Cell-Type Annotation with a Fine-Tuned Foundation Model

Challenge

Annotating cell types across a large PBMC dataset without leaning on manual marker-gene curation for every cluster.

Approach

A scGPT-style single-cell foundation model, fine-tuned on a public 10x Genomics PBMC reference and validated against manual annotation.

What we found

Predicted labels agreed with expert-curated labels on most of the cells checked. The held-out set and the expert labels used for that comparison aren't published alongside the figure yet — see the caveat on the metric above.

Limits of this demonstration

This is an R&D demonstration on a public 10x Genomics PBMC reference, not a client engagement. No repo or notebook is linked yet, which is why the metric above is flagged TODO-VERIFY. Rare cell populations and disease-specific states are not represented in this reference, so agreement here says nothing about performance on those.