Researchers at MIT and the Whitehead Institute say their machine-learning model can estimate which chemical signals developing cells encountered from their gene activity. The team trained IRIS on human stem cells and applied it to mouse-embryo data. To test one prediction, they cultured mouse-embryo foregut tissue outside the embryo: activating the WNT pathway expanded two markers associated with lung-related mesenchyme.
Cells coordinate development through chemical messages from neighboring cells. These messages trigger signaling pathways, step-by-step routes that change which genes are active. Researchers usually have to test pathways separately in different cell types to learn how each one responds.
IRIS uses stem-cell training to estimate signals in mouse embryos
The MIT-led team trained IRIS on measurements from human embryonic stem cells exposed to combinations of six major signaling pathways. The model learned patterns of gene activity linked to each pathway. The Nature Methods paper reports that these response patterns could transfer across cell types, allowing IRIS to estimate pathway activity in cell types outside its training examples.
The researchers then applied IRIS to single-cell data from mouse embryos during gastrulation, an early stage when cells rapidly take on different developmental roles. MIT says the model recovered signaling histories in well-studied lineages and predicted activity in less-studied cell types. This test connects human-cell training data to mouse-embryo analysis; it does not show that the same signals have been mapped in human embryos.
One lung-development prediction was tested in embryos
IRIS predicted that WNT signaling helps specify respiratory mesenchyme, cells involved in lung development. To test that prediction, researchers cultured foregut tissue dissected from E9.0 mouse embryos for 24 hours with either a WNT activator or inhibitor. They then used a fluorescence-based method to measure two markers, Tbx4 and Foxf1. WNT activation expanded both markers near the putative lung region, while inhibition eliminated Tbx4 expression. This supports a role for WNT in this developmental process; the experiment measured marker expression in cultured tissue, not lung function or a treatment outcome.
The Nature Methods paper says the team then used earlier WNT activation to improve a human stem-cell protocol for producing respiratory mesenchyme. The researchers say this could help organoid research, but it has not been tested as a treatment or in patients.
MIT published its account on Oct. 15. The peer-reviewed Nature Methods paper appeared on Sept. 8, 2026.
Sources
- MIT News: AI model decodes the language cells use to communicate (Oct. 15, 2026)
- Nature Methods: “Reconstructing signaling histories of single cells via perturbation screens and transfer learning” (Sept. 8, 2026)
- Image: Karl Marquez / UCSanDiegoStemCellProgram, “Microscopic image of stem cells, Hues 9 stained with DAPI (blue),” Wikimedia Commons, CC BY-SA 3.0. The unchanged contextual image does not show IRIS data or the reported experiments.
.jpg){kind=link}