Shaping the zebrafish notochord

Cellular reorganization based on active cell motility and adhesion is the major means of morphogenesis in animal embryos, and is particularly elaborated in vertebrates. Nowhere is this fact more clearly shown than in the dorsal mesoderm of gastrulating zebrafish embryos. Here, occurring in the company of both cellular internalization at the blastoderm margin(Carmany-Rampey and Schier,2001) (R. J. A., D. Faruque and M. L. Concha, unpublished) to form the mesoderm and an equally prominent spreading of the blastoderm by epiboly(Warga and Kimmel, 1990),there is a rapid and massive cellular reorganization. The reorganization underlies mesodermal convergence, meaning the narrowing of the tissue with respect to the embryonic axis (the anteroposterior, or AP axis), and mesodermal extension — the lengthening of the tissue with respect this axis.

Our current view of the cellular behaviors that drive these tissue-level shape changes comes primarily from a series of studies of explantedXenopus dorsal mesoderm (Shih and Keller, 1992a; Shih and Keller, 1992b; Keller et al.,2000). Molecular and genetic studies (including studies in zebrafish and other species) have enriched our current understanding(Solnica-Krezel, 1999;Tada and Concha, 2001;Wallingford and Harland, 2001;Myers et al., 2002a;Myers et al., 2002b). The explant studies in Xenopus show that dorsal mesodermal converges and extends largely by cell rearrangement within the tissue, and without dependence upon an external substrate(Shih and Keller, 1992a). Furthermore, from this work has evolved the concept of a single but complex cellular behavior, termed mediolateral intercalation behavior (MIB), which underlies the rearrangements (Fig. 1). The MIB hypothesis is elegant because a single force-generating cellular machine, distributed across a field of cells,produces both convergence and extension, both narrowing and elongation of the field. By the MIB hypothesis, as applied particularly to the domain of notochord-forming cells within dorsal mesoderm, motile and adhesive cells become polarized along one particular axis, the mediolateral (ML) axis. The polarity may depend on, and be coordinated within the field, by a noncanonical Wnt signaling planar polarity pathway (Choi and Han, 2002; Heisenberg et al., 2000). The cells take on a bipolar shape, elongating along the ML axis. This process requires that the individual cells all correctly orient actin-based cytoskeletal machinery that mediates motility, and perhaps also orient associated adhesion complexes on their plasma membranes(Montell, 1999;Zalik et al., 1999). Localized release of intracellular Ca²⁺, via connexin 43 channels(Essner et al., 1996), and activation of Rho GTPases may be crucial in such polarized cellular morphogenesis (Choi and Han,2002; Hall and Nobes,2000; Smith et al.,2000). The cells intercalate mediolaterally. To accomplish this,they all protrude filopodial processes both medially and laterally that extend between immediate cellular neighbors, and new cell-cell contacts are made,perhaps involving adhesion molecules of the cadherin/protocadherin superfamily localized to their filopodial tips (Keller et al., 2000). The newly contacting cells then contract their processes. They exert traction upon one another and pull together. Such events, which occur across the entire field of cells, narrow the field(convergence). The resulting intercalations push previously neighboring cells apart along the AP axis, lengthening the field along this axis(extension).

Previous work suggests that MIB may also underlie convergence and extension in the zebrafish gastrula (Concha and Adams, 1998; Heisenberg et al., 2000; Kimmel et al.,1994; Solnica-Krezel et al.,1996; Warga and Kimmel,1990; Myers et al.,2002a; Myers et al.,2002b). The cells disperse along the AP axis by intercalating with neighbors to form a discontinuous AP string, as expected if MIB mediates extension (Kimmel and Warga,1986; Kimmel et al.,1994). However, it is not known to what extent the phenomenon is coupled to convergence as the MIB hypothesis predicts it should be. Additionally, other more coherent cellular flows, e.g. migrations that do not involve cellular intercalations, might also occur to increase the AP length of the developing notochord.

To test the MIB model in zebrafish, we used a confocal microscope to image and make time-lapse recordings of the dorsal mesoderm in dye-labeled intact wild-type (WT) and no tail (ntl) mutant embryos(Cooper et al., 1999b) (N. S. Glickman, PhD Thesis, University of Oregon, 2000). ntl is the zebrafish homolog of the mouse T-box gene Brachyury(Schulte-Merker et al., 1994). Our recordings begin during gastrulation, shortly after mesodermal internalization at the blastoderm margin, and we followed the movements of most of the cells in the field of view during convergence and extension. The method differs significantly from most previous studies in that confocal imaging, 4D recording and cell-by-cell analysis of these records allows us to keep track of, and quantify, the behaviors (e.g. movement velocities and intercalation behaviors) of a substantial fraction of the dorsal mesodermal cells rather than just a small sample. We use this method to measure rate constants for convergence and extension, providing for meaningful comparison between tissues or embryos.

We find that the rate of convergence within the notochord domain is comparable with that in the somite, whereas extension is several-fold higher in the notochord than in the somite, suggesting that the cells are reorganizing differently in the two domains. Quantitative features of convergence and extension and local cellular reorganizations within notochord domain are as expected from the MIB hypothesis. Our data suggest furthermore that for the WT notochord domain, MIB quantitatively accounts for the observed morphogenesis. We show that ntl is a key regulator of convergence of the notochord domain; in ntl mutants, convergence is severely disrupted, and our analyses suggest that MIB is initiated but cannot be completed. Surprisingly, in the near absence of early convergence inntl mutants, the dorsal mesoderm can still extend. Hence, extension does not require MIB. We discuss the meaning of these finding for understanding mesodermal morphogenesis and its regulation.

Adult zebrafish (Danio rerio) and embryos were reared at 28.5°C, as described by Westerfield(Westerfield, 1995), and staged in accordance with Kimmel et al.(Kimmel et al., 1995). We performed time-lapse recordings with wild-type (strain AB) embryos and with the mutant and wild-type siblings of ntl^b195(Halpern et al., 1993). Thentl^b195 allele is caused by an insertion, resulting in a stop codon and a truncated non-functional protein(Schulte-Merker et al.,1994).

