Tri-omic single cell mapping of 3D epigenome and transcriptome in whole mouse brains throughout lifespan

Abstract

Exploring the genomic basis of transcriptional programs has been a longstanding research focus. Here, we report a single-cell method, ChAIR, to map chromatin accessibility, chromatin interactions, and RNA expression simultaneously. After validating in cultured cells, we applied ChAIR to whole mouse brains and delineated the concerted dynamics of epigenome, 3D genome, and transcriptome during maturation and ageing. Particularly, gene-centric chromatin interactions and open chromatin states provided 3D epigenomic mechanism underlying cell type-specific transcription and revealed spatially-resolved specificity. Importantly, the composition of short-range and ultra-long chromatin contacts in individual cells is remarkably correlated with transcriptional activity, open chromatin state, and genome folding density. This genomic property, along with associated cellular properties, differs in neurons and non-neuronal cells across different anatomic regions throughout lifespan, implying divergent nuclear mechano-genomic mechanisms at play in brain cells. Our results demonstrate ChAIR’s robustness in revealing single-cell 3D epigenomic states of cell type-specific transcription in complex tissues.

Introduction

It has been viewed that the three-dimensional (3D) genome folding, along with epigenomic states, plays a key role in shaping genome functions including transcription^1-3. However, much of the existing knowledge has been gleaned largely from bulk-cell data, which reveals essential but averaged views. Recent developments in single-cell technologies, especially multi-omic approaches, have significantly advanced the exploration of 3D genome, epigenome, transcriptome, and their intricate interplays. However, current single-cell multi-omic approaches face several limitations. For example, methods^4-6 combining scRNA-seq with scATAC-seq cannot capture chromatin connectivity, while others^7-10 that combine scRNA-seq with scHi-C lack the ability to identify cis-regulatory elements (CREs) and struggle to map enhancer-promoter interactions due to low resolution and high background noise. Therefore, it is highly desirable to develop new approaches capable of simultaneously detecting the specific interplays between regulatory elements, which are essential for understanding 3D epigenomic mechanisms of transcription regulation.

In this study, we report ChAIR, a droplet-based high-throughput, single-cell tri-omic approach that simultaneously maps transcriptome, epigenome, and 3D genome. We validated the ChAIR protocol in K562 (a human myelogenous leukemia cell line) and Patski (a hybrid mouse fibroblast cell line) cells and applied ChAIR to mouse brain cells throughout lifespan, tracking the concurrent dynamics of transcription and remodeling of 3D chromatin folding and epigenomic states during differentiation, maturation, and ageing. Our analyses provided evidence at single-cell level supporting a temporal, sequential 3D epigenomic mechanism that modulates transcription activation. We also discovered that the proportion of ultra-long-range chromatin megacontacts reflects key genomic properties, including genome folding density, transcriptional activity, and are highly associated with nuclear volume and cellular specificity, which suggests a potential mechano-genomic mechanism in shaping nuclear morphology, cellular activity, and brain cell functions.

RESULTS

ChAIR captures tri-omic data in single cells

To simultaneously profile chromatin interactions and epigenomic states, we previously established a bulk-cell multi-omic method ChIATAC¹¹. It efficiently maps chromatin accessibility and interactions between open chromatin loci while preserving higher-order chromatin folding architectures in same samples. By adopting ChIATAC to a microfluidic system, we developed a protocol to capture Chromatin Accessibility, Interaction, and RNA profiles (ChAIR) simultaneously in single cells. In ChAIR, crosslinked cells are manipulated through in situ restriction digestion, proximity ligation, and Tn5 tagmentation. The processed cells are then individually encapsulated into droplets using the 10x Genomics Multiome system, which barcodes DNA and RNA molecules simultaneously. The corresponding DNA and RNA information is obtained through DNA sequencing (Fig. 1a). Each ChAIR experiment in one microfluidic channel can generate single-cell datasets from 6,000–10,000 cells, and at full capacity, a 16-channel device can generate datasets from more than 100,000 cells. Each dataset offers tri-omic information: ChAIR-RNA for gene expression, ChAIR-ATAC for chromatin accessibility, and ChAIR-PET (paired-end tags) for chromatin contacts.

Fig. 1: ChAIR captures comprehensive 3D epigenome and transcriptome landscapes in single cells.

a, Schematics of the ChAIR procedure to generate ChAIR-RNA, ChAIR-ATAC and ChAIR-PET datasets in single cells. E, enhancer; P, promoter. b, A browser-based visualization of K562 ChAIR single-cell tracks, ensemble ChAIR data and bulk-cell ChIATAC and RNA-seq data. The genes in green are oriented in the 5′–3′ direction, while the genes in blue are in the 3′–5′ direction. The top 800 cells of single-cell data with the most chromatin contacts and also with all three modalities available within the displayed region are visualized. The cells were sorted in order based on the the hierarchical clustering of the ChAIR-PET data. The color assigned to each cluster was for distinguishing nearby clusters. c, The PCA (PC1 and PC2) plots of Patski ChAIR-RNA data showing cell-cycle specificity (left) and the 42 metacells ordered by cell-cycle pseudotime (right). d, A characterization of S-specific (top) and G2/M-specific genes (bottom) in Patski cells by normalized ChAIR-RNA data, ChAIR-ATAC at promoters and ChAIR-PET associated with promoters. e, Example views of 3D models of genome folding architectures, highlighting chromatin compartments A and B, open chromatin, gene expression and haplotype-resolved nuclear positioning of Xist and Flna in chrXa and chrXi, respectively. Panel a created with BioRender.com.

In a barnyard experiment using K562 and Patski cells, we demonstrated the effectiveness of ChAIR for generating high-quality tri-omic single-cell data (Supplementary Table 1; Supplementary Fig. 1a) with high reproducibility (Supplementary Fig. 1b-c). While the ChAIR-RNA and ChAIR-ATAC data showed similar levels of robustness in capturing gene transcripts and mapping open chromatin sites to related data, the number of chromatin contacts per cell in ChAIR-PET data was lower than that in scHi-C assays (Extended Data Fig. 1a-e). This is anticipated, as ChAIR is designed to target open chromatin regions, which make up approximately 1-5% of the genome. Indeed, over 40% of the chromatin contacts in ChAIR-PET data were connected to open chromatin loci (Extended Data Fig. 1f), indicating a notable enrichment for transcription-associated interactions.

