Multifactorial profiling of epigenetic landscapes at single-cell resolution using MulTI-Tag

Abstract

Chromatin profiling at locus resolution uncovers gene regulatory features that define cell types and developmental trajectories, but it remains challenging to map and compare different chromatin-associated proteins in the same sample. Here we describe Multiple Target Identification by Tagmentation (MulTI-Tag), an antibody barcoding approach for profiling multiple chromatin features simultaneously in single cells. We optimized MulTI-Tag to retain high sensitivity and specificity, and we demonstrate detection of up to three histone modifications in the same cell: H3K27me3, H3K4me1/2 and H3K36me3. We apply MulTI-Tag to resolve distinct cell types and developmental trajectories; to distinguish unique, coordinated patterns of active and repressive element regulatory usage associated with differentiation outcomes; and to uncover associations between histone marks. Multifactorial epigenetic profiling holds promise for comprehensively characterizing cell-specific gene regulatory landscapes in development and disease.

Fig. 1. MulTI-Tag directly identifies user-defined chromatin targets in the same cells.

a, Schematic describing the MulTI-Tag methodology. (1) Antibody–oligonucleotide conjugates are used to physically associate forward-adapter barcodes with targets and are loaded directly into pA–Tn5 transposomes for sequential binding and tagmentation. (2) pA–Tn5 loaded exclusively with reverse adapters are used for a secondary CUT&Tag step to efficiently introduce the reverse adapter to conjugate-bound loci. (3) Target-specific profiles are distinguished by barcode identity in sequencing. b, Genome browser screenshot showing individual CUT&Tag profiles for H3K27me3 (first row) and RNA PolIIS5P (second row) in comparison with MulTI-Tag profiles for the same targets probed individually in different cells (third and fourth rows) or sequentially in the same cells (fifth and sixth rows). c, Heat maps describing the enrichment of H3K27me3 (red) or RNA PolIIS5P (blue) signal from sequential MulTI-Tag profiles at CUT&Tag-defined H3K27me3 peaks (left) or RNA PolIIS5P peaks (right). d, Genome browser screenshot showing H3K27me3 (red), H3K4me2 (purple) and H3K36me3 (teal) MulTI-Tag signal from experiments in H1 hESCs using an individual antibody (rows 1, 3 and 5) or all three antibodies in sequence (rows 2, 4 and 6). e, Normalized CUT&Tag (light colors) and MulTI-Tag (dark colors) enrichment of H3K27me3, H3K4me2 and H3K36me3 across genes in H1 hESCs. RPM, reads per million.

Fig. 2. MulTI-Tag in single cells.

a, Schematic describing single-cell MulTI-Tag experiments. H1 hESCs (fuschia) and K562 cells (gold) were profiled separately or in a mixture of the two cell types in bulk, and then cells were dispensed into nanowells on a Takara ICELL8 microfluidic device for combinatorial barcoding via amplification. b, Genome browser screenshot showing aggregated single-cell MulTI-Tag data (rows 2, 4, 6 and 8) in comparison with ENCODE ChIP-seq data (rows 1, 3, 5 and 7) profiling H3K27me3 (rows 1, 2, 5 and 6) and H3K36me3 (rows 3, 4, 7 and 8) in K562 (rows 1–4) and H1 (rows 5–8) cells. All single-cell MulTI-Tag data are from cells co-profiled with H3K27me3 and H3K36me3. c, Connected UMAP plots for single-cell MulTI-Tag data from H1 and K562 cells. Projections based on H3K27me3 (left), H3K36me3 (right) or a WNN integration of H3K27me3 and H3K36me3 data (center) are shown. NMI of cell type cluster accuracy is denoted for each projection. Lines are connected between points that represent the same single cell in different projections. d, WNN UMAP projections with MulTI-Tag enrichment scores plotted for POLR3E (top left), HOXD3 (top right), HOXB3 (bottom left) and SALL4 (bottom right). The balance of enrichment between H3K36me3 and H3K27me3 in each cell is denoted by color, and the total normalized counts in each cell are denoted by the transparency shading.

