Spatial-CUT&Tag: Spatially Resolved Chromatin Modification Profiling at Cellular Level

Abstract

Spatial omics emerged as a new frontier of biological and biomedical research. Here we present spatial-CUT&Tag for spatially resolved genome-wide profiling of histone modifications by combining in situ CUT&Tag chemistry, microfluidic deterministic barcoding, and next generation sequencing. Spatially resolved chromatin states in mouse embryos revealed tissue type-specific epigenetic regulations in concordance with ENCODE references and provide spatial information at tissue scale. Spatial-CUT&Tag revealed epigenetic control of the cortical layer development and spatial patterning of cell types determined by histone modification in mouse brain. Single-cell epigenomes can be derived in situ by identifying 20μm pixels containing only one nucleus using immunofluorescence imaging. Spatial chromatin modification profiling in tissue may offer new opportunities to study epigenetic regulation, cell function and fate decision in normal physiology and pathogenesis.

Chromatin state determines genome function and is regulated in a cell type-specific manner (1, 2). Despite recent breakthroughs in single-cell sequencing (3-5) that has enabled the profiling of epigenome in single cells (6-11), it remains challenging to integrate its spatial information of individual cells in the tissue of origin. Here, we report spatial-CUT&Tag for spatial histone modification profiling, which combines in tissue deterministic barcoding (12, 13) with the cleavage under targets and tagmentation (CUT&Tag) chemistry (14, 15) (Fig. 1A and Fig. S1). First, antibody against the target histone modification was added to a fixed tissue section, followed by a secondary antibody binding to enhance the tethering of pA-Tn5 transposome. After activating the transposome, adapters containing a ligation linker were inserted to genomic DNA at the histone mark antibody recognition sites. Then, a set of DNA barcode solutions Ai (i = 1-50) were flowed over the tissue surface via microchannel-guided delivery to perform in situ ligation to the adapters. Afterwards, a second set of barcodes B_j (j = 1-50) were flowed over the same tissue surface through microchannels perpendicular to those in the first barcoding step. These barcodes were then ligated at the intersections, resulting in a two-dimensional (2D) grid of tissue pixels, each of which contains a distinct combination of barcodes A_i and B_j (i = 1-50, j = 1-50). The tissue section was imaged after these steps to correlate the tissue morphology with the spatial epigenomics map. Finally, DNA fragments were collected by cross-link reversal to complete library construction (see Supplemental Materials and Methods).

Fig. 1. Spatial-CUT&Tag: design and validation.

(A) Schematic workflow. (B-D) Comparison of unique fragments, FRiP, and mitochondrial reads between our spatial method and other non-spatial methods. (E) H&E image from an adjacent tissue section of E11 mouse embryo and a region of interest for spatial mapping. (F) Spatial distribution and (G) UMAP of unsupervised clustering analysis. (H) Projection of ENCODE ChIP-seq data onto the spatial-CUT&Tag embedding.

We performed spatial-CUT&Tag with antibodies against H3K27me3 (repressing loci), H3K4me3 (activating promoters) and H3K27ac (activating enhancers and/or promoters) in E11 mouse embryos. With 50 μm pixel size, we obtained a median of 9,788 (H3K27me3), 16,777 (H3K4me3), or 19,721 (H3K27ac) unique fragments per pixel of which 16% (H3K27me3), 67% (H3K4me3), or 16% (H3K27ac) of fragments fell within peak regions, indicating high coverage and low background (as reference, fraction of reads in peaks or FRiP of bulk CUT&Tag of E11 mouse embryo with H3K27me3 was ~24%) (Fig. 1B&C). In addition, the proportion of mitochondrial fragments is low (a median of 0.16% (H3K27me3), 0.13% (H3K4me3), or 0.01% (H3K27ac) of fragments was from mitochondrial reads) (Fig. 1D). With 20 μm pixel size (cellular level), we obtained a median of 10,064 (H3K27me3), 7,310 (H3K4me3), or 13,171 (H3K27ac) unique fragments per pixel of which 20% (H3K27me3), 37% (H3K4me3), or 12% (H3K27ac) of fragments fell within peak regions (Fig. 1B&C). The fractions of read-pairs mapping to mitochondria are 0.01% (H3K27me3), 0.02% (H3K4me3), or 0% (H3K27ac) (Fig. 1D). Additionally, the fragment length distribution was consistent with the capture of nucleosomal and subnucleosomal fragments (the subnucleosomal fragments may represent background signal from untethered Tn5) (Fig. S2). To measure the extent of tagmentation by free Tn5, we compared the spatial-CUT&Tag H3K27me3 signals to reference ChIP-seq and ATAC-seq (2). The results showed that around ~11.5% of peaks that did not overlap with ChIP-seq were observed in ATAC-seq (Fig. S3), which may correspond to the Tn5 insertion events unrelated to the histone mark (16).

