Omicverse Single Cell Annotation

OmicVerse Single-Cell Annotation

Goal

Annotate one clustered AnnData object with one OmicVerse backend at a time. Treat celltypist, gpt4celltype, and scsa as alternative branches that share the same input/output contract, not as sequential stages.

Quick Workflow

1. If the input is raw counts, hand it off to the OmicVerse preprocessing path first; this skill starts from annotation-ready AnnData with a cluster column such as leiden.
2. Create ov.single.Annotation(adata) once.
3. Choose exactly one backend:
   • celltypist when you already have a CellTypist model path or want to resolve one with query_reference(source="celltypist").
   • gpt4celltype when you want LLM-based cluster labels and can provide AGI_API_KEY.
   • scsa when you want marker-database scoring from clustered cells and have or can point to a local SCSA database.
4. Pass only the backend-specific kwargs for the selected branch.
5. Validate the expected adata.obs label column and any branch-specific side products before plotting or downstream analysis.

Interface Summary

• Annotation.annotate(method='celltypist', cluster_key='leiden', **kwargs) dispatches across celltypist, scsa, gpt4celltype, harmony, scVI, and scanorama.
• Annotation.query_reference(source='cellxgene', data_desc=None, llm_model='gpt-4o-mini', llm_api_key='sk*', llm_provider='openai', llm_base_url='https://api.openai.com/v1', llm_extra_params=None) selects reference resources; source branches are cellxgene and celltypist.
• Annotation.download_reference_pkl(reference_name, save_path, force_download=False) and Annotation.add_reference_pkl(reference) support the CellTypist branch.
• pySCSA(adata, foldchange=1.5, pvalue=0.05, output='temp/rna_anno.txt', model_path='', outfmt='txt', Gensymbol=True, species='Human', weight=100, tissue='All', target='cellmarker', celltype='normal', norefdb=False, cellrange=None, noprint=True, list_tissue=False, tissuename=None, speciename=None) is the SCSA constructor.
• pySCSA.cell_anno(clustertype='leiden', cluster='all', rank_rep=False) and pySCSA.cell_auto_anno(adata, clustertype='leiden', key='scsa_celltype') expose the SCSA annotation flow.
• gptcelltype(input, tissuename=None, speciename='human', provider='qwen', model='qwen-plus', topgenenumber=10, base_url=None) is the LLM marker-to-label helper; provider branches are openai, kimi, and qwen.

Read references/source-grounding.md before documenting more detailed parameter behavior.

Branch Selection

• Use celltypist when you want pretrained model annotation and already know the .pkl model, or when query_reference(source="celltypist") should choose one for you.
• Use gpt4celltype when you want prompt-driven annotation from cluster markers and can provide an API key.
• Use scsa when you want marker-database scoring from clustered cells and a local database path.
• Use query_reference(source="celltypist") only to choose a CellTypist model; do not treat it as a required stage.
• Treat harmony, scVI, and scanorama as source-present but out-of-scope extensions unless the user explicitly asks for transfer-learning annotation.

Input Contract

• Start from AnnData that already has normalized expression and a cluster column such as leiden.
• For marker-based branches, keep cluster_key aligned with the label column you want to map.
• For CellTypist, you need a valid model path or a successful download step.
• For gpt4celltype, set AGI_API_KEY before the call; if you only want the generated prompt, call gptcelltype(...) directly.
• For SCSA, prefer a local database path instead of notebook-local absolute paths.
• If you still need raw preprocessing, run the notebook's QC, preprocess, neighbors, leiden, and umap handoff first.

Minimal Execution Patterns

For CellTypist model selection and annotation:

import omicverse as ov

obj = ov.single.Annotation(adata)
res = obj.query_reference(source="celltypist", data_desc="human PBMC scRNA-seq")
model_name = res.iloc[0]["model"]
obj.download_reference_pkl(model_name, save_path="models/celltypist.pkl")
obj.add_reference_pkl("models/celltypist.pkl")
obj.annotate(method="celltypist")

For LLM-based annotation:

import os
import omicverse as ov

os.environ["AGI_API_KEY"] = "your-key"
obj = ov.single.Annotation(adata)
obj.annotate(
    method="gpt4celltype",
    cluster_key="leiden",
    tissuename="PBMC",
    speciename="human",
    model="gpt-5-mini",
    provider="openai",
    topgenenumber=5,
)

For prompt-only LLM smoke checks, call the helper directly:

import os
from omicverse.single._gptcelltype import gptcelltype

os.environ["AGI_API_KEY"] = ""
prompt = gptcelltype({"0": ["MS4A1", "CD79A"]}, tissuename="PBMC", speciename="human")

For SCSA annotation:

import omicverse as ov

obj = ov.single.Annotation(adata)
obj.add_reference_scsa_db("models/pySCSA.db")
obj.annotate(
    method="scsa",
    cluster_key="leiden",
    foldchange=1.5,
    pvalue=0.01,
    celltype="normal",
    target="cellmarker",
    tissue="All",
)

Validation

After celltypist, check:

• adata.obs["celltypist_prediction"]
• adata.obsm["celltypist_decision_matrix"]
• adata.obsm["celltypist_probability_matrix"]

After gpt4celltype, check:

• adata.obs["gpt4celltype_prediction"] when the backend returned a real mapping
• the direct gptcelltype(...) helper returns a prompt string when you run the prompt-only path

After scsa, check:

• adata.obs["scsa_prediction"]

Before any branch, check:

• cluster_key exists in adata.obs
• the data object is already annotation-ready
• downstream plotting keys such as X_umap exist if you plan to render embeddings

Constraints

• Do not use notebook-local absolute paths like /scratch/... in reusable docs or examples.
• Do not assume celltypist is installed in every environment.
• The prompt-only fallback lives in gptcelltype(...), not in Annotation.annotate(method="gpt4celltype").
• The live source exposes extra method branches harmony, scVI, and scanorama; they are not part of the notebook spine.
• For SCSA, keep the key you validate aligned with the branch you run: pySCSA.cell_auto_anno(..., key='scsa_celltype') uses its own default, while Annotation.annotate(method='scsa', ...) writes scsa_prediction.

Resource Map

• Read references/backend-selection.md when choosing between the three annotation backends.
• Read references/source-notebook-map.md to trace notebook cells into this reusable skill.
• Read references/source-grounding.md for inspected signatures, source-level branch behavior, and compatibility notes.
• Read references/compatibility.md when notebook prose, parameter names, or current runtime behavior diverge.