Fig. 3. Coordinated multifactorial analysis in the same cells using MulTI-Tag.

a, Schematic describing a three-antibody MulTI-Tag experiment. b, Connected UMAP plots for single-cell MulTI-Tag data from H1 and K562 cells. Projections based on H3K27me3 (top), H3K4me2 (left), H3K36me3 (right) or a WNN integration of H3K27me3, H3K4me2 and H3K36me3 data (center) are shown. Lines are connected between points that represent the same single cell in different projections. c, Violin plots describing the distribution of the proportions of MulTI-Tag H3K27me3 (red), H3K4me2 (purple) or H3K36me3 (teal) unique reads out of total unique reads in individual H1 (left) or K562 (right) cells. d, Schematic describing coordinated multifactorial analysis strategy for MulTI-Tag. Genes in individual cells are analyzed for the enrichment of all MulTI-tag targets, and gene–cell target combinations are mapped onto a matrix for clustering and further analysis. e, Top: heat map describing co-occurrence of MulTI-tag targets in six genes of interest in each of 373 H1 cells and 372 K562 cells. The balance of enrichment between H3K4me2/H3K36me3 and H3K27me3 in each cell is denoted by color, and the total normalized counts in each cell are denoted by the transparency shading. Bottom: Instances of ‘bivalent’ enrichment of H3K27me3 and H3K4me2 or H3K36me3 in the same gene in the same cell are highlighted, with color reflecting normalized counts. f, WNN UMAP projection with cells colored by the sum of all counts occurring in a ‘bivalent’ context (that is, H3K27me3 and H3K4me2/H3K36me3 enrichment in the same gene). g, Violin plots describing calculated Cramér’s V of association between target combinations listed at bottom in individual H1 (fuschia, n = 373) or K562 (gold, n = 372) cells.

Fig. 4. MulTI-Tag profiling of continuous developmental trajectories.

a, Schematic describing differentiation of H1 hESCs (black) into three germ layers—Ectoderm (blue shading), Endoderm (red shading) and Mesoderm (green shading)—followed by MulTI-Tag profiling of H3K27me3, H3K4me1 and H3K36me3. b, Connected UMAP plots for single-cell MulTI-Tag data from H1 hESCs differentiated to three germ layers. Projections based on H3K27me3 (left), H3K36me3 (right) or a WNN integration of H3K27me3 and H3K36me3 data (center) are shown. Lines are connected between points that represent the same single cell in different projections. c, Violin plot showing the distribution of inferred pseudotimes derived from a WNN integration of H3K27me3 and H3K4me1 data for each cell type profiled. Number of cells profiled for each cell type is denoted at left. d, WNN UMAP projection colored by percent H3K27me3 as a proportion of total unique reads in each single cell. User-defined cell type clusters are denoted by dashed lines, and computationally derived pseudotemporal trajectories are denoted by solid lines and user-classified by color. e, Heat map describing co-occurrence of MulTI-tag targets in selected genes of interest whose RNA-seq expression increases (top) or decreases (bottom) during differentiation from hESC to mesoderm in 4,754 single cells classified as hESC or different stages of differentiated mesoderm. Heat maps are sorted left to right by increasing pseudotime in the mesendoderm/mesoderm trajectory. The balance of enrichment between H3K4me1/H3K36me3 and H3K27me3 in each cell is denoted by color, and the total normalized counts in each cell are denoted by the transparency shading. f, hESCs plotted to the WNN UMAP projection and colored by predicted H3K27me3 percent as a proportion of total unique reads (Methods). hESCs adjacent to the ectoderm trajectory or the mesendoderm trajectory are denoted by arrows. g, Heat maps denoting H3K27me3 enrichment in ‘high-H3K27me3’ and ‘low-H3K27me3’ hESCs (left); log fold change (LFC) in enrichment (center); and −log₁₀(P value) of differential enrichment (right) for select genes colored by their function in hESCs (black), mesendoderm (gray), endoderm (red), mesoderm (green) or ectoderm (blue).

