What we said the genes would do, before we looked

A model of an embryo implies things about its genes. If cells divide on a particular schedule, the genes that run cell division should follow that schedule. So we wrote down eight predictions with hours attached, committed them, and only then opened the data — 62 single embryos, one per hour, sequenced whole. Two held. Five failed. The biggest failure rebuilt the model, which is the best thing a failed prediction can do.

The clock they were predicted against

Hours after the egg is laid, on a logarithmic scale because the interesting things crowd into the first day. The hatched band is a gap in the data: no embryo was sampled between 20 and 32 hours, so nothing can be said about what happens in there.

Each prediction, over what the genes actually did

The shaded bar on each chart is the window the prediction named. The line is the median expression of that set of genes, on a log scale, and the faint lines behind it are the individual genes. Open a card for what we said, what happened, and what changed because of it.

What this was worth

Five failures out of eight looks bad and is not. A prediction written after seeing the data cannot fail, and so cannot teach anything; these could, and one of them did something no amount of careful modelling would have.

The cell-cycle genes are the case. Our model had cells stopping division when the outer layer becomes an epithelium, at about 16 hours. The genes that run division stay at full cleavage level until about 38 hours, and only fall away between 38 and 50. That is not a detail — it changes how the animal gets from five hundred cells to its final thousand or so, and it moved the whole late half of the embryo onto a different schedule. We would not have found it by looking harder at the model.

Three of the failures say the same thing in different words. The polarity and junction genes never rise, because they are already there from the two-cell stage. The cells that move inside at 12–15 hours never transcribe the factors that normally license that move. The germline pool is not diluted but cleared, in one step, at 15–17 hours. The early embryo is not building what it needs as it goes: it is running on what the egg was loaded with, and the first large transcriptional change comes later than any of us expected.

One prediction was a fork rather than a guess. Nobody has published when the gut lumen opens, and our model can open it whenever we like — which makes it an assumption, and assumptions in this project have to be labelled. So we named two possibilities and the measurement that separates them. The ion-transport genes that would pump a lumen open are at their lowest exactly when the early hypothesis needs them high. The late opening it is.