Scalable single-cell profiling of chromatin modifications with sciCUT&Tag

Abstract

Cleavage under targets and tagmentation (CUT&Tag) is an antibody-directed in situ chromatin profiling strategy that is rapidly replacing immune precipitation-based methods, such as chromatin immunoprecipitation-sequencing. The efficiency of the method enables chromatin profiling in single cells but is limited by the numbers of cells that can be profiled. Here, we describe a combinatorial barcoding strategy for CUT&Tag that harnesses a nanowell dispenser for simple, high-resolution, high-throughput, single-cell chromatin profiling. In this single-cell combinatorial indexing CUT&Tag (sciCUT&Tag) protocol, lightly cross-linked nuclei are bound to magnetic beads and incubated with primary and secondary antibodies in bulk and then arrayed in a 96-well plate for a first round of cellular indexing by antibody-directed Tn5 tagmentation. The sample is then repooled, mixed and arrayed across 5,184 nanowells at a density of 12-24 nuclei per well for a second round of cellular indexing during PCR amplification of the sequencing-ready library. This protocol can be completed in 1.5 days by a research technician, and we illustrate the optimized protocol by profiling histone modifications associated with developmental gene repression (H3K27me3) as well as transcriptional activation (H3K4me1-2-3) in human peripheral blood mononuclear cells and use single-nucleotide polymorphisms to facilitate collision removal. We have also used sciCUT&Tag for simultaneous profiling of multiple chromatin epitopes in single cells. The reduced cost, improved resolution and scalability of sciCUT&Tag make it an attractive platform to profile chromatin features in single cells.

Fig. 1:. Overview of the sciCUT&Tag experimental workflow.

a, Schematic of the sciCUT&Tag sample processing procedure. Cells are first incubated with primary and secondary antibodies in bulk, then arrayed in 96- well format for tagmentation with barcoded pA-Tn5. The 96-well plate is then pooled and split with a TaKaRa ICELL8 robot to the 5,184-well chip for PCR amplification that adds a second barcode before sequencing. Cells from different samples are indicated in blue, yellow and red; chromatin epitope conditions are hashed and unhashed; nucleosomes are colored fuschia with yellow star-shaped epitopes; pA-Tn5 is colored peach and green; barcoded indices are multicolored line segments next to black genomic template, b, Schematic of the informatics approach for sciCUT&Tag data processing. For mixed-donor samples, collisions are first detected by de novo genotyping to identify and remove barcodes with SNPs from multiple genotypes. Dimensionality reduction is then performed on a matrix of cells by genomic bins before graph-based clustering to identify subpopulations within a sample. These subpopulations than can then be assigned a celltype annotation based on gene coverage. Distinct SNPs at the same locus are green and pink within black cells. See text for details of the procedure.

Fig. 2:. Validation of a de novo genotyping strategy for SNP-based collision removal in sciCUT&Tag of mixed-donor PBMCs.

a, Fragment size distributions for H3K27me3 (blue) and H3K4mel-2–3 (yellow) sciCUT&Tag profiles, b, Reads/cell distributions for H3K27me3 (blue) and H3K4mel-2–3 (yellow) sciCUT&Tag profiles. A black-dashed line marks 500 reads/cell and barcodes with less than 500 reads are excluded from further analyses, c, Nextfloŵ pipeline schematic built on Souporcell^ for de novo genotyping and SNP-based collision removal. sciCUT&Tag profiling of multiple epitopes are the blue, red and yellow paths that are merged into one purple path before de novo genotyping with Souporcell. Input files are genomic alignments and the reference genome used for alignment. The cell barcode is expected in the read name and is parsed to the cell barcode (CB) tag of the genomic alignment. Outputs are barcoded bed files, PCA of Souporcell cluster assignment probabilities, an ArchR^ object annotated with genotype assignments and tabulated cell metadata. The software used in the pipeline are shown in green and the pipeline terminus is a black line, d, Cell counts versus the log differential of the genotype assignment cluster probabilities colored by donor for H3K27me3 (top) and H3K4mel-2–3 (bottom), e, The two genotype assignment cluster probabilities plotted against each other and colored by genotype assignment for H3K27me3 (top) and H3K4mel-2– 3 (bottom), f, The percentage of cells per condition (donor x dispense target) for each genotype assignment for H3K27me3 (top) and H3K4mel-2–3 (bottom).

