Those of us who build tools have a weakness for ever-growing inventories. Give us a new instrument and we will count something with it, and then we’ll count it again, more cheaply and in more samples. Spatial biology is the latest beneficiary, and it may also be the most expensive way yet devised of not asking a question.
We can now say, with high granularity and at great expense, which cells are in a tissue, where they sit and whom they sit next to. It’s essentially a biological Google Map. But a map is not a story. A tissue is a city where millions of cellular citizens decide when to divide, differentiate, wander off or die, and interpreting it means reconstructing those decisions, not taking a census.
So, what informs their decisions? Some of it is memory: lineage. Some of it is what the cell is doing at that moment: its metabolism, its stress, its shape, the pathways that are firing. The rest arrives from outside. Location, neighbors, matrix, vessels, immune cells and mechanical forces all act on it, from a few microns out to several millimeters.
Early scientists asked how cells coordinate their decisions to give rise to spatial patterns in developing tissues. Alan Turing gave us morphogens (1), and Lewis Wolpert gave us the French Flag (2). Both assumed that every cell reads a signal in exactly the same way: so much morphogen, so much cell fate. Cells have never read these papers. Development works not because signals are clean but because cells are good at interpreting dirty ones.
Think of the location services on your phone. It combines GPS, Wi-Fi and cell towers to work out where you are, trusts whichever signal is most reliable at the moment, and falls back on the others when GPS drops indoors. Cells must run something similar, weighing intrinsic and extrinsic signals, and the algorithm is written in the genome, although we have yet to decipher it. Maybe AI will, but only if we feed it clean signals. (Sadly, our phone has more sophisticated ideas about position than most of our models of development.)
Cancer is the obvious case. Epithelial cells normally grow and mature inside well-defined zones, and the ones that wander usually get rid of themselves. Tumor cells lose their GPS, or keep navigating with a map of a neighborhood that no longer exists. Lost in an unfamiliar environment, they revert to their infancy, becoming stem-like, invasive and very hard to evict.
What would we need to measure? History means clonal relationships read from somatic mutations or barcodes, along with epigenetic marks, and in intact human tissue this is the measurement we still mostly cannot make. State is more tractable: identity markers, signaling proteins in their active forms, where the transcription factors are, and the cell’s cycle, metabolism and stress.
Surroundings are not one measurement. Neighbors and what they are doing, the ligands and receptors in play, the matrix, the gradients and the forces: each must be taken at more than one distance, and no single scale will give you all of them. Time must be recovered from serial samples or from disease stages arranged within one specimen.
How are we doing? We are counting better than in 2014 when in situ sequencing was brand new. Spatial multi-omics now measures cell state directly, and whole slides can be imaged at single-cell resolution, so multiple scales are finally within reach. Machine learning is beginning to extract rules rather than merely labels. But the data remain noisy, fragmented, shallow or underpowered. We have built a very large machine for counting things imperfectly.
Pouring ever more AI onto messy measurements is not a cure; it is a way of making the mess look tidy. A model trained on noise will learn the noise, confidently. In the future post we will ask whether computational denoising quietly erases the very things worth finding: the rare, the local and the therapeutically relevant.
References
-
A. M. Turing, Philos. Trans. R. Soc. Lond. B 237, 37 (1952).
-
L. Wolpert, J. Theor. Biol. 25, 1 (1969).