Dechorionated embryos (3.3-3.5 hpf)(Kimmel et al., 1995) were incubated in embryo rearing medium plus 10 mM HEPES buffer (pH 7.2)(Cooper et al., 1999a;Cooper et al., 1999b;Westerfield, 1995). Embryos were stained in 100 μM BODIPY FL C5-Cer/C5-DMB-Cer(N-(4,4-difluoro-5,7-dimethyl-4-bora-3a,4a-diaza-s-indacene-3-pentanoyl)sphingosine (in 1% DMSO), here after referred to as BODIPY ceramide (Molecular Probes) (Cooper et al., 1999a;Cooper et al., 1999b).

Embryos were mounted in a viewing chamber(Westerfield, 1995) in 3%methylcellulose and 0.3% agar, dorsal side upwards. We examined seven wild-type embryos and four ntl mutant embryos. Images were taken with a Zeiss 310 upright confocal microscope with a Kr/Ar laser. We imaged 12 focal planes, at 3 μm intervals for each time point, set 2 minutes apart. The 4D confocal images were compiled into a 4D movie for playing in a modified version of NIH image. Cell locations over time were collected using a modified version of NIH Image (Wasband, NIH) customized to animate 3D time-lapse image series. Each cell was traced over time by manually marking its approximate geometric center in three dimensions through each time frame, recording all cell divisions and final cell fate from its location. All data produced in this way were further analyzed by routines written in the analysis environment IDL (Research Systems Inc, CO). Final presentations were rendered by code generated for the ray-tracer POVRay (POVray).

The precise records of cell positions over time allow us to generate metrics that describe the patterns of motion of morphogenesis. The raw cell tracks were rotated in space to align the AP axis vertically. AP location and velocities can then be expressed in terms of y coordinate and velocities calculated as dy/dt. Similarly, mediolateral movements proceed in x with velocities of dx/dt. Cell velocity and speed were calculated as changes in location about the time frame of interest e.g. dx/dt at time t is calculated from the displacement x[t+I] —x[t-I] for a short period I, usually ±8 minutes. The area of the entire tracked field and the tracked axial domain was measured by drawing a convex hull around all the tracked cells. The width of the field of tracked cells is the mean width of the lateral edges of the hull. The 2D density of the tracked field was approximated by dividing the area contained by the convex hull by the number of cells within the box. This measure is approximate and assumes cells are uniformly tracked through time, to reduce possible errors the relative change in density rather than absolute measures were used in these analyses.

In order to analyze average local cell rearrangement, we calculated for each cell in turn the relative location and movement of each other cell in its vicinity. Then a smoothed map of all accumulated relative cell movements was used to generate a vector field to visualize the mean local tissue shape change during a short epoch.

Specific cell rearrangement motifs can be looked for in the cell tracks by calculating a connectivity matrix that describes adjacency of neighboring cells based upon their geometry in space (2D or 3D). Neighbor changes will reflect, for example, intercalation events because as two cells move towards each other mediolaterally, they become new neighbors and displace their anterior and posterior neighbors. The analysis of neighbor changes are limited to the wild-type and mutant notochord/axial domains, where the optics are best and we have most completely tracked the cell populations. We calculated the immediate cell neighbors of all traced cells for each time frame, from a complete 3D Delaunay triangulation. The program Qhull(Barber et al., 1996) was built as a library callable from within IDL for this purpose. Each vertex corresponds to the location of a cell. The algorithm assumes that the local field has been completely tracked, that cells are compact in shape and are in close contact with their neighbors — as is evident from inspection of the field (see below, Fig. 3A). To exclude erroneous apparent neighbor changes between non-contiguous cells, a conservative maximum cell spacing of 18 μm was used. Any cells separated by an edge longer than this threshold were not considered to be neighbors for this analysis. (In practice, the patterns of change seen in connections were relatively insensitive to variation in this threshold, despite significant changes in the total number of permitted connections.) Changes in neighbor connections over time were detected from the changes in connectivity in the triangulation. Changes due to cell divisions, cells moving into and out of the traced field, and transient make-break events were excluded from the analysis. Hand-checking a set of such identified intercalations, in the original 4D record set, showed the accuracy of the computer-assisted method to be at least 90%. The angle at which cells make and break connections reflects the angle of the interaction event. The distributions of angles were tested for uniformity using a Watson U² test and any samples found to be non-random(P<0.05) were tested for their mean axial alignment. Variations in alignment over time were visualized by a moving sample window through the duration of the experiment.

For the purposes of comparison between experiments, an equivalent single rate constant equal to the mean rate over that interval can be used. The same calculations were made for an extension rate constant k_E, but in this case, the change in the AP position (dy/dt) of each cell was calculated.

We made a simulation to show how an idealized 2D field of cells would reorganize if converging with a linear velocity gradient in space and extending with an equal gradient of opposite sign (thus preserving cell density), see Fig. 2. A hexagonal array of cells was incrementally constricted and stretched to demonstrate the resulting affect. Globally, the field narrowed exponentially and extended exponentially, as expected. Individual cell reorganization relative to adjacent neighbors was constant and reiterated across the field.

In the simplest version of the MIB hypothesis, cells pull together by intercalation along the mediolateral (ML) axis and simultaneously are pushed apart along the AP axis. The hypothesis makes a number of verifiable predictions, illustrated by an animation of hypothetical `cells' undergoing MIB (Fig. 2A-D; movies for this and other figures are available as supplemental information atSupplementary Information). During the sequence illustrated, the field of cells narrows to half of its original width (convergence — along the horizontal inFig. 2), and elongates to twice its original length (extension — along the vertical). The amount of narrowing equals the amount of lengthening, i.e. convergence equals extension.

