Embryo

Roundworm

The roundworm is one published recording: the nuclei files of the WormGUIDES atlas (Santella et al. 2015, BMC Bioinformatics 16:189), read from its public repository at commit 17be611. It is one individual Caenorhabditis elegans embryo (strain JIM113, image series DISPIM_061416, imaged uncompressed on a light-sheet microscope, one frame a minute) whose nuclei were tracked automatically and then curated by the atlas’s authors.

Cell names follow Sulston et al. (1983, Developmental Biology 100:64–119). Fates come from two files in the same repository, the parts list and the list of programmed cell deaths, both after Sulston. The faint outline is the atlas’s own “Embryo Outline” mesh.

The source repository is licensed GPL-3.0 and declares no separate data license. The files Embryo loads are derived from it, so they are published under GPL-3.0 as well, with a notice beside them naming the source, the commit and what was changed.

The published files are content-addressed: build a829bdbfd6c7, published 2026-09-18, 1.6 MB in all. A new export is a new build with a new address, so a shared link never silently changes what it shows.

Sea squirt

The sea squirt is one published recording: embryo Phmamm-1 (ASTEC-Pm1), one wild-type Phallusia mammillata embryo imaged live on a light-sheet microscope, from 64 cells to the tail bud. Its cell shapes are the per-cell surface meshes published with Guignard et al. (2020, Science 369:eaar5663). Of the three levels of detail in that pack, Embryo draws the middle one.

Cell names and tissue fates come from the 2025 curation of the same embryo (Biasuz et al. 2025, bioRxiv, doi 10.1101/2025.10.13.681089), joined to the shapes by cell id and by nothing else. Both are read from the embryo’s figshare record (doi 10.6084/m9.figshare.8223890.v6).

Both works are published under CC BY 4.0, so the files Embryo loads are shared under CC BY 4.0 too, with a notice beside them naming the record, the two works to credit and what was changed.

One thing is still untraced. The published table gives every cell a tissue fate, but the authors’ documentation does not say which fate-map work those fates rest on. Until we have traced it, Embryo calls them the published table’s fates, and nothing more.

The published files are content-addressed here too: build 0bfe8e50886d, published 2026-09-19, 21.2 MB in all, streamed in 15 pieces, so the first view needs 1.3 MB.

Roundworm cells

The roundworm’s cells come from the CShaper morphological atlas (Cao et al. 2020, Nature Communications 11:6254), embryo Sample04, the one the paper itself works through. It is a single Caenorhabditis elegans embryo on a confocal microscope at 0.25 µm per voxel, about a minute and a half between timepoints, whose cell membranes the authors segmented with a neural network and then curated.

Cell names follow Sulston et al. (1983, Developmental Biology 100:64–119), joined through the atlas’s own name dictionary. Every segmented region of every timepoint resolves to a name, so no cell on screen is unnamed and none had to be guessed.

This recording carries no fates. The plain-English fate sentences the roundworm beside it shows come from a GPL-3.0 parts list, and joining them here would put a copyleft license on a data set that is otherwise cleanly CC BY 4.0. The founding families and what they become are read from the name after Sulston 1983, which is a paper rather than a code base, so those stay.

The atlas is published CC BY 4.0. Its bytes are not held by the archive itself: the record is a link to a public file share, which we mirrored before building anything on it. The files Embryo loads are derived from that mirror and published CC BY 4.0 as well, with a notice beside them naming the work, the license and every step below.

The published files are content-addressed: build 8a2615fd6242, published 2026-09-20, 8.5 MB in all, of which 0.2 MB before the first view can draw. A new export is a new build with a new address.

Roundworm

360 recorded minutes, from 4 cells to 580. Over the whole recording 1,330 cells are tracked, at 106,275 measured positions: one per cell for every minute it exists, with no gaps.

663 cells are seen dividing, each into exactly two observed daughters born the minute after the mother’s last. 113 tracked cells are on the published list of programmed cell deaths.

Of the 580 cells alive at the last minute, the published parts list names a fate for 553. The rest are shown without one.

Sea squirt

180 recorded frames, from 64 cells to 702. Over the whole recording 1,342 cells are tracked, as 58,336 cell shapes: one published mesh per cell for every frame it exists, with no gaps.

639 cells are seen dividing, each into exactly two recorded daughters born the frame after the mother’s last.

The table names 732 of those cells in Conklin’s scheme. No cell born after frame 125 has a name, so of the final 702 cells only 95 do. Every cell carries a tissue fate at every frame, 21 distinct labels in all.

Time is the table’s own clock: it puts the first frame at 4.1 hours after fertilization and gives 1.5 minutes per frame, which puts the last at 8.4 hours.

Roundworm cells

150 timepoints, about 1.5 minutes apart, from 4 cells to 334. Over the whole recording 803 cells are tracked, at 21,017 segmented cell-timepoints with a surface each.

