechanisms enable cells to invade the dense extracellular matrix (ECM), mesoderm and other cell types characteristic of the embryonic microenvironment. The complexity of the embryonic microenvironment means that invading cells must rapidly change cell shape, form and sustain protrusions that penetrate different sized gaps, and attach to and remodel the ECM. Thus, there is a tremendous need to identify and test the function of molecules that are crucial to embryonic cell migration, and to better understand their mechanistic basis.

Neural crest cell migration is one of the most prevalent examples of how cells efficiently distribute throughout the growing vertebrate embryo to precise targets. In the head, cranial neural crest cells must invade through dense ECM, loosely connected mesoderm and migrating endothelial cells. Yet it has remained unclear how the migrating neural crest cells that first encounter the embryonic microenvironment penetrate small gaps between mesodermal cells and degrade the ECM to move in a directed manner to peripheral targets. By combining dynamic in vivo imaging and super resolution microscopy with gain- and loss-of-function experiments, we are poised to examine the function of genes presumed to be crucial to neural crest cell migration. Thus, the embryonic neural crest is an attractive in vivo model for studying the function of cell invasion genes in mechanistic detail.

Using single cell RT-qPCR and transcriptome profiling, we discovered the enhanced expression of several genes in the most invasive chick cranial neural crest cells, including Aquaporin 1 (AQP-1) (McLennan et al., 2015a; Morrison et al., 2017a). AQP-1 is a transmembrane channel protein that facilitates the flux of water across the plasma membrane (Agre et al., 1993) and is one member of a family of at least 13 aquaporins (Ishibashi et al., 2011). AQP-1 has been detected in several aggressive human cancers and its expression correlates with poor disease prognosis (Tomita et al., 2017; De leso and Yool, 2018). However, the mechanistic basis of AQP-1 function is still unclear because studies have been limited to analyzing cell behaviors using in vitro assays. This has led to the generation of several distinct hypothetical mechanisms for AQP-1 function. For example, AQP-1 is thought to allow cells to rapidly change cell volume to form thin filopodial protrusions that squeeze in between neighboring cells (Papadopoulos et al., 2008; Verkman, 2009; Karlsson et al., 2013) or to collapse a cell protrusion and cause it to retreat from a repulsive signal (Cowan et al., 2000). Alternatively, AQP-1 may function to stabilize a filopodia as aquaporins have been shown in vitro to localize to the front end of migrating cells (Saadoun et al., 2005) and are speculated to function as a cell motility engine (Condeelis, 1993; Stroka et al., 2014). Thus, a better understanding of the in vivo function of AQP-1 and its connection to cell guidance signaling would be beneficial to studies in cell migration and invasion in cancer and developmental biology.

In this study, we examined the expression and function of AQP-1 during chick cranial neural crest cell migration and possible upstream guidance and downstream AQP-1 effectors. We first characterized AQP-1 mRNA and AQP-1 protein expression within migrating cranial neural crest cells using three dimensional (3D) confocal and super resolution microscopy. We examined the role of AQP-1 in vitro and in vivo by measuring changes to the neural crest cell migratory pattern, individual cell behaviors, and filopodial dynamics after gain- and loss-of-function of AQP-1. To test our hypothesis of a ‘bulldozer’ role for AQP-1-expressing neural crest cells, we analyzed focal adhesion activity, integrin localization and ECM degradation after AQP-1 manipulation. We examined the possible colocalization of Eph guidance receptor expression with AQP-1 in the same migrating neural crest cells using multiplexed fluorescence in situ hybridization. Finally, we use computer simulations to predict what AQP-1 related functions are likely to enhance cell migration. Together, our data demonstrate a crucial role for AQP-1 during cranial neural crest cell migration and offer a mechanistic basis for in vivo AQP-1 function to promote cell invasion.

To detect the expression of AQP-1 mRNA in vivo specifically within migrating cranial neural crest cells, we took advantage of our recently optimized integrated protocol combining RNAscope multiplexed fluorescence in situ hybridization, immunohistochemistry and tissue clearing with FRUIT (Morrison et al., 2017b). This enabled the detection of AQP-1 mRNA expression in migrating HNK-1-labeled neural crest cells (Fig. 1A-C). By spot counting the number of AQP-1 transcripts per neural crest cell, we found that AQP-1 expression was visible in migrating cranial neural crest cells observed at Hamburger and Hamilton stage (HH) 13 (Hamburger and Hamilton, 1951) en route from rhombomere 4 (r4) to the second branchial arch (BA2) (Fig. 1C; neural crest cells are color-coded to show number of AQP-1 counted transcripts). In a typical r4 neural crest cell migratory stream, the highest numbers of detected transcripts of AQP-1 were found to be in the cell subpopulation at the front of the migratory stream (Fig. 1C). We determined this by quantifying the fluorescence signal of AQP-1 mRNA expression as a function of distance along the migratory stream (from the neural tube towards BA2) in a large number of cells/embryos (Fig. 1D). This confirmed that lead neural crest cells showed a significantly higher level of AQP-1 expression compared to cells further back in the invading stream (Fig. 1D).

Whether AQP-1 protein is expressed within subregions of individual migrating neural crest cells is difficult to determine because it is challenging to resolve fluorescence signal in thin (1-2 µm wide) protrusions. Structured illumination microscopy (SIM) has emerged as an excellent tool for resolving diffraction-limited issues by increasing optical resolution (Gustafsson et al., 2008). When we applied SIM to visualize AQP-1 protein location within migrating neural crest cells in vitro, we discovered AQP-1 was present on the cell membranes, including the tips of filopodia (Fig. 1E). To determine AQP-1 protein expression in vivo, we analyzed AQP-1 protein by immunohistochemistry within HNK-1-labeled migrating neural crest cells and confirmed AQP-1 expression throughout the migratory stream (Fig. 1F,G; HH11-15). Furthermore, by quantifying fluorescence intensity in individual leader and follower neural crest cells, we determined that there is a higher level of AQP-1 protein in the lead neural crest cells in vivo (Fig. 1H). These data suggest that AQP-1 may be influencing dynamic cellular shape changes by acting as a water channel at the leading edge of the migratory stream.