Fig. 3:. Systematic investigation of dimensionality reduction parameters for sciCUT&Tag.

a, H3K27me3 profiling in PBMCs was binned at 500 bp, 5 kb and 50 kb genomic intervals, then input to iterative LSI— for dimensionality reduction. Proportional variance (blue) and correlation to reads/cell (red; Pearson’s r) was calculated across the first 45 LSI dimensions. Gray lines mark dimensions 15 and 30. b, For H3K27me3, UMAPs were computed from the first 15,30 and 45 LSI dimensions and colored by reads/cell, c, H3K4mel-2–3 profiling in PBMCs was binned at 500 bp, 5 kb and 50 kb genomic intervals, then input to iterative LSI for dimensionality reduction. Proportional variance (blue) and correlation to reads/cell (red; Pearson’s r) was calculated across the first 45 LSI dimensions. Gray lines mark dimensions 15 and 30. d, For H3K4mel-2–3, each bin width, UMAPs were computed from the first 15,30 and 45 LSI dimensions and colored by reads/cell.

Fig. 4:. Tunable combinatorial indexing in sciCUT&Tag with the ICELL8 system.

a, Readout from the TapeStation for sciCUT&Tag profiling of H3K27me3 and H3K4mel-2–3 in PBMCs showing fragment size distribution, b, sciCUT&Tag barcoding schema, c, A 96-well plate heat map shaded per well by unique barcode combinations with greater than 100 reads recovered by sequencing. Marginal bar plots show the mean number of unique combinations recovered per s5 and s7 index and are shaded accordingly. Marginal bar plots for the s7 indices are labeled by profiled epitope. Error bars are standard deviation, d, 5,184-well chip shaded by unique barcode combinations with greater than 100 reads recovered by sequencing per well. Marginal bar plots show the mean number of unique combinations recovered per column and row and are shaded accordingly. Marginal bar plots along thê-axis are labeled by the cell/well dispense target. Error bars are standard error.

Fig. 5:. Visual inspection of genomic coverage and peak-calling analysis.

a, Genomic coverage at the HOXA locus on chromosome 7 for H3K27me3 (red) and H3K4mel-2–3 (teal) aggregated profiles. Peaks detected by SEACR^ (black) are plotted below their corresponding track, b, Peak width comparison between H3K27me3 and H3K4mel-2–3 peak sets (independent f-test). Box plots show the median with first and third quartiles and outliers, c, Comparison of H3K27me3 and H3K4mel-2–3 reads per million in the two respective peak sets. Raw reads per peak were normalized by total fragment counts per million. Box plots show the median with first and third quartiles and outliers, d, Quantification of overlapping peaks across the H3K27me3 and H3K4mel-2–3 peak sets.

Fig. 6:. sciCUT&Tag supports graph-based clustering and cell type annotation of human PBMCs.

a,UMAP embedding for H3K27me3 profiling in PBMCs colored by sciCUT&Tag cell type annotation based on gene coverage. Each point represents one cell. b,Reads/cell distribution per population in the H3K27me3 dataset. Black-dashed line is the dataset median. Density plot heights are normalized, box plots show the median with first and third quartiles. Each point represents one cell, c, Heat map of gene coverages per population for H3K27me3 with a log fold- change greater than 0.5 under a false discovery rate (FDR) of 0.01. d, Heat map of gene coverages per population for H3K27me3 with a log fold-change less than 0.5 under an FDR of 0.01. e, UMAP embedding for H3K4mel-2–3 profiling in PBMCs colored by sciCUT&Tag cell type annotation based on gene coverage. Each point represents one cell, f, Reads/cell distribution per population in the H3K4mel-2–3 dataset. Black dashed line is the dataset median. Density plot heights are normalized, box plots show the median with first and third quartiles. Each point represents one cell, g, Heat map of gene coverages per population for H3K4mel-2–3 with a log fold-change greater than 0.5 under a FDR of 0.01. h, Heat map of gene coverages per population for H3K4mel-2–3 with a log fold-change less than 0.5 under an FDR of 0.01. Heat map data are row-normalized, binary sorted by row and clustered by column. Mem, memory; NK, natural killer; Mono, monocyte.

Fig. 7:. Orthogonal validation of sciCUT&Tag cell-type annotation by bulk projection of public data.

a, UMAP embeddings for H3K27me3 (top) and H3K4mel-2–3 (bottom) profiling in PBMCs colored by cell type annotation based on gene coverage in sciCUT&Tag. Each point represents one cell, b, Bulk projections of ENCODE ChlP-seq into sciCUT&Tag UMAP embeddings. Counts per genomic bin (50 kb for H3K27me3,500 bp for H3K4mel and H3K4me3) were randomly split into 500 pseudo-cells per bulk dataset, then projected via sciCUT&Tag LSI representations. H3K27me3 profiles for FACS-purified populations were projected into H3K27me3 sciCUT&Tag (top) and H3K4mel (bottom left) and H3K4me3 (bottom right) profiles were projected into H3K4mel-2–3 sciCUT&Tag. Th, T helper; Treg, T regulatory; C-Mem, central memory; E-mem, effector memory.