LangCell

Use This Skill When

Use this skill for the local LangCell project at /DATA/disk0/zhaosy/home/LangCell. It is the right choice when the task involves:

• zero-shot cell identity or cell type annotation from tokenized single-cell data
• few-shot cell type annotation with very limited labels
• finetuning only the LangCell cell encoder (LangCell-CE)
• preprocessing AnnData into the tokenized format expected by LangCell
• preparing text descriptions for candidate cell identities
• understanding how LangCell combines cell embeddings and text embeddings

Do not use this skill for ordinary Scanpy analysis that does not depend on LangCell.

Start Here

1. Confirm whether the user wants zero-shot annotation, few-shot annotation, or cell-encoder-only finetuning.
2. Check that the input is already tokenized, or route through preprocessing first.
3. Check whether the required external assets exist: checkpoints, tokenized dataset, ontology / text-description JSON.
4. Prefer the zero-shot path first if the user is exploring LangCell rather than benchmarking a supervised baseline.

Choose A Path

Zero-shot annotation

Start here for most LangCell usage. The defining behavior is:

• encode cells with cell_bert + cell_proj
• encode candidate texts with text_bert + text_proj
• score cell-text matches with ctm_head
• combine similarity and matching scores for final predictions

Use LangCell-annotation-zeroshot/zero-shot.ipynb as the primary reference path.

Few-shot annotation

Use LangCell-annotation-fewshot/fewshot.py when only a tiny labeled support set is available and the user still wants the multimodal LangCell path.

LangCell-CE finetuning

Use LangCell-CE-annotation/finetune.py when the user wants a standard supervised classifier on top of the pretrained cell encoder.

Preprocessing

LangCell does not take raw .h5ad directly for these downstream scripts. First:

• read AnnData with scanpy
• add obs["n_counts"]
• ensure var["ensembl_id"] exists
• tokenize with LangCellTranscriptomeTokenizer
• save with save_to_disk(...)

Guardrails

• Do not claim raw .h5ad can be passed directly into LangCell inference; tokenization is required first.
• Do not assume checkpoints live in the repo. The official repo expects external downloads.
• Do not treat zero-shot prediction as plain nearest-neighbor on cell embeddings; the project combines similarity and cell-text matching.
• Do not assume label columns are named consistently. The repo checks several alternatives such as celltype, cell_type, str_labels, and labels.
• If new cell types are needed, prepare textual descriptions carefully instead of inventing bare labels with no ontology context.

Read More Only If Needed

• Read references/local-usage.md for entry points, asset requirements, and practical checks.
• Read references/model-notes.md when the task depends on LangCell's multimodal design or paper-level motivation.