To begin to determine the function of AQP-1 in cranial neural crest cell migration, we explanted cranial neural tubes in an in vitro assay and measured changes in cell migratory behaviors with confocal time-lapse microscopy (Fig. 2A,B). We took advantage of acetazolamide (AZA), a chemical inhibitor of AQP-1 function (Bin and Shi-Peng, 2011; Zhang et al., 2012; Cai et al., 2018; Ameli et al., 2012). Although AZA can also affect AQP-4 function, AQP-1 is the only aquaporin significantly expressed by the invasive front of the migratory stream and therefore the only aquaporin significantly affected by AZA in this assay (Huber et al., 2007; McLennan et al., 2015a; Morrison et al., 2017a). When we inhibited AQP-1 function in vitro, we observed that the migrating neural crest cells moved significantly more slowly than control neural crest cells (in the presence of DMSO) (Fig. 2B,C; Movie 1). Representative tracks show that neural crest cells exposed to AZA remain closer to the neural tube explant and to other cells when compared with controls (Fig. 2E).

To determine changes in cell migration after AQP-1 gain of function, we overexpressed AQP-1 in premigratory neural crest cells [by transfection with AQP-1 full-length construct (FL)] and again explanted neural tubes in culture (Fig. 2A,B). AQP-1 overexpression resulted in neural crest cells that moved significantly faster when compared with non-transfected cells in the same culture (Fig. 2B,C; Movie 1). Neural crest cells overexpressing AQP-1 were also significantly faster when compared with EGFP-only (pMES) transfected cells from different cultures that were prepared and imaged concurrently (Fig. S1D). Typical cell tracks from gain-of-function of AQP-1 experiments confirmed that neural crest cells moved further away from the neural tube explant in the same amount of time (Fig. 2F). Interestingly, when AQP-1 was overexpressed, cells migrated with increased straightness compared with their controls (Fig. 2D). These in vitro data suggest that AQP-1 significantly influences the speed at which neural crest cells migrate.

To determine the in vivo function of AQP-1, we first knocked down AQP-1 function in premigratory cranial neural crest cells by morpholino (MO) transfection. Similar to in vitro experiments, AQP-1 MO-transfected neural crest cells did not travel as far as control MO-transfected neural crest cells (Fig. 3A,C,D). In AQP-1 MO-transfected embryos, we found fewer migrating transfected neural crest cells in BA2 compared with control MO-transfected embryos (Fig. 3B). To further verify the in vivo loss of AQP-1 function in neural crest cell migration, we microinjected AZA into the paraxial mesoderm adjacent to r4, prior to neural crest cell exit from the dorsal neural tube. The AZA-injected sides of embryos showed fewer migrating neural crest cells moving into BA2 compared with the control sides of the same embryos (Fig. 3F) and control DMSO-injected embryos, as seen with immunolabeling of neural crest cells using HNK-1 (Fig. 3G). We also measured fewer neural crest cells present in BA2 when injected with AZA, as quantified by the percentage of HNK-1 immunostaining signal in BA2 (Fig. 3E). The furthest distance migrated after AZA injection was the same as control, presumably due to cells escaping AZA, to dilution/degradation of AZA over time or to differences in how AZA affects AQP-1 activity when compared with a morpholino (Fig. S1E). There was no significant difference in HNK-1 immunostaining in the control embryos when comparing the control side to the DMSO-injected side (Fig. S1F). These data suggested that AQP-1 expression, and therefore function, promotes efficient neural crest cell invasion in vivo.

To further investigate the role of AQP-1 in neural crest migration in vivo, we overexpressed AQP-1 (via transfection with AQP-1 FL) in premigratory cranial neural crest cells. Overexpression of AQP-1 caused neural crest cells to migrate further than control pMES-EGFP transfected neural crest cells at both 12 h (Fig. 3H,I) and 16 h (Fig. 3J,K) after transfection. Measurements confirmed a significant increase in the distance migrated by AQP-1 FL transfected neural crest cells when compared with pMES transfected neural crest cells analyzed at the same time points (Fig. 3L,M). To determine whether the increased distance migrated when AQP-1 was overexpressed was due to an increase in cell speed, we performed in vivo time-lapse confocal imaging of the migratory streams at 5 min intervals after control (Gap43-YFP) and AQP-1 FL transfections (Fig. 3N,O; Movie 2). Cell tracking confirmed that neural crest cells migrated further and faster within the invasive front (front 20% of the migratory stream) when transfected with AQP-1 FL than controls (Fig. 3P,S; Movie 2). Interestingly, when neural crest cells were transfected with AQP-1 FL, they also displayed a decrease in cell directionality when compared with pMES transfected neural crest cells (Fig. 3Q). To indicate whether AQP-1 may be playing a role in delamination, the number of transfected cells was counted at 16 h. Although there were statistically more AQP-1 transfected neural crest than pMES, the range of cell numbers varied greatly and may be due to differences in transfection efficiency as opposed to a role in delamination (Fig. 3R). Together, these data suggest that AQP-1 is crucial to the directed neural crest cell invasion but they do not reveal the mechanistic basis underlying this AQP-1 function.

To begin to address the mechanistic basis of AQP-1 function during cranial neural crest cell migration, we first perturbed AQP-1 function and used spinning disk confocal microscopy on the lead neural crest cells in vivo to observe any rapid changes in cell morphology and behavior. Specifically, we collected z-stacks at 20 to 30 s intervals of neural crest cells transfected with Gap43-mTurquoise2 as well as AQP-1 MO (knockdown), AQP-1 FL (overexpression) or pMES (control) (Fig. 4A,B,J; Movie 3). Images of transfected cells were first processed by creating a binary mask of single cells in ImageJ using the membrane-bound Turquoise2 signal. We then used the CellGeo software (Tsygankov et al., 2014) and additional calculations to identify and measure filopodia length, position, angle and survival time (Fig. 4C-E). With knockdown of AQP-1, we find a significant reduction in the number and length of filopodia in migrating neural crest cells (Fig. 4F,G). In addition, filopodia in cells lacking AQP-1 retracted at significantly faster rates than control cells, leading to less stable filopodia with a shorter survival time (Fig. 4H,I). In contrast, lead neural crest cells in which AQP-1 was overexpressed have enhanced stable filopodia with reduced protrusion and retraction rates, and a longer survival time (Fig. 4H,I). The average number of filopodia per cell was reduced when AQP-1 was overexpressed but their length was unchanged (Fig. 4F,G).

