Fig. 1. Automated annotation of immune cells across human tissues using CellTypist.

(A) Schematic showing sample collections of human lymphoid and non-lymphoid tissues and their assigned tissue name acronyms. (B) Schematic of single-cell transcriptome profiling and paired sequencing of αβ TCR, γδ TCR and BCR variable regions. (C) Workflow of CellTypist including data collection, processing, model training and cell type prediction (upper panel). Performance curves showing the F1 score at each iteration of training with mini-batch stochastic gradient descent for high- and low-hierarchy CellTypist models, respectively (lower panel). The black curve represents the median F1 score averaged across the individual F1 scores of all predicted cell types. (D) UMAP visualization of the immune cell compartment colored by tissues. Note jejunum samples in (A) were further split into epithelial (JEJEPI) and lamina propria fractions (JEJLP). (E) As in (D), but colored by predicted cell types using CellTypist.