During MIB cells reorganize, as illustrated by close-up views near the field center (^*) at the beginning(Fig. 2E) and end(Fig. 2F) of the same sequence. A single cell (^*) retains several of its original neighbors(numbered for identification) but changes others. Cells that start out a rank apart along the ML axis become new neighbors (e.g. cells 2 and 6). Conversely,cells that were neighbors along the AP axis become separated by a rank (e.g. cells 1 and 10). This local reorganization is uniform across the field.

Strikingly, however, movement velocities of the cells are not uniform because of the incremental nature of MIB. The cell at the center(^*; which we shall see represents a point along the midline of the embryo) does not move. Cell 8 moves inward a single rank, and cell 9, located three times as far away along the ML axis, moves inward three ranks(Fig. 2E,F), moving three times faster because it is pulled by its own intercalations and also pulled by inward-moving more medial cells. `Converging' cells move inwards with a velocity proportional to their distance from the midline, such that a linear gradient of velocity with position is present across the field(Fig. 2H). This position dependence on the rate of movement can also be seen by examining the paths the cells take during a given time interval(Fig. 2G). The paths of the cells at more lateral positions are longer than the paths of cells close to the center, reflecting their rate of travel.

Similarly, along the AP axis the rate of outwards movement depends on AP position. Cell 11 is three times farther from the center than cell #1 and it moves three times faster (Fig. 2E,F). Again, the relationship between cellular position in the field and velocity is linear (Fig. 2I). The slopes of the lines inFig. 2H,I give convenient measures (rate constants) for convergence and extension (k_c and k_E, respectively). The convergence rate constant is negative,reflecting inward movement, and the extension rate constant is positive,reflecting outward movement. Their absolute values are equal.

The MIB hypothesis thus predicts linear gradients of velocity with position. Furthermore, it follows from the same arguments that the shape changes of the field should not be constant with time. As the field narrows during convergence, only smaller ML positional values remain (the larger ones are lost as the population moves closer the midline) and only lower velocities should be observed. Hence, the rate of change in field width should decrease with time. As the field lengthens during extension new, larger AP positional values are established and velocities are correspondingly higher. Hence, the rate of change in field length should increase with time. The kinetics are exponential in both cases (Fig. 2J,I), and the exponential curves have the same slopes as the gradients described above: k_c and k_E.

A perhaps surprising feature of MIB is that a large change in shape of the cellular field is produced by relatively small local cellular rearrangements. This is another prediction of the hypothesis.

Convergence and extension in embryos has usually been measured by tracking the movements of one or more labeled cells or small cell groups toward the dorsal midline, and along the AP axis. The above analyses show clearly that this method is unreliable when used to make interpretations about the movement of the entire field of cells. Even where convergence and extension are uniform across the field and during time, the ML and AP cellular velocities are not uniform with time, and depend on cell position within the field. Hence, we used a labeling and recording method(Cooper et al., 1999a;Cooper et al., 1999b), and analytical tools (Concha and Adams,1998) (see Materials and Methods) that allowed us to keep track of all or most of the cells being recorded, and we tracked the field of developing cells as fully as we could. Fig. 3 illustrates the results for a wild-type embryo. Cell outlines are apparent in the time series of original recordings (A), and the series of records with respect to both time and depth (z-axis sectioning with the confocal microscope) intervals were close enough to one another to permit us to follow unambiguously individual cells in the field of view, no matter how the cells moved, and during cell divisions. The field is approximately centered on the dorsal midline, and the particular z-level illustrated in Fig. 3A is at an intermediate depth through the internalized prospective mesoderm (or hypoblast) at the center of the field. The blastoderm margin is evident at first (Fig. 3A1), and then,because of epiboly (Kimmel et al.,1995; Solnica-Krezel et al.,1996), it moves off towards the vegetal pole (downwards in the figure), out of the field being recorded. Brachet's cleft(Kimmel et al., 1995), the boundary between ectoderm and mesoderm, is evident (arrows inFig. 3A2).

Animating a recording such as used forFig. 3A (see Movie 2)reveals that the overall cellular flows are as expected of MIB(Fig. 2G). Movement pathways of lateral cells are predominantly ML, whereas those of medial cells are predominantly AP (see Myers et al.,2002b). The animation also allows visualization of another predicted feature of MIB: the position-dependent rates of cell movement(Fig. 2H,I). Lateral cells move towards the midline faster than cells already close to the midline. Cells near the top and bottom of the field move faster along the AP axis than those near the middle of the field.

Within the mesoderm at 9.3 hpf, boundaries between the prospective notochord and presomitic mesoderm become clearly visible (notochord-somite or axial-paraxial boundaries; arrowheads, Fig. 3A3). By animating the recordings, and playing Movies 2 and 3 backwards (seehttp://dev.biologists.org/supplemental/),one can see that these boundaries arise in a patchy (or piecemeal) way; they can be vaguely recognized in Fig. 3A2 (8.8 hpf), farther apart than at later stages, thus revealing that the notochord domain is converging. However, the boundaries are not at all visible at the initial time point (7.3 hpf, midgastrula period). At the last time point, 4 hours later, the notochord-somite boundaries have straightened out and come together to enclose a notochord domain that is now only two or three cells wide (Fig. 3A4). The boundaries between the most anterior somites are also just becoming visible (arrows, Fig. 3A4).

Once such landmarks become visible, we could assign cells particular fates(e.g. position of a cell within the area demarcated by the notochord boundaries defines a notochord cell). Fig. 3B shows a time series from the same records, where we represent tracked cells as spheres color-coded according to these eventual fates. We find, in this and in other wild-type embryos similarly analyzed, that the notochord domain is coherent from outset of the recording; i.e. the domain of prospective notochord cells is spatially separate from the somitic domains(and also from overlying floorplate-forming cells,Fig. 3C), before overt boundaries arise. During convergence, the notochord domain narrows from 10-14 cells across at the time our recording begins to two or three cells across. This kind of decrease is expected from sustained MIB.