We also examined the direction in which the filopodia extended relative to the direction of migration of the cell. A vector was created from the base of the filopodia to the tip and the angle to the direction of motion was calculated (Fig. S2C). The control filopodia were roughly bi-polar in distribution, pointing both towards and away from the direction of motion, as shown on the rose plot, although the Rayleigh test does indicate a uniform distribution (P=0.2) (Fig. S2A). With overexpression of AQP-1, the filopodia were found to point randomly in all directions (P=0.22) (Fig. S2A). After knockdown of AQP-1 by MO, filopodia on average were directed −60° rostral from the direction of motion roughly towards branchial arch 3 (BA3) (P=4.8×10⁻⁶) (Fig. S2A). We considered the possibility that the longest filopodia may be more important for directional migration by comparing the length of the filopodia found in a 30° window around the direction of motion to the length of filopodia in all other directions. For control cells, the filopodia in the direction of migration were significantly longer than the remaining filopodia (P=5.9×10⁻⁹) (Fig. S2B). However, this association is lost in both overexpression and knockdown of AQP-1, and the filopodia in the direction of motion are no longer than filopodia elsewhere in the cell (P=0.31 and P=0.2, respectively) (Fig. S2B). These data strongly suggest that AQP-1 functions to stabilize filopodia in migrating neural crest cells; results in the literature suggest that this is likely due to water flux (Papadopoulos et al., 2008; Karlsson et al., 2013; Saadoun et al., 2005).

To further explore the role of AQP-1 in neural crest cell invasion, we examined whether AQP-1 is involved in two important aspects of cell migration: cell–ECM attachment and ECM degradation. First, we overexpressed AQP-1 in migrating neural crest cells, isolated those cells from developing embryos and performed RNAseq. The sequencing results were compared with those from neural crest cells transfected with pMES only. Significant pathways affected by AQP-1 expression include many guidance signaling pathways, such as semaphorins, FGFs, neuregulin and ephrins, as well as the actin cytoskeleton (Fig. 5A). Members of the AQP family have been shown to promote migration by facilitating the recycling of integrins and therefore the turnover of focal adhesions that are necessary for cell invasion (Chen et al., 2012). When AQP-1 is overexpressed, integrin A8 and integrin B5 were significantly increased (Table S1). Therefore, we performed immunohistochemistry for HNK-1, AQP-1 and phosphorylated focal adhesion kinase (pFAK) on cranial neural crest cells in migrating r4 neural crest in vivo. Confocal imaging with Airyscan detection revealed a colocalization between the puncta of pFAK and AQP-1 (Pearson's coefficient 0.74), indicating a very close physical proximity of these two proteins (Fig. 5C). To further investigate the possible link of AQP-1 with pFAK and integrins, we performed immunohistochemistry for HNK-1, pFAK and integrin B1 after transfection with AQP-1 FL (overexpression) or pMES (control) in vitro. Using Airyscan confocal microscopy and spot detection analysis, AQP-1 overexpression resulted in fewer pFAK puncta, as well as fewer integrin B1 puncta on or very near neural crest cell membranes. RNAseq analysis showed that integrin B1 RNA levels did not change when AQP-1 was overexpressed (Table S1). Therefore, this result suggests increased turnover of integrin B1 via internalization, which is consistent with previous literature (Chen et al., 2012), or dissociation of the integrin heterodimers as well as internalization or loss of phosphorylation of FAK (Fig. 5D).

From the RNA-seq data, we also focused our attention on the ECM-related genes, as AQP-1, AQP-3 and AQP-4 have all been shown to be positive regulators of the ECM-degrading matrix metalloproteases (MMPs) (Chen et al., 2015; Ding et al., 2011; Jiang et al., 2014; Wei and Dong, 2015; Xiong et al., 2017; Xu et al., 2011). When AQP-1 is overexpressed in migrating neural crest cells, MMP9 as well as members of the ADAMs, including ADAM33, were significantly increased (Fig. 5B). To test whether AQP-1 is upstream of MMP activity, we measured MMP activity from neural crest cells using an in-gel zymography assay. Specifically, neural tube cultures were exposed to either AZA (an AQP-1 inhibitor), GM6001 (MMP inhibitor) or used as controls. The medium from these cultures was then extracted to determine MMP activity. Neural crest cells exposed to AZA exhibited a reduction of MMP activity of 73% at a size that most likely matches MMP2 (Fig. 5E; Fig. S1G). In comparison, neural crest cells exposed to GM6001 exhibited a reduction of MMP activity of 83%, which was our positive control to show that the assay was functional (Fig. 5E; Fig. S1G). This result was verified with an in vitro degradation assay, where neural tubes (transfected with AQP-1 FL or pMES) were plated on top of fluorescently labeled gelatin (Garmon et al., 2018). After 40 h of incubation, the assays were imaged, and average fluorescent intensity of the gelatin was quantified as a read out of degradation (Fig. 5F). Neural tubes transfected with pMES significantly degraded gelatin compared with control wells where no neural tubes were plated as expected (Fig. 5F,G). Transfection with AQP-1 FL resulted in even higher gelatin degradation, which is consistent with the in-gel zymography (Fig. 5F,G). These data suggest that AQP-1 is involved in the promotion of integrin turnover and ECM degradation.

Previous work in commissural axon guidance in mouse has shown evidence for the co-immunoprecipitation of AQP-1 and Eph receptors, namely EphB2, to form a stable complex within a cell (Cowan et al., 2000). As Eph receptors and ephrins are implicated in axon pathfinding, Cowan and colleagues speculated that aquaporins might function in the growth cone to help integrate the guidance information elicited by Eph-ephrin clustering (Cowan et al., 2000). Based on this study, we asked whether the Eph receptors EphB1 and EphB3, previously shown to be highly expressed in the most invasive cranial neural crest cells (McLennan et al., 2015a), were colocalized with AQP-1 in the same migrating cranial neural crest cells. To address this, we used multiplexed fluorescence in situ hybridization (RNAscope) to label AQP-1, EphB1 and EphB3 mRNA expression, and HNK-1 immunohistochemistry to quantify co-expression of these three mRNAs in cranial neural crest cells in vivo (Fig. 6A,B). As before, the HNK-1 channel was used to segment r4 neural crest from whole-mount HH13 embryos (white outlines, Fig. 6A,B) and spots of AQP-1, EphB1 and EphB3 mRNA transcripts were counted per cell. When we quantified the mRNA expression, we found that AQP-1 mRNA was only found in cells also expressing EphB1 and/or EphB3 mRNA although there were many EphB1- and/or EphB3 mRNA-positive cells without AQP-1 mRNA (Fig. 6C). When we quantified the average number of detected transcripts in migrating neural crest cells, we found that cells expressing AQP-1 mRNA had a higher number of EphB1 and EphB3 mRNA transcripts than their AQP-1 negative counterparts (Fig. 6D). By immunohistochemistry, we could also verify that AQP-1 protein and EphB1/B3 protein co-label the same neural crest cells in vivo (Fig. 6E-H).

