A model that reads the genome one cell at a time

Arc Institute, with Michael Greenberg's lab at Harvard Medical School, posted a model this week called Bolero.

Give it a stretch of DNA and a representation of which cell you mean, and it predicts how active the genes are and how open the DNA is in that specific cell type. It can even estimate what a sequence would do in a cell state nobody has profiled yet.

It shipped with its own training data: an atlas of 10.8 million cells, spanning human brain and body, across development and aging, and across species.

This is a preprint, so it hasn't been peer reviewed. The behavioral experiments could only run in model organisms, so the reads on human brain tissue lean on cross-species transfer. The authors are careful about that.

For years the field has been building cell atlases. The Human Cell Atlas, single-cell sequencing, map after map of which cells exist and what they're doing. Building the atlas was the triumph. Reading it, asking what a given piece of DNA actually does in this cell and not that one, stayed slow and manual.

A model that reads regulation per cell state is the tool that turns those atlases from a catalog into something you can question. The maps were step one. This is what they were for.

Read the primary source →