We compared spatial-CUT&Tag at 20 μm pixel size to published scCUT&Tag also on mouse brain at the same sequencing depth (8) and found that spatial-CUT&Tag detected more unique fragments (H3K27me3: 9,735, H3K4me3: 3,686) than scCUT&Tag (H3K27me3: 682, H3K4me3: 453) (Fig. 1B). We also isolated tissue pixels containing single nuclei that showed similar unique fragment counts compared to scCUT&Tag. However, the FRiP from spatial-CUT&Tag (H3K27me3: 10%, H3K4me3: 53%) is lower than the scCUT&Tag (H3K27me3: 24%, H3K4me3: 82%) (Fig. 1C), presumably due to the use of frozen sections that has been reported to affect chromatin structures and generate higher background noise (17).

The Pearson correlation coefficient from different spatial-CUT&Tag experiments was above 0.95 (Fig. S4, A to F), demonstrating a high degree of reproductivity. Spatial-CUT&Tag also reproduced the chromatin state pattern and peaks (Fig. S4, G to I). Furthermore, the peaks called from spatial-CUT&Tag aggregate data were consistent with the ENCODE ChIP-seq data (Fig. S5A). Spatial-CUT&Tag also yielded high-quality profiles in the liver comparable to the reference data (Fig. S5B).

Identification of cell types de novo by chromatin states was conducted. Mapping the clusters back to spatial locations identified spatially distinct patterns that agreed with the tissue histology in a hematoxylin and eosin (H&E)-stained adjacent tissue section (Fig. 1 E to G, Fig. S6). Cluster 1 of H3K27me3 and cluster 6 of H3K4me3 correspond to the heart. Cluster 2 of H3K27me3 and H3K4me3 and cluster 4 of H3K27ac correspond to the liver region. Cluster 8 of H3K27me3), cluster 3 (H3K4me3) and cluster 1 (H3K27ac) are associated with the forebrain, while cluster 9 (H3K27me3), cluster 5 (H3K4me3) and cluster 3 (H3K27ac) with the brainstem, including midbrain. Cluster 11 (H3K27me3), cluster 8 (H3K4me3) and cluster 2 (H3K27ac) are present in more posterior regions of the central nervous system (CNS) like the spinal cord.

To benchmark spatial-CUT&Tag, we projected the organ-specific ChIP-seq data onto our UMAP embedding (2, 18). Overall, cluster identification matched well with the ChIP-seq projection (Fig. 1, G and H) and distinguished major cell types in E11 mouse embryo. To identify spatial patterning during development, we examined cell type-specific marker genes. For H3K27me3, chromatin silencing score (CSS) was calculated to predict the gene expression (9), and active genes should have a low CSS due to the lack of H3K27me3 repressive mark (Fig. 2A and Fig. S7A). For example, Hand2, which is required for vascular development and plays an essential role in cardiac morphogenesis (19), showed a lack of H3K27me3 enrichment in the heart (C1). For H3K4me3 and H3K27ac, gene activity score (GAS) was used since they are related to active genes (Fig. 2B, Fig. S8A and Fig. S10A). For example, Nfe2, which are essential for regulating erythroid and hematopoietic cell maturation and differentiation (19), were active in liver and to some extent in the heart (C2 and C6 for H3K4me3). We further analyzed Gene Ontology (GO) pathways and the results agreed with the anatomical annotation (Fig. S7B, S8B, and S10B). To understand which regulatory factors are most active across clusters, we calculated transcription factor (TF) motif enrichments in H3K4me3 and H3K27ac modification loci (Fig. S9 and S11). As expected, the most enriched motifs in liver correspond to GATA transcription factors including Gata2, while Mef2a was enriched in the heart region. To predict gene regulatory interactions of enhancers and their target genes across clusters, we correlated gene expression from scRNA-seq (5) and H3K27ac modifications at the candidate enhancers (Fig. 2C). The correlation-based map successfully predicted some enhancer-gene interactions that have been indeed experimentally validated. For example, the predicted enhancers of Ascl1 and Kcnq3 were enriched in the CNS (C1,C2, C3 and C6), which are in agreement with the VISTA validated elements (20).