To explore the hypothesis that AQP-1 is downstream of neural crest cell guidance, we performed in vitro cultures where migrating neural crest cells were exposed to BA2 tissue, a known source of guidance signals (Fig. 6I; Movie 4). When AQP-1 was overexpressed in neural crest cells, they were less directed than untransfected controls in the same cultures or in pMES-only transfected neural crest cells (Fig. 6I, J; Movie 4). This finding is consistent with the loss of directionality observed when AQP-1 was overexpressed in neural crest cells in vivo (Fig. 3Q) and with the loss of polarized filopodia in the direction of migration (Fig. S2B). One possible explanation for this result is that by overexpressing AQP-1, motility is increased due to more AQP-1 channels, but these channels are no longer focused by complexing with guidance receptors. Overall, these data suggest that AQP-1 is involved in the fast response of cells to guidance factors, possibly by complexing with guidance receptors, including EphB receptors.

To test hypothetical mechanistic scenarios by which AQP-1 functions during neural crest cell invasion, we used a hybrid computational model of cranial neural crest invasion. The model consists of a discrete, off-lattice model for the cell dynamics that is coupled to a continuum, reaction-diffusion model of the dynamics of a known cranial neural crest cell chemoattractant (McLennan et al., 2010; VEGF) on a growing domain. The model is based on our previous studies (McLennan et al., 2015a,b, 2012, 2017) and is fully described in the supplementary Materials and Methods. Briefly, we used a two-dimensional approximation of the neural crest cell migratory domain and incorporated finite size effects by considering cells as hard disks that are not allowed to overlap (Table S2; Fig. S3). We used a fixed time step model (with a time step of 1 min) during which a cell senses its environment and moves accordingly. We considered two subpopulations of cells, namely leaders and followers; leaders undertake a biased random walk up a cell-induced gradient of chemoattractant (Table S2).

In this study, we modified five parameters, or features, of our model to simulate the mechanisms by which AQP-1 influences cell migration: cell speed, filopodia stability, filopodia polarity, filopodia number and ECM degradation (Fig. 7A). Filopodia stabilization is realized by a leader sampling the microenvironment only every three time steps and moving persistently in between, as opposed to sampling and potentially moving in a different direction at each time step. Filopodia polarization means that a leader only stabilizes its filopodium when a cell makes an informed movement towards a higher concentration of VEGF as opposed to a random movement. Filopodia number is the number of random directions in which a leader samples the concentration of chemoattractant per time step. As our experimental results showed that overexpression of AQP-1 induced an increase in MMP activity that resulted in ECM degradation (Fig. 5E-G), we introduced into the model the creation of tunnels in the ECM (Fig. 7B). This was implemented by recording the histories of leader positions and defining them as tunnels. Finally, the invasion of cells was assumed to occur on a uniformly growing rectangular domain, the growth of which was based on biological measurements (McLennan et al., 2015a, 2012).

In our computational model (supplementary Materials and methods), we initially quantified changes in the dynamics of cells when we recapitulated the effects of overexpression and downregulation of AQP-1 by modulating relevant model parameters. Different combinations of parameter perturbations are labeled as Model x (where x can be from 1 to 11); the values of parameters for different models can be found in Fig. 7A. In model simulations, we measured the density of cells along the two-dimensional migratory domain and the likelihood of migratory stream breakage, defined as the proportion of cells not in chains or tunnels at the end of a simulation (Fig. 7D). In the control case, the cell speed was set to 62 µm/h, filopodia were not stabilized and there was a representative number of three filopodia per cell (Model 1). To simulate AQP-1 overexpression, the cell speed was increased by ∼30% to 82 µm/h, filopodia were stabilized and there were two filopodia per cell (Model 2). To simulate inhibition of AQP-1, the cell speed was decreased to 57 µm/h, cells had only a single filopodium, which was not stabilized (Model 3). Initially (Models 1-9), we did not incorporate ECM degradation (i.e. we did not include the tunneling mechanism). For each model, simulations were run ten times with each parameter set, and the results at t=24 h were averaged to account for random variations. As expected, cells at the front traveled farther when their speed was increased (Fig. 7C; Model 2). Cells traveled shorter distances when their speed was reduced to simulate AQP-1 knockdown (Fig. 7C; Model 3). We did, however, observe an unexpected phenomenon; when the cell speed was increased, the simulated cell migratory streams broke apart more frequently and the model did not sustain collective cell migration (Fig. 7D). Migratory stream breakage is not observed in vivo and therefore this observation led us to further investigate cell speed versus filopodia dynamics and the effects upon collective cell migration in silico.

In our simulations, when cell speed was increased, with or without filopodia stabilization, cells at the front traveled farther compared with the control case (compare Model 4 and Model 5 with Model 1) but stream breakage was much more likely to occur (Fig. 7D,E). The number of filopodia (two or three) affected the likelihood of stream breakage at control cell speeds, with more filopodia resulting in a higher likelihood of breakage but with little impact upon the distance migrated (compare Model 6 with Model 7) (Fig. 7D,E). At higher speeds, whether cells had two or three filopodia did not significantly affect the probability of stream breakage or the distance the cells migrated (compare Model 2 with Model 5) (Fig. 7C-E). Stabilization of filopodia increased the farthest distance traveled by cells regardless of the cell speed. It also noticeably reduced the likelihood of stream breakage at control cell speeds, with a smaller effect at enhanced speeds (compare Model 1 with Model 6, and Model 4 with Model 5) (Fig. 7D,E). Additionally, polarization of filopodia had an insignificant effect on the migration pattern (compare Model 7 with Model 8 and Model 2 with Model 9) (Fig. 7D,F).

