scTab

scTab scales cross-tissue cell-type annotation to 22.2 million cells, showing that accurate pan-tissue annotation requires nonlinear models and that performance keeps improving with both dataset and model size — a scaling result for single-cell classification. It uses a tabular data-augmentation scheme over a large scRNA-seq corpus.

From Fabian Theis's group (joint→Helmholtz Munich); published in Nature Communications (2024).