Fig. 4: Zero-shot cell type annotation performance of each model.

Heatmaps illustrating (a) classification accuracy and (b) Macro-F1 scores of each model across intra-datasets. Red indicates lower performance, and blue represents the highest performance. c The river plot of CellFM illustrates the predicted cell types and their relationships to the actual cell types on the Immune dataset. d The river plot of scGPT illustrates the predicted cell types and their relationships to the actual cell types on the Immune dataset. Heatmaps show (e) classification accuracy and (f) Macro-F1 scores of each model across inter-datasets, with each value representing the average accuracy calculated from five independent runs using different random seeds. Red indicates lower performance, and blue represents the highest performance. g The river plot of CellFM illustrates the predicted cell types and their relationships to the actual cell types on the hPancreas dataset. h The river of scGPT plot illustrates the predicted cell types and their relationships to the actual cell types on the hPancreas dataset. Source data are provided as a Source Data file.