As our experimental findings showed that increasing AQP-1 activity led to an increase in MMP activity, and thus to an increase in degradation of the ECM, we included the formation of tunnels by leader cells in silico. We found this had no significant effect on distance the cells migrated; however, it did dramatically decrease the percentage of follower cells not in chains to zero (compare Model 1 with Model 10) (Fig. 7D,G). In contrast, to simulate overexpression of AQP-1 in the model and include the effects of increased MMP activity, we increased cell speed, filopodia were stabilized and decreased in number from three to two, and the leaders generated tunnels for the followers. In this scenario, our model predicted that the stream is unlikely to break up and invasion is very robust (compare Model 2 with Model 11) (Fig. 7D,G). Overall, these simulations support the experimental observations that AQP-1 influences cell migration and invasion in multiple ways (Fig. 7H; Movie 5).

Our discovery of high AQP-1 expression within lead cells of the chick cranial neural crest cell migratory stream (McLennan et al., 2015a, 2017a) led us to examine the function of this water channel protein. We initially confirmed that AQP-1 mRNA and protein are both higher in lead cranial neural crest cells in vivo, and we used state-of-the-art imaging to show that AQP-1 expression is localized to neural crest cell membranes, including cell filopodia. We tested the hypothesis that AQP-1 regulates cell migration to promote neural crest cell invasion using gain and loss of function of AQP-1, and quantification of cell dynamics obtained from time-lapse imaging sessions. We elucidated four important aspects of AQP-1 function that support its crucial in vivo role in neural crest cell migration.

First, we discovered that modulation of AQP-1 expression altered neural crest cell motility and invasive ability. An increase in AQP-1 expression in premigratory cranial neural crest cells resulted in higher cell speeds in vitro and enhanced invasion in vivo (Figs 2 and 3). In contrast, the knockdown of AQP-1 function by MO transfection or microinjection of the chemical blocker AZA into the neural crest cell migratory pathway resulted in slower cell speeds in vitro and reduced invasion in vivo (Figs 2 and 3). These results are consistent with previous data that imply a role for AQP-1 in cell migration and invasion across a wide variety of adult mouse and human cancer cell types (Saadoun et al., 2005; Chen et al., 2012, 2015; Wei and Dong, 2015; Xiong et al., 2017; Cao et al., 2006; Hu and Verkman, 2006; Klebe et al., 2015).

Second, we found that AQP-1 promotes neural crest cell invasive ability by stabilizing filopodia, supporting a function for AQP-1 in enabling cells to ‘bulldoze’ through the embryonic microenvironment. Initially, we hypothesized that AQP-1 functions to promote neural crest cell invasion by: (1) rapidly changing the cell shape to permit the cell to infiltrate between gaps in loosely connected mesoderm and dense ECM through which cells travel; and/or (2) stabilizing cell filopodia to allow them to break down and displace the surrounding tissue. In support of (2), we found a significant reduction in the number and length of neural crest cell filopodia after AQP-1 knockdown (Fig. 4). Furthermore, neural crest cell filopodia retract faster and survive for shorter durations in cells with reduced AQP-1 function (Fig. 4).

We did not observe overall cell volume changes in migrating neural crest cells in vivo by dynamic analysis of three-dimensional cell volumes measured from time-lapse imaging sessions (data not shown). Given that changes in cell size that are induced by alterations in external osmolarity have been linked to AQP-1 function in human vascular smooth muscle cells in culture (Shanahan et al., 1999), we cannot completely rule cell volume changes out because detecting changes in the dynamic filopodia volume of migrating neural crest cells may require higher resolution microscopy. Furthermore, a similar volume of water may flow into or out of control and AQP-1-perturbed neural crest cells, as perturbations are not localized to distinct regions of individual cells.

Third, we learned that AQP-1 directly influences neural crest cell adhesion and ECM degradation, suggesting a mechanistic basis for neural crest cell ‘bulldozing’ through the microenvironment. In support of this, we showed that AQP-1 and pFAK are colocalized in migrating neural crest cells, and upon AQP-1 overexpression, fewer pFAK and integrin B1 puncta are present on cell surfaces (Fig. 5). This implicates AQP-1 in the integrin-mediated focal adhesion signaling pathways previously shown to be important for neural crest cell migration (Desban and Duband, 1997; Desban et al., 2006; Parsons, 2003). Furthermore, it has been previously shown that AQP-1 regulates FAK expression in bone marrow mesenchymal stem cells (Meng et al., 2014). FAK binds directly to integrin B1, and AQP-2 has been shown to internalize integrin B1 (Chen et al., 2012; Lechertier and Hodivala-Dilke, 2012). Therefore, in migrating neural crest cells, less integrin B1 and pFAK on cell surfaces when AQP-1 is overexpressed may indicate increased integrin turnover and less cell adhesion (Fig. 5). AQP-1 perturbation affects the expression and activity of MMPs in migrating cranial neural crest cells when measured by RNA-seq profiling, in-gel zymography and a degradation assay (Fig. 5). Whether pFAK colocalization with AQP-1 in migrating neural crest cells also plays a role in initiating downstream intracellular signals, including Eph-ephrins, is unclear (Carter et al., 2002; Miao et al., 2000) and will be the focus of future studies. Together, these data clearly demonstrate that AQP-1 promotes neural crest cell invasion by influencing FAK activity, integrin turnover and ECM degradation.

Fourth, we showed that AQP-1 is involved in directed neural crest cell migration, suggesting that AQP-1 is downstream of the ability of a cell to readout guidance factors in the local microenvironment. AQP-1 and EphB receptors have previously been shown to be in the invasive front of the neural crest migratory stream (McLennan et al., 2015a), and here we show that they colocalize in the same migrating neural crest cells at both the RNA and protein levels (Fig. 6). Neural crest cell directionality in response to endogenous guidance cues is reduced when AQP-1 is overexpressed both in vitro and in vivo, as presumably the majority of the AQP-1 water channels are no longer downstream of guidance signals (Figs 3 and 6). Furthermore, AQP-1 overexpression led to neural crest cell filopodia extending in random directions (Fig. S2). It is not clear whether AQP-1 is enriched in a directional manner, because although there is higher AQP-1 protein expression in lead migrating neural crest cells in vivo (Fig. 1H), our super-resolution microscopy that revealed AQP-1 protein expression in neural crest cell filopodia was performed in vitro in the absence of endogenous guidance cues (Fig. 1E). Future experiments that dissect the AQP-1 and EphB signaling relationship and explore other guidance receptors we have previously identified by profiling migrating cranial neural crest cells (McLennan et al., 2015a,b) will help to shed light on this.

