Cell Annotation

scpFormer was especially strong for rare or phenotypically similar hematopoietic populations. scpFormer achieved a macro F1 of 0.978 on the Levine benchmark, compared with 0.868 for Seurat and 0.920 for CellTypist. In the Hao PBMC cross-batch setup, scpFormer achieved accuracy 0.918 and macro F1 0.789, outperforming Seurat and CellTypist metrics reported in the article. In the Levine human bone marrow benchmark, scpFormer outperformed Seurat and CellTypist across accuracy, precision, recall, and macro F1. The cross-batch annotation gain is interpreted as evidence that scpFormer captures batch-invariant co-expression structure instead of batch-specific artifacts. For cell type annotation, scpFormer is fine-tuned on labeled reference data and evaluated on held-out query data.