Omicverse Reference Label Transfer

OmicVerse Reference Label Transfer

Goal

Map labels from a reference AnnData onto a query AnnData by building a shared integrated space, training a weighted kNN transfer model, and writing predicted labels plus uncertainty scores back to the query object.

Treat harmony, scVI, and scanorama as alternative integration branches. They share the same query/reference contract, so keep them in one skill rather than splitting into thin wrappers.

Quick Workflow

1. Start from a query AnnData and a reference AnnData that share gene names.
2. Make sure the reference has a cell-type column such as celltype in adata_ref.obs.
3. If the data are already log-normalized, recover or restore counts before concatenating.
4. Construct ov.single.AnnotationRef(adata_query, adata_ref, celltype_key=...).
5. Run preprocess(mode=...) on the combined object.
6. Choose one integration backend with train(method=...).
7. Transfer labels with predict(method=...).
8. Validate the method-specific prediction keys before plotting or writing outputs.

Interface Summary

• AnnotationRef(adata_query, adata_ref, celltype_key='celltype') checks shared genes, concatenates query and reference, and stores the integrated object on self.adata_new.
• AnnotationRef.preprocess(mode='shiftlog|pearson', n_HVGs=3000, batch_key='integrate_batch') prepares the concatenated object for transfer.
• AnnotationRef.train(method='harmony', **kwargs) computes the integrated latent space used for transfer.
• AnnotationRef.predict(method='harmony', n_neighbors=15, pred_key=None, uncert_key=None) transfers labels with weighted kNN.
• The current source also exposes batch_correction(..., methods='harmony') with additional methods such as combat, scanorama, scVI, CellANOVA, and Concord; the notebook uses only the harmony and scVI transfer path.

Read references/source-grounding.md before documenting narrower parameter behavior than the live source supports.

Branch Selection

• Use harmony when you want the notebook's default transfer path and a fast, interpretable integration backend.
• Use scVI when you want the scVI latent-space branch and the environment has scvi-tools.
• Use scanorama when you explicitly want Scanorama integration and its dependencies are available.
• Keep harmony as the default unless the user asks for a different integration backend.
• Keep mode='shiftlog|pearson' as the default preprocessing choice unless the user wants a different normalization/HVG strategy.

Input Contract

• Query and reference must share feature names before AnnotationRef can concatenate them.
• The reference must carry a cell-type label column in adata_ref.obs.
• If the notebook reference was stored with a feature_name column, align adata_ref.var_names first.
• If adata_query.X.max() or adata_ref.X.max() suggests the data are already log-normalized, recover counts before transfer.
• AnnotationRef.preprocess(...) uses the combined query/reference object, so treat the pair as one integrated preprocessing unit.

Minimal Execution Patterns

For harmony-based transfer:

import omicverse as ov

ref_anno = ov.single.AnnotationRef(adata_query, adata_ref, celltype_key="celltype")
ref_anno.preprocess(mode="shiftlog|pearson", n_HVGs=3000, batch_key="integrate_batch")
ref_anno.train(method="harmony", n_pcs=50)
ad_pred = ref_anno.predict(method="harmony", n_neighbors=15)
ov.pp.mde(ad_pred, use_rep="X_pca_harmony_anno")
ov.pl.embedding(ad_pred, basis="X_mde", color="harmony_prediction")

For scVI-based transfer:

import omicverse as ov

ref_anno = ov.single.AnnotationRef(adata_query, adata_ref, celltype_key="celltype")
ref_anno.preprocess(mode="shiftlog|pearson", n_HVGs=3000, batch_key="integrate_batch")
ref_anno.train(method="scVI", n_layers=2, n_latent=30, gene_likelihood="nb")
ad_pred = ref_anno.predict(method="scVI", n_neighbors=15)
ov.pl.embedding(ad_pred, basis="X_mde", color="scVI_prediction")

For prompt-driven label transfer around the reference selection step, keep the notebook's reference selection logic separate from the transfer logic; this skill starts once you already have the paired query/reference objects.

Validation

After preprocess(...), check:

• self.adata_new has integrate_batch in obs
• self.adata_new.obsm["X_pca"] exists
• self.adata_new.var["highly_variable_features"] exists

After train(method="harmony"), check:

• self.adata_query.obsm["X_pca_harmony_anno"] exists
• self.adata_ref.obsm["X_pca_harmony_anno"] exists

After train(method="scVI"), check:

• self.adata_query.obsm["X_scVI_anno"] exists
• self.adata_ref.obsm["X_scVI_anno"] exists

After train(method="scanorama"), check:

• self.adata_query.obsm["X_scanorama_anno"] exists
• self.adata_ref.obsm["X_scanorama_anno"] exists

After predict(...), check:

• harmony_prediction and harmony_uncertainty for the harmony branch
• scVI_prediction and scVI_uncertainty for the scVI branch
• scanorama_prediction and scanorama_uncertainty for the Scanorama branch

Before any plotting, check:

• the prediction columns are present in ad_pred.obs
• the integrated embedding key you pass to ov.pp.mde(...) exists

Constraints

• Do not bake notebook-specific download URLs, file names, or absolute paths into the reusable workflow.
• Do not assume the reference dataset is already in the correct gene-name convention; align feature names first when needed.
• Do not treat scVI as free: it depends on scvi-tools and can be materially heavier than the harmony path.
• Do not collapse the notebook into one monolithic "load data and run" step; the reusable spine is query/reference construction, preprocessing, backend selection, prediction, and validation.
• Treat scanorama as a source-present branch even though the notebook focuses on harmony and scVI.

Resource Map

• Read references/integration-method-selection.md when choosing between harmony, scVI, and scanorama.
• Read references/source-notebook-map.md to trace notebook sections 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, input assumptions, or current runtime behavior diverge.