Single-cell chromatin state analysis with Signac

Key results

1. ShownCorrelating expression with accessibility of peaks within 500 kb of a gene's transcription start site (TSS), relative to matched background peaks, gave mostly positive, often distal links in peripheral blood mononuclear cell (PBMC) multiome data that agreed with fine-mapped eQTL genes above chance.

   37,424 links (one-sided z-test p < 0.05, |r| > 0.05); 89% positively correlated; 58% spanned >100 kb from the TSS; per gene mean 6.37 linked peaks (SD 7.09), per peak mean 1.57 genes (SD 1.26). For linked peaks overlapping GTEx v8 whole-blood CAVIAR-fine-mapped eQTLs, the eQTL gene equaled the linked gene in 52.6% of cases vs 13.4% expected from randomized links. · Analysis of multimodal human PBMC data; Peak-to-gene linkage; Fine-mapped eQTL analysis

2. ShownA modified TF-IDF for LSI, with IDF = N/ni and TF-IDF = log(1 + TF × IDF × 10^4), kept PBMC cell types separable at 20% of the original counts, whereas the commonly used IDF = log(1 + N/ni) did not.

   Counts per cell downsampled to 80%, 60%, 40% and 20% (DropletUtils); k-nearest-neighbor (KNN) purity (k=100, LSI components 2 to 40, cell types from the RNA assay) declined as counts fell for both methods, more steeply for the original; separation judged from UMAP, with purity values given only graphically. · Comparison of LSI dimension reduction methods; Latent semantic indexing; Count downsampling analysis

3. ShownQuantification, quality-control and dimension-reduction steps processed the Brain Initiative Cell Census Network (BICCN) adult mouse brain scATAC-seq dataset of 733,526 cells, with runtime and memory growing approximately linearly with cell number.

   Data span 45 brain regions with 315,334 unified Genrich peaks; cells subsampled from 733,526 down to 50,000 and FeatureMatrix, NucleosomeSignal, TSSEnrichment, RunTFIDF and RunSVD timed in triplicate; FeatureMatrix also run on 1, 2, 4 or 8 cores with maximum resident memory via GNU time; absolute values shown only in figures. · Scalable analysis of single-cell chromatin data; Raw data processing; Cell downsampling analysis

4. ShownCalling peaks separately for each RNA-annotated cell type with MACS2 recovered rare-population peaks lost in whole-dataset calls, and comparison with 10x Cellranger showed many Cellranger peaks spanning multiple distinct MACS2 peaks but never the reverse.

   Reference mapping of the RNA assay annotated 23 PBMC cell types, including gamma delta T cells; 15,777 cases of one Cellranger peak merging multiple MACS2 cell-type-specific peaks, none in the opposite direction; rare-population peak loss shown graphically without counts in text. · Analysis of multimodal human PBMC data; Gene expression data preprocessing and cell…; DNA accessibility data processing

5. SuggestedMitochondrial variant calling and allele-frequency clustering in a colorectal cancer scATAC-seq sample separated tumor-epithelial clones from clones spread across immune cell types (consistent with a shared hematopoietic progenitor) and enabled clone-resolved differential accessibility.

   Informative variants required variance-mean ratio > 0.01, strand concordance >= 0.65 and detection in >=5 cells; 16147C>T reached nearly 100% allele frequency in tumor epithelial cells; clustering (k=10, cosine distance) gave 10 clones, with clones 1, 2 and 4 epithelial-specific and others dispersed across immune types; data from a single patient. · Joint analysis of DNA accessibility and…; scATAC-seq data processing; Mitochondrial variant detection; Clonal clustering