omicverse-cross-modal-celltype-transfer

Cross-modal cell-type annotation transfer uses a shared embedding to map labels from a reference AnnData to a query AnnData via weighted KNN. This enables labeling an ATAC query from an RNA reference without manual annotation.

Weighted KNN transfer trains a model on the reference embedding and applies it to the query embedding to predict cell-type labels. It calculates prediction uncertainties and writes transf_celltype and transf_celltype_unc back to the query object.

Cell-type transfer works with any shared embedding stored in .obsm, such as X_glue. You can select the target annotation column using label_keys prefix matching to apply the weighted KNN model.

Converting OmicVerse notebooks into reusable workflows requires calling weighted_knn_trainer on the reference embedding and weighted_knn_transfer on the query. The process writes predictions and uncertainties directly to the query AnnData object.

Cross-modal transfer supports optional visualization via the embedding() function after writing transf_celltype and transf_celltype_unc to query.obs. This allows quick validation of transferred cell-type labels and uncertainties.

Weighted KNN cross-modal transfer depends on a pre-existing shared embedding between the reference and query AnnData. It cannot transfer labels if the modalities lack an aligned embedding space in .obsm.