Last, in support of AQP-1 function, our computational model simulations identified cell speed, filopodia stabilization and ECM degradation as key parameters for controlling neural crest cell invasion (Fig. 7). When we increase or decrease cell speed to simulate AQP-1 gain or loss of function, respectively, we observe the intuitive changes in the maximum distance traveled by cells (Fig. 7; Models 2, 3; Movie 5). However, unexpected breakdown of the cell migratory stream occurs in silico when the cell speed is increased; a pattern that is not observed in vivo (Fig. 3). Examination of cell speed versus filopodial dynamics revealed that stabilization of filopodia rather than filopodia number or polarity increased the furthest distance traveled by cells (regardless of speed) and reduced the likelihood of stream separation (Fig. 7). Thus, our computational model simulations support a role for AQP-1 in promoting neural crest cell invasion through ECM degradation, stabilization of filopodia and regulation of cell speed.

In summary, our findings implicate four AQP-1-related processes in neural crest cell migration: modulation of cell speed, stabilization of filopodia, cell adhesion and ECM degradation, and co-localization with cell guidance receptors (Fig. 8). Whether these functions are dependent or independent of AQP-1 functioning as a water channel is currently unknown; however, given the vast recent literature on AQP-1 it is likely to be related to its water transport functions. By stabilizing filopodia, increasing MMP-mediated ECM degradation and controlling adhesions in response to guidance receptors, lead neural crest cells ‘bulldoze’ through the embryonic microenvironment in a directed manner (Fig. 8). Later-emerging neural crest cells are able to follow and collective migration is maintained. Computational modeling has identified cell speed, filopodia stabilization and ECM degradation as key parameters for promoting neural crest cell migration. The combination of computational modeling with an in vivo dynamic imaging platform with single cell resolution provides a powerful tool with which to examine both the upstream regulation of AQP-1 activity and the downstream signaling events in directed cell migration. With the established high correlation of AQP-1 expression and cancer cell aggressiveness (Tomita et al., 2017; De leso and Yool, 2018), our data relating to in vivo AQP-1 function during embryonic neural crest cell migration and the mechanistic basis of cell bulldozing offer further details on, and illustrate the importance of, the function of aquaporins in development and cancer.

White Leghorn, fertilized chicken eggs (Centurion Poultry, Lexington, GA, USA) were incubated in a humidified incubator at 38°C to the desired developmental stage (Hamburger and Hamilton, 1951).

RNAscope on whole chick heads was performed as previously described (Morrison et al., 2017b). Briefly, HH13 chick embryos were harvested and fixed in 4% paraformaldehyde for 2 h. Following a dehydration gradient of methanol, embryos were stored for 4 days at −20°C before the RNAscope protocol began. After rehydration, embryos were digested in a 0.1× dilution of protease solution in PBS-Tween for 6 min at 22°C with gentle agitation. AQP-1, EphB1 and EphB3 RNAscope probes were designed by Advanced Cell Diagnostics against the relevant gene (GenBank Accession Numbers NM_001039453.1, NM_205035.1 and XM_422762.4, respectively). AQP-1 was labeled with Cy3, EphB1 with Atto488, and EphB3 was labeled with Cy5 dyes. As a control, embryos were processed as described but no probes were used (Fig. S1A). After RNAscope, immunohistochemistry was performed with a 4% molecular biology-grade BSA blocking buffer using HNK-1 primary antibody (TIB-200 hybridoma cell line, ATCC Cell Lines) at a 1:25 concentration (overnight incubation at 4°C). Secondary antibody Alexa Fluor 594 goat anti-mouse IgM (A21044, ThermoFisher Scientific) was used at 1:500 overnight at 4°C. The embryos were then cleared with FRUIT clearing buffer (Hou et al., 2015) at 60% followed by 80% overnight at 4°C.

Immunohistochemistry on cryosections or whole-mount chick heads was performed with either anti-Phospho-FAK (Tyr861) (44-626G, ThermoFisher Scientific) at 1:200, anti-AQP-1 (Millipore, AB3272) at 1:50 or anti-EphB1/3 (sc-926, discontinued product, Santa Cruz Biotechnology) at 1:100. Generally, tissue was fixed in 4% PFA overnight at 4°C. After washing in PBS, tissue was blocked with 4% BSA, 0.1% Tween block for 2 h at room temperature followed by overnight incubation with primary antibodies at 4°C. After washing in 0.1% Tween block, secondary antibodies were applied at 1:500 overnight at 4°C. After a final wash with PBS, whole heads were mounted for imaging and cryosections were coverslipped using Vectashield HardSet Antifade Mounting Medium with DAPI (H-1500, Vector Laboratories). The z stacks of the collected images were loaded into Imaris 9.20. The images were segmented using the surface function with HNK-1 setting (settings: grain size at least 0.439 µm, largest sphere 1.65 µm, threshold automatically detected at upper end, lower end 110, estimated diameter 6-7 µm, voxels above 1000). Cells were then selected at the back and front, respectively, of the r4 stream. The mean intensity of the HNK-1 was recorded. The HNK-1 intensity value was normalized from a spot detected outside the cells on the same slide to account for background. The cell intensity was averaged and plotted on a box plot via Origin for the front and the back of the stream.

Secondary antibodies used were: goat anti-rabbit IgG (H+L) (CM405F, Biotium) for pFAK, F(ab’)2-goat anti-rabbit IgG (H+L) Cross-Adsorbed Secondary Antibody, Alexa Fluor 546 (A-11071, ThermoFisher Scientific) for AQP-1, Alexa Fluor 488 goat anti-rabbit 488 (A-11008, ThermoFisher Scientific) for EphB1/3 and Alexa Fluor 647 goat anti-mouse IgM (A-21238, discontinued product, ThermoFisher Scientific) for HNK-1. All secondary antibodies were used at 1:500.

Pearson's coefficients for colocalized fluorescence were calculated in Imaris 9.0.0 using the Coloc module on 3D datasets. Images of pFAK, HNK-1 and AQP-1 immunohistochemistry from different regions of the r4 migratory neural crest stream were created using a 40×1.2NA objective with Airyscan imaging on a LSM 800 (Zeiss), with the exception of Fig. 1E, which was acquired on a DeltaVision OMX structured Illumination microscope (GE Lifesciences) with a 60×1.4NA objective. Post-Airyscan processing, a mask was created with the HNK-1 channel to isolate neural crest and thresholds for pFAK and AQP-1 channels were created to eliminate background. Pearson's coefficient for n=5 embryos was recorded and averaged.

