Where the nucleus isn't

To model an embryo as cells packed together, you need a starting point for each cell. Almost everyone uses the nucleus, because in the animal these models are tested against, the nucleus is what the microscope gives you. We measured how wrong that is, and found something more useful than we expected: the error is not random, and neither of the two obvious ways to fix it works.

The shortcut

If you know where cells are, you can work out which ones touch by dividing the space between them — each cell gets the region closest to its own point. It is a good trade, because contacts are hard to measure and positions are not. The reference map of which cells touch which in the nematode C. elegans was built this way, and so were several since.

But in that construction the point you supply is also, silently, the centre of the cell. A nucleus is not the centre of its cell. It is an organelle with its own position, pushed around by its own mechanics, and in an early embryo it is often conspicuously off to one side.

How far off, measured

There is a public dataset where the same embryos have both: membranes segmented cell by cell, and the tracked nuclei. So we measured the distance between each cell's nucleus and the true centre of that cell, for every cell in every frame of two embryos.

That last number is the one that matters. If nuclei were scattered randomly inside their cells, the error would be noise and would partly cancel out across a tissue. It isn't. Nuclei sit further out than their own cells' centres — away from the middle of the embryo — in about two thirds of cells, and the size of that lean follows the same law in both embryos, agreeing to within 4%. That makes it a bias, not noise.

It also does not go away as the embryo divides. Cells get smaller, and the offset shrinks at almost exactly the same rate, so as a fraction of the cell it stays put.

What it costs

We ran one model twice over the same embryo, changing nothing but where each cell starts: once at the nuclei, once at the cells' measured centres. Seeding at the centres is better in 37 of 38 frames, and we then ran the whole thing again on a second embryo and got the same answer, better in 48 of 53. The penalty grows as the embryo crowds — below fifty cells the two are almost identical; above two hundred and fifty the model seeded at nuclei starts missing contacts that are really there.

The part we expected to be a footnote

A result like this is only useful if something follows from it, and for most studies nothing does: if you had the cell outlines, you wouldn't be using nuclei in the first place. So we tested the two corrections anyone would reach for, expecting to report that they help.

So the honest answer is the unhelpful-sounding one

The cost of starting from nuclei is not a calibration error that a constant can absorb. It is the part of a cell's shape that the nucleus does not know about, and the only way to get it is to measure the cell. If you have segmentations, use them; it is free and it is better almost every time. If you don't, the honest budget for your contact map is the larger error, not the smaller one.

And read whatever error you get against the grey line rather than against zero. A model that has reached the disagreement between two real embryos of the same species has reached the end of what comparing it to one specimen can tell you.