Build Single-cell Reference Data (Optional)

Purpose

Use scimilarityMakeRef to construct the SCimilarity single-cell reference database. This is only applicable to human samples.

Usage

Model download: https://zenodo.org/records/10685499

After downloading, extract the model folder and specify the --model_dir parameter.

SDAS cellAnnotation scimilarityMakeRef -o ./scimilarity_ref --reference sc.h5ad --label_key annotation \
--model_dir ./model_v1.1 \
--remove_tmp

Input Parameter Description

Parameter	Required	Default	Description
-o / --output	Yes		Output folder
--reference	Yes		Single-cell ref h5ad, must contain the raw expression matrix
--label_key	Yes		Name of the column in single-cell ref h5ad.obs indicating cell type
--ref_layer	No		Layer in single-cell ref h5ad storing raw counts
--ref_gene_symbol_key	No	_index	Name of the column in single-cell ref h5ad.var indicating gene symbol (_index means using h5ad.var.index)
--filter_rare_cell	No	100	The minimum cell count for a cell type to be included
--seed	No	42	Random seed
--model_dir	No	./model_v1.1	Path to the SCimilarity model folder
--ef_construction	No	1000	ef_construction for HNSW KNN algorithm; higher values make the database more accurate but more time-consuming
--M	No	80	M for HNSW KNN algorithm, controls the number of connections in the nearest neighbor graph
--remove_tmp	No		Whether to delete the CellArrDataset format data and model-generated embeddings in the output folder. Note: the original assays, cell_metadata, gene_annotation, sample_metadata, and cellsearch folders in the output will all be deleted
--gpu_id	No	-1	ID of the GPU to use. If -1, use CPU
--n_threads	No		Number of threads to use in CPU mode, defaults to all CPUs

Output Results

The constructed single-cell reference data will be stored in the scimilarity_ref folder. The directory structure and key files are as follows:

./scimilarity_ref
├── annotation
│   ├── labelled_kNN.bin
│   └── reference_labels.tsv
├── assays
├── cell_metadata
├── cellsearch
│   └── cell_embedding
├── gene_annotation
└── sample_metadata

Result File	Description
annotation/labelled_kNN.bin	KNN data calculated from single-cell ref embeddings, used for cell type search
annotation/reference_labels.tsv	Cell type label for each cell in the single-cell ref
assays, cell_metadata, gene_annotation, sample_metadata	CellArrDataset format data converted from single-cell ref h5ad, can be deleted. Use --remove_tmp to delete automatically after the program finishes
cellsearch/cell_embedding	Embeddings of the single-cell ref calculated by the model in tileDB format, can be deleted. Use --remove_tmp to delete automatically after the program finishes