Pan-human Azimuth

Single-cell atlases now span most human tissues, but each study labels its cells with its own vocabulary at its own resolution, so annotations rarely line up across datasets. Pan-human Azimuth targets that harmonization problem: a supervised neural network that places every cell in a human scRNA-seq or snRNA-seq experiment into one hierarchical, organism-scale cell typology.

It was developed by the Satija Lab at the New York Genome Center through the NIH Human BioMolecular Atlas Program (HuBMAP), with collaborators contributing cell ontology curation and kidney pathology expertise. It is the organism-scale successor to the original per-tissue Azimuth reference-mapping application (Hao et al., Cell 2021), which required choosing a tissue-specific reference before mapping — a constraint that made cross-tissue comparison awkward and left cells from unrepresented tissues unmappable. Pan-human Azimuth replaces that with one classifier spanning 23 human tissues and 381 cell types at its finest resolution, organized on an adapted version of the DISCO reference typology.

The design point differs from self-supervised single-cell foundation models such as scGPT and Geneformer, which learn general-purpose representations and are then fine-tuned for annotation. Pan-human Azimuth is trained end-to-end as a hierarchical classifier and applied by direct inference — closer to atlas-scale annotation with SCimilarity, except that labels come from a calibrated classifier rather than nearest-neighbor retrieval.

#Key Features

• One organism-wide reference: A single checkpoint annotates any covered tissue, so cross-tissue and cross-study comparisons share identical label definitions instead of reconciling tissue-specific references after the fact.
• Hierarchical, calibrated output: Labels are emitted at broad, medium, and fine granularity alongside a finest-resolution call, each carrying a confidence score from trained calibrators so that scores track empirical accuracy.
• Cell Ontology interoperability: A built-in crosswalk maps predicted labels to standard Cell Ontology terms and identifiers, letting annotations feed ontology-aware atlases and HuBMAP infrastructure without manual translation.
• Spatial transcriptomics annotation: The same classifier annotates spatial data, recovering canonical kidney cortical structures and distinguishing glomerular states consistent with expert pathology review.
• Commodity-hardware throughput: Inference runs in minibatches to bound peak memory and processes roughly 50,000–60,000 cells per minute on a standard laptop, with optional GPU acceleration and splittable datasets for parallel runs.

#Technical Details

Pan-human Azimuth is a supervised neural network classifier trained on a curated corpus of scRNA-seq and snRNA-seq data from 23 human tissues across multiple sequencing technologies, with cancer datasets deliberately excluded and uniform annotations and stringent quality control enforced during curation. Query data is normalized and aligned to the model's feature panel; inference returns per-level calibrated probabilities, a 128-dimensional cell embedding for visualization and downstream analysis, and optional Cell Ontology identifiers. Two versioned checkpoints exist, v0 and the current default v1, which adds a pulmonary ionocyte label, an "Unassigned" class that absorbs low-UMI cells and ambient RNA profiles, and recalibrated confidence scores; the documentation recommends discarding predictions below a confidence of 0.5.

Weights ship as Keras model files of roughly 84 MB each, archived on Zenodo under CC BY 4.0 and downloaded automatically into a local cache on first use rather than fetched by hand, with either version selectable through a model_version argument. Two MIT-licensed interfaces are released: the Python package panhumanpy, exposing the AzimuthNN classes and an ANNotate workflow over AnnData objects, and the R package AzimuthAPI, offering CloudAzimuth hosted inference with no local setup or local execution through reticulate. The describing preprint is on bioRxiv under CC BY-NC and has not yet been peer reviewed.

#Applications

The immediate use is standardized annotation of new single-cell experiments without choosing a tissue-specific reference or hand-curating marker lists, which matters most for multi-tissue studies and consortium data that must interoperate. Because one label space covers the organism, the model also supports meta-analysis: the authors mapped tens of millions of cells in the Tabula Sapiens and scBaseCamp repositories, compared composition across thousands of samples, and identified pronounced tissue specialization among fibroblast states. The spatial extension applies the same typology to tissue sections, and the ontology crosswalk makes outputs usable by atlas builders. Non-cancer disease samples are annotated by mapping cells to their most similar healthy types, though diseased states may map imperfectly.

#Impact

Pan-human Azimuth reframes cell type annotation as one organism-scale classification problem rather than a per-tissue reference-mapping exercise, and its calibrated confidences and explicit "Unassigned" class make the output auditable in a way that raw label transfer usually is not. Delivered through HuBMAP with Python, R, and hosted-cloud interfaces and ontology-mapped labels, it reads as infrastructure for consortium-scale atlas curation as much as a tool for single experiments. The caveats are real: the reference covers 23 tissues rather than the whole body, cancer is excluded by construction, disease annotation quality depends on similarity to healthy counterparts, and the primary description remains a preprint under non-commercial terms even though both packages are permissively licensed.