To visualize the ChAIR data of individual cells, we developed ChAIR-viewer (seeMethods) to display ChAIR-RNA, -ATAC, and -PET data in three single-cell tracks (Fig. 1b). Clear variations were observed, especially in ChAIR-PET data across clusters, reflecting temporal heterogeneity of cells that may correspond to distinct cell-cycle phases. Furthermore, the ensemble ChAIR data showed profiles consistent with the bulk-cell data (Fig. 1b; Supplementary Fig. 2).

The ChAIR-RNA reads were primarily mapped to exons and showed a bias toward the 3’ end of transcripts (Extended Data Fig. 2a). The ensemble ChAIR-ATAC data exhibited similar signals at open chromatin loci and transcription start sites (TSSs) to those observed in ChIATAC¹¹, ATAC-seq¹², and scATAC-seq⁶ data (Extended Data Fig. 2b). Notably, sci-Hi-C data¹³ showed no signal enrichment at both ATAC peaks and TSSs, highlighting the effectiveness of ChAIR data in mapping open chromatin. Furthermore, ChromHMM¹⁴ analysis showed that ChAIR-identified open chromatin loci and chromatin loops shared the same epigenomic properties with those observed in ATAC-seq and ChIATAC, specifically showing the enrichment for promoter-associated features (Extended Data Fig. 2c-d). The ensemble ChAIR-PET data and the ChIATAC contact data shared the same chromatin loop profiles (Extended Data Fig. 2e). Moreover, the 2D contact profiles (Supplementary Fig. 3) showed that the higher-order chromatin structures in the ChAIR-PET and ChIATAC data were highly similar to Hi-C data. Notably, ChAIR-PET data exhibited a stronger signal at chromatin loops between open chromatin loci compared to Hi-C-based single-cell method^10,13 (Supplementary Fig. 4). Thus, our results demonstrated the superior capability of ChAIR in identifying CREs and the chromatin interactions linking gene promoters to CREs, while preserving the higher-order profiles observed in Hi-C data.

3D epigenome-transcription interplay during cell cycle

The cell cycle is a fundamental cellular process, consisting of G1, S, G2, and M phases, and phase-specific genes have been well-characterized¹⁵. Although the dynamics of higher-order chromatin structures during cell cycle have been extensively investigated^16-18, the exact interplays between phase-specific transcription and chromatin folding dynamics remain poorly understood. Therefore, the tri-omic ChAIR data derived from K562 and Patski cells provided an opportunity to address this question.

Based on phase-specific genes¹⁵ in ChAIR-RNA data (Extended Data Fig. 3a), we annotated cells in G1, S, and G2/M phases (Fig. 1c, left; Extended Data Fig. 3b-f). We then performed principal component analysis (PCA), and grouped closely positioned cells with similar RNA profiles into metacells, which were sequentially ordered based on cell-cycle pseudotime (See Methods) across the G1, S, and G2/M phases (Fig. 1c, right; Extended Data Fig. 3b, right). The chromatin contacts within each metacell provided robust pseudobulk data to characterize chromatin folding during cell-cycle progression while minimizing stochasticity from individual cells. In addition, the ChAIR-PET data can also be directly used for single-cell analysis to assess cell-cycle chromatin contact profiles in individual cells following established method¹⁹. As anticipated, five cell-cycle phases (post-M, G1, early S, late S-G2, and pre-M) were identified in both K562 and Patski cells (Supplementary Fig. 5).

To investigate the dynamic relationship between transcription, chromatin folding, and chromatin accessibility throughout the cell cycle, we analyzed Patski metacells along the cell-cycle pseudotime from G1 to S, and then to G2/M, assessing S-specific genes for their expression (RNA levels), promoter accessibility (ATAC signals), and promoter-associated chromatin interactions (PET counts) (Fig.1d, top). The RNA level remained at a basal level throughout the most of G1, began to increase in late G1, accelerated after entering S until peaked at the end of S, and then started to decline moving into G2/M phase. In contrast, the PET counts rose rapidly from early G1 through mid-S, and began to drop entering G2/M. Similarly, though delayed, the ATAC signals exhibited a slow, gradual increase in G1, followed by an accelerated increase in S phase, and declined in G2/M. The same pattern was also observed for G2/M-specific genes (Fig.1d, bottom). These temporal and sequential dynamics suggested a stepwise causal mechanism of transcription activation, where distal enhancers and promoters first establish contacts, potentially facilitated by pioneering factors²⁰, which then lead to promoter opening and ultimately transcription activation.

Taking advantage of Patski cells derived from an F1 hybrid mouse with a high frequency of heterozygous SNPs²¹, we obtained haplotype-resolved ChAIR data and validated the results by analyzing chrXi-specific Xist and chrXa-specific Flna (Supplementary Fig. 6). Using haplotype-resolved ChAIR-PET data in metacells, we reconstructed 3D genome models throughout the cell cycle (Supplementary Fig. 7) and revealed the typical whole genome conformation with the euchromatin toward the nuclear interior and heterochromatin on the periphery (Fig. 1e). Both of the 3D models and the 2D contact matrices from different cell-cycle phases revealed distinct chromatin conformations, in line with previous findings^8,19. These conformational variations over the three cell-cycle phases were further supported by marked changes in genome folding density and volume (See Methods) as estimated by 3D modeling²² (Supplementary Fig. 7c-d). Together, we demonstrated the power of ChAIR in enabling efficient single-cell analysis to delineate the genomic framework underlying transcription regulation.

Tri-omic landscapes in mouse brain cells

Although gene expression studies have provided significant insights into the complex cellular composition and functions of the mouse brain²³, the genomic frameworks underlying its diverse cellular properties remain poorly understood. This gap makes the mouse brain an ideal system for ChAIR to uncover the underlying genomic regulatory mechanisms of cell type-specific transcription. To gain a comprehensive understanding of how the cell type-specific 3D epigenome is coordinated with transcription, we analyzed whole mouse brains rather than discrete brain sections to comprehensively profile compositional changes of cells across brain regions throughout the lifespan. We selected five critical developmental stages from infancy through young adult to aged mice: postnatal days 2 (P2), 11 (P11), 95 (P95), 365 (P365), and 730 (P730) (Fig. 2a).

