Figure 1. Garnett accurately classifies peripheral blood mononuclear cells.

A) Overview of Garnett algorithm. See Methods for algorithmic details. Briefly, Garnett takes as input a marker file that defines cell types using marker genes, and builds a cell type hierarchy that can include cell subtypes. Next, Garnett trains a classifier using elastic net multinomial regression¹⁷ at each node beginning at the root of the tree by comparing cell type representative cells. Lastly, Garnett hierarchically classifies all cells and optionally provides a second cluster-extended classification. B) t-SNE plots of 10x Genomics’ 100,000 cell PBMC dataset (n = 94,571 cells). The first panel is colored by cell type based on FACS sorting, the second panel is colored by cluster-agnostic cell type according to Garnett classification, and the third panel is colored by the Garnett cluster-extended type, which labels cells based on the composition of their cluster or community. C) A heatmap of data in (B) comparing the labels based on FACS (rows) with the cluster-agnostic (left) and cluster-extended (right) cell type assignments by Garnett (columns). Color represents the percent of cells of a certain FACS type labeled each type by Garnett. D) t-SNE plots of 10x Genomics V2 chemistry applied to PBMCs from a healthy donor (n = 8,381 cells). The first panel is colored by type determined manually using known gene markers. The second and third panels are colored by Garnett cluster-agnostic and cluster-extended cell type assignments by a classifier trained on the data shown in panels (B) and (C). E) Similar to panel (C), a heatmap of data in (D).