What switches on when
Four hundred and nine tardigrade genes, placed next to the genes whose expression rises and falls with theirs. Nothing here was drawn by hand: the positions come from the measurements, and so does the order. Drag the slider and the embryo's four days run past — watch the chromosome machinery go dark, then a tight knot of patterning genes fire, then the synapses build for the rest of development.
The pale vertical bands are gaps where nobody sampled — the widest runs twelve hours, from 20 to 32. A gene whose real peak falls inside a band is recorded at whichever edge sample is highest, which is why points stack in columns on the right-hand edge of each one. Sixty-three of these genes “peak” at exactly 41 hours, the first sample after a six-hour hole. That is the sampling, not the animal.
Lines join genes whose profiles across the 62 embryos are in each other's top eight and correlate above 0.55. 77 of the 409 genes have no partner that close and are left out of the drawing rather than floated around its edge — they are still on the curves page. Point size is how much of that gene is present at the moment on the clock, not its total. Position is a force layout of those same lines, so distance on screen is not a measurement.
Start at 2 hours and press play. The first thing that happens is a shutdown. The condensin machinery — the protein complexes that fold chromosomes for division — is loaded into the egg by the mother and is already there in the first embryo sampled. Somewhere around 41 hours the condensin II subunits go out together, all three of them, while condensin I stays on to the end. Nothing we know of tells them to stop at the same moment; they simply do.
Then the knot in the middle lights up. Patterning transcription factors peak between 41 and 53 hours — a twelve-hour window in a four-day development — and around them, by co-expression alone, sit cadherins, immunoglobulin adhesion molecules and a Celsr/Flamingo, the cadherin that lays out axon tracts in a fly. A POU4 gene, the tardigrade counterpart of the unc-86 that specifies neuron identity in a nematode, comes on in the same window after being undetectable in almost every earlier embryo.
And then the long tail. The synaptic genes do not pulse. They climb from about 45 hours and are still climbing when the animal hatches at around 101. The pattern is laid down fast and the wiring is built slowly, and here you can watch the difference rather than take it on trust.
What a line does and does not mean
This is not a regulatory network, and it would be easy to mistake it for one. Nothing here observes regulation. Two things are measured, and both are descriptions of expression: when a gene switches on, taken as the first time it passes a fifth of its own maximum, and how closely two genes' profiles track each other across the 62 embryos.
Where two connected genes switch on more than three hours apart, the line carries a direction. That direction means precedes. It does not mean regulates. A preceding B is equally consistent with A driving B, with something else driving both, and with coincidence — and 156 of the 570 lines here are oriented at all, so for most of them we cannot even say that much. The page that collects the regulatory relationships somebody has actually demonstrated, from in situ hybridisation and knockdowns, is the one next door, and it is much smaller. That difference in size is an honest picture of the field.
How this page was built
The expression comes from Levin et al. 2016's 62 single timestamped embryos, as counts per million, smoothed with a six-hour sliding window — a window in hours rather than in samples, because the embryos are not evenly spaced in time. The clock reads hours post laying, which for this dataset is 2 × the supplement's minutes / 60 + 2: those minutes are half of real time, and the zero is the two-cell stage rather than laying. We had that wrong on this site until 24 September 2026, and the hour-by-hour page explains how it was settled.
The 409 genes are every gene our knowledge graph has named and can place on the genome, plus the condensin subunits, plus the three nearest unnamed profiles to each of those — which is why most of the points have no label. The lines are mutual top-eight neighbours correlating above 0.55. To check that this structure is not an artefact of the method, the same construction was run 100 times on profiles shuffled independently gene by gene: it produced 0 ± 1 lines, never more than 4, against the 570 here.
The code is tools/brain/regulatory_network.py. One thing it does is worth copying:
it measures when each gene stops as well as when it starts. Every condensin subunit is
maternally loaded and present in the first embryo, so all their onsets are identical and say
nothing — and the one worked example this page was built around would have been invisible in
a view that only knew about switching on.