Extended Data Fig. 1. Design and validation of MulTI-Tag.

a) Schematic of protocol variations tested for distinguishing CUT&Tag targets by sequencing barcode. Top: Approaches for pairing barcodes with antibodies, either by pre-incubation of barcoded pA-Tn5 with a secondary antibody (‘Pre-incubation’, left), or covalent conjugation of barcode-containing adapters to secondary (‘2° conjugate’, center) or primary (‘1° conjugate’, right) antibodies. Bottom: Approaches for tagmenting multiple targets, either in separate cells (‘Individual’, left), in the same cells simultaneously (‘Combined’, center), or in the same cells sequentially (‘Sequential’, right). b) Scatterplots describing the enrichment of H3K27me3 (X-axis) and PolIIS5P (Y-axis) in H3K27me3 (red points) or PolIIS5P (blue points) peaks for combinations of experimental conditions described in 2a. Pearson’s R² of all data points is denoted for each of the nine protocol conditions. c) Genome browser screenshot showing individual CUT&Tag profiles for H3K27me3 (first row) and RNA PolIIS5P (second) in comparison with MulTI-Tag profiles for the same targets probed individually in different cells (third and fourth rows secondary conjugate MulTI-Tag; seventh and eighth rows primary conjugate MulTI-Tag) or sequentially in the same cells (fifth and sixth rows secondary conjugate MulTI-Tag; ninth and tenth rows primary conjugate MulTI-Tag). d) Violin plot describing distribution of fraction of on-target reads in peaks, defined as the percentage of reads corresponding to the same target for which the peak was called, from CUT&Tag (columns 1 and 5), single-antibody MulTI-Tag (2 and 6), sequential MulTI-Tag with H3K27me3 tagmented first (3 and 7), or sequential MulTI-Tag with PolIIS5P tagmented first (4 and 8). All calculations are based on peaks called from H3K27me3 (red) and PolIIS5P (blue) ENCODE ChIP-seq data. e) Top: Schematic of MulTI-Tag with additional CUT&Tag step, in which 1° antibody conjugates are loaded into pA-Tn5 along with free i5 adapter (left), and secondary antibody and pA-Tn5 loaded only with i7 adapter are added before tagmentation (right). Bottom: TapeStation HSD1000 trace describing DNA size and enrichment from libraries produced from CUT&Tag (lanes 1 and 2), "standard" MulTI-Tag with conjugate-only tagmentation (3 and 4), or MulTI-Tag with a secondary CUT&Tag step as described in methods (5 and 6), targeting H3K27me3 (1, 3, and 5) or H3K36me3 (2, 4, and 6) in K562 cells. f) Boxplots describing Fraction of Reads in Peaks (FRiP) score, defined as the fraction of a single target’s total unique reads mapping to peaks called for that target, calculated for H3K27me3 (red) or PolIIS5P (blue) ENCODE ChIP-seq peaks for four biological replicates each from CUT&Tag or sequential MulTI-Tag. Chi-square test p-values are denoted above comparisons.

Extended Data Fig. 2. MulTI-Tag profiling in bulk H1 hESCs.

a) Heatmaps describing the enrichment of H3K27me3 (red), H3K4me2 (purple), or H3K36me3 (teal) signal from H1 cell MulTI-Tag profiles using single antibodies (left) or three antibodies sequentially (right) in H3K27me3 (top), H3K4me2 (middle), or H3K36me3 (bottom) peaks. b) Table describing Fraction of Reads in Peaks (FRiP) score in ENCODE ChIP-seq peaks for H3K27me3, H3K4me2, and H3K36me3 for CUT&Tag and MulTI-Tag experiments in H1 cells. c) Heatmaps describing comparative enrichment of H3K27me3 in bivalent (top) vs. non-bivalent (bottom) enriched regions in CUT&Tag (left) or MulTI-Tag (right) experiments. d) Heatmaps describing the same as c) for H3K4me2.

Extended Data Fig. 3. Combinatorial indexing for single-cell MulTI-Tag.

