Single-cell foundation models vs. marker-based annotation: what actually changes

The pitch for single-cell foundation models is usually framed as an accuracy upgrade over marker-gene annotation. That framing skips the more useful question: what kind of mistake does each approach make when it's wrong, and which of those mistakes can you actually catch?

What marker-based annotation gets right

Marker-based annotation is transparent by construction. A cluster is labeled a given cell type because it expresses a specific, named set of genes, at a threshold someone chose and can defend. When it's wrong, you can usually see why — a marker panel that doesn't distinguish two closely related subtypes, a threshold that's too strict for a noisy sample. The failure is legible, which means it's correctable by a domain expert without retraining anything.

Where it struggles: rare or genuinely novel cell states that don't match any curated marker set, and batch effects that require re-tuning thresholds per dataset by hand.

What changes with a foundation model

A model like scGPT learns a representation from a large reference corpus instead of a hand-written rule set, which is genuinely useful when you don't have — or don't want to write — a marker panel for every dataset you touch. But the failure mode inverts: a misannotation isn't a legible rule that failed, it's a confident label with no visible reasoning behind it. Catching that requires the same domain review a marker-based pipeline needs, just applied after the fact instead of encoded up front.

It also isn't free in the way it's sometimes presented. Fine-tuning on your reference, validating against it, and maintaining that pipeline is real, ongoing work — not a one-time swap.

Where we land on this

Not either/or. A foundation model is a reasonable accelerant for the first pass across a large dataset, especially one without a mature marker panel. Marker-based review stays as the check on its output, particularly for any cluster near a decision boundary or any population the reference corpus is unlikely to have seen well. Treat the model's output as a fast draft, not a finished annotation.

Our own cell-type annotation case study follows roughly this shape — a fine-tuned foundation model validated against manual annotation on a public PBMC reference. The agreement figure reported there is still flagged TODO-VERIFY: the held-out set and the expert labels it's compared against aren't published alongside it yet, for the same reasons laid out in our benchmarking post on binding-affinity metrics — an agreement percentage without a documented comparison set is a claim, not a result.