Fig. 1

The accesson matrix constructed from the sparse fragment count matrix improved the clustering of scATAC-seq data. a Step-by-step workflow of APEC. Peaks were grouped into accessons by their accessibility pattern among cells with the K-nearest-neighbor (KNN) method. b t-Distributed Stochastic Neighbor Embedding (tSNE) diagrams of the hematopoietic single-cell dataset based on the dimension-transformed matrices from different algorithms, i.e., APEC: accesson matrix; cisTopic: topic matrix; LSI: LSI matrix; chromVAR: bias-corrected deviation matrix; and Cicero: aggregated model matrix. The cells are FACS-indexed human hematopoietic cells, including HSCs (hematopoietic stem cells), MPPs (multipotent progenitors), LMPPs (lymphoid-primed multipotential progenitors), CMPs (common myeloid progenitors), CLPs (common lymphoid progenitors), pDCs (plasmacytoid dendritic cells), GMPs (granulocyte-macrophage progenitors), MEPs (megakaryocyte-erythroid progenitors), and UNK (unknown type) cells. c The ARI (Adjusted Rand Index) values for the clustering of the human hematopoietic cells by different algorithms. The same as the two normalization methods applied in cisTopic, we normalized the accesson matrix in APEC based on probability (P) and z-score (Z). Center line, median; box limits, upper and lower quartiles; whiskers, 1.5× interquartile range; points, outliers. d Three CMP subtypes identified in APEC and the motifs enriched in each cell subtype