a) Schematic describing single cell MulTI-Tag species mixing experiments. Human K562 cells (red) and mouse NIH3T3 cells (blue) were mixed and profiled in bulk, then cells were dispensed into nanowells on a Takara ICELL8 microfluidic device for combinatorial barcoding via amplification. b) Barnyard plots describing the number of unique fragments exclusively mapping to the hg19 genome build (X-axis) vs. mm10 (Y-axis) in all cells with greater than 100 unique reads for each of the denoted experiments. Points are colored by the cell identity as human (red; > 90% of unique reads mapping to hg19), mouse (blue; >90% mapping to mm10), or mixed (magenta; < 90% mapping to either), and collision rate, defined as the percentage of cells classified as ‘mixed’, is denoted for each experiment. c) Violin plots describing distributions of unique reads per cell in K562 cells (left), H1 cells (center), or the K562-H1 cell mixed population (right). Median values for total unique reads (black), H3K27me3 unique reads (red), or H3K36me3 unique reads (teal) are displayed at the top of each violin. Number of cells described is displayed at top of each cell type group. d) Violin plot describing distribution of fraction of on-target reads in peaks, defined as the percentage of reads corresponding to the same target for which the ENCODE ChIP-seq peak was called, in H3K27me3 (red) and H3K36me3 (teal) peaks from single cell MulTI-Tag in H1 cells (left) and K562 cells (right). Number of peaks is displayed above each violin. e) Violin plots describing Fraction of Reads in Peaks (FRiP) score in ENCODE ChIP-seq peaks for H3K27me3 (red) or H3K36me3 (teal) data from single cell CUT&Tag (white) or sequential single cell MulTI-Tag (grey). Number of cells described and number of peaks used is displayed below each violin. f) Jittered scatterplot describing the number of counts mapping to each single cell within each of the indicated genes in single cell CUT&Tag (black) vs. single cell MulTI-Tag (grey). The percentage of cells with non-zero counts for each locus and assay are denoted at the bottom. g) Table describing comparative metrics for MulTI-Tag (this study) in comparison with scMulti-CUT&Tag, scCUT&Tag^,, and scChIP-seq.

Extended Data Fig. 4. MulTI-Tag across diverse target combinations.

a) Schematic describing single cell MulTI-Tag profiling different combinations of targets in the same combinatorial indexing experiment. One of four targets (PolIIS5P, H3K9me3, H3K4me1, or H3K36me3) was tagmented in sequence with H3K27me3 in bulk, then arrayed in a 96 well plate as displayed for i7 tagmentation (Methods). b) Violin plots describing distributions of unique reads per cell in K562 cells (left), H1 cells (center), or the K562-H1 cell mixed population (right) for the experiments described in a). Median values for H3K27me3 unique reads (red), PolIIS5P unique reads (blue), H3K9me3 unique reads (magenta), H3K4me1 unique reads (orange), or H3K36me3 unique reads (teal) are displayed at the top of each violin. Number of cells described for each cell type-target combination is displayed at the bottom of each violin. c) Connected UMAP plots for single cell MulTI-Tag data from experiments described in a). Projections based on H3K27me3 (center), PolIIS5P (top left), H3K9me3 (bottom left), H3K4me1 (top right), or H3K36me3 (bottom right) are shown. Total cells represented and normalized mutual information (NMI) of cell type cluster accuracy are denoted for each projection. Lines are connected between points that represent the same single cell in different projections.

Extended Data Fig. 5. Cross-target analysis in scMulTI-Tag.

a) Violin plots describing distributions of unique reads per cell in H1 cells (left) or K562 cells (right) for experiments described in Fig. 3. Median total unique reads (black), H3K27me3 unique reads (red), H3K4me2 unique reads (purple), or H3K36me3 unique reads (teal) are displayed at the top of each violin. Number of cells described is displayed at top of each cell type group. b) Heatmaps describing the enrichment of H3K27me3 (red), H3K4me2 (purple), or H3K36me3 (teal) signal from K562 cell profiles using single antibodies in bulk MulTI-Tag (left) or three antibodies sequentially in aggregate single cell MulTI-Tag (right) in H3K27me3 (top), H3K4me2 (middle), or H3K36me3 (bottom) peaks as called from bulk MulTI-Tag data. c) Heatmaps describing the same as b) for H1 hESCs. d) Violin plots describing the distribution of the fraction of on-target reads in peaks, defined as the percentage of reads corresponding to the same target for which the ENCODE ChIP-seq peak was called, in H3K27me3 (red, n = 74079), H3K4me2 (purple, n = 65388), and H3K36me3 (teal, n = 93085) peaks from bulk individual MulTI-Tag (white) vs. sequential single cell MulTI-Tag (grey) in K562 cells. e) Violin plots describing the same as d) for H1 hESCs (H3K27me3 n = 39290, H3K4me2 n = 119250, H3K36me3 n = 198078). f) Violin plots describing the distributions of proportions of each co-occurrence state as described below the plot in individual H1 (fuschia, n = 373) or K562 (gold, n = 372) cells, with points denoting individual cell values. The last four co-occurrence states are rescaled and inset at top right; p-values derived from two-sided student’s t-test comparing distributions between cell types are listed above violins (not corrected for multiple hypothesis testing).