In vitro neural tube cultures were performed as previously described (McLennan et al., 2010) using Ham's F-12 Nutrient Mix media (11765054, ThermoFisher Scientific). For the initial experiments to examine cell speed, DMSO was added to the media for controls for the AZA (A6011, Sigma-Aldrich) experiments because AZA was solubilized in DMSO. To transfect some cells in the neural tube cultures with AQP-1 FL, dorsal neural tubes were electroporated in ovo as previously described (McLennan and Kulesa, 2007) and allowed to recover for 2-4 h before being removed and plated in vitro. Alternatively, neural tubes were removed from embryos and then constructs or morpholinos were electroporated into the tissue using petri dish platinum electrode for tissue chamber (45-0505, BTX), 5 pulses, 60 V. Neural tubes were allowed to recover in the chamber for 10 min and were then plated. The AQP-1 FL construct was designed and built in the lab, using pMES as a backbone. Prior to neural crest experiments, AQP-1 FL was tested in a chick cell line and shown to overexpress AQP-1 (Fig. S1B,C). For the directionality towards the BA2 tissue experiments, neural tubes were soaked in Hoechst (1:200 dilution for 5 min prior to plating) and BA2 tissue was soaked in DiI (V22889, ThermoFisher Scientific) (1:30 dilution for 30 min). Time-lapses of cultures were taken on a LSM 710, LSM 780 or LSM 800 (Zeiss) at 2.5 min intervals (single z-planes). For immunohistochemistry on in vitro cultures, the same antibodies and protocols were followed as for cryosections and whole-mount embryos; however, the incubation times were shorter. In vitro cultures were fixed with 4% paraformaldehyde for 20 min. Cultures were incubated with primary and secondary antibodies for 1 h each. Images were taken using confocal microscopy and Airyscan detection with a 40×1.2NA water immersion objective. Owing to the length of time taken to collect Airyscan z stacks and to reduce possible photobleaching, this imaging was not performed for the entire cell but focused on the cell membranes against the fibronectin-coated dish surface.

The z stacks of the collected images were loaded into Imaris 9.20. The cells were segmented using the surface function to manually outline the shape of the cell. Spots were then detected on each integrin and pFAK channel. The integrin channel spots included a diameter of 0.6 µm spot with background subtraction and a quality filter was picked using a statistical upper and lower threshold break in the data. The pFAK spots were detected using 0.4 µm spots, and background subtraction with the quality filter was picked using a statistical upper and lower threshold break in the data. The number of spots for each cell was recorded for pFAK and integrin channels, then an average was calculated for each data set. The data were then plotted on a box plot via Origin.

For the neural tube cultures with DMSO, AZA or AQP-1 FL, cells were manually tracked using ‘Spot’ detection in Imaris for ≥6 h. The spots were set to ≥9 µm in size. The mean speed and displacement of each track was calculated using Imaris (Bitplane). Straightness equals displacement of track divided by length of track. The box plots in each figure were generated by using the values from each dataset indicated. Boxes and whiskers indicate the quartiles (25-75%) and range (minimum to maximum values apart from outliers as determined by the software), respectively, of each dataset. P-values were calculated using a standard Student's t-test or paired t-test. Data distribution was assumed to be normal, but this was not formally tested.

AZA (500 µM) was injected into multiple sites of the mesoderm on one side adjacent to the hindbrain region of HH8-9 embryos. For control injections, the same injections were performed using DMSO. After 24 h reincubation, embryos were fixed in 4% paraformaldehyde for 2 h at room temperature and then immunohistochemistry was performed for HNK-1 as described above. Each cranial region of injected embryos was then cut down the midline and each half mounted in a glass slide as previously described (Teddy and Kulesa, 2004). Both halves of the cranial region of each embryo were imaged on a LSM 800 (Zeiss) so that injected and control halves of each embryo could be compared. In ovo electroporations were performed as previously described (McLennan and Kulesa, 2007). Fluorescein-tagged AQP-1 and control morpholinos were designed by, and obtained from, Gene Tools. H2B mCherry (2.5 µg/µl) and morpholinos (0.5 mM) were injected and electroporated at HH8-9 and reincubated for 20 h before being fixed, processed for HNK-1 and imaged as described above. An empty EGFP vector, pMES, was used as a control for AQP-1 FL, as AQP-1 FL was inserted into the pMES vector. pMES and AQP-1 FL were either injected and electroporated with DiI before being reincubated for 12 h or injected and electroporated without DiI before being reincubated for 16 h. Time-lapses were performed as previously described (McKinney et al., 2013). We calculated the percentage of area covered using the ‘Surfaces’ function of Imaris (Bitplane) to create a surface mask by manually drawing the outline of the whole branchial arch. Next, we calculated the area of the HNK-1 fluorescence signal using the masked arch surface. We set a consistent intensity threshold to the same value for each dataset, a surface grain size of 1 μm was set, the diameter of the largest sphere was set to 1 µm and then the automatic ‘Surfaces’ function was applied. We calculated the percentage of the front of arch the HNK-1 signal covered by comparing the two values in the front 50% of the stream. For electroporated embryos, cells were automatically detected using the ‘Spot’ function in Imaris. The spots were set to ≥9 µm in size. The spots were counted in the front 50% of the stream. The percentage of total spots versus cells in the whole stream was calculated. The percentage of distance the cells migrated was calculated by measuring the total distance of the migratory route and measuring what distance the transfected cells migrated. This was calculated on 3D z-stacks using the measurement tool in Imaris. The box plots in each figure were generated by using the values from each dataset indicated. Crosses indicate outliers; boxes and whiskers indicate the quartiles and range, respectively, of each dataset. P-values were calculated using a standard Student's t-test or paired t-test. Data distribution was assumed to be normal, but this was not formally tested.

