A cell deciding what to become is reading its mail. Stem cells send and receive chemical signals from their neighbours, and those signals — relayed through multi-step signalling pathways — switch genes on and off, determining whether a cell ends up brain, liver or muscle.
Retracing the sequence of instructions a given cell received would be an extraordinary tool for understanding how tissues form and how that goes wrong in disease. It has not been done at scale, and the reason is an assumption.
The assumption that was blocking the field
Scientists have long held that the effects of a signalling pathway vary widely across cell types. If true, mapping signalling histories means mapping every pathway separately in every cell type — a combinatorial explosion of experiments, each painstaking.
Work led by Pulin Li, MIT assistant professor of biology and member of the Whitehead Institute for Biomedical Research, with graduate student Nicholas Hutchins, found otherwise. Each pathway leaves a distinctive “fingerprint” — a characteristic pattern of gene activity reflecting the signals a cell encountered — and crucially, those fingerprints are consistent across different cell types for the same pathway.
That single empirical finding is what makes everything downstream possible. If the signature of a pathway is stable, you do not need to map each cell type. You need to learn the fingerprints once and then look for them anywhere.
IRIS, and the transfer-learning argument
The model that reads them is IRIS, a neural network that examines a cell’s overall gene activity and estimates which pathways were likely active at specific points during development — including for cell types it has never encountered.
Li’s analogy for why that works is the clearest statement of the method:
“Think of voice recognition systems like Siri, which are trained mainly in English, but then use that training to help them recognize other languages. This is called transfer learning, and this is why IRIS can work across many different cell types.”
Training data: a large experimental dataset measuring how thousands of human embryonic stem cells responded to dozens of combinations of six major signalling pathways at multiple developmental stages — an atlas of how signal combinations shape cell behaviour.
Test: single cells from mouse embryos during gastrulation, the stage when cells rapidly commit to divergent fates. IRIS accurately predicted when and where specific pathways would activate in cells destined for heart, gut, muscle and spinal cord tissue.
Note what that test actually demonstrates — different species, different experimental modality, cell types outside the training distribution. The generalisation claim is doing real work, not being asserted.
The findings are published open access in Nature Methods, dated 8 September.
Why this is in a creative-technology publication
Two reasons, and the second is the substantive one.
The first is the method. This is a textbook case of transfer learning paying off because of a structural fact about the domain, not because the model was large. The pathway fingerprints are consistent; that consistency is the shared representation the transfer exploits. The lesson generalises: transfer learning works when there is an invariant underneath the surface variation, and the research contribution was finding the invariant, not training the network.
Anyone who has fine-tuned a model on a small custom dataset has relied on the same bet — that the base model learned something general enough to carry. This is what it looks like when someone identifies and verifies what that general thing is.
The second is what it is for. Li points to the applied end: once you know the pattern of signals that drives a stem cell towards a specific type, you can recreate those signals to direct cell fate in a lab. The paper’s stated applications are stem cell engineering for regenerative medicine and better organoids — miniature 3D models that mimic real organs, used for studying disease and testing drugs.
Directed self-assembly of living structure from a known recipe of signals is, among other things, a fabrication technique. Artists and designers working with living materials — bio-art, mycelium, grown structures — have been constrained by exactly the gap this closes: you can cultivate a biological material, and steering what it becomes is mostly empirical. A model that reads and predicts the instruction set is the beginning of a design tool, several steps away and pointed in that direction.
The honest scope
This is a paper about reading signalling histories in embryonic development, validated on mouse gastrulation. It is not a design tool, and the organoid and regenerative-medicine applications are prospects the authors name rather than results they report.
What is established is narrower and genuinely surprising: a thing the field assumed was cell-type-specific is not, and a network trained on one species’ stem cells can identify the signals in another’s embryo. The assumption was the obstacle, and removing it is the contribution.
Sources
- AI model decodes the language cells use to communicate — MIT News
- The paper in Nature Methods (open access)
- Pulin Li — Whitehead Institute
- MIT Department of Biology