Fig. 2: Single-cell 3D epigenomic and transcriptomic landscapes of whole-brain cells during mouse maturation and aging.

a, The study design of collecting whole-brain tissues at various postnatal days. b, A UMAP plot of the combined ChAIR-RNA data highlighting major brain cell classes in various colors. Ast, astroependymal cells; VAS, vascular cells. c, The UMAP plots of ChAIR-RNA data from five age points. The cell numbers (n) are provided. d,e, Cell group-specific signal matrices calculated by transcriptomic (d) and 3D epigenomic features (e) across 23 cell groups that belonged to seven cell classes in ChAIR data. The signal was normalized by calculating the ratio of the cell group-specific feature signal to the total signal of all cell group-specific features. A CIS was then provided to evaluate cell group-specific feature’s effectiveness in distinguishing different cell groups. f, A browser view of ensemble ChAIR data at the Cbln1 locus showing Cbln1-associated enhancers (Es), chromatin loops and expression specifically in CBGRC cells relative to other neuronal cells. g, Left: the detailed browser views of ChAIR ensemble and single-cell data at the Etv1 locus and associated enhancers (Es). Top right: one CBGRC-specific enhancer was validated by VISTA32. Right bottom: the corresponding 2D contact heat map of ChAIR-PET and ChAIR-ATAC data at the Etv1 locus. h, The spectrum of chromatin contact distance and boxplots of the megacontact percentage, global chromatin accessibility, global transcriptional activity and genome folding density across 17 specific cell groups, including ExNs, InhNs and non-neurons. The center line denotes the median, the box represents the interquartile range (25th–75th percentiles) and the whiskers extend to 1.5× interquartile range. i, Examples of reconstructed 3D genome architectures representing nuclei among various brain cells, with MGL having the highest chromatin density and TEGLU cells having the lowest. Panel a created with BioRender.com.

With ChAIR's high-throughput capacity, we generated highly reproducible ChAIR data with both biological and technical replicates from mouse brains (Supplementary Table 1; Supplementary Fig. 8). Unsupervised dimensionality reduction analysis using ChAIR-RNA and ChAIR-ATAC data effectively separated major brain cell types. Although the ChAIR-PET data were less effective, likely due to data sparsity, the ChAIR-PETs that were associated with active genes showed significantly improved cell clustering results. In addition, the integration of tri-omic ChAIR data further enhanced cell type distinction (Extended Data Fig. 4a-g). Despite these improvements, scRNA-seq remains the most convenient dataset for cell clustering and annotation.

Through rigorous quality control (See Methods), we collected high-quality ChAIR data from a total 222,698 cells (Supplementary Table 1-2; Fig. 2b-c). By integrating with a widely used mouse brain scRNA-seq dataset²⁴, the ChAIR-RNA data identified 121 cell types characterized by the expression of canonical marker genes of the 199 well-defined mouse brain cell types²⁴ (Extended Data Fig. 4h; Supplementary Fig. 9). These cells belong to 45 cell groups of 7 major brain cell classes: excitatory neuron (ExN), inhibitory neuron (InhN), astroependymal (Ast), oligodendrocyte progenitor cell (OPC), oligodendrocyte (Oligo), microglia (MGL), and vascular cell (VAS) (Fig. 2b). Meanwhile, the ChAIR-ATAC and -PET data closely matched related datasets, showing expected epigenomic state and chromatin structure^25,26 (Extended Data Fig. 4i; Supplementary Fig. 10-11).

To analyze the correlation between ChAIR-RNA, -ATAC, and -PET data, we focused on 23 out of the 45 cell groups that had sufficient tri-omic data for in-depth analysis. Given that scRNA-seq data effectively defined cell identity using cell type-specific marker genes, we reasoned that the 3D epigenomic features associated with the marker genes would exhibit a comparable degree of specificity. To quantitatively evaluate this, we formulated a ‘cell identity score’ (CIS) to measure the distinctiveness of various molecular features in cell identities. These features include transcription (RNA), gene activity (ATAC signal on gene body), promoter strength (ATAC at TSS), enhancer strength (ATAC at enhancer site), and chromatin loop (PET). Essentially, CIS is calculated by dividing the unique molecular signals of a specific cell type against the background signals from other cell types (See Methods). All mono-modal CIS displayed observable cell type-specificity, with the transcription CIS and chromatin loop CIS at the same level, whereas ATAC-related CISs lagged behind. Notably, although the transcription CIS and chromatin loop CIS showed only minor difference, their signal patterns across the 23 cell groups were visually distinct, suggesting that chromatin contacts might provide a perspective on cell identity that is somewhat independent from gene expression. Furthermore, integrating chromatin loop specificity with ATAC-based open chromatin specificity, collectively referred to as 3D epigenomic specificity, significantly enhanced the individual CISs (Fig. 2d-e). For instance, in OPC, MFOL, and MOL, the composite 3D epigenomic CISs (4.04, 12.18, and 6.40, respectively) were significantly higher than the transcription CISs (2.17, 2.38, and 3.56, respectively). Together, our results highlighted that the 3D epigenomic landscape can provide a more robust distinction of cell identities than gene transcription.

Cell type-specific enhancers of brain cells

Identification of cell type-specific enhancers in the brain has tremendous value in functional neuroscience studies and therapeutic development^27-29. The 3D epigenomic landscape defined by the combination of ChAIR-ATAC and -PET data greatly facilitated the identification of cell type-specific enhancers that are associated with specific marker genes in brain cells. By applying stringent criteria to a large set of distal CRE candidates linked to marker genes (See Methods), we identified 562 cell type-specific enhancers (Supplementary Table 3) and systematically analyzed their characteristics (Extended Data Fig. 5). One example involved 12 notable enhancers specifically connected to the promoter of Cbln1 (Fig. 2f; Supplementary Fig. 12), which encodes Cerebellin-1 protein that is essential for synaptic integrity and plasticity specifically in cerebellar granule cells (CBGRC)³⁰. Another example was the enhancers that are connected to Etv1 (Fig. 2g). Etv1 is known to play a key role in orchestrating the neural activity-dependent gene regulation for terminal maturation of CBGRC³¹. Intriguingly, one of the enhancers was validated in mouse embryos by LacZ reporter gene assays, showing specificity in mouse hindbrain³². Additional examples include enhancers for Epha4, Gabra6, and a few other genes, were shown in Supplementary Fig. 13. Furthermore, despite the Epha4-specific P-E interaction was reported in a recent study¹⁰, where the GAGE-seq data (scRNA+scHi-C) had to incorporate external scATAC-seq³³ data to assist the detection of this linkage. In contrast, the tri-omic ChAIR data directly detected this interaction, highlighting its advantage in identifying cell type-specific enhancers. Intriguingly, the Gabra6-specific enhancers were age-specific, emerging only after adulthood (P95-P730), coinciding with increased Gabra6 expression in aged mice. Thus, the cell type-specific enhancers of mouse brain cells identified by ChAIR data represent a valuable resource for further exploration.

