Omicverse Cross Modal Celltype Transfer

OmicVerse Cross-Modal Cell Type Transfer

Goal

Transfer cell labels from a reference AnnData to a query AnnData using a shared latent space such as X_glue, then write predicted labels and uncertainty scores back to the query object. Keep this as one skill: the notebook's plots are optional presentation on top of the same transfer job, not a separate capability.

Quick Workflow

1. Confirm the query and reference expose the same embedding basis in .obsm, or fall back to X.
2. Confirm the reference has the target label column in .obs.
3. Train ov.utils.weighted_knn_trainer(...) on the reference embedding.
4. Transfer labels with ov.utils.weighted_knn_transfer(...).
5. Store the first label column and uncertainty column in query_adata.obs.
6. Plot on an existing basis for smoke validation; use mde only if pymde is available.

Interface Summary

• weighted_knn_trainer(train_adata, train_adata_emb, n_neighbors=50) builds the reference KNN model.
• weighted_knn_transfer(query_adata, query_adata_emb, ref_adata_obs, label_keys, knn_model, threshold=1, pred_unknown=False, mode='package') returns prediction and uncertainty DataFrames.
• embedding(adata, basis, ..., show=None, save=None, return_fig=None, ...) plots any existing basis.
• mde(data, device=None, **kwargs) is optional; it depends on pymde and is only for visualization.

Branch Selection

• Use mode='package' for transfer. The current source only implements this branch; mode='paper' is documented in the docstring but raises in the live code.
• Leave pred_unknown=False for the default label-assignment path.
• Set pred_unknown=True only when you want threshold-based Unknown labels.
• Use mde only as an optional visual branch. If pymde is missing, skip it and plot on X_glue or another precomputed basis instead.

Input Contract

• train_adata and query_adata must expose the same latent basis name in .obsm, or use X.
• ref_adata_obs[label_keys] must exist; the function selects columns whose names start with label_keys.
• If you want one label column, make the prefix unique.
• The notebook's transf_celltype and transf_celltype_unc are the reuse convention for storing outputs on the query object.

Minimal Execution Patterns

Core transfer:

knn = ov.utils.weighted_knn_trainer(train_adata=rna, train_adata_emb="X_glue", n_neighbors=15)
labels, uncert = ov.utils.weighted_knn_transfer(
    query_adata=atac,
    query_adata_emb="X_glue",
    ref_adata_obs=rna.obs,
    label_keys="major_celltype",
    knn_model=knn,
    mode="package",
)
atac.obs["transf_celltype"] = labels.loc[atac.obs.index, "major_celltype"]
atac.obs["transf_celltype_unc"] = uncert.loc[atac.obs.index, "major_celltype"]

Optional smoke plot:

ov.utils.embedding(
    atac,
    basis="X_glue",
    color="transf_celltype",
    show=False,
)

Validation

• labels and uncert must both be DataFrames indexed by query_adata.obs_names.
• The expected label column should appear in both outputs.
• query_adata.obs["transf_celltype"] and query_adata.obs["transf_celltype_unc"] should exist after writeback.
• If you call mde, verify pymde is installed first; otherwise skip it.
• For plots, confirm the basis you pass actually exists in .obsm.

Resource Map

• Read references/integration-method-selection.md for skill boundary and branch selection.
• Read references/source-notebook-map.md to map notebook cells to the reusable workflow.
• Read references/source-grounding.md for inspected signatures and source behavior.
• Read references/compatibility.md for known runtime caveats.