Differential movements can be detected between regions adopting different fates. For example, we observe slipping apart or shear between cells the notochord domain and the immediately overlying floor plate-forming cells in the prospective neural plate. As is evident inFig. 3C, the notochord and floor plate-forming cells start out at the same AP position but after undergoing marked convergence and extension, the final position of the floor plate-forming cells is posterior relative to the notochord cells. Tracking other cells in the epiblast (i.e. in the prospective neural plate outside of the medial floor plate domain; data not shown) reveals that this shearing is not unique to the notochord and floorplate, but is a general feature of epiblast cellular movement relative to that in the internalized mesoderm of the hypoblast. We can account for the shear by considering the combined morphogenetic movements of epiboly and internalization of cells by ingression at the blastoderm margin (see Discussion).

The notochord domain also shears with the neighboring somite domains(Fig. 3D), as previously reported at later stages in zebrafish by Devoto et al.(Devoto et al., 1996). In this case, as we describe next, we can account for the shearing by a difference in the rate in which the two domains extend. The presentation inFig. 3D reveals that extension of the notochord domain is dramatically greater than extension of the somite mesoderm. The same difference is also present in Xenopus(Keller et al., 1989;Keller et al., 2000).

We quantitatively analyzed the kinetics of the cellular flows in our records to explore further the similarities and differences in convergence and extension in the notochord and somite domains. We observed that during gastrulation and continuing into the early segmentation period of development,the width of the notochord domain decreases with the kinetics of exponential decay (Fig. 4A), as predicted by the MIB hypothesis (Fig. 2J). This finding suggests that if MIB, or a similar mechanism underlies convergence, cellular interactions (including rates of ML intercalations) remain approximately constant during the recording period.

Plotting ML velocities of all the cells tracked in the field as a function of their ML position, at any time point, reveals the linear gradient underlying convergence (Fig. 4B) predicted by the model(Fig. 2H). As shown inFig. 4B, the data points for the velocities of cells in left and right somite domains (red) fall on the same line as those for cells in the notochord domain (green). This finding shows that the rates of convergence in the three domains are similar. Calculating the rates separately for each of the domains and integrating (or averaging: see Materials and Methods) across all time points confirms this conclusion; k_c=-0.0066, -0.0065 and -0.0057 for the notochord, and left and right somite domains, respectively. Evidently, the rate of convergence is slightly higher in the notochord domain (∼8% higher that the average of the somite domains in this example), a conclusion substantiated by similar estimates in a second wild-type embryo. Convergence in the notochord domain estimated by this method is in excellent agreement with that measured directly from its decreasing width with time(Fig. 4A; k_c=-0.0064 compared with -0.0066).

Kinetic analysis also confirms the relatively higher extension of the notochord domain when compared with the somite domain, noted above(Fig. 3D). Cells in the notochord domain have much higher AP velocities than somite domain cells beginning at the same AP position. The relationship between AP velocity and AP position, within either the notochord or somite domain, is linear(Fig. 4C). As we see fromFig. 4C, two separate lines describe the separate rates of extension of the two domains; the slope(k_E) is substantially higher for the notochord. Over all time points, extension in the notochord domain is about three times higher than in the somites in this embryo; k_E=+0.0064 for the notochord domain compared with +0.0023 for average of both somite domains.

The rates of convergence and extension are expected to be equal if MIB is pushing cells apart exclusively along the AP axis (as in the model inFig. 2) and if no factors,other than MIB, are contributing to convergence and extension. In this example, as predicted by this extreme model, convergence of the notochord domain almost exactly equals extension (|k_c|=0.0066 and |k_E|=0.0064; 3% difference). In a second wild-type embryo examined in this way, |k_c| is 35%higher than |k_E|. By contrast, extension is much lower than predicted by the model in the somite domain, occurring, as we have seen at about one third of the predicted rate. An explanation that we favor for why convergence and extension are more in balance in the notochord domain than in the somite domain is that MIB is driving convergence at nearly the same rate in the two domains, but that during MIB, cells are rearranging themselves differently in the two domains (see Discussion).

The overall results of our kinetic analyses in wild-type embryos are in accordance with the MIB hypothesis. In particular, the kinetics of cell movement are approximately as predicted. Thus, MIB may account for all (or most) of the cellular movements within the wild-type notochord domain. To test this conclusion more rigorously, we ask if the velocity gradients underlying the cell movements, presumably generated by MIB, can account for the all observed convergence and extension of the notochord domain. We do this by using the values of the convergence and extension rate constants to subtract the cell movement due to MIB. If MIB accounts for all cell movement, then this procedure should subtract all motion (other than random noise) from the field of cells. This is essentially what we observe, as illustrated for the notochord domain in Fig. 5. The purple lines indicate the observed locations and movements of the notochord cells at 7.3 hpf (Fig. 5A) and 4 hours later (Fig. 5B). Convergence and extension of the tissue are clearly observed. However, after removing the calculated components of the cell movements resulting from MIB,little convergence and extension is observed after 4 hours (yellow lines inFig. 5). The field has approximately the same shape at the end as at the beginning. This analysis strongly suggests that we are not missing some major factor contributing to the change in shape of the field.

By the MIB model, active convergence drives extension in the dorsal mesoderm, and our analyses in the wild-type notochord and somite domains (see also below for cell neighbor analyses in the notochord) are consistent with this hypothesis. We reasoned that if convergence were abolished in the notochord domain, then so too should extension.

We used ntl mutants to carry out this experiment. The ntlgene encodes a T-box transcription factor (the zebrafish ortholog of mouseBrachyury) (Schulte-Merker et al., 1994) that is broadly expressed in nascent mesoderm and then maintained specifically in the gastrulating axial mesoderm that develops as notochord (Schulte-Merker et al.,1994). In ntl loss-of-function mutants, notochord development fails and gene expression analysis suggests that, within thentl^- axial domain, convergence fails during gastrulation(Melby et al., 1997).