Megacontacts associated with nuclear properties of different brain cells

Recognizing the potential implication of ultra-long megacontacts (>2Mb) in revealing cell type-specific genome folding profiles in mouse and human brains^26,34,35, we systematically investigated how megacontacts may impact global chromatin accessibility, transcriptional activity, and genome folding density in post-mitotic mouse brain cells of adult mice (P95). We selected 17 cell types representing major cell groups in ExNs, InhNs and non-neuronal cells with sufficient ChAIR data for in-depth analysis. Overall, ExNs had the lowest megacontact proportions (25-30%), with CBGRC as an outlier. InhNs had slightly higher proportions, while the non-neuronal cells, particularly MOL and MGL, had the highest proportions (Fig. 2h). The short-range contacts (20kb-1Mb) correlated with increased global chromatin accessibility, higher transcriptional activity, and lower genome folding density. In contrast, higher megacontact ratios were linked to decreased global chromatin accessibility, reduced transcriptional activity, and increased genome folding density (Fig. 2h), consistent with previous studies^26,34,35. Our observation suggested that the increased megacontacts were linked to genome compaction and transcription reduction. Importantly, the estimated genome folding density was closely correlated with the ratio of megacontacts (Fig. 2h-i). Overall, neurons, except CBGRC, exhibited lower megacontact ratios, greater chromatin accessibility, and higher transcriptional activity compared to non-neuronal cells.

We then investigated whether genome conformation revealed by mapping data can reflect nuclear volume. To explore this, we compared our mapping-based findings with imaging-based measurements. Notably, previous mouse brain microscopy data showed that both MOL and CBGRC had smaller nuclei size compared to dentate gyrus granule cells (DGGRC)^36-38, suggesting that cells with greater genome compaction tend to have smaller nuclear volumes and vice versa. This was further supported by recent sequential DNA/RNA FISH results^39,40 (Extended Data Fig. 6). Convincingly, we established a strong megacontact-associated correlation between nuclear volume, global transcriptional activity, and chromatin accessibility across different brain cells, using mapping-based ChAIR data. This finding was validated by multiple lines of imaging-based measurements.

Spatially-resolved 3D epigenomic specificity

Although ChAIR effectively generates tri-omic single-cell data, it lacks spatial resolution. To address this, we merged the ChAIR data with the spatial transcriptomic data (Stereo-seq) from an adult mouse brain section⁴¹, using ChAIR-RNA data as a bridge (see Methods). We linked 20 anatomical regions from the Stereo-seq data to 19 cell types defined by ChAIR-RNA data with high specificity (Fig. 3a). This ChAIR-Stereo data alignment was verified by assessing the spatial distribution of canonical marker genes in the histological staining of adjacent tissue section in Allen Brain Atlas (atlas.brain-map.org) (Extended Data Fig. 7a-d). As a result, this integration provided spatially resolved details to ChAIR data, especially assigned the anatomic specificity to the corresponding 3D epigenomic states in ChAIR-ATAC and -PET data.

Fig. 3: Regional-specific 3D epigenomic features in mouse brain cells revealed by ChAIR.

a, An integration of Stereo-seq (bin50) data with ChAIR-RNA data, connecting regions in a coronal hemibrain section defined by Stereo-seq (left) with the corresponding cell types characterized by ChAIR (right). b, A spatial distribution of specific cell types and locations (first column) and the corresponding signal intensities of marker gene transcription (second column), marker gene-associated chromatin loop (third column) and 3D epigenomic features (fourth column) derived from ChAIR data. The signal was normalized by calculating the ratio of the cell-type-specific feature signal to the total signal of all cell-type-specific features. A CIS was provided to evaluate cell-type-specific feature’s effectiveness in distinguishing different cell types.

Strikingly, as shown by TEGLU cells in neocortex (Fig. 3b, 1^th columns), both the chromatin loops and the composite 3D epigenomic features exhibited significantly higher signal specificity as measured by CISs than those represented by transcription (Fig. 3b, 2-4^th columns; Extended Data Fig. 7e), consistent with our earlier observations in characterizing ChAIR data (Fig. 2d-e). For example, the marker gene transcription profiles in TEGLU7 (specific cells in cortex layer 2/3) showed signal enrichment in both cortex layer 2/3 and dentate gyrus, while the marker gene-associated chromatin loop and particularly the 3D epigenomic features exhibited much elevated signal intensity and specificity in cortex layer 2/3 but not in dentate gyrus. The same patterns were also observed in TEGLU4/8 (layer 4) and TEGLU2/3/10 (layer 5/6). Furthermore, the marker genes of DGGRC cells of dentate gyrus showed slightly higher expression signals in dentate gyrus than in the cortex layers, while the associated chromatin loops and 3D epigenomic features displayed markedly high specificity to dentate gyrus. Additional examples and single-cell-resolved browser views of 3D epigenomic features associated with cell type-specific marker genes were presented in Supplementary Figs. 14-15.

Thus far, our results demonstrated that the integration of spatial transcriptomic data with ChAIR, using ChAIR-RNA as a bridge, can provide spatial resolution to the 3D epigenomic states captured in ChAIR data. Additionally, our observation of increased specificity in 3D epigenomic features associated with marker genes provided convincing evidence in supporting a strong link between the 3D genome folding, epigenomic state, and gene transcription.

Concerted interplays between 3D epigenome and transcription during brain cell differentiation

It has been shown that mouse brain cells undergo substantial changes in transcription and genome architecture during neuronal development²⁶. However, most of our prior knowledge was derived from studies that integrated unrelated single-cell mono-modal datasets with prior assumptions. Here, we investigated the concurrent relationship of transcriptome, chromatin accessibility, and 3D genome structure across different cell lineages using the single-cell tri-omic ChAIR data.

Building on our observation that the distance spectrum of chromatin contacts provided a distinctive metric for assessing genome properties across brain cells (Fig. 2h), we investigated the well-documented cell lineage from OPCs to mature Oligos²⁴. All OPCs and Oligos cells were ordered according to the proportion of megacontacts from the least to the most in individual cells (Fig. 4a, top). In this arrangement, OPC primarily featured with short-range contacts, while most MOL1 cells were characterized by abundant megacontacts (Fig. 4a, bottom). This ordering recapitulated the differentiation trajectory from OPC to MOL1, revealing a striking concordance between RNA pseudotime and the sequential order of megacontacts as ‘megacontact pseudotime’ (Fig. 4b, top). This observation was supported by marker gene expression patterns and the correlation between megacontact pseudotime with chronological age (Fig. 4b, bottom). As expected, the 3D modeling results showed that MOL1 exhibited a significantly more compact genome conformation compared to OPC (Fig. 4c and Supplementary Fig. 16).

