Annotating Cell Types at Scale with Foundation Model Representations

Annotation is the part of single-cell analysis that looks like it should be automatable but has resisted complete automation for years. The basic problem is that identifying what type of cell you are looking at requires knowing what genes should be on or off for that cell type in that tissue in that disease context. That knowledge is not uniform. It is distributed across thousands of papers, expert conventions that have shifted over time, and marker gene lists that overlap in complicated ways across tissues.

The scale mismatch has only gotten worse. A typical scRNA-seq study now generates hundreds of thousands of cells. Studies intended for atlas-building or foundation model training run into tens of millions. Manual annotation at those scales is not feasible. The question is not whether to automate annotation but how to do it without losing the resolution that makes single-cell data worth collecting in the first place.

Why Reference-Based Annotation Has Limits

The most common automated approach is reference-based annotation: build a classifier on an annotated reference dataset, apply it to new query data. Tools like SingleR, Seurat label transfer, and scANVI all work on this principle. For coarse cell type labels, they perform well. For fine-grained subtypes, the performance depends heavily on whether those subtypes are represented in the reference.

The core limitation is distributional shift. A reference dataset and a query dataset are often processed differently, from different donors, from slightly different tissue regions. Even if the same cell type exists in both datasets, the expression profile can shift enough that a nearest-neighbor classifier fails to retrieve the correct label. Rare subtypes are particularly vulnerable because there are few reference examples to anchor the decision boundary, and the classifier defaults to the most common neighboring cell type.

There is also a circularity problem in reference construction. Large reference atlases aggregate multiple studies, and annotation from those studies is passed along as ground truth. In practice, earlier studies used cruder annotation methods, and inconsistencies accumulate. A classifier trained on this reference can learn to reproduce those inconsistencies rather than reflect genuine biology.

What Foundation Model Representations Change

A model pretrained on a broad distribution of cell profiles has already encountered most common cell types and many rare ones during training. When you encode a new cell through that model's encoder, the resulting representation places the cell in a high-dimensional space that reflects the shared biology it shares with the training distribution.

For annotation, this changes the task from "find the closest reference cell" to "find where this cell sits in a shared representation space." The distinction matters when the reference and the query are from different processing protocols or tissue sources. The shared representation abstracts over some of that technical variation, because the model learned to represent cell-type programs that persist across contexts.

In practical terms, this means that annotation transfer using foundation model embeddings is more robust to protocol differences than direct count-level nearest-neighbor transfer. When we process a query dataset through a pretrained encoder and then retrieve the nearest annotated reference cells in embedding space, the matches are more biologically coherent even when the datasets were processed years apart with different library preparation kits.

Rare Subtypes and the Coverage Problem

Rare subtypes are the critical test. A cell type that represents one percent of a tissue sample might appear in only a few hundred cells across a large study. If those cells are not in the training data, no amount of model sophistication recovers them. The foundation model approach does not solve the coverage problem, it shifts it.

Instead of needing the rare subtype explicitly in a reference dataset, you need the model to have seen the constituent gene expression programs that make up the subtype's identity during training. If the rare subtype is a novel disease-associated cell state that has no precedent in any healthy reference, the foundation model will still fail. It will assign the cell to the closest known type, which may be biologically misleading.

This is the honest boundary of the approach: foundation model annotation works well when the query contains known cell types or variants of known types. It works poorly when the query contains genuinely novel biology with no analogue in training data. For disease target discovery, this matters because some of the most interesting biology is the disease-specific cell state that has not been characterized before.

The practical response is to treat foundation model annotation as a first-pass assignment, not a final label. High-confidence assignments for common cell types can be accepted automatically. Low-confidence assignments, and the cells that end up in ambiguous or mixed neighborhoods in embedding space, require human review. The goal is to use automation to handle the easy 80 to 90 percent and concentrate expert attention on the cells where uncertainty is highest.

Building a Confidence-Stratified Annotation Workflow

The workflow we have converged on for large datasets uses a tiered strategy. The first pass uses foundation model embeddings to assign cell type labels and compute a confidence score based on the distance to the nearest annotated cluster and the homogeneity of the neighborhood. Cells with high confidence and consistent neighbors get assigned automatically. Cells with low confidence or conflicting neighbor labels are flagged for review.

The second layer looks at the flagged cells using marker gene expression profiles. Rather than assigning these cells based purely on the embedding, the reviewer looks at the top differentially expressed genes for the ambiguous cluster and uses biological knowledge to decide whether it is an unannotated subtype, a doublet, or a cell in a transient state that does not belong to any stable type.

This tiered approach keeps total expert annotation time manageable even for large datasets. For a dataset of 500,000 cells, the first-pass assignment might handle 85 to 90 percent automatically, leaving 50,000 to 75,000 cells that need review. That is still a lot of cells, but it is tractable with a focused review process, and the flagged cells tend to be the biologically interesting ones.

Annotation Consistency Across Studies

One underappreciated benefit of foundation model-based annotation is consistency across studies. When you annotate two studies using the same embedding space and the same reference labels, the resulting annotations are comparable in a way that independently annotated studies are not. Cell types with the same name in two different papers often do not reflect the same cell state, because the original authors used different marker gene thresholds and different annotation conventions.

When the annotation is anchored to a shared representation space rather than to study-specific conventions, the labels are defined by position in that space. This makes cross-study comparisons more interpretable. A disease-associated subtype identified in one dataset can be asked whether it corresponds to a specific annotated cluster in another dataset, and the comparison has a concrete computational meaning rather than relying on subjective label matching.

For target discovery specifically, this cross-study comparability is one of the more valuable properties of foundation model annotation. A target hypothesis that rests on a disease-associated cell state is stronger if the same cell state is identifiable across independent datasets from different labs. Foundation model annotation makes that cross-study comparison tractable at scale.