Recordings made in the ntl mutant begin the same way as in the wild type. Cells ingress at the blastoderm margin in the early gastrula, and we made our recordings of the internalized dorsal mutant `mesodermal' layer present during gastrulation. However, as gastrulation continues and the notochord-somite boundary becomes visible in the wild type, no such boundary is at first apparent in the mutant (Fig. 6A,E). Later, by the one-somite stage, a pair of boundaries appears in the mutant, encompassing an `axial' domain only slighter broader than the wild-type notochord domain at this stage(Fig. 6B,F). We could use these boundaries in the mutant to score our tracked mesoderm-derived cells as`axial' cells (i.e. cells enclosed by the boundaries) versus somitic cells(cells outside of the boundaries). The axial/somite boundaries appear not to enclose a mutant notochord domain, however, but to enclose the ventral part of the primordium of the spinal cord (termed the neural keel at this stage)(Papan and Campos-Ortega,1994). We infer this from the cellular morphology at early somite stages: in the wild-type notochord domain, the cells become highly organized. They take on a wedge-shape and are significantly elongated along the ML axis[Fig. 6C,D (parts 4 and 5)]. At a more dorsal focal plane in the wild type, just before coming to the floor plate one encounters the notochord-floor plate boundary, a flat, thin zone diffusely labeled with BODIPY-ceramide, where cell outlines are difficult to discern (Fig. 6D, part 3). Dorsal to this boundary the cells of the floor plate region are, at the three-somite stage, disorganized and irregular in shape(Fig. 6D, parts 1 and 2). In the ntl mutants at the same stage, the cellular morphology at any focal plane is irregular (Fig. 6F), like that of the wild-type floor plate. Furthermore, a notochord-floor plate boundary cannot be discerned.

The width of the axial domain in ntl mutants, as judged from the positions of the axial-somite boundaries, becomes very similar to that of the wild-type notochord domain, clearly revealing that the mutant axial domain has undergone convergence by the three-somite stage. In the wild-type embryo, as we have seen, the narrowing occurs during much of the gastrulation and early segmentation periods. By contrast, convergence of the axial domain in thentl mutant is at first nearly absent. For example, the field of tracked mutant cells does not appreciably narrow during the 1.5 h interval during gastrulation shown as cell pathways inFig. 7A, and by computed width of the field in Fig. 8A. Then,at postgastrula stages, just preceding the appearance of the boundaries surrounding the mutant axial domain, convergence of the mutant axial domain dramatically picks up (Fig. 8A;time points after about 9.5 hpf).

Notably, at same stages when convergence of the mutant axial field is severely perturbed, extension is occurring at a high rate. The field of tracked cells is lengthening quite prominently inFig. 7A, andFig. 8B shows that in this example, extension of the mutant axial domain is substantially higher than the wild-type notochord domain during the gastrula period. Without substantial convergence, seeing any extension at all during this period was unexpected. In the wild type, because of the close balance between convergence and extension,the cross-sectional (or planar) area of the notochord remains nearly unchanged(Fig. 8C; first a 20% increase then a 20% decrease over the recorded period). In the mutant, with marked extension but not convergence, the computed area of the domain expands markedly, about threefold, implying that an active process of extension or stretching is taking place (Fig. 8C).

A possible explanation for observing extension in the mutant comes from understanding that ntl^- midline axial `mesoderm' joins the ventral neural tube, and expresses the floor plate fate(Amacher et al., 2002;Halpern et al., 1993). We could suppose that, in the mutant, the mutant axial cells leave the axial mesodermal domain and move into the overlying epiblast-derived floor plate. These cells would then undergo rounds of cellular intercalation within the epiblast. Intercalations of axial mesodermal cells within the epiblast ofntl mutants may contribute to the observed extension of the axial domain. To examine this hypothesis, we tracked some epiblast cells overlying the midline axial domain in a ntl mutant. We observe that, as in the wild type, the two sets of cells do not mix(Fig. 7B). Hence, `ectopic'intercalations between midline hypoblast and epiblast do not appear to account for the observed extension. Rather, during extension in the mutant, the cells seem to thin out, forming a layer one-cell thick(Fig. 7B,C). As the cells thin out, they cover a greater area, resulting in extension of the axial domain. The overall density of cells in the ntl mutant axial domain decreases, while the density of cells in the wild-type notochord increases(Fig. 8D). The observed decrease in cell density in ntl mutants also helps explain the increase in area, and extension, of the axial mesoderm(Fig. 8C).

Our analyses reveal cellular correlates of reduced convergence ofntl mutant axial domain cells during gastrulation. ML movements are present in wild-type embryos, but most sustained movements in mutants are AP(vertical versus horizontal tracks, Fig. 7A). In mutants, cells appear to mix abnormally between the axial domain and somite-forming domains (Fig. 7A; red and green interspersed in ntl^-). Wild-type cells show only limited mixing with medial and lateral neighbors during convergence and extension (Fig. 9A). Of course, we expect limited mixing because constrained cell mixing is the basis of convergence and extension by MIB. However, ML mixing is prominent in ntl mutants (Fig. 9A), a circumstance where convergence is greatly reduced and mesoderm thins out. Cells in the wild type move towards the dorsal midline, as expected during MIB, but movement with respect to the midline is scrambled inntl mutants (Fig. 9B).

These data show that the ntl mutant cells intermix far more extensively than do wild-type cells, but do not reveal how the intercalations are oriented. To address this, we examined cell neighbor exchanges. In cellular fields undergoing intercalations, new neighbor pairs will be generated and old ones lost in a reciprocal fashion. By the MIB model,neighbor gains should be oriented mediolaterally, and neighbor losses should be oriented anteroposteriorly (Fig. 2E,F; Fig. 10A). We detect large numbers of both neighbor gains and losses in the axial domains of wild-type and ntl^- embryos, and in the wild type they are strongly oriented as predicted (Fig. 10B). Surprisingly, in ntl mutants, neighbor changes are still made with a significant but broadened tendency to align as in MIB(Fig. 10B). A most interesting change underlies the broadening. Analysis of the time course of the neighbor changing events (Fig. 10C)shows that the wild-type axial domain is able to maintain oriented intercalation activity nearly throughout the recording period. However, in thentl mutant, only pulses of ordered neighbor gains or losses occur,interspersed with generally longer periods of disorganization.