Fig. 2. Spatial mapping and integrative analysis of mouse embryos.

(A and B) Genome browser tracks (left) and spatial mapping (right) of gene silencing by H3K27me3 and gene activity by H3K4me3 modification. (C) Predicted enhancers of Ascl1 (chr10: 87,463,659–87,513,660; mm10) (left) and Kcnq3 (chr15: 66,231,223–66,331,224; mm10) (right) from H3K27ac profiling. Cluster of each track corresponds to Fig.1F. Enhancers validated by in vivo reporter assays are shown between main panels. (D and F) Integration of scRNA-seq (5) and spatial-CUT&Tag data. (E and G) Spatial mapping of selected cell types identified by label transferring. (H) List of cell types in scRNA-seq. (I) Refined clustering of radial glial enabled identification of sub-populations. Scale bar, 1 mm.

We then integrated scRNA-seq data (5) with spatial-CUT&Tag data to identify cell types (Fig. 2, D to H, Fig. S12). Spatial tissue pixels were found to conform well into the clusters of single-cell transcriptomes. Several organ-specific cell types were detected (Fig. 2, E and G). For example, the definitive erythroid lineage cells were exclusively enriched in the liver, which is the major hematopoietic organ at day 11.5 (19). Cardiac muscle cell types were observed only in the heart region in agreement with the anatomical annotation. Chondrocytes & osteoblasts were observed widely in the embryonic facial prominence. Inhibitory interneurons were highly enriched in the brain stem. Postmitotic premature neurons were observed extensively in the spinal cord region. A high-resolution clustering further identified sub-populations of developing neurons with distinct spatial distributions and chromatin states (Fig. 2I, Fig. S12B). For instance, the H3K27ac radial glia could be further divided to three clusters. Genes related to stem cell maintenance in the CNS (e.g. Sox1) had higher GAS in subcluster 2, which was enriched along the ventricles in the developing brain stem and spinal cord. In contrast, subcluster 3 were in the spinal cord parenchyma, while subcluster 1 were mainly outside the CNS, and thus might represent neural crest progenitors (e.g. active Sox10) (Fig. 2I). Additionally, two subclusters were found in the chondrocytes & osteoblasts, and the genes related to developing teeth (e.g. Barx1) had higher GAS in subcluster 2 (Fig. S12B).

The data were further used to examine the developmental process from radial glia to excitatory neurons via postmitotic premature neurons as the immediate state and the cells were ordered in pseudo-time. Spatial projection of each pixel’s pseudo-time revealed the spatially organized developmental trajectory (Fig. S13). Cells early in differentiation were enriched around the ventricles in the developing brainstem whereas those farther away exhibited a more differentiated phenotype (21) (Fig. S13B). We then identified changes in gene activity based on H3K4me3 across this developmental process, and many genes recovered are important in neuron development, including Pou4f1and Car10 (19, 22) (Fig. S13, C and D).

