One coordinate system for cells across species.

The human problem

Cell data come from different tissues, experiments, and species, making biological comparison difficult.

The prior constraint

Many analysis systems are trained for one dataset or species and require extensive harmonization before comparison.

AI’s actual role

The model learned a common numerical representation of cells from large single-cell datasets.

The documented result

The Nature paper reports a corpus of 36 million cells spanning more than 1,000 cell types and eight species, with benchmarked downstream analyses detailed in the paper.

Why it may matter

A shared representation could help researchers compare biological systems and transfer hypotheses, while experiments remain the test of what those similarities mean.

Limitations

An embedding compresses data and can hide important biological variation.

Training-data coverage and annotation quality shape what transfers.

Cross-species similarity is not proof of identical function.

Unresolved questions

Which rare cell types are poorly represented?

How stable are results across sequencing technologies and laboratories?

Which proposed relationships survive experimental validation?