A signalling pathway can leave a recognisable pattern of gene activity across different cell types. Researchers at MIT and the Whitehead Institute used that shared signature to train IRIS, a model that infers which chemical instructions developing cells have received.[1]
The work, led by Pulin Li and Nicholas Hutchins, was published in Nature Methods on 8 September and described by MIT on 15 October. Its central result is transfer: patterns learned from controlled stem-cell experiments helped interpret cells in a developing mouse embryo, including cell types absent from training.[1]
A chemical message does not come with a readable label
Cells communicate through signalling pathways such as WNT, BMP, fibroblast growth factor and other molecular systems. When a pathway is activated, it can change which genes are transcribed. Measuring the genes currently active in a cell gives researchers a snapshot of those downstream consequences, but not an unambiguous record of every signal that produced them. Multiple pathways can overlap and cells can be at different developmental stages.[1]
Individual marker genes can be difficult to interpret because responses vary and pathways overlap. IRIS instead uses a cell’s broader gene-activity pattern, drawing on experimentally established examples of how signalling changes that pattern.[1]
From controlled treatment to a biological prediction
| Stage | Evidence | What it contributes |
|---|---|---|
| Stem-cell screen | Known signalling conditions and measured gene activity | Labelled examples for model training |
| Mouse embryo analysis | Inferred histories across more than 40 cell types | Tests whether pathway signatures transfer |
| Lung-development follow-up | Embryonic experiments supported a predicted signal | Checks a biological implication of the model |
The training atlas measured thousands of human embryonic stem cells exposed to dozens of combinations of six major signalling pathways at different developmental stages. Knowing which treatments were applied gave the model a reference for connecting gene activity to signalling conditions.[1]
The surprising property is transferability
IRIS connects patterns of gene expression with pathway activity, then applies those patterns to other developmental contexts. A transferable signature reduces the need to perform every possible signalling experiment separately in every cell type.[1]
In developing mouse embryos, the model reconstructed signalling activity across more than 40 defined cell types, including cells following divergent routes towards tissues such as heart, gut, muscle and spinal cord. It also revealed variation in signalling conditions within a nominal cell type. That additional detail is a reminder that development is not a series of uniform steps: cells can share a label while having slightly different experiences or futures.[1]
The team used the model to identify a signalling environment predicted to favour a cell identity important in lung development. Experiments in mouse embryos supported that prediction. This links a computational reconstruction to a biological change researchers could test directly.[1]
A more focused route to laboratory tissues
Scientists developing organoids and other laboratory tissues need reliable ways to direct stem cells towards a desired identity. IRIS could help select signalling conditions for targeted experiments, reducing the number of combinations that must be explored. Its demonstrated evidence is developmental and experimental; applications to patient treatment remain prospective.[1]
The model’s value comes from connecting three kinds of evidence: controlled signalling perturbations, gene activity in developing cells and an experimentally tested prediction. That sequence gives researchers a way to investigate the biological mechanism behind a computational pattern.[1]
Sources & further reading
- Reconstructing signaling histories of single cells via perturbation screens and transfer learning Nature Methods · Accessed 24 Oct 2026
- AI model decodes the language cells use to communicate MIT News and Whitehead Institute · Accessed 25 Oct 2026