This disorderly behavior suggests that MIB can be initiated but not sustained in ntl mutants, and prompted us to examine how cells are actually moving with respect to their local neighbors over both short and long time periods. We have shown that in a cellular field uniformly undergoing MIB,ML cell movement is towards the field center and AP movement is away from the field center, with velocity gradients accompanying both. If local cellular reorganizations are in fact uniform and if we now hold constant the position of any cell in the field, we expect ML neighbors to move towards that cell and AP neighbors to move away. Strikingly, we observe these predicted local correlates of both convergence and extension in the notochord domain of wild-type embryo (Fig. 11A). The ML inward movements are comparable in magnitude to the AP outward movements, reflecting the close balance of convergence and extension. Relative velocity gradients are evident (from the arrow lengths) along both axes,mirroring the global velocity gradients across the field. Furthermore, as shown by comparing the series of panels inFig. 11A, the wild-type embryo is able to maintain these highly oriented local behaviors, i.e. the pattern is unchanging over significant time courses.

However, we see this same pattern only occasionally in the ntlmutant (Fig. 11B). Behavior is wild type-like at the first time point selected (first panel). Afterwards,whereas the AP movements tend to remain outward from the reference cell, the ML movements are scrambled, first inward (first panel), then outward (last panel in Fig. 11B). Disorganized local ML convergence of cells toward one another underlies the global disruption in convergence of the axial field.

Mediolateral intercalation behavior (MIB) may be the basic means by which dorsal mesoderm undergoes convergence and extension to establish the embryonic axis (AP axis) in any chordate (Keller et al., 2000; Munro and Odell,2002a; Munro and Odell,2002b). In this report, we have defined the global reshaping and local interactions we expect to see in a cellular field uniformly undergoing convergence and extension by MIB. We observe that both narrowing (convergence)and lengthening (extension) of the dorsal mesoderm occur as predicted by the MIB hypothesis. MIB is further strongly supported by our analyses of how local cellular neighborhoods are reorganizing during convergence and extension, and of the magnitudes and direction of movements of cells relative to their neighbors. The behaviors occur very uniformly across the entire notochord domain.

Our studies also reveal complexity in the behaviors that underlie convergence and extension. The notochord domain and adjacent domains of somite-forming cells converge at a similar rate, but they extend at different rates. In mutants that lack function of a key mesodermal patterning gene (the T-box gene no tail)(Schulte-Merker et al., 1994),convergence of the notochord domain is substantially blocked during gastrulation, but unexpectedly, the domain extends rapidly. A likely explanation for these findings, that we explore below, is that in addition to MIB, other morphogenetic force-generating processes are shaping the dorsal mesoderm during this crucial period of early development.

A striking finding of our study is that linear relationships, gradients,between cellular movement velocities and cellular position underlie both convergence and extension. These relationships arise because of the incremental and additive nature of ML intercalation events across the field:during convergence, cells located laterally will be displaced inward towards the center of the field (the dorsal midline) not only by their own intercalations with their own neighbors, but because they are connected by a chain of adhesions to more medial cells, and cells all along the chain are undergoing MIB. The resulting velocity gradient along the chain will be linear, as we observed, if the local cellular behavior along the chain (i.e. MIB) is, on average, uniform. Exactly the same kind of explanation works to explain the AP gradient underlying extension.

These simple relationships between velocity and position are extremely useful in practice. If we plot cellular velocity versus cellular position,from our recordings of the cell movement pathways, the slopes of the lines are straightforward to determine. The slopes serve as rate constants(k_c, k_E) that reveal how rapidly the shape of the field is changing. It follows from the linear relationships that if MIB is constant over time, then the kinetics of the shape change are exponential (first order), as we confirmed experimentally.

From our estimate of k_c=-0.006 min^-1, halving of the width of the notochord domain occurs within 2 hours. The notochord domain is about 20 cells wide in the early gastrula (as determined from expression of the Not gene floating head at 6 hpf) (seeMelby et al., 1996;Talbot et al., 1995). If the rate of convergence is constant, the domain will narrow to a domain only about a single cell wide during about 8 hours.

From our data, we propose that MIB is not only a prominent feature in dorsal mesoderm, but that this behavior quantitatively accounts for the shape changes of the field. We performed a severe test of this proposition for the notochord domain: From our measured velocities and locations of each cell in the field, we determined the overall values of k_c and k_Efor the domain at each time point in our recording. Then, we computationally subtracted from the movement pathways of each cell those components (the ML and AP relocations) of the movements that were accounted for by the MIB. As predicted, the resulting movement pathways, with these components absent,exhibited no apparent trends. The cells jitter about in a possibly random fashion, and there was no prominent change in the shape of the field. Hence,MIB appears to be the single cellular morphogenetic behavior rearranging cells and shaping the dorsal mesoderm, during the mid- and late gastrula period. This finding is in marked contrast to descriptions of teleost dorsal mesodermal morphogenesis in literature from a century ago (e.g.Morgan, 1895). At that time,internalized dorsal mesodermal cells were proposed to migrate, with cell intercalations playing no role at all. By that model there would be no gradients, the migrating cells would all be moving with the same velocities. We observed mesodermal migrations without gradients on the ventral side of the zebrafish gastrula (data not shown) (Myers et al., 2002a). We observed no such cell migration in the dorsal trunk-forming mesoderm of the embryo.