Fig. 4: Chromatin architecture reorganization during cell differentiation.

a, The spectrum of chromatin contact distance of OPCs and Oligos. Top: individual cells (vertical lines, n = 23,904) were sorted by the extent of megacontacts from the least to the most. Bottom: OPCs and subtypes of Oligo in coordination with the sorting of chromatin contact distances. Contact frequencies were normalized to 0–1. b, UMAP plots of ChAIR data (from left to right): the cell lineage trajectory from OPCs to Oligos, RNA pseudotime, megacontact pseudotime, OPC-specific Pdgfra profile, MOL1-specific Mal profile and cell distribution of age points. c, The example views of reconstructed 3D genome models of OPC and MOL1 metacells. The nuclear positions of specific marker genes and chromosomes are indicated. The sizes of the embedded balls reflected the relative levels of gene expression. d, The normalized signals (scale ranging from 0 to 1) of gene expression (RNA), promoter chromatin accessibility (ATAC) and chromatin connectivity (PET) in individual cells during the transition. The dotted line indicates the turning point from OPC to MOL. The cells were sorted by the extent of megacontacts same as in a, from lowest to highest. e, The spectrum of chromatin contact distance (top) and cellular composition (bottom) of neuroblast DGNBL1/2 to DGGRC1/2 (n = 8417) in dentate gyrus. The contact frequencies were normalized to 0–1. f, UMAP plots of ChAIR data (from left to right) on the differentiation trajectory from DGNBL1/2 to DGGRC1/2, RNA pseudotime and megacontact pseudotime. g, The reconstructed 3D models of nuclear architectures from DGNBL1, DGNBL2 and DGGRC2 metacells. The nuclear positions of specific marker genes are indicated. The sizes of the embedded balls reflected the relative levels of gene expression. h, The normalized signals (scale ranging from 0 to 1) of gene expression (RNA), promoter chromatin accessibility (ATAC) and chromatin connectivity (PET) in individual cells. The cells were sorted by the extent of megacontacts, from lowest to highest, same as in e.

Next, we sought to examine the interplay between gene transcription, epigenomic state, and 3D genome folding during differentiation from OPC to MOL. Specifically, we measured the expression of MOL-specific marker genes (n=50) (Supplementary Table 4), ChAIR-ATAC signals at the gene promoter regions, and chromatin interactions connected to the marker genes along the differentiation trajectory and megacontact pseudotime (Fig. 4d). The results showed a steady increase in all three modalities, from the same starting point to a plateau. Notably, at the point of transition from OPC to MOLs, the ChAIR-PET signals were higher than -ATAC and -RNA signals, indicating that the MOL-specific chromatin interactions were established before the promoter opening and the activation of MOL-specific transcription.

In addition to non-neuronal cells, we also analyzed the megacontacts of excitatory neurons in dentate gyrus (DG), tracing their differentiation from neuroblasts (DGNBL1/2) to mature granule cells (DGGRC1/2)⁴². Intriguingly, the chromatin contacts in most DGNBL1 cells predominantly exhibited high level of megacontacts, while DGNBL2 positioned in the middle, and DGGRC1/2 exhibited mostly short-range contacts (Fig. 4e), contrasting to the pattern observed in non-neuronal cells during the OPC-to-MOL differentiation. We validated this trajectory by demonstrating a high correlation between the RNA and megacontact pseudotimes (Fig. 4f). The 3D modeling results further revealed that DGNBL, with a higher megacontact ratio, exhibited a more compact genome. In contrast, DGGRC cells, which predominantly displayed short-range contacts, showed a more relaxed genome conformation, suggesting a larger nuclear size (Fig. 4g). At the point of DGNBL differentiating into DGGRC, chromatin interactions associated with DGGRC-specific genes showed higher signal levels than those of promoter accessibility and gene expression of the marker genes (n=51) (Fig. 4h), similar to the observation in OPC-to-MOL differentiation. We also extended our surveys of megacontact dynamics to cellular differentiation from CBNBL to CBGRC in cerebellum and OBNBL to OBINH in olfactory bulb (OB) (Extended Data Fig. 8).

Together, we revealed that neurons located in different brain regions appeared to display the same patterns of chromatin remodeling during differentiation, progressing from predominant megacontacts in progenitor cells to mostly short-range contacts in mature neurons. Remarkably, neurons and non-neuronal cells seemed to possess different genomic properties and appeared to implement divergent genomic mechanisms to regulate their differentiation and support their diverse functions.

Chromatin rewiring and transcriptional changes across lifespan

Brain cell functions decline with age, which contributes to the development of degenerative neuronal disorders⁴³. Most studies have focused on isolated brain regions, lacking a whole-brain perspective. Using the ChAIR data derived from whole mouse brains, we explored the dynamic shifts of cellular composition across lifespan. Notably, progenitor cells of both neuronal and non-neuronal cells were more abundant in early life, while the mature excitatory neurons (mExNs) gradually became more predominant than mature inhibitory neurons (mInhNs) (Fig. 5a). Furthermore, the dynamic shift in cell composition also occurred in different brain regions, with the most notable change observed in ExNs from the telencephalon (TE) and cerebellum (CB). Strikingly, we found that ExNs in CB at P2 accounted for only 10% of all ExNs, but this proportion rapidly increased to 40% by P11 and eventually exceeded 80%, making them the predominant ExNs in later stages of life. In contrast, ExNs in TE continuously declined from more than 50% at P2 to less than 25% in adult (P95) and aged mice (Fig. 5b, left). Conversely, the proportion of all InhNs remained relatively stable (Fig. 5b, right). This dramatic shift in ExNs from TE and CB may imply an important structural and functional transition over the course of brain development and ageing.

Fig. 5: The remodeling of chromatin folding in mouse brain cells throughout the lifespan.

a,b, The dynamic shift of cellular composition of major brain cell classes and their progenitors during maturation and aging from infant (P2/P11) to adult (P95) and aged stages (P365 and P730). Cellular composition of major brain cell types across age points (a) and cellular composition of ExNs (left) and InhNs (right) from distinct anatomic regions across age points (b). mExN, mature ExNs; ExNBL, excitatory neuroblast; mInhN, mature inhibitory neurons; InhNBL, inhibitory neuroblast; Ast, astroependymal cells; VAS, vascular cells; OB, olfactory bulb; DE, diencephalon; DG, dentate gyrus; HB, hindbrain (pons and medulla); ME, mesencephalon. c, The distance spectrum of chromatin contacts in different cells across five age points (top) along with aggregated strength of A/B compartment and TAD signals (bottom). The lines were smoothed by generalized linear models, with shading indicating the confidence interval. d, The example views of 2D contact heat maps and eigenvector plots derived from pseudobulk analysis of ChAIR-PET data in TEGLU (top) and CBGRC (bottom) cells. Ank1 (1) and Zmat4 (2) gene loci are indicated in the boxed areas. e, The single-cell browser views showing individual CBGRC cells with ChAIR-PET data connecting A–A (top) and B–B (bottom) compartments across infancy (P2/P11), adult (P95) and aged (P365 and P730) stages. Middle: the eigenvector values are also shown for reference. The numbers (n) of cells (normalized by total number of cells per each age point) with intercompartment contacts at each age point are provided. Notably, a compartment shift from A to B occurred in the region in which Zmnt4 resides.