Extended Data Fig. 6. Verification of H3K27me3-H3K36me3 co-enrichment in MulTI-Tag.

a) Genome browser screenshot showing H3K27me3 (red) and H3K36me3 (teal) enrichment from ENCODE ChIP-seq (rows 1, 2, 5, and 6) or bulk MulTI-Tag (rows 3, 4, 7, and 8) in K562 cells (rows 1-4) or H1 hESCs (rows 5-8) at the PCSK9 gene. Colored boxes indicate co-enrichment of H3K27me3 and H3K36me3 in the same gene in H1 hESCs. b) Heatmaps describing the enrichment of H3K27me3 (red) and H3K36me3 (teal) signal from ENCODE ChIP-seq (left) or bulk MulTI-Tag (right) in H1 hESCs in 86 genes for which 1) a MulTI-Tag H3K27me3 peak overlapped a 2 kb window surrounding the TSS, and 2) a MulTI-Tag H3K36me3 peak overlapped the gene body. Selected genes of interest, including those involved in metabolic and developmental signaling, are highlighted at right. c) Violin plots describing the number of normalized counts for H3K27me3 (red) and H3K36me3 (teal) mapping to the top 100 genes as classified by the percentage of single H1 hESCs enriched with H3K27me3 (left), H3K36me3 (right), or co-enriched for H3K27me3 and H3K36me3 (center) in the genes in question. ENCODE ChIP-seq (white), CUT&Tag (light grey), bulk MulTI-Tag (medium grey) and aggregate single cell MulTI-Tag (dark grey) counts are displayed for each category. P-values derived from student’s t-tests are listed above violins. d) Violin plots describing ENCODE RNA-seq counts mapping to the top 100 genes as classified by the percentage of single H1 hESCs enriched with H3K27me3 (left), H3K36me3 (right), or co-enriched for H3K27me3 and H3K36me3 (center) in the genes in question.

Extended Data Fig. 7. Clusters derived from scMulTI-Tag in hESC trilineage differentiation.

a) Violin plots describing distributions of unique reads per cell in H1 hESCs (left), endoderm (center-left), mesoderm (center-right), or ectoderm (right) for all cells with at least 100 unique reads originating from each of the three targets used in the experiments described in Fig. 4. Median values for total unique reads (black), H3K27me3 unique reads (red), H3K4me1 unique reads (orange) or H3K36me3 unique reads (teal) are displayed at the top of each violin. Number of cells described is displayed at top of each cell type group. b) UMAP plot for single cell MulTI-Tag data from projection of H3K36me3 data, with cells colored by Seurat cluster (left) or cell type (right). c) UMAP plots for single cell MulTI-Tag data from projection of H3K27me3 data (center), H3K4me1 data (right), or a weighted nearest neighbor integration of H3K27me3 and H3K4me1 data (left). Cells are colored by Seurat clusters. For each plot, four groups of representative clusters are highlighted with quadrants describing the fraction of H1 (top left), ectoderm (top right), endoderm (bottom left), or mesoderm (bottom right) cells contained in the highlighted clusters as a proportion of the total cells from each cell type contained in the experiment. Quadrants are colored based on the proportion of the maximum value in the quadrant.

Extended Data Fig. 8. Pseudotemporal trajectories derived from scMulTI-Tag in hESC trilineage differentiation.

a) UMAP plot for single cell MulTI-Tag data from projection of H3K27me3 data, with monocle3-derived pseudotemporal trajectories overlaid. Cells are colored by inferred pseudotime. b) UMAP plot describing the same as b) for H3K4me1 data. c) UMAP plot describing the same as a) and b) for a weighted nearest neighbor integration of H3K27me3 and H3K4me1 data. d) monocle3-derived pseudotemporal trajectories for H3K27me3 data, colored by manual annotation of likely correspondence to known differentiation trajectories. e) Same as d) for H3K4me1 data. f) same as d) and e) for a weighted nearest neighbor integration of H3K27me3 and H3K4me1 data. g) Violin plots showing the distribution of inferred pseudotimes derived from H3K27me3 (left) or H3K4me1 (right) data for each cell type profiled. Number of cells profiled for each cell type is denoted at left. h) Pseudotime-ordered heatmaps describing the cell types of the cells assigned to each manually curated trajectory derived from different MulTI-Tag data. Data used to derive each trajectory is displayed at left. For each trajectory, cells are colored by color intensity based on the real assayed differentiation time ranging from hESC (black) to the terminal cell type (mesoderm = green; endoderm = red; ectoderm = blue). Cells assigned to the inferred trajectory that belong to a different trajectory (‘incorrect’) are colored white. For each trajectory-data source combination, inversion rate, defined as the fraction of cell pairs in the trajectory for which the real differentiation time is out of order, and incorrect rate, defined as the fraction of cells assigned to an incorrect trajectory, are displayed at right.