400 cells are seen dividing. Ancestry is read from the name rather than from the tracking, so a cell’s mother, sisters and descendants are exact even where the segmentation stumbles.

The segmentation does stumble, and this is the recording’s weakest point. 1,589 cell-timepoints inside a cell’s run are missing, 7.0 percent of the total, spread over 270 cells. Loss is near zero until about 190 cells and climbs steeply after, so by the last timepoints roughly a sixth of the cells are absent.

A missing cell-timepoint is left missing. It keeps its slot in the track with a radius of zero and is simply not drawn, so a cell can vanish for a stretch and come back. The caption says how many timepoints a cell lost.

Roundworm

Each ball is a cell’s nucleus, at the position and diameter the atlas records for that minute. Nothing is interpolated, smoothed or imputed. The horizontal position and the diameter are stored exactly, and the depth to within 1/512 of a voxel.

Units are the microscope’s voxels. The source declares no physical size, so the roundworm shows no scale bar and states no length.

The view is turned so that up is up and the embryo’s left stays its left. Nothing is mirrored, and the head-end and tail-end labels follow the axis the source declares.

Sea squirt

Each shape is one cell’s surface as published: a closed mesh per cell per frame. Embryo switches meshes from frame to frame and never morphs one into the next.

To fit in a browser, a cell mesh with more triangles than its frame’s budget is simplified. That applies to 37,229 of the 58,336 shapes, mostly the larger cells of the early frames, and takes 10.5 million published triangles down to 6.8 million. As cells get smaller and more numerous, fewer need it. Every other shape is the published mesh, untouched. Vertex positions are kept to within 0.07 µm.

Units are real here: we measured the mesh frame at 0.5 µm per unit against the table’s cell centers, so the cut’s readout and the embryo’s length are in micrometres. A handful of early shapes sit well away from where the 2025 table puts the same cell. We draw them as published and have not traced the cause.

A rotation measured from the data turns the view into the body’s own frame: long axis from head to tail, up toward the side that folds in. Nothing is mirrored or re-registered. That frame places the camera, the two end labels and the cutting directions, and nothing else.

Roundworm cells

The atlas publishes label volumes rather than meshes, so the surfaces on screen do not exist in the source. We built them. Each timepoint’s volume is downsampled by a stride of 2 to 5 voxels, chosen per timepoint so the few huge cells of the opening are not sampled at a hundred times the detail their budget can hold.

For each cell we then extract an isosurface over its own voxels by marching tetrahedra, smooth it with a Taubin low-pass that does not shrink the surface, and reduce it by vertex clustering to a budget of at most 240 triangles. 78 million extracted triangles become the 2.8 million the browser loads.

Reduction pulls a surface inward, so each cell is finally scaled about its own center until it encloses the volume the source measured. That is the only number in this export that is adjusted at all, and it changes a cell’s size rather than its shape. Measured against the atlas’s published volume table, the cells on screen sit within about one percent.

Positions are in real micrometers, 0.25 µm per source voxel. The view is turned by a rotation measured from the cells themselves: the long axis runs head to tail, signed by descent, since the AB lineage is the embryo’s front and the P1 lineage its back. Nothing is mirrored.

Roundworm

The recording is one frame a minute. Between two frames Embryo draws a straight line from one measured position to the next, and for a division from the mother’s last position to each daughter’s first. That in-between motion is our drawing, and the readout under the timeline says so while it plays. The view settings can turn it off, and the embryo then steps from minute to minute.

A cell whose track ends shrinks out over its last minute. Every such cell is on the published list of programmed deaths.

The outline switches from one mesh to the next each recorded minute, with no blending. The repository doesn’t say how it was traced, so Embryo calls it a modeled outline. It holds still until minute 252 and then begins to bend.

Sea squirt

Nothing is drawn between frames: the sea squirt steps from one published frame to the next, and its shapes are never blended.

While a frame’s shapes are still downloading, each cell is drawn as a bead at the center of its published mesh, sized to the mesh’s volume. The readout says that beads are standing in, and the shapes replace them as they arrive.

Cutting the embryo open removes whatever lies on one side of a flat plane. What shows on the plane is each cell’s own inside wall, painted flat in the cell’s color. That paint is a drawing aid, and the slice itself was never measured.

See-through skin thins the skin-fated cells to a veil so that the tissues under them show. Following a kept cell only turns the camera. Playing backwards shows the same frames in reverse order.

Roundworm cells

Every drawn timepoint is a segmented one. Shapes are switched from one timepoint to the next and never blended, so there is no in-between motion to explain here.

While the surfaces load, each cell is drawn as a ball at the center of its segmented volume, sized to match it. The view settings can leave it that way, which takes the download from 8.5 MB to 0.2.

