After running spaceranger count or spaceranger annotate, cell type annotations can be found in the segmented_outs/cell_types directory. For example:

Copy
segmented_outs/
├── cell_types
│   ├── 10x_Cloud
│   │   ├── cell_annotation_differential_expression.csv
│   │   ├── cell_annotation_results.json.gz
│   │   └── cell_types.csv
│   └── Azimuth
│       ├── cell_annotation_differential_expression.csv
│       └── cell_types.csv

The contents of these files are detailed below.

File Name: cell_annotation_sample_cloupe.cloupe

Description: A new .cloupe file is generated, which includes coarse cell types in the "Custom Groups" section. By default, this group is labeled "Cell Types," but the name can be customized during the annotation analysis setup.

File Name: cell_annotation_results.json.gz

Description: 10x Cloud models only. This file is a compressed JSON containing a list of dictionaries. Each element in the list represents the annotation results from a single cell, derived from the cell annotation model.

For each cell, the corresponding dictionary includes the top 500 matches obtained using an approximate-Nearest Neighbor (ANN) lookup. These matches are summarized for the total number of occurrences for a given cell type. While more cells supporting a particular annotation can increase your confidence in the annotation, occasionally the most common nearest-neighbor cell type can have a low number of supporting cells because the nearest-neighbors are split amongst several highly similar cell types (e.g., 'Cd16-Negative, Cd56-Bright Natural Killer Cell, Human' and 'Cd16-Negative, Cd56-Dim Natural Killer Cell'). The dataset_id corresponds to the Chan Zuckerberg CELL by GENE (CZ CELLxGENE) study from which the annotation was derived. To view this study, insert the id into this URL: https://cellxgene.cziscience.com/e/{dataset_id}.cxg/.

An example output is shown below:

Copy
{
        "barcode": "AAACCAAAGAATGCAA-1",
        "matches": [
            {
                "cell_count_in_model": 32,
                "cell_type": "monocyte",
                "dataset_ids_with_counts": [
                    {
                        "count_per_dataset": 30,
                        "dataset_id": "87ce26ed-e5d1-44b4-81cc-cc5b709a169f"
                    },
                    {
                        "count_per_dataset": 2,
                        "dataset_id": "b0e547f0-462b-4f81-b31b-5b0a5d96f537"
                    }
                ]
            },

File Name: cell_types.csv

Description: This file contains the cell type annotation for each cell and can be used to import the fine-scale cell type annotations directly into Loupe Browser.

The Pan-Human Azimuth model CSV file contains these columns:

• barcode: the segmented cell or bin being annotated.
• broad_cell_type: the high-level annotation of the cell type. For cells with low UMI counts (< 100), this field is set as Low UMI Barcode for filtering out annotations with low UMI support.
• coarse_cell_type: the mid-level annotation of the cell type. Those coarse cell types are the display nodes we manually curated. For cells with low UMI counts (< 100), this field is set as Low UMI Barcode for filtering out annotations with low UMI support.
• fine_cell_type: the original annotation derived from the model based on the most common cell type amongst the 500 nearest-neighbors. Note: This may be the same as coarse_cell_type if the original reference was only annotated to that level of detail.
• full_hierarchical_labels: a concatenation of the broad, course, and fine cell types, separated by pipes (|).
• final_level_softmax_prob: a probabilistic estimate of how correct the cell annotation is.
• coarse_cell_type_unfiltered: the high-level annotation of the cell type (e.g., T Cell, B Cell, Neutrophil, etc.). Those coarse cell types are the display nodes we manually curated. Not subject to filtering by UMI count.
• umi_count: the number of UMIs associated with the cell.

Here is an example:

Copy
barcode,broad_cell_type,coarse_cell_type,fine_cell_type,full_hierarchical_labels,final_level_softmax_prob,coarse_cell_type_unfiltered,umi_count
cellid_000000023-1,Epithelial cell,Epithelial cell of lung,Club cell,Epithelial cell|Epithelial cell of lung|Club cell,0.72484344,Epithelial cell of lung,1404
cellid_000000024-1,Epithelial cell,Epithelial cell of lung,Club cell,Epithelial cell|Epithelial cell of lung|Club cell,0.46240065,Epithelial cell of lung,1482
cellid_000000026-1,Epithelial cell,Epithelial cell of breast,Mammary luminal cell,Epithelial cell|Epithelial cell of breast|Mammary luminal cell|PIP mammary luminal cell,0.9022441,Epithelial cell of breast,1206

The 10x Cloud model CSV file contains these columns:

• barcode: the segmented cell or bin being annotated.
• coarse_cell_type: the mid-level annotation of the cell type. Those coarse cell types are the display nodes we manually curated. For cells with low UMI counts (< 100), this field is set as Low UMI Barcode for filtering out annotations with low UMI support.
• fine_cell_type: the original annotation derived from the model based on the most common cell type amongst the 500 nearest-neighbors. Note: This may be the same as coarse_cell_type if the original reference was only annotated to that level of detail.
• cell_count_in_model: the number of cells in the model that support the given fine_cell_type annotation, with a maximum of 500 cells.
• coarse_cell_type_unfiltered: the high-level annotation of the cell type (e.g., T Cell, B Cell, Neutrophil, etc.). Those coarse cell types are the display nodes we manually curated. Not subject to filtering by UMI count.
• umi_count: the number of UMIs associated with the cell.

Here is an example:

Copy
barcode,coarse_cell_type,fine_cell_type,cell_count_in_model,coarse_cell_type_unfiltered,umi_count
cellid_000000054-1,epithelial cell,luminal cell of prostate epithelium,500,epithelial cell,1094
cellid_000000055-1,epithelial cell,luminal cell of prostate epithelium,488,epithelial cell,788
cellid_000000057-1,epithelial cell,luminal cell of prostate epithelium,486,epithelial cell,1084

The number shown for cell_count_in_model reflects the level of support for a cell's annotation. The algorithm identifies the 500 most similar cells in the reference set using embeddings from both the query dataset and the reference database. The cell type assigned is the annotation that appears most frequently among these 500 nearest neighbors. For example, if 400 of the nearest neighbors are labeled as T cells and 100 as lymphocytes, the cell will be annotated as a T cell, and the cell_count_in_model will be 400. The maximum possible value for this metric is 500.

This number should be interpreted with caution as an indicator of confidence in the model’s assignment. A high fraction of nearest neighbors supporting a fine-level cell type can suggest greater confidence in the annotation. However, a low value does not necessarily indicate low confidence in the coarse cell type assignment. For example, a T cell coarse-level annotation might be supported by different T cell subtypes, each represented by relatively few cells, but with all 500 nearest neighbors still classified as T cells based on Cell Ontology terms. This nuance highlights two key points: (1) the cell_count_in_model should not be treated as a confidence metric, and (2) there is no threshold that can reliably serve as a confidence cutoff for this number.

File Name: cell_annotation_differential_expression.csv

Description: This file contains the results of a differential expression analysis conducted between coarse cell types. These differentially expressed genes can be used to check that the cell type contains the expected marker genes. The pipeline uses the same algorithm employed in Space Ranger and Loupe Browser to calculate fold changes and p-values, ensuring consistency within these platforms.