Premigratory neural crest cells were transfected with pMES, AQP-1 FL or AQP-1 MO via in ovo electroporation as described above. All perturbations were co-transfected with Gap43-mTurquoise for visualization of membrane dynamics. HH13 chick embryos were mounted dorsal side down on glass-bottomed dishes with a grease-sealed Teflon membrane over the top to preserve humidity (Rupp and Kulesa, 2007). Samples were placed inside a heated chamber around a spinning disk confocal microscope (PerkinElmer, Ultraview), allowed to acclimate to the chamber, and two-channel imaging proceeded until any sign of phototoxicity was observed. Cells in the front 10% of the migratory r4 neural crest stream were chosen for imaging. Images were collected with a 40×1.2NA water immersion objective in 1 µm z-steps for up to 15 µm total depth in either 20 or 30 s increments. Time-lapse data were analyzed using a combination of Imaris (Bitplane AG) and ImageJ. If cells were overlapped in z, Imaris 3D imaging was used to create a mask of the cell of interest which was then projected onto 2D. A projected image of the cell was imported into ImageJ and after smoothing and background subtraction, an 8-bit binary mask was created using the Gap-mTurquoise2 fluorescence. Minor adjustments to the mask were made by hand for extremely thin filopodia or touching cells. The time-lapse masked image was imported into a modified version of the CellGeo software (Tsygankov et al., 2014) that did not use the java-enabled GUI but allowed for parameter adjustment using the same algorithms. BisectoGraph and FiloTrak modules were used and parameters of 20 smooth, 1 CrR, 1 CutOff, 9 critical length and 5.5 critical width were used. Filopodia were tracked manually. The remainder of the analysis was written in MATLAB (Mathworks) to collect filopodia data from multiple cells under each condition. A total of seven cells from four embryos were used in control, six cells from four embryos for AQP-1-FL, and six cells from three embryos for AQP-1 MO experiments. Statistical significance was determined using an ANOVA with Tukey's HSD post-hoc calculation and P-values reported.

Premigratory neural crest were electroporated with either AQP-1 FL or pMES and eggs re-incubated for 24 h. The neural crest stream adjacent to rhombomere 4 was isolated from healthy ∼HH15 embryos. Five pMES transfected embryos were pooled for each of the three biological replicates (n=15 total). Four AQP-1 FL embryos were pooled for each of the three biological replicates (n=12 total). Tissue was dissociated as previously described (McLennan et al., 2015a). Cells were isolated by FACS, which included forward scatter, side scatter, pulse width, live/dead stain (7AAD) and YFP gates, as previously described (McLennan et al., 2015a; Morrison et al., 2017a). Cells were sorted directly into 7 μl of Clontech lysis solution containing 0.05% RNAse inhibitor. Following lysis for 5 min at room temperature, lysates were immediately frozen on dry ice and stored at −80°C. Bulk RNA-seq lysates were thawed on ice. cDNA synthesis and library preparation were performed with SMART-seq v4 Ultra Low Input RNA-seq (634891, Takara, Kusatsu) and Nextera XT DNA sample prep and indexing library preparation kits, as recommended by the manufacturer (FC-131-2001, FC-131-2004 and FC-131-1096, Illumina). Resulting short fragment libraries were checked for quality and quantity using a Bioanalyzer (Agilent) and Qubit Fluorometer (Life Technologies). Libraries were pooled, re-quantified and sequenced as 50 bp single reads on two lanes of the Illumina HiSeq 2500 in High Output mode using HiSeq Control Software v2.2.58 (Illumina). Following sequencing, Illumina Primary Analysis version RTA v1.18.64 and bcl2fastq2 v2.18 were run to demultiplex reads for all libraries and generate FASTQ files. More than 3 million total alignments were produced per sample. Single-end 51-base reads were aligned to the chicken genome galGal4 from UCSC with annotations from Ensembl 84 using STAR (2.5.2b) with options –alignEndsType EndToEnd and sjdbScore 2. Downstream analysis was carried out in R (3.4.1). Differential expression analysis was performed using the edgeR package (3.18.1). Genes were indicated as differentially expressed if they had a P-value less than 0.05 and a fold change greater than 1.5-fold (absolute value). Gene ontology enrichment was carried out using a hypergeometric test on lists of differentially expressed genes.

To achieve high levels of MMP secretion from neural crest cells grown in culture, neural tubes were isolated, plated onto the bottom of four-well dishes (176740, ThermoFisher Scientific) with no coating and covered in 250 µl of Ham's F-12 media. Thirty-nine neural tubes were exposed to 100 µM AZA, 39 neural tubes were exposed to DMSO (0.05 µl in 1 ml) as control for AZA, 32 neural tubes were exposed to 50 µmol/l GM6001 (Anderson et al., 2006) and 32 neural tubes were exposed to DMSO (1 µl in 1 ml) as control for GM6001. After 24 h of incubation, the media were harvested. Protein in media was concentrated using 3K Pierce concentrator (88512, ThermoFisher Scientific) and then used in-gel zymography following the supplier's protocol (ZY00100BOX, LC2670, LC2671, LC2675, LC2676, ThermoFisher Scientific). The substrate was gelatin. Importantly, an image of the gel prior to colloidal blue staining (LC6025, ThermoFisher Scientific) was taken so that the protein ladder could be clearly seen for determining protein size after staining. Quantification of the bands on the gel image were performed using ImageJ and a protocol from Protocol Place (protocol-place.com/assays/gelatin-zymography/).

A degradation assay was performed using QCM Gelatin Invadopodia Assay (ECM671, Millipore). Cy3-labeled gelatin-coated wells in 8-well chambered slides (80821, Ibidi) were used. The protocol was slightly modified by heating all the gelatin to 60°C so that there was no issue with room temperature and warm gelatin not mixing fully. To optimize cell adhesion, fibronectin was then added over the gelatin as previously described (Garmon et al., 2018). Fifteen neural tubes (transfected with either pMES or AQP-1 FL) were plated into each well with 500 µl of media and incubated for 40 h. For controls, some wells received no neural tubes, but were still incubated for 40 h with 500 µl of media. Cultures were fixed with 4% paraformaldehyde for 20 min and then imaged. MMP2 and MMP9 are secreted proteases and possible gelatin degradation will occur throughout each well, not just where neural crest cells are migrating. Therefore, Cy3-labeled gelatin was imaged at five randomly chosen areas of each well using a 40×1.2NA objective, n=3 wells per condition. The z-stack images were loaded into Imaris 9.20. A surface rendering was created on the whole Cy3 gelatin channel throughout the whole image with the automatic settings used to detect the surface using voxel totals above 10. The average intensity across the whole image for the Cy3 channel was recorded. An average was taken from each condition and a box plot was made using Origin.

We thank members of the Microscopy, Histology and Molecular Biology core facilities, and our Scientific Illustrator, Mark Miller, at the Stowers Institute for Medical Research.

The peer review history is available online at https://dev.biologists.org/lookup/doi/10.1242/dev.185231.reviewer-comments.pdf.