A cut is a clip plane through the whole embryo, and the flat faces it exposes are painted in each cell’s own color so the interior reads as a solid. Those painted faces are a drawing aid and show nothing the recording measured.

Roundworm

A cell’s name is its ancestry. Every letter is one division: a and p for the front and back daughter, l and r for left and right, d and v for top and bottom. That is how Embryo finds a cell’s ancestors, its descendants and its relatives, by reading names rather than guessing from positions.

Color is the founder family: the 6 founding cells AB, MS, E, C, D and P4, plus a gray for the few early cells that exist before the founders have split off. What each family goes on to become, as the legend words it, is after Sulston et al. 1983. That comes from what is known about the species. This recording doesn’t show it.

A fate is the published parts list’s own wording, copied as is and only for a cell the list names. It is never inferred, and never passed from a cell to its relatives. For a cell that divides, the caption counts the published fates of its final descendants.

Several family colors are hard to tell apart under some kinds of color blindness, which is why the legend and the caption always name the family in words as well.

Sea squirt

Color is the published tissue fate of each cell, what its descendants are recorded as becoming, and the caption quotes the table’s label word for word. For a cell that divides, the caption counts the labels of the cells it has become by the last frame.

We grouped the labels into the legend’s 7 tissues ourselves. Each of the 21 labels goes to the first of skin, nerves, notochord, tail muscle, gut and germ line whose word it contains (Epidermis, Neural, Notochord, Muscle, Endoderm, Germ), and the rest are “other tissues”. The plain words beside them are general descriptions, and labels within one tissue differ slightly in shade.

A cell whose labels span two of those tissues is drawn striped in both: 18 cells, none after frame 108.

Names come from the table, in Conklin’s scheme. Where the table has none, the cell has none. Ancestry comes from the recorded mother-to-daughter links, so relatives reach back only as far as the first frame.

Here too, some tissue colors are hard to tell apart under some kinds of color blindness. The legend names every tissue and the caption repeats it in words.

Roundworm cells

A cell’s name is its ancestry, exactly as on the roundworm’s beads. Every letter is one division: a and p for the front and back daughter, l and r for left and right, d and v for top and bottom. Names here run to 11 characters.

Color is the founder family: the 6 founding cells AB, MS, E, C, D and P4, plus a gray for the few early cells that exist before the founders have split off. This is the same palette and the same rule as the beads, on purpose, so the two recordings of this animal can be read against each other.

What each family goes on to become, as the legend words it, is after Sulston et al. 1983. That is general knowledge about the species and not something this recording shows.

No cell here carries a published fate sentence. The recording stops at 334 cells, where most cells are still a division or two short of the body parts they are named for, and the parts list that supplies those sentences elsewhere is licensed in a way we will not import onto this data set. The caption therefore says where a cell came from and where its line goes, and stops there.

Roundworm

• It ends at minute 360, with 580 cells and the first bend of what will become a worm. Soon after, the embryo starts to move inside its shell, which is what stops this kind of tracking.
• It starts at 4 cells. The first divisions happened before recording began, so the earliest ancestors are named but never seen.
• It shows only nuclei, with no membranes or cell shapes, so a tissue reads as loose beads instead of a surface.
• It has no scale. Positions and sizes are in voxels of a recording that declares no physical size.
• The positions are the atlas’s curated tracking. We did not re-check them against the images.
• Fates are the parts list’s vocabulary, written for anatomists, and 553 of the final 580 cells carry one.

Sea squirt

• It ends at 8.4 hours, at the tail bud: 702 cells, a notochord that has only begun to line up, and no tadpole yet.
• It starts at 64 cells. Everything before that is unrecorded, and relatives cannot be traced past it.
• The shapes are the published segmentation, errors included. We did not re-check them against the images, and the large cells of the early frames are simplified to fit a browser.
• Late cells have a fate but often no name: 95 of the final 702 are named.
• Fates are broad tissue labels. Which fate-map work they rest on is still untraced.
• The painted slices of a cut are only a drawing aid.

Roundworm cells

• It ends at 334 cells, where the roundworm’s beads reach 580. This is about the first two thirds of that recording, and it stops before most cells arrive anywhere.
• It starts at 4 cells. The first two divisions happened before recording began, so the earliest ancestors are named but never seen.
• The segmentation loses cells as the embryo fills: 1,589 cell-timepoints are missing across 270 cells, and roughly a sixth of the cells are absent by the end. Those gaps are left visible rather than filled.
• We built the surfaces. The atlas publishes labeled voxels, which we downsampled, then extracted an isosurface from, smoothed and reduced to a triangle budget. Fine detail at the scale of a few voxels is gone.
• The outline never changes. A roundworm embryo develops inside a rigid eggshell, so the silhouette is the same at the last timepoint as at the first. What changes is everything inside it.
• It carries no fates at all, by choice. The sentences the beads show come from a copyleft parts list that would change this data set’s license, and we left them out.