Extended Data Fig. 3. Utility of NO for cell-typing.

(a) Cell-typing based on NO at gene bodies (AUC = 1). Epi1: RPE-1 replicate 1 (79 cells); Epi2: replicate 2 (77 cells); LCL1: HG01573 (46 cells); LCL2: HG02018 (50 cells), LCL3: NA19036 (50 cells); LV: latent variable. (b) UMAP visualization of Strand-seq libraries based on NO at gene-bodies (normalized by segmental ploidy status²⁴). (c) We also explored dimensionality reduction of Strand-seq libraries based on DNA motif accessibility. Using the chromVAR package¹¹⁰, single-cell NO profiles for 2 kb DNase I hypersensitive sites (DHSs) were transformed into a deviation Z-score, which measures how likely a certain motif accessibility would occur when randomly sampling sets of peaks with similar GC content and read depth. For each single-cell, the deviation Z-score was calculated for 870 human TF motifs from the cisBP database¹¹¹. These dimensionality reduction plots suggest that batch effect within the same cell type (three individuals in LCL, and two batches in RPE-1 sequenced separately) is minimal, and far less than the cell-type dependent variability. (d) UMAP using scMNase-seq²⁶, including 45 NIH3T3 cells and 272 murine naive T cells, based on NO at the gene-bodies. (e) UMAP of RPE-1 (the originally commercially available cell line) and its transformed derived³⁷ cell lines (BM510 and C7). Two biological replicates were sequenced for each cell line. (f) Receiver operating characteristic (ROC) using the PLS-DA based classifier. AUC for classifying each cell line was 0.9614, 0.9694, and 0.9892 for RPE-1, BM510, and C7 respectively. (g-h) Cell-typing for LCL, RPE-1, skin fibroblast, AML, T-ALL, and umbilical cord blood cells (g), and ROC curve depicting classification performance (overall AUC = 0.998) (h). (i-j) Cell-typing in five RPE-1 derived cell lines³⁷ (RPE-1, BM510, C7, C29, and C11) (i), and ROC curve depiciting classification performance (overall AUC = 0.9648) (j).