Appearance
CellTypist Local Annotation
CellTypist helps annotate single-cell RNA-seq datasets by comparing cells to trained reference models. In Liatir, it is used for local .h5ad / AnnData workflows.
What it does
CellTypist takes an AnnData file and predicts likely cell labels. It can also apply majority voting to smooth labels across nearby cells when that option is enabled.
When to use it
Use this model when you have a single-cell dataset and want a first-pass cell type annotation. It is useful for exploration, quality checks, and preparing a dataset for deeper single-cell analysis.
Inputs in Liatir
- AnnData
.h5adfile. - CellTypist model name, such as
Immune_All_Low.pkl. - Optional majority voting.
CellTypist expects normalized single-cell expression data. If the matrix is raw counts, the run may fail or produce misleading labels.
Outputs
Liatir can produce:
- predicted labels;
- confidence-like scores where available;
- CSV/JSON summaries;
- output provenance with model, parameters, input file, and runtime metadata.
How to interpret results
Read the label distribution first. If the top label covers nearly all cells, check whether that matches the biology of the dataset.
CellTypist labels are reference-based suggestions. A wrong tissue, species, assay, or preprocessing method can produce confident but misleading labels.
Hardware and installation
CellTypist runs on CPU and does not require a GPU. Memory use depends mostly on the size of the AnnData matrix.
Liatir installs CellTypist in an isolated managed Python runtime.
Limits and cautions
Cell annotation depends on the reference model. A label can be wrong if the dataset, species, tissue, assay, or preprocessing does not match the reference well. Treat the result as an annotation aid, not as a final biological claim.