We next combined spatial-CUT&Tag with immunofluorescence staining in the same tissue section (Fig. 3). We stained an olfactory bulb (OB) with 4’,6-diamidino-2-phenylindole (DAPI) for nuclear DNA (Fig. 3, A and B) and next performed spatial-CUT&Tag against H3K27me3 with 20 μm pixel size, which distinguished the major cell types, including glomerular (C1) and granular layer (C2) (Fig. 3). Spatial patterns of H3K27me3 modification were validated by in situ hybridization (Fig. 3D). With DAPI staining for nucleus, we could select the pixels of interest such as those containing single nucleus or those showing specific histone modifications, allowing for extracting single-cell epigenome data without tissue dissociation (Fig. 3, E to I).

Fig. 3. Spatial mapping of an immunofluorescence-stained mouse olfactory bulb tissue section.

(A) H&E image from an adjacent tissue section and a region of interest for spatial mapping. (B) Fluorescent image of nuclear staining with DAPI. (C) Unsupervised clustering analysis and spatial distribution of each cluster (20 μm pixel size). (D) Spatial mapping (left) of H3K27me3 modification for selected marker genes. In situ hybridization (right) and expression images (middle) of corresponding genes are from the Allen Mouse Brain Atlas. (E) Fluorescent images of selected pixels containing single nucleus (DAPI). (F) Heatmap of chromatin silencing score of selected pixels. (G) Comparison of unique fragments in pixels with non-single nucleus (>1 nucleus) and single nucleus. (H) Unsupervised clustering of selected pixels containing single nuclei, and (I) colored by CSS for selected genes.

We also conducted spatial-CUT&Tag with 20 μm pixel size to analyze the brain region of an E11 mouse embryo (Fig. S14) and observed distinct spatial patterns. H3K27me3 yielded most spatial clusters (Fig. S14, A to C). Clusters identified agreed with the projection of ENCODE organ-specific bulk ChIP-seq data in the UMAP embedding (Fig. S14, C and D). We further surveyed H3K27me3 modifications and observed distinct modification patterns across clusters (Fig. S14E and Fig. S15A). Cfap77 was repressed extensively except in a portion of the forebrain. Six1, which is involved in limb development, had low CSS in Cluster 5. Although Sfta3-ps and Rhcg lack H3K27me3 enrichment in the forebrain, they had distinct spatial patterns. Pathway analysis of marker genes revealed that cluster 1 was involved in the forebrain development, cluster 2 corresponded to anterior/posterior pattern specification, and cluster 4 was associated with heart morphogenesis, all in agreement with anatomical annotations (Fig. S14A and Fig. S15B). Next, we sought to improve the clustering resolution by integrating data across H3K4me3 and H3K27ac histone marks at gene resolution. The granularity of two-dimensional representation of the data obtained from the integrated analysis was further improved (Fig. S14F). To assign cell types to each cluster, we integrated the spatial-CUT&Tag data (H3K4me3 and H3K27ac) with the mouse embryo cell atlas from scRNA-seq (5) (Fig. S14, G to K). For example, chondrocytes & osteoblasts were mainly in the embryonic facial prominence, and radial glia and inhibitory neuron progenitors were observed in the forebrain (Fig. S14, H and J). Although H3K4me3 and H3K27ac had fewer clusters than H3K27me3 at the 20 μm resolution,we found that the clusters that appeared to be homogenous could be further deconvoluted into sub-populations.

Finally, spatial-CUT&Tag with 20 μm pixel size was applied to the P21 mouse brains, and unsupervised clustering revealed distinct spatial features (Fig. 4, A to C). We set out to explore the spatial patterns of specific marker genes to distinguish cell types and compared them to the gene expression pattern in the single-cell transcriptomic atlas (23) (Fig. 4, D and E, Fig. S16 and Fig. S17). For example, Sox10 showed high GAS in cluster 2 of H3K4me3, and Itpr2 had low CSS in cluster 6 of H3K27me3, indicating these clusters were enriched with oligodendrocyte lineage cells. Cells of these clusters were particularly enriched in a stripe-like structure that corresponds to the corpus callosum (Fig. 4, D and E). For cluster 3 of H3K4me3 and H3K27me3, Adcy5 was activated and Rbms3 was repressed, suggesting medium spiny neurons were enriched in these clusters. Some clusters that appeared to be homogenous could be further deconvoluted into sub-populations with distinct spatial distributions (Fig. 4F). For example, cluster 2 of H3K27me3 could be further divided into two clusters. Cux2, a marker of the superficial cortical layers 2 and 3, had lower H3K27me3 signal in subcluster 1. In contrast, Blc11b, a marker of the deeper cortical layers 4-6 presented higher H3K27me3. While Polycomb has been previously shown to play a role in the establishment of cortical layers at embryonic stages (24, 25), our data suggests that H3K27me3 is also involved in maintaining cortical layer identities at postnatal stages. To examine the interplay between active and repressive marks and infer the potential H3K4me3/H3K27me3 bivalency, we identified all active promoters specific to individual populations marked by H3K4me3 and plotted the signals of H3K4me3 and H3K27me3 (Fig. S18). As expected, H3K27me3 signals were depleted when the promoter is enriched in H3K4me3 in the respective population. However, H3K27me3 signals were also observed around few marker genes in oligodendrocytes and medium spiny neurons.

