Whitehead’s IRIS Predicts a Lung-Development Signal in Mouse Embryos

Whitehead Institute researchers have published IRIS, a neural-network model trained on signaling experiments in human pluripotent stem cells and applied to single-cell data from mouse embryos. The team reported that IRIS predicted a signal that would encourage lung-specific development, then confirmed that prediction in mouse-embryo experiments.

Learning signals from gene activity

Cells respond to combinations of chemical signals as they develop. But measuring the response to every pathway across every cell type is not practical. The Nature Methods study instead asked whether pathways leave response patterns that can be recognized across different kinds of cells.

IRIS was trained on a signaling-perturbation atlas: thousands of human stem cells exposed to combinations of six major developmental pathways at multiple stages. It uses a cell’s broad gene-activity profile to infer which pathways were active. The study found that diverse cell types share conserved response signatures, giving each pathway a transferable fingerprint rather than requiring a separate map for each cell type.

A developmental map meets an experiment

Applied to mouse-embryo single-cell atlases, IRIS mapped signaling states in more than 40 defined cell types and reconstructed signaling histories along developmental lineages. The researchers also reported predictions of pathway activity in cells headed toward heart, gut, muscle and spinal-cord tissue.

The more pointed test involved organ-specific mesenchyme, connective tissue involved in organ development. IRIS predicted that activating a particular pathway would favor lung-specific development. Mouse-embryo experiments confirmed the prediction, according to the researchers.

A way to focus the search

That validation does not turn IRIS into an automatic recipe for directing cell fate. Its demonstrated role is to infer likely signaling histories and reduce the combinations researchers must test when building human stem-cell differentiation protocols. The authors say that could support stem-cell engineering, organoid development, disease modeling and drug testing.