We then examined the distribution of chromatin contacts across mouse brain cells at five age points. Despite mature neurons and non-neuronal cells at P95 typically showed mostly short-range contacts and megacontacts, respectively (Fig. 2h-i), they primarily exhibited short-range contacts at infancy (P2/11). However, they both showed a significant increase in megacontacts overtime. In contrast, InhNs maintained short-range contacts over time (Fig. 5c; Supplementary Fig. 17a-c). Further analysis revealed that changes in chromatin contact ranges in ExNs were specific to the CB, while those in TE remained unchanged. It is worth noting that although transcriptome stabilized after P95, there was an increase in megacontacts with advancing age (Supplementary Fig. 17e-g), suggesting a strong correlation between megacontact formation and ageing in CBGRC. This indicates that even without transcriptional changes, chromatin folding continues to remodel in response to ageing, consistent with a recent report in human cerebellar granule cells⁴⁴. Among non-neuronal cells, only Oligos shifted from predominantly short-range contacts in early life to megacontacts in later stages (Fig. 5c). In contrast, OPCs consistently exhibited short-range contacts, while VASs and MGLs maintained stable megacontacts over time (Supplementary Fig. 17a-c). Additional analyses revealed that cells with chromatin contacts predominantly in short-range were positively correlated with active transcription and negatively with trans-contacts, while cells with extensive megacontacts showed the opposite trend (Supplementary Fig. 17d), consistent with our previous findings on megacontacts.

To characterize chromatin conformational changes in contact distance with further details, we examined chromatin compartments, TADs, and loops using ChAIR-PET data from different cell groups (Fig. 5d). For example, TEGLU (ExN in TE) maintained short-range chromatin contacts and stable TAD structures during ageing, while CBGRC in CB transitioned from short-range chromatin contacts to megacontacts and shifted from a TAD-dominant to a compartment-dominant state with clear plaid-like patterns. Our findings indicated that short-range chromatin contacts typically correlated with a predominance of TAD structures, while megacontacts were associated with increased compartmentalization in brain cells as they age. More specifically, the chromatin conformation in ExNs from cerebrum showed consistent TAD and compartment signals, whereas the cerebellar ExNs were characterized by a notable transition from a TAD-dominant state to a state with increased compartmentalization over time (Supplementary Figs. 18-19). Furthermore, our 3D modeling results indicated that TEGLU consistently exhibited lower genome folding density and lager nuclear volume than CBGRC. Furthermore, while TEGLU's genome folding density and transcriptional activity remained constant, CBGRC experienced a progressive increase in genome folding density accompanied by reduced transcriptional activity during ageing (Extended Data Fig. 9).

Megacontact is a reliable indicator of ageing

Motivated by the observation that the megacontact proportions in a genome varied across cell groups throughout the lifespan (Fig. 5), we investigated their correlation with chronological age. Notably, ExNs (CBGRC), InhNs (CBINH) in cerebellum, and non-neuronal ACMB displayed a strong positive correlation between megacontact and chronological age, which was further supported by the transcriptomic age analysis using the SCALE model⁴⁵ (Fig. 6a-c). Similar trends were also observed in other cell groups (Supplementary Fig. 20). Overall, 80% (36 of 45) of analyzed cell groups showed positive correlations between megacontact and both chronological and transcriptomic age, with most (29 of 36) exhibiting significant positive correlations (Pearson correlation coefficient>0.5) (Fig. 6d). Of the cells that showed negative correlation, most of them were progenitor cells like CBNBL, OBNBL, and OPC, as well as cholinergic neurons like HBCHO and TECHO, likely due to their prevalence in early life. Intriguingly, while two types of neocortical neurons (TEGLU and TEINH) showed no ageing-related changes, many other neurons, primarily those from evolutionarily older regions (e.g., OB, DE, DG, ME, HB, and CB), exhibited strong correlations with ageing. This difference potentially suggested that ageing-related megacontact dynamics may be more pertinent to ancient brain structures than to the neocortex.

Fig. 6: Chromatin megacontacts in brain cells as a marker of aging.

a–c, Example views of single-cell megacontact analysis examining the correlation between the megacontact ratio (Mega, in blue) and both chronological age (CA, in purple) and transcriptomic age (TA, in orange) in CBGRC (a), CBINH (b) and ACMB (c) (top). The cells were sorted by megacontact ratio from low to high. The fitted curves for the signals of the megacontact ratio, chronological age and transcriptomic age are shown (bottom). The numbers (n) of single cells are also provided. d, A Pearson correlation between the megacontact ratio, chronological age and transcriptomic age across 45 cell groups. The shaded area represents data points exhibiting limited correlation (from +0.5 to −0.5) between megacontact ratio and both chronological and transcriptomic age. e, The zoomed-in views of 2D contact heat maps of CBGRC-specific ChAIR-PET data in Fig. 5d, showing an expanded compartment A where Ank1 (green dot 1) resides and a compartment switch from A to B along with the dissolution of the TAD structure around Zmat4 (green dot 2) during aging. The numbers (n) of CBGRC cells for each age point are provided. f, The single-cell browser views of ChAIR-PET data in individual CBGRC cells along with gene expression of the ensemble ChAIR-RNA data for Ank1 (top) and Zmat4 (bottom) at infant (P2/P11), adult (P95) and aged (P365 and P730) stages. The numbers (n) of cells (normalized by random sampling of an equal number of cells investigated in each age point) exhibiting three modalities of ChAIR data associated with examined genes are provided. g, A generalized model depicting the roles of chromatin folding profiles, including short-range contacts and ultralong megacontacts, in different mouse brain cells and during differentiation and aging. Panel g created with BioRender.com.

