What the genes are doing, hour by hour
Sixty tardigrade embryos, each killed and sequenced at a known moment — one every hour from the two-cell stage to about a hundred hours, when the animal hatches. That is 9,069 genes measured across four days of development. Drag the slider and watch which ones switch on.
Start with condensin II, because it answers a question we published a paper about.
Condensin II is the machine that folds chromosomes. In 2021 a Science paper reported that tardigrades had lost it — no CAP-H2, no CAP-D3, no CAP-G2 — and used that to explain their unusual genome organisation. We found CAP-H2 in the genome, modelled how its clasp grips the ring, and published it.
Here is the part no paper shows. Scrub the first half-day and CAP-H2 is running throughout, swinging between about 50 and 290 counts per million across every one of the fast early divisions. Scrub past 48 hours and it is gone: a thirtyfold drop, and it never comes back while the embryo grows limbs, claws, a gut and a nervous system.
And it does not go alone. Switch to the cell-cycle set. Histone H3 falls ninetyfold over the same span, cyclin B eightysixfold. Condensin II is not shutting down on its own schedule — it goes down with the whole replication machinery, at the moment the embryo stops dividing fast and starts making organs.
Correction, 22 September 2026. This paragraph used to end: “Then switch to the nervous-system set: pax6 and elav do the opposite, and rise.” That was wrong, and re-running the analysis for the preprint found it. pax6 does rise, but as a pulse — near zero through cleavage, 230 cpm at 47 hpl, 444 at 53 hpl, then back to tens of cpm; across the whole series it has no trend at all (Spearman ρ = +0.08, p = 0.57). And elav does not rise: it falls 6.2-fold (ρ = −0.67), reading as zero in every embryo after about 58 hpl. For a pan-neural gene in an embryo building a nervous system that is odd enough that we treat it as a statement about this dataset rather than about the animal.
So "when does condensin II activate" has an answer, and it is not the one the question expects. It never switches on, because it arrives switched on, supplied in the egg. What it does is switch off, with the rest of the division machinery.
So we tested it: drop those nine and recompute. The ratios do not move at all — 32×, 92× and 86× with or without them — because they sit in the gap rather than on either side. The collapse is real. What we cannot say is exactly when inside that window it happens, because the only embryos in it are the flagged ones. The honest statement is "between 32 and 48 hours after laying", not a number.
One more note on the clock, because this page had it wrong until 24 September 2026. Levin's supplementary table gives sample times that are half of real time, and his zero is the two-cell stage rather than laying; the public GEO labels for the same embryos are twice the supplement's values. We settled which is right on 18 September by matching every embryo against a second dataset's daily egg samples: only on the doubled axis do they fall onto egg days one to five in order. So the clock above reads 2 × minutes / 60 + 2, and the series ends at about 101 hours, at hatching. For six days the curves here ran at half speed against stage markers that were already in real hours, which made genes look as though they peaked a whole stage earlier than they do. The question is closed; the correction is in the build log.
What you are actually looking at
Each track is one gene. The line is its abundance across development, in counts per million — the normalisation matters, because the embryos were sequenced to different depths (1.6 to 4.3 million reads) and raw counts would make a deeply sequenced embryo look like a busy one. The vertical line is where the slider is; the number on the right is that gene's level at that moment.
How this page was built
The counts come from Levin et al. 2016, who sequenced single embryos once per hour for four and a
half days; 62 passed quality filtering and 60 carry a usable timestamp. We mapped their transcripts
onto genes in the H. exemplaris reference assembly
(tools/genome/layer2.py), normalised per embryo and aligned the result to the staged
timeline the embryo simulation runs on
(tools/genome/expression_timeline.py).
Genes are found by their annotation rather than by an accession written into the code. The first version of this page hardcoded five gene identifiers, every one of them was wrong, and it produced a panel with three silently empty tracks. Looking a gene up by what the genome says it is fails loudly instead. Where several genes carry the same annotation — four features are annotated CDK1 — the one actually expressed is shown and the others are counted beside it.
Of 9,069 genes, 429 are here: the named ones, plus those whose level actually changes across development. A gene that is flat, or never rises above noise, has no story to tell on a timeline and would only make the list harder to search.