Organism-scale annotation with Pan-human Azimuth

Area of Science:

• Single-cell genomics
• Computational biology
• Human tissue atlases

Background:

• Existing single-cell atlases lack consistent annotations across studies, hindering their use as a unified reference.
• This inconsistency limits the ability to perform large-scale, cross-tissue analyses and comparisons.

Purpose of the Study:

• To develop a standardized, organism-wide cell typology for human tissues.
• To create a supervised neural network model for mapping diverse single-cell data onto this unified typology.
• To enable robust cross-dataset and cross-tissue single-cell analyses.

Main Methods:

• Development of Pan-human Azimuth, a supervised neural network.
• Training the model on a curated corpus maximizing tissue and technology diversity with uniform annotations.
• Application of the model to millions of cells from repositories like Tabula Sapiens and scBaseCamp.
• Extension of the model to annotate spatial transcriptomic data.

Main Results:

• Successful mapping of tens of millions of human cells onto a single, hierarchical, organism-scale typology.
• Identification of significant tissue specialization in fibroblast states.
• Accurate annotation of spatial transcriptomic data, including kidney structures and glomerular states.
• Demonstration of cross-tissue comparisons across thousands of samples.

Conclusions:

• Pan-human Azimuth provides a unified reference for human single-cell data, overcoming annotation inconsistencies.
• The tool facilitates large-scale single-cell analysis, cross-tissue comparisons, and discovery of cell specialization.
• Pan-human Azimuth is a valuable resource for standardizing single-cell and spatial transcriptomic analyses, with accessible interfaces.