Next, we sought to explore the underlying 3D epigenomic framework that regulates ageing-related genes identified in this study (Supplementary Table 5). Generally, the 3D epigenomic states were highly associated with the transcriptional activity of ageing-related genes. For example, Ank1, a CBGRC-specific gene encoding ankyrin-1, which is crucial for maintaining cell stability⁴⁶, remained in an active compartment across the five age points, with its TAD structure expanding and forming distinct stripes after maturation, particularly in aged CBGRC (Fig. 6e). Concurrently, Ank1’s expression and promoter-associated chromatin contacts steadily increased over time (Fig. 6f, top). In contrast, the nearby Zmat4 gene, known for its role in Alzheimer’s disease^47,48, switched from compartment A (active) in early life to compartment B (repressive) in later stages. This transition was accompanied by the dissolution of the TAD structure and the loss of promoter-associated chromatin interactions as well as the diminished expression (Fig. 6f, bottom). These chromatin conformational changes demonstrated a likely genomic mechanism by which alterations in higher-order chromatin landscape and local chromatin interactions modulate gene expression during ageing. Collectively, our results revealed that the dynamic changes in the megacontact ratio can be used as a measurement to monitor the cellular ageing progression.

Discussion

In this study, we introduced ChAIR, a tri-omic single-cell method that simultaneously maps chromatin accessibility, chromatin interaction, and transcription profiles. We demonstrated its application in cell lines and primary mouse brain cells to investigate the concurrent dynamics of 3D genome folding, epigenome states, and transcription modulation during cell growth, differentiation, maturation, and ageing.

ChAIR offers unique technical advantages over the current bi-omics single-cell methods^4-10 in several aspects. Besides the tri-omic nature, ChAIR specifically enriches chromatin interactions between open chromatin loci, thus facilitating the identification of CREs in accessible regions and capturing chromatin interactions involved in active transcription. Furthermore, the integrated 3D epigenomic features from the ChAIR-PET and -ATAC data associated with marker genes improve the ability to define cellular specificity. Additionally, ChAIR’s high-throughput capacity enables the analysis of a broad range of cell types from complex tissues. Furthermore, integrating ChAIR data with spatial transcriptomic data through ChAIR-RNA adds anatomic specificity to the ChAIR data.

One technical limitation in the current version of ChAIR is the relatively lower numbers of chromatin contacts per cell compared to scHi-C data, primarily due to its focus on open chromatin regions (~1-5% of the genome). However, the gene-centric ChAIR-PET data associated with active transcription enhances the ability to distinguish different cell types. Furthermore, the high-throughput capacity of ChAIR compensates for fewer contacts per cell by merging similar cells for pseudobulk analysis. With further improvements, we envision that ChAIR could possibly target both open chromatin and heterochromatin, with specificity for each region, thus enabling a more comprehensive genome coverage.

Applying ChAIR, we tackled long-standing questions in transcription regulation, particularly the causal relationship between genome structure and gene transcription. Recent functional perturbation experiments depleting cohesin diminished global genome folding structures, with minimal impact on transcription⁴⁹, raising questions about how genome structure influences transcription. Our single-cell tri-omic ChAIR data provide evidence supporting a temporal and causal mechanism of transcription activation during cell growth and differentiation. This process begins with the establishment of chromatin interactions between gene promoters and distal regulatory elements, followed by chromatin opening at the promoter regions, and ultimately, the activation of gene transcription.

We have also uncovered that the relative abundance of short-range and ultra-long-range chromatin contacts is highly associated with nuclear and cellular properties. The proportion of megacontacts in a genome serves as an effective measure of the genomic state across different cell types and developmental stages (Fig. 6g). In mouse brain cells, neurons tend to have lower megancontact ratios, larger nuclear volumes, increased chromatin accessibility, and higher transcriptional activity, while non-neuronal cells possess the opposite pattern. This difference suggested that neurons and non-neuronal cells may employ different mechano-genomic mechanisms to organize their nuclear architecture and regulate genomic functions, ultimately shaping their unique cellular properties. Moreover, the content of megacontacts in most brain cells exhibited a significant increase, which was associated with the reduced transcriptional activity during cellular ageing, making it a potential hallmark of ageing.

Methods

Cell culture

K562 cells were cultured in RPMI-1640 (ThermoFisher) with 10% FBS (Gibco). Patski cells are fibroblasts derived from the embryonic kidney from a cross between a BL6 female mouse carrying an Hprt^BM3 mutation and a Mus spretus male. The cells were selected in HAT media so that the BL6 X chromosome is constantly inactive⁵⁰. Patski cells were cultured in Dulbecco’s modified Eagle medium (DMEM) with 10% FBS (Gibco) and 1% penicillin/streptomycin. All cells were cultured at 37 °C and 5% CO₂.

Animals

C57BL/6J mice at postnatal day 2, 11, 95 (n = 2), 365 (n=2) and 730 were purchased from Shanghai SLAC Laboratory Animal Company. The animal room has controlled temperature (18-23 °C), humidity (40-60%), and a 12 light/12 dark cycle. All experimental procedures were approved by the Zhejiang University Animal Care and Use Committee. Whole Mouse brains were dissected and immediately frozen in liquid nitrogen and kept in −80 °C.

ChIATAC (bulk-cell)

ChIATAC libraries were generated using 50,000 FA-EGS-crosslinked K562 and Patski cells following the established protocol¹¹. The ChIATAC libraries were sequenced by paired-end sequencing (2 × 150 bp) on an Illumina NovaSeq 6000.

ChAIR (more details in Supplementary Information)

