IRIS traces developmental signals from single-cell gene activity

Following the effects of a signal

As tissues develop, neighboring cells exchange chemical messages that influence what each cell becomes. A signaling pathway is the chain of reactions that carries such a message into changes in gene activity. The researchers found that a pathway's response pattern can remain recognizable across different cell types, allowing IRIS to identify signals in unfamiliar cells.

The training material came from controlled experiments on human pluripotent stem cells, which can develop into many cell types. Researchers varied combinations and sequences of signals and recorded how the cells responded. The public experimental record describes using this atlas to train a neural network, a system that learns patterns from data, before applying it to gene-activity measurements from individual mouse-embryo cells. This also exposed differences in signaling states within the same cell type.

A prediction tested in developing lung tissue

The study, published in Nature Methods on September 8, reconstructed signaling patterns along heart, gut, muscle and neural-development lineages in mouse embryos. Its experimental figures also describe testing the WNT signaling pathway in cultured mouse foregut tissue. Researchers then adjusted a human stem-cell protocol to start WNT stimulation earlier and maintain it longer. This improved the production of respiratory mesenchyme, the supporting tissue involved in lung development, as assessed using characteristic gene markers. These are developmental and laboratory results; they do not establish a clinical diagnostic or treatment.

A research workflow others can inspect

The authors' public code accepts single-cell RNA-sequencing data, which measures RNA molecules to estimate gene activity. It expects the measurements and their annotations in an AnnData object, a data container used in single-cell analysis, including cell-type, experimental-batch and known pathway-state labels. Researchers can load pretrained models or train their own, then visualize inferred states. The team recommends GPU access for the machine-learning stages and supplies example notebooks and figure-reproduction scripts, making the computational part available for other laboratories to examine.