Fig. 4. Spatial mapping and integrative analysis of mouse brain.

(A) Image of mouse brain tissue section and the region of interest for spatial mapping. (B and C) Unsupervised clustering analysis and spatial distribution of each cluster (20 μm pixel size). (D and E) Spatial mapping of gene activity by H3K4me3 and gene silencing by H3K27me3 modification for selected marker genes. (F) Refined clustering identified sub-populations in neurons with distinct spatial distributions and marker genes. (G and H) Integration of scCUT&Tag (8), scRNA-seq (23) and spatial-CUT&Tag data. (I) List of cell types in scRNA-seq. (J) Spatial mapping of selected cell types identified by label transferring. Scale bar, 500 μm.

To further identify cell types, we integrated spatial-CUT&Tag data with the published scCUT&Tag (8) and scRNA-seq dataset (23). It revealed that microglia, mature oligodendrocytes, medium spiny neurons, astrocytes, and excitatory neurons were enriched in cluster 1, 2, 3, 4, and 7 respectively in the H3K4me3 dataset, and furthermore sub-populations of neurons could be identified (Fig. 4, G to J, Fig. S16 and S17). Moreover, the integration of spatial-CUT&Tag with scRNA-seq or scCUT&Tag could allow for predicting which region a specific cell type in scRNA-Seq or scCUT&Tag is localized in (Fig. 4 G to J). We identified that mature oligodendrocytes (MOL1) are abundant in the corpus callosum, while medial spinal neurons (MSN2) are present in the striatum, and TEGLU3 excitatory neurons in deeper cortical layer 6, in agreement with previous results (23), but determined herein by epigenetic modification states. TEGLU8 excitatory neurons have been shown to populate cortical layer 4 (23), and indeed we observed that the corresponding epigenetic state of this neuronal population is distributed in a more superficial cortical layer than TEGLU3 (Fig. 4J). We found that a subpopulation of non-activated microglia (MGL1) mainly populates the striatum, but not the corpus callosum or cortex. In addition, the epigenetic state associated with protoplasmic astrocytes (ACTE2) is mainly localized in the corpus callosum although also observed in the cortex and striatum at lower frequencies.

Our study demonstrated the profiling of chromatin states in situ in tissue sections with high spatial resolution. Although spatial-CUT&Tag focused on the tissue mapping of histone modifications, integration with other assays such as transcriptome and proteins is feasible with our microfluidic in tissue barcoding approach by combining reagents for DBiT-seq (12) and spatial-CUT&Tag to achieve spatial multi-omics profiling. Moreover, the mapping area could be further increased by using a serpentine microfluidic channel or increasing the number of barcodes (e.g. 100 × 100). spatial-CUT&Tag is an NGS-based approach, which is unbiased for genome-wide mapping of epigenetic mechanisms in the tissue context.

One Sentence Summary.

Spatial-CUT&Tag enables genome-wide tissue mapping of chromatin modification states at cellular level.

Data and materials availability

The sequencing data reported in this paper are deposited in the Gene Expression Omnibus (GEO) with accession code GSE165217. Code for sequencing data analysis is available on Zenodo (26).