Every cell in an embryo becomes what it is partly because of the chemical signals it received along the way, but those signals leave no obvious record. Working out which ones acted, and when, normally means slow experiments in animals, one pathway and one cell type at a time, and it is close to impossible in human tissue. A team at MIT's Whitehead Institute has built a neural network, IRIS, that reads that history from a cell's gene activity instead. The work was published in Nature Methods on 8 September 2026.
The key finding is that each signalling pathway leaves a recognisable pattern in gene activity that looks much the same across very different cell types. That runs against a widely held belief that responses are specific to each cell type, and it is what makes transfer possible. The team, led by Pulin Li with graduate student Nicholas Hutchins, exposed thousands of human embryonic stem cells to dozens of combinations of six major developmental pathways, including WNT, BMP, FGF, retinoic acid and Hedgehog, at several stages, and trained IRIS on the resulting atlas. Li compares it with a voice assistant trained mostly in English that can still pick up other languages.
The tests are what make the paper stand out. Applied to single cells from gastrulating mouse embryos, IRIS correctly placed when and where pathways switched on in cells heading towards heart, gut, muscle and spinal cord. It also predicted that WNT signalling helps steer early mesenchymal cells towards a lung-supporting identity, and the team confirmed this in cultured mouse foregut tissue. They then used that prediction to adjust a published stem cell recipe for respiratory mesenchyme, a step towards lung organoids with branching structure.
For readers who follow AI in science, this is a useful counterweight to the large general models we usually cover. IRIS is a small, purpose-built network whose value comes from a carefully designed experiment that generates its training data, and it was checked against real embryos rather than benchmarks. It sits in the same broad trend as the agent-driven biology work from Google DeepMind and OpenAI, but its claims are narrower and better tested.
There are limits. The predictions cover six pathways, and the authors themselves found older gene-marker methods worked poorly for some of them, a sign that some pathways are harder to read than others. Changing WNT alone, they note, is unlikely to produce lung cells fully equivalent to those in the body. The training data are public, but neither the MIT article nor the paper text we read points to a code release, which will decide how quickly other labs can use it. Mapping whole human embryos, the team's stated goal, is still ahead.