We show that in the wild-type embryo, convergence of the notochord-forming mesoderm occurs at about the same rate as the adjacent somitic mesoderm. This finding suggests that during gastrulation, all of the trunk dorsal mesoderm is behaving as a single unit with respect to axis narrowing, i.e. there appears to be mechanical continuity across the entire field. This is an extremely interesting result because the field becomes subdivided during the same period by prominent notochord-somite boundaries. From cell lineage analyses we know that from the onset of gastrulation, within the dorsal mesoderm, the domain of prospective notochord cells is already a lineage compartment separate from the somitic domains (Kimmel and Warga,1986; Melby et al.,1996). Our recordings also show that the domains are spatially separate during and after gastrulation: even as intercalations occur within the notochord and somite domains, they apparently do not occur across the notochord-somite boundaries, or between the two domains before the boundaries become recognizable. Moreover, by the stage our recordings begin, the notochord and somite domains are already expressing different patterns of developmental regulatory genes; genes encoding transcription factors,signaling molecules and adhesion molecules(Kodjabachian et al., 1999). Hence, it would be naive to imagine that the molecular basis of convergence of dorsal mesoderm is uniform.

Whereas the rates of convergence are similar for the notochord and somite domains, extension is much higher in the notochord domain, more than three times higher than in the somite for the wild-type example shown inFig. 4C. If MIB is driving convergence at the same rate in both domains, then why are the rates of extension different? One possibility is that the cellular populations are rearranging differently during MIB. In the notochord, convergence and extension are in relatively close balance. We understand this to mean that essentially all of the ML intercalations are pushing cells apart along the AP axis, as in the model shown in Fig. 2. Hence the lengthening of the axis is proportional to its narrowing, and the area of the domain examined along the plane of the intercalations is preserved (Fig. 8A-C). However, in the wild-type somite domain, the intercalations might thicken the tissue as well as lengthen it; i.e. the ML intercalations would push some cells apart along the dorsoventral axis. Our data are consistent with this possibility (e.g. Fig. 7C), but our recording procedures were not optimized to quantify the thickening in the somitic domains. Whether the difference we observe is due to somitic thickening or something else, these findings reveal complexity of MIB, suggesting convergence and extension are separately regulated.

Mediolateral cell intercalations underlie convergence and extension of the notochord in Xenopus (Keller et al., 1989) in a way similar to what we observe in zebrafish. By midgastrula stages, the notochord-forming cells have a distinct bipolar shape and exhibit protrusive activities at their tips(Keller et al., 1989;Keller et al., 1992). Then they become wedge-shaped as the notochord narrows to a domain only one or two cells wide. During early morphogenesis, convergence of the Xenopusnotochord is at least roughly balanced by extension(Keller et al., 1989). Convergence and extension also occurs in the somitic mesoderm inXenopus (Keller et al.,2000). Extension is more rapid in the notochord, and shearing between the notochord and somite is observed. Intercalating somitic cells that do not contribute to extension may thicken the somites dorsoventrally (seeKeller et al., 2000). Finally,disruption of the no tail ortholog in Xenopus, Xbra,disrupts convergence (Conlon and Smith,1999).

In each of these features, the convergence and extension behavior and regulation appear similar in zebrafish and Xenopus, suggesting that the basic cellular machinery of MIB has been highly conserved between the two species.

One difference between zebrafish and Xenopus is that in zebrafish the yolk syncytial layer (YSL) provides a substrate for the mesendoderm,rather than a blastocoele roof present in amphibians. YSL nuclei undergo convergence and extension, and their movements are similar to those of the overlying mesendodermal cells (D'Amico and Cooper, 2001). This observation raises the possibility that the zebrafish cell movements are, at least in part, passively imposed by the YSL. However, passive convergence and extension of the zebrafish notochord-forming cells seem quite unlikely. Examined in detail, the blastoderm cell movements generally occur faster than the YSL nuclear movements(D'Amico and Cooper, 2001),arguing that convergence and extension are autonomous in the YSL and cellular blastoderm.

As expected from previous studies of gene expression of ntlmutants (Melby et al., 1997),we see a prominent defect in convergence of the gastrula-stage axial mesoderm— essentially the field does not narrow(Fig. 8A). The defect appears largely limited to the axial mesoderm and is transient, observed only during the gastrula period. Later, after epiboly is completed, convergence of the domain in the mutant begins at a rapid rate, and the domain width becomes similar to that of the WT notochord by the three-somite stage. It may not be coincidental that the time when convergence initiates in ntl mutants is when we see morphological boundaries enveloping the axial domain together with the ventral neural keel (including the floor plate) that undergoes pronounced convergence and extension. Notably, wild-type floor-plate cells do not express ntl at these stages: its axial expression is limited to the notochord. Hence, the ntl^- cells, as they join the floor plate, may acquire the floor plate, ntl-independent, program of regulation of morphogenesis.

How does ntl regulate notochord morphogenesis, such that lack of function produces such a major defect? A candidate implicated inXenopus is Xwnt11 (Tada and Smith, 2001). This gene is a direct transcriptional target of the ntl ortholog Xbra. Dominant-negative constructs ofXwnt11 (Tada and Smith,2000), its receptor Frizzled-8(Wallingford et al., 2001b),or a key downstream target Dishevelled/Xdsh(Wallingford et al., 2000) all block convergence in Xenopus, apparently acting via a non-canonical planar polarity Wnt/Ca²⁺ pathway(Choi and Han, 2002;Tada and Concha, 2001;Wallingford et al., 2001a). Mutational analyses involving silberblick/wnt11(Heisenberg et al., 2000)clearly reveals a role in early morphogenesis of the Xwnt11 ortholog in zebrafish. However, key aspects of Wnt11 regulation differ between the two species. slb/wnt11 is first expressed in the early blastoderm margin,and this expression domain is independent of ntl(Makita et al., 1998). Midlinentl-dependent expression becomes strong only later, reaching highest levels at the 5- to 12-somite stages. Hence slb/wnt11 is an unlikely candidate for mediating ntl-dependent MIB in the early axial mesoderm in zebrafish.

