VarnaOps

Reference atlas report: Human skin v1

Atlas

Atlas	Value
Species	Human (homo_sapiens)
Cells	475,731
Cell types	26
Tissues	20
Source studies	8 (see Source studies)
Source datasets	8
Donors	157
Genes	15,294
Condition	Adult, healthy donors (donor_status = normal)
Build	e9d75017acc8-475731
Cell-type labels	Cell Ontology (CELLxGENE Census), rolled up to subclass level
Data source	CZI CELLxGENE Census, version 2025-11-08
License	CC BY 4.0

Sequencing technology

Technology	Category	Cells	Share
Whole cell (scRNA-seq)	Suspension	470,444	98.9%
Single nucleus (snRNA-seq)	Suspension	5,287	1.1%
10x 3' v2	Assay	328,554	69.1%
10x 3' v3	Assay	125,045	26.3%
10x 3' v1	Assay	9,222	1.9%
MARS-seq	Assay	7,704	1.6%
10x 5' v1	Assay	2,885	0.6%
Smart-seq2	Assay	2,029	0.4%
Seq-Well S3	Assay	292	<0.1%

Measured from the built reference (obs suspension_type and assay, CZ CELLxGENE Census schema), not declared. Nuclei carry less cytoplasmic RNA and less ambient contamination than whole cells, so the suspension mix affects how expression should be read.

^* Unknown — cells whose Cell Ontology label is coarser than our subclass anchors (e.g. a coarse lineage label) or unannotated in the source study. Not a deployed cell type and excluded from the count. They still shape the latent space during semi-supervised training, but they are never assigned a type that is stored or served, and they are excluded from every metric here.

Reference UMAP

Cell-type similarity — dendrogram

Hierarchical clustering of per-cell-type centroids in the deployed scANVI latent (Euclidean distance, average linkage, optimal leaf ordering). Transcriptionally similar types sit adjacent; Unknown excluded. Same colors as the UMAP.

Source studies

This reference is built from 8 studies across 8 CZ CELLxGENE Census datasets (platform version 2025-11-08), plus the Census platform itself. The data is CC BY 4.0; that is CZI's platform-wide submission condition, not 8 independent per-dataset determinations. Each study, with its DOI and licence, is listed at varnaops.com/attribution.html#datasets. Cite the studies, not this report, when using the reference.

Marker validation

Each cell type's curated panel is scored on its own cells and ranked against all the others: top-1 counts the types whose own panel ranks first, top-3 those landing in the first three, over the 26 types with enough cells to score. Your run's report shows the same pair from the same code. These labels are the atlas's own curated types rather than predictions, so this is the ceiling — a query scoring near it agrees as closely as the markers allow.

Left: pick a type to colour the atlas UMAP by its score_genes panel score; right: cell-type × panel heatmap (z-scored per row, so the diagonal should be reddest), on a balanced ~3k-cells/type subsample. Panels come from CZ CELLxGENE CellGuide, resolved per cell type by Cell Ontology term and never hand-picked. Gene membership is canonical (literature / HuBMAP ASCT+B, CC BY 4.0) for 97 of 182 panels, ordered by how strongly each gene separates that cell type from the others in skin of body, skin; the remaining 85 come from that contrast ranking alone, used where canonical markers were too few. Closely related sister lineages share curated panels and score against each other, so an off-diagonal neighbour reflects the limits of literature markers, not a mislabelled population.

Model

Model configuration	Value
Architecture	scVI → scANVI
Likelihood	negative binomial · per-gene dispersion
Latent dimensions	30
Hidden dimensions	128
Hidden layers	2
Optimizer	Adam
Learning rate	0.001
KL weight	0.1 constant · 89-epoch warmup on scVI
Classifier loss	class-balanced cross-entropy (effective-number, β = 0.999999)
scVI deploy epoch	399
scANVI deploy epoch	19
scVI selection	minimum validation ELBO
scANVI selection	minimum validation balanced cross-entropy

Training & validation losses

Evaluation

Integration quality — scib-metrics

Classification performance

Split	Cells	Macro-F1	Accuracy
Training set	406,198	0.951	97.4%
Validation set (held-out)	45,097	0.936	96.7%

Attribution

Modifications. Relative to the source data we applied: healthy donors only; adult stages only; the tissues this atlas covers; primary tissue only, organoids and cell culture excluded; cells duplicated between an integrated atlas and the study it re-publishes removed; genes restricted to those shared across contributing datasets, then a highly-variable subset for training; study labels remapped through Cell Ontology to a coarser vocabulary, so the labels here are ours; cells concatenated into one atlas with derived representations (latent embedding, UMAP) not present in the source. Cite the source studies, not this report — they are listed under Source studies above, and full detail for every resource is at varnaops.com/attribution.html.