Logistic regression, but for naming your cells

What it does CellTypist takes a raw count matrix from single-cell RNA-seq and assigns a cell type to each cell. It does this with logistic regression classifiers trained by stochastic gradient descent, using either built-in models (currently focused on immune sub-populations) or models you train yourself. Results come back as predicted labels plus decision and probability matrices you can export or fold into an AnnData object.

The interesting bit The models are tiny — about 1 MB each — so grabbing all of them is cheap, and they’re pickled classifiers rather than heavyweight deep learning. There’s also a prob match mode that handles the awkward cases: cells that match nothing (novel types get flagged as Unassigned) or match several types at once (multi-label output).

Key highlights

• Pre-trained models with an immune focus, downloadable and updatable from within the tool
• Two prediction modes: strict best match or threshold-based prob match for ambiguous cells
• Accepts plain count tables (.csv, .tsv, .mtx, etc.) — no AnnData required to start
• Outputs integrate with the AnnData ecosystem, including confidence scores and probability matrices
• Colab tutorials cover classification, multi-label, and cross-dataset label transfer

Caveats

• Requires raw count matrices (reads or UMIs); the README is explicit that non-expressed genes should be included, so pre-processed data won’t fly
• Built-in models lean heavily toward immune cell types — other tissues mean training your own

Verdict Worth a look if you do scRNA-seq analysis and want a fast, transparent alternative to eyeballing clusters. If you’re outside single-cell biology, there’s nothing here for you.