Extended Data Fig. 9. Analysis of scMulTI-Tag changes across pseudotime.

a) Violin plots describing H3K27me3 (red), H3K4me1 (orange), and H3K36me3 (teal) single cell MulTI-Tag enrichment in genes that decline in expression as defined by RNA-seq during differentiation from hESCs to mesoderm (top, n = 29), endoderm (middle, n = 20), or ectoderm (bottom, n = 19). Enrichment is partitioned by pseudotime quartile (1=lowest, 4=highest). P-values of Wilcoxon Rank Sum test between quartile 1 and all other quartiles for each target are displayed above violins. b) Violin plots describing same as a) for genes that increase in expression as defined by RNA-seq . Mesoderm n = 54, Endoderm n = 35, Ectoderm n = 36. P-values less than 0.05 are highlighted in red. c) Heatmaps describing co-occurrence of MulTI-tag targets in selected genes of interest whose RNA-seq expression increases (top) or decreases (bottom) during differentiation from hESC to endoderm in 3626 single cells classified as hESC or different stages of differentiated mesoderm. Heatmaps are sorted left-to-right by increasing pseudotime in the mesendoderm/endoderm trajectory. The balance of enrichment between H3K4me1/H3K36me3 and H3K27me3 in each cell is denoted by color, and the total normalized counts in each cell are denoted by the transparency shading. d) Same as c) for ectoderm trajectory.

Extended Data Fig. 10. Trajectory-specific H3K27me3 dynamics uncovered by MulTI-Tag.

a) Scatterplot showing single cells plotted by increasing pseudotime on the X-axis, increasing fraction of H3K27me3 as a proportion of total unique reads on the Y-axis, and colored by trajectory to which they belong (Ectoderm = blue, Mesendoderm = grey, Mesoderm.= green, Endoderm = red). LOESS smoothing curves describing average results for each trajectory are overlaid. b) Violin plots describing the distribution of the proportions of MulTI-Tag H3K27me3 (red), H3K4me1 (orange), or H3K36me3 (teal) unique reads out of total unique reads in individual H1 hESC (left, n = 1750) endoderm (center-left, n = 1167), mesoderm (center-right, n = 1693), or ectoderm (right, n = 485) cells. c) Volcano plot showing all human transcription factors plotted by log fold change in H3K27me3 MulTI-Tag normalized enrichment between ‘H3K27me3-low’ and ‘H3K27me3-high’ H1 hESCs on the X-axis, and negative log10 Wilcoxon Rank-Sum p-value of the comparisons on the y-axis. Genes for which the total normalized counts are greater than 20 and the p-value is less than 0.05 are highlighted in red. d) Genome browser shots showing aggregate H3K27me3 MulTI-Tag enrichment in ‘H3K27me3-high’ (red) and ‘H3K27me3-low’ (dark red) cells at the HOXA (left) and TBXT (T, right) loci. e) Gene Ontology analysis of transcription factors with a statistically significant reduction in H3K27me3 in ‘H3K27me3-low’ hESCs as compared to all human TFs, with p-values, length of bars and reported values at right of bars corresponding to negative Log10(p-value) for each category displayed. Bars are colored by FDR and p-value thresholds as denoted. f) Violin plots describing calculated Cramér’s V of association between H3K27m3 and H3K4me1 in individual H3K27me3-high hESCs (black), H3K27me3-low hESCs (grey), endoderm (red), mesoderm (green), and ectoderm cells. Wilcoxon Rank-Sum p-values of comparisons between ‘H3K27me3-high’ hESCs and other cell types are displayed at top. P-values less than 0.05 are highlighted in red.