Label Transfer

The quality of label transfer is entirely contingent on the reference atlas. The reference must be comprehensive, containing all cell types expected in the query, and meticulously annotated. A reference missing a rare population will force the algorithm to incorrectly map those query cells to the closest available label, a phenomenon known as classification by exclusion. High-quality references, like the Human Cell Atlas, provide standardized ontologies that ensure consistent nomenclature across studies.

The core algorithmic mechanism relies on identifying anchors—pairs of cells from the reference and query that are mutual nearest neighbors in a shared low-dimensional space. This reciprocity filter is critical for rejecting spurious matches. The process typically involves:

• Canonical Correlation Analysis (CCA) to align datasets in a correlated subspace.
• L2-normalization of embeddings to mitigate sequencing depth artifacts.
• A shared nearest neighbor (SNN) graph to score anchor consistency before propagating labels.

Unlike traditional classifiers that require retraining for every new dataset, label transfer operates in a zero-shot manner. The reference model, often a pre-trained foundation model like Geneformer or scGPT, generates contextualized embeddings for the query cells without any fine-tuning. The query cells are projected into the reference's latent space, and labels are assigned based on proximity to annotated reference neighborhoods, enabling instant annotation of millions of cells.

Modern label transfer implementations provide a prediction score for each transferred label, quantifying the confidence of the mapping. This score is typically derived from the proportion of nearest reference neighbors sharing the majority label. Low-confidence mappings often indicate:

• Novel cell states absent from the reference.
• Doublets or low-quality cells.
• Intermediary differentiation states that sit between discrete reference labels. Filtering by a confidence threshold prevents the propagation of erroneous annotations.

Label transfer is inherently a data integration task. The query and reference often originate from different laboratories, protocols, and donors, introducing severe batch effects. Algorithms like Seurat v5 and scArches decouple biological variation from technical noise by learning a conditional latent space. This ensures that a T-cell from a 10x Genomics query correctly maps to a T-cell from a Smart-seq2 reference, rather than clustering by technology.

Advanced label transfer respects the cell ontology tree. Instead of forcing a flat classification, algorithms can propagate labels at multiple resolutions—assigning a broad lineage label (e.g., Lymphoid) when fine-grained mapping (e.g., CD4+ Naive T-cell) is uncertain. This hierarchical approach, implemented in tools like CellTypist, provides a structured annotation that reflects the nested nature of cell identity and avoids high-confidence errors at the terminal leaf nodes.

The core computational framework that enables label transfer. A query dataset is projected onto a pre-built reference atlas using dimensionality reduction and nearest-neighbor search. The algorithm identifies the most transcriptionally similar reference cells for each query cell, then propagates annotations based on a majority vote or similarity-weighted consensus. This avoids computationally expensive full dataset reintegration and allows rapid, scalable annotation of new single-cell experiments against curated, gold-standard references like the Human Cell Atlas.

The computational alignment of multiple single-cell datasets into a shared latent space, correcting for batch effects while preserving genuine biological variation. Methods like Harmony and scVI are critical precursors to label transfer, ensuring that cells of the same type from different experiments occupy overlapping positions in the embedding. Without proper integration, batch-specific technical noise can dominate the transcriptional similarity search, leading to systematic misannotation when labels are transferred across studies.

The process of assigning biological identities to single-cell clusters by comparing their gene expression signatures to known marker genes or reference databases. Label transfer automates this traditionally manual curation step. Key approaches include:

• Marker-based: Manual inspection of canonical gene expression
• Correlation-based: Comparing cluster averages to bulk reference profiles
• Classifier-based: Training supervised models on labeled reference data
• Reference mapping: Projecting query cells onto annotated atlases for direct label propagation

Large-scale pretrained transformer models like Geneformer and scGPT that learn universal cell representations from massive single-cell corpora. These models are revolutionizing label transfer by providing context-aware embeddings that capture nuanced cell states beyond simple marker gene matching. A foundation model can perform zero-shot annotation by attending to gene-gene interaction networks learned during self-supervised pretraining, enabling accurate label transfer even for rare or transitional cell populations that lack clear marker signatures.

A preprocessing step that removes technical variation introduced by different experimental batches, sequencing platforms, or laboratory protocols. Effective batch correction is a prerequisite for successful label transfer, as uncorrected technical artifacts can dominate the similarity search and cause systematic misclassification. Methods range from linear approaches like ComBat to nonlinear techniques like Mutual Nearest Neighbors (MNN) and deep generative models. The goal is to align cells of the same biological type across batches while preserving genuine condition-specific differences.