Pan-human Azimuth combined data from 23 tissues into a single cell type hierarchy

Pan-human Azimuth combined data from 23 tissues into a single cell type hierarchy

On 21 July, the HuBMAP team released the Pan-human Azimuth preprint. The neural network annotates individual human cell profiles using a shared hierarchy. The authors trained it on 27,04 million profiles and made both a cloud service and tools for R and Python publicly available.

Single-cell atlases show which genes are active in each cell. Biologists use this pattern to determine the cell’s type and state. However, one atlas may assign one name to a cell, while another uses a different name. When organs are compared, differences in terminology can easily be mistaken for biological differences.

In the preprint by Sourav Sarkar and colleagues, the original labels from different datasets were manually mapped onto a single cell type tree. The authors reviewed uncertain assignments using gene activity and trained a classifier on the resulting annotations. It reads an RNA activity table, assigns each cell to one of eight levels in the hierarchy, and reports its confidence in the result. Empty droplets and ambient RNA were assigned a separate class. As a result, similar cells from different organs are now named according to the same rules.

The authors evaluated the classifier on 1,1 million profiles from Tabula Sapiens v2. The data came from 24 donors and covered 28 tissues. The model had not encountered about 600 thousand profiles from nine of the donors during training. Using the same method, it annotated 85,9 million cells from scBaseCamp. In spatial kidney sections, the classifier reproduced the pattern of the renal cortex and distinguished healthy from sclerotic glomeruli in agreement with pathologists’ annotations.

A shared hierarchy makes it possible to query a specific cell state across several tissues and individuals at once. For example, researchers can test whether the same fibroblast, immune cell, or epithelial state recurs in different organs. An atlas of 7 million mouse cells has already revealed coordinated age-related changes across organs. Pan-human Azimuth gives human datasets a common language for conducting the same kind of analysis.

The project documentation describes this method as mapping a new dataset onto an assembled reference. Pan-human Azimuth turns inconsistent cell naming into a reproducible procedure. Once the cells have been annotated in this way, cross-tissue comparisons can be tested across tens of millions of profiles.