ArchR: An integrative and scalable software package for single-cell chromatin accessibility analysis

Key results

1. 01

   ShownArchR was faster and leaner in memory than SnapATAC and Signac in every benchmark comparison, and completed import, dimensionality reduction and clustering of ~220,000 real and over 1.2 million simulated cells on eight-core, 32 GB hardware.

   70,000-cell datasets finished in under an hour with 32 GB RAM and eight cores (n = 3 runs). At ~70,000 peripheral blood mononuclear cells (PBMCs), SnapATAC ran out of memory with 128 GB/20 cores, and both SnapATAC and Signac did so at 32 GB/eight cores. Runs took less than 3 h for ~220,000 cells and under 8 h for over 1.2 million simulated PBMCs split into 200 samples. · ArchR enables efficient and comprehensive…; ArchR enables comprehensive analysis of…; Figure legends

2. 02

   ShownProjecting synthetic doublets into the low-dimensional embedding let ArchR flag genotype-confirmed doublets accurately, modestly beating Scrublet on scATAC-seq and agreeing with RNA-based doublet calls on paired Multiome cells.

   On 38,072 cells from ten pooled human cell lines with demuxlet genotypes as ground truth, ArchR reached ROC AUC 0.918 versus 0.641 for total fragment count; 27,220 cells remained after doublet removal. On 10x Multiome PBMCs, high-confidence ArchR calls agreed with Scrublet's scRNA-seq-based calls (AUC = 0.921). · ArchR identifies putative doublets in scATAC-seq…; Figure legends

3. 03

   ShownOf 56 gene-score models tested against matched scRNA-seq, models combining gene-body signal with distance-weighted distal elements bounded by neighboring genes predicted expression best; model 42 ranked top in both PBMC and bone marrow data and was adopted in ArchR.

   Pearson r between gene scores and expression linked via Seurat canonical correlation analysis, per gene and per 100-cell group (500 groups), for top 1,000 differentially expressed and 2,000 variable genes. Model 42 adds bidirectional exponential decay from the transcription start site (TSS; extended 5 kb upstream) and gene end; bulk and Multiome data confirmed this class. · Improved inference of gene scores enables…; Figure legends

4. 04

   SuggestedArchR's iterative latent semantic indexing (LSI) clustered simulated single cells by cell type, whereas SnapATAC and Signac split them by sample, and landmark-based estimated LSI preserved data structure from very small subsets.

   Bulk hematopoietic ATAC-seq (14 cell types, on average five replicates) was downsampled to 1,000 ± 500, 5,000 ± 1,000 and 10,000 ± 2,500 fragments per cell; ArchR gave a higher adjusted Rand index at each depth. Estimated LSI recovered clusters with as few as 50 landmark cells in PBMC (n = 27,845) and bone marrow (n = 26,748) data. · ArchR provides high-resolution and efficient…; Figure legends

5. 05

   ShownOn a ~220,000-cell multi-laboratory hematopoiesis compendium, ArchR resolved 21 clusters including rare plasma cells, built reproducible peak and peak-to-gene link sets, and traced B cell regulatory dynamics; on a subset, its trajectories agreed with Slingshot and Monocle 3.

   49 samples; plasma cells ~0.1% (265 cells); 396,642 reproducible peaks, 215,916 differentially accessible; scRNA-seq alignment in less than 1 h; 70,239 peak-to-gene links (R > 0.45, FDR < 0.1). On a 10,251-cell subset, B cell and monocyte trajectories matched Slingshot and Monocle 3 (r^2 > 0.96); the full data gave 11,999 B cell trajectory peak-to-gene links (r > 0.5). · ArchR enables comprehensive analysis of…; ArchR enables integration of matched scRNA-seq…; Figure legends