FGFs are attractive alternative candidates. FGFs are regulated by T-box genes, and are strongly implicated in the control of mesodermal morphogenesis,convergence and extension in particular. FGFs act via downstream targets including ephrins (Chong et al.,2000), and, like Wnt11, intracellular Ca²⁺ release(Nutt et al., 2001). Again there are key regulatory changes between Xenopus and zebrafish: inXenopus the key Xbra-dependent FGF in axial mesoderm is eFGF, but in zebrafish other FGFs, including FGF8 and FGF3, play at least part of the role of this gene in dorsal mesodermal morphogenesis during gastrulation (B. Draper, personal communication).

Current understanding is that both noncanonical Wnt signaling and FGF signaling are regulating cell polarity as would be required for the cells to orient MIB properly. Interestingly, our analyses indicate that the major defect in ntl mutants may not be loss of cell polarity: We find that cells in the mutant can apparently initiate correctly oriented MIB. They frequently move towards their neighbors(Fig. 11), and make and break contacts with their neighbors (Fig. 10) in the same orientation as in the wild type. However, although many ML contacts are made in the mutant, the cells appear unable to exert the necessary tension to pull together — rather, they frequently slip apart. This proposal explains the ML cellular intermixing, as well as the randomized ML movement of cells with respect to the midline(Fig. 8) and one another(Fig. 11).

Loss of the tension necessary to pull the tissues together may reflect a requirement for an adhesion molecule missing in ntl mutants. Studies of the paraxially expressed T-box gene spadetail (spt;tbx16 — Zebrafish Information Network)(Griffin et al., 1998) provide a clue about what might be missing from the ntl^-convergence machinery. spt function is required for proper paraxial cell behaviors (Ho and Kane,1990; Kimmel et al.,1989) and is hierarchically upstream of a protocadherin that is critical for proper mesodermal cell-cell adhesion during gastrulation(Kim et al., 1998;Yamamoto et al., 1998). Lack of a corresponding adhesion molecule in the ntl^- axial domain (Kuroda et al., 2002)might underlie the defects we observe.

Dorsal mesoderm efficiently extends in ntl mutants. It extends without significant convergence in the axial domain, clearly revealing that a mechanism other than MIB can drive extension in this domain. Furthermore,somitic mesoderm in the mutant also appears to extend more rapidly in the wild type, where, as discussed above, convergence normally seems to be driving thickening as well as extension.

The result comes as a surprise because there was no hint from our data, or from the literature, that anything other than MIB drives dorsal mesodermal extension. Finding that extension occurs in the mutant without convergence may well be providing a clue about how normal development works. For example, we do not understand what normally orients the AP pushing apart of cells during MIB such that convergence and extension are normally in balance in the notochord domain. The convergence-independent extension in the mutant may be uncovering something about this mechanism.

We examined specifically whether abnormal radial intercalation between axial hypoblast and epiblast might account for the observed extension in the mutant, and observed that it did not (Fig. 7B). However, radial intercalations are very likely to be occurring within (not between) these cell layers, particularly in the dorsal mesoderm. We infer this because by late in gastrulation in ntlmutants (before convergence picks up in the mutant), both axial and paraxial mesoderm have thinned to a layer only a cell or two thick(Fig. 7C). During the same period, the cellular density of the mutant axial domain is substantially decreasing (Fig. 8D). Rather than packing together as in the wild-type notochord (which eventually becomes very cell-dense; Fig. 8D), the cells of the mutant axial domain are spreading apart predominantly in the AP direction, and this spreading accounts for the extension of the domain.

The cellular spreading observed in ntl mutants suggests that blastoderm epiboly could be driving extension in the mutant. Epiboly occurs simultaneously with convergence during the stages we studied. During epiboly,blastoderm cells intercalate radially(Kimmel and Law, 1985;Wilson et al., 1993) and spread apart along the AP axis (Concha and Adams, 1998; Warga and Kimmel,1990). Solnica-Krezel et al.(Solnica-Krezel et al., 1996)previously suggested that epiboly contributes to extension in volcanomutants, and epiboly is underlain by an AP velocity gradient (R. J. A., D. Faruque and M. L. Concha, unpublished). We emphasize that our proposals for epiboly driving extension in the ntl mutant does not in any way argue against MIB as the key mechanism in the wild type. Indeed, in other mutants with disrupted convergence such as knypek and trilobite, or the knypek-trilobite double mutants, extension is also disrupted,even though epiboly is intact (Myers et al., 2002a). These genes function in the Wnt planar polarity pathway (Myers et al., 2002b)and the results, combined with our findings with ntl, might be revealing that Wnt signaling drives extension that is closely coupled to convergence, but that ntl^- disrupts this relationship leaving other, parallel, mechanisms able to extend the axis. Furthermore, in epiboly mutants (Kane et al.,1996) or in embryos treated with a teratogen(Bauman and Sanders, 1984),both convergence and extension of the notochord can occur in the absence of epiboly, the opposite of the findings with the Wnt signaling mutants. Hence,it is likely that MIB normally functions redundantly with other cellular mechanisms to produce extension of the dorsal mesoderm. This proposition might be examined with new genetic analyses, e.g. in double mutants in which epiboly, Wnt signaling and ntl function are disrupted in combination. With the approach and quantitative methods we have developed for this study,we can now account for the individual reorganizations of populations of cells and distinguish between competing models for the mechanism of morphogenesis and its control.

We thank Charline Walker and Amy Biggs for technical assistance. Ray Keller, Peter B. Armstrong, Mark Cooper, Lilianna Solnica-Krezel, and members of the Institute of Neuroscience, University of Oregon, provided thoughtful comments on a draft of the paper. The research was supported by NIH grant HD22486 and an MRC Senior Research Fellowship to R. J. A.