50,000-300,000 FA-EGS-crosslinked cells or nuclei were used as starting material to perform the ChAIR assay. Cells were first lysed by 100 μl 0.1%SDS FA buffer (50 mM HEPES-KOH, pH 7.5, 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, 0.1% Sodiumdeoxycholate, 0.1% SDS) at 4 °C for 1 h (omitting this step for nuclei). Cells/nuclei were spun down at 600 × g for 5 min at 4 °C and permeabilized by 10 μl 0.1% SDS at RT for 2 h. The reaction was quenched by adding 2.5 μl 20% Triton X-100 and incubated at 37° C for 20 min. Cells/nuclei were then in situ digested by AluI (NEB; R0137L) or AluI+HpyCH4V (NEB; R0620L) restriction enzymes at 37 °C for at least 2 h or overnight by adding 2 μl AluI/ 1 μl AluI + 1 μl HpyCH4V, 5 μl 10× Cutsmart buffer (NEB; B7204S), 25.5 μl ddH₂O. Restriction enzyme digested chromatin DNA was A-tailed by 0.6 μl 1 mM dATP, 1 μl Klenow Fragment (3′→5′ exo-) (NEB; M0212M), and 1 μl recombinant albumin (2mg/ml) (NEB; B9200S) at 37 °C for 1 h. The reaction was stopped by incubating at 65 °C for 20 min. In situ ligation is then performed by the addition of 20 μl 5× Quick ligation buffer (NEB; B6058S), 1 μl T4 DNA ligase (NEB; M0202L), 3 μl (2 ng/μl) bridge linker (Forward strand: 5'-/5Phos/CGTGATATT/iBIOdT/CACGACTCT-3', reverse strand: 5'-/5Phos/GAGTCGTGAAATATCACGT-3'), and 23.4 μl ddH₂O for at least 4 h at room temperature or overnight at 16 °C. Cells were spun down at 500 × g for 5 min in 4 °C and washed once with ATAC-RSB buffer with RNase inhibitor (10 mM Tris- HCl pH 7.4, 10 mM NaCl, 3 mM MgCl₂, 0.1% Tween-20, 1% BSA). Cells/nuclei were centrifuged at 500 × g for 5 min at 4 °C and resuspended in 1x resuspension buffer provided by 10x Genomics. The cells/nuclei were then sonicated for 1 second to break up the cell clusters. After filtration through a 20 μm pluriStrainer to remove aggregates, cells/nuclei were counted to confirm their concentration. Next, samples were tagmented for 1 h at 37 °C and loaded into the 10x Genomics Chromium Controller with the Chromium Next GEM Single Cell Multiome ATAC+Gene Expression Reagent Kits (10x Genomics; 1000283), following manufacturer’s instructions. Upon emulsion breaking, 10 μl proteinase K (ThermoFisher; AM2548) was added to the mixture and incubated at 65 °C for 2 h for decrosslinking. After nucleic acid recovery, M280 beads were used to pulldown the biotinylated dsDNA and the supernatants were recovered for purifying the cDNAs. Both dsDNA and cDNA generated from the captured mRNA were processed to construct ChAIR DNA and RNA libraries. These libraries were prepared and sequenced in accordance with the manufacturer's guidelines. Sequencing reads were converted from BCL to FASTQ format and demultiplexed using bcl2fastq.

Nuclei isolation

Nuclei used for mouse brain ChAIR data were isolated from mouse brain samples according to Corces et al.⁵¹ with modification. In brief, brain tissues were shredded into small pieces in a liquid nitrogen environment to maintain tissue integrity, then fixed with 2% formaldehyde at room temperature for 20 minutes. Crosslinked tissues were processed for nuclei isolation through mechanical disruption via dounce homogenization and subsequent density gradient centrifugation. The isolated nuclei underwent further crosslinking with 2 mM Ethylene Glycol-bis (Succinimidyl Succinate) (EGS) at room temperature for 45 minutes, and were stored at −80 °C.

ChAIR data processing

In this study, hg38 and mm10 were used as reference genomes. ChAIR-RNA data was processed by 10x Cell Ranger to generate the RNA count matrix, with each row representing a cell and each column representing a gene. In parallel, ChAIR-DNA data was processed with our in-house pipeline ChAIR-PIPE available at https://github.com/fengchuiguo1994/ChAIR-PIPE. ChAIR-PIPE started with the bulk processing of DNA reads from ChAIR via ChIA-PIPE⁵², which included steps such as linker trimming and mapping. Subsequently, single-cell identifiers were appended to each reads using cell barcodes derived from 10x Cell Ranger ARC. De-duplication was done based on the mapping coordinate and the cell barcode. For ChAIR-ATAC data, peaks from ensemble data were called by MACS2⁵³. The ChAIR-ATAC information was compiled into a fragment matrix with each row representing a cell and each column representing a peak. For ChAIR-PET data, each paired-end-tag (PET) read was identified, and a contact matrix with each row representing a cell and columns with the number of total PETs, and counts of intra-chromosomal PETs were categorized into specified contact distance intervals: 1-20 kb, 20 kb-1 Mb, 1-2 Mb, and over 20 Mb. The percentage of intra-chromosomal PETs in total counts was calculated as a general library quality assessment. The RNA count matrix and the DNA fragment matrix were used as inputs for further analysis with Seurat⁵⁴ and Signac⁵⁵, respectively. For the species-mixing experiment, we identified cells as doublets if less than 80% of their mapped reads from ChAIR-RNA, -ATAC, and -PET data were attributed to either hg38 or mm10. We used the following cutoffs for cell calling:

K562: nCount_RNA > 500 & nCount_ATAC > 200 & mitochondrial percentage (percent.mt) < 20% & ribosomal RNA percentage (percent.ribo) < 50%.

Patski: nCount_RNA > 1200 & nCount_ATAC > 1200 & percent.mt < 20% & percent.ribo < 50%.

Mouse brain P2, P11: nCount_RNA > 800 & nCount_ATAC > 600 & percent.mt < 20% & percent.ribo < 50%.

Mouse brain P95, P365, P730: nCount_RNA > 200 & nCount_ATAC > 100 & percent.mt < 20% & percent.ribo < 50%.

Potential doublet cells were identified by DoubletFinder⁵⁶ with parameters (pN = 0.25, pK = 0.09, nExp = 0.054, PCs = 1:30) and removed from downstream analysis.

ChAIR data visualization

ChAIR data visualization tool ChAIR-Viewer was developed in-house and available at https://github.com/fengchuiguo1994/ChAIR-Viewer. ChAIR-viewer enabled the visualization of ensemble ChAIR-PET, -ATAC, and -RNA data together with ChromHMM state, and gene annotation track. Single-cell visualization of ChAIR-PET, -ATAC, and -RNA data can also be realized with four visualization modes.

ALL mode: Display unselected DNA, RNA, and PET signals within a specified region, providing a holistic overview of genomic activity.

CLEAN mode: Focus on displaying all RNA, ATAC, and a subset of PET signals related to the promoter region (TSS±5kb), for a more targeted visualization.

GENE mode: Focus on showing a subset of RNA and ATAC, and PET signals only associated to a specific gene, offering a gene-centric view of genomic interactions.

REGION mode: Display all PET signals for a predefined list of regions within a specific area, filtering out unrelated PET signals, which is useful for focused studies on specified genomic locations.

Normalization for interaction frequency of chromatin loops defined by ChAIR-PET data

For ensemble ChAIR-PET data, we identified significant chromatin loops with at least 3 PETs and supported by ChAIR-ATAC peaks. To normalize them in consideration of the sequencing depth, we applied VC_sqrt method to normalize chromatin loops based on the ChAIR-PET loops ensemble data in consideration of the sequencing depth using the formula:

$$VC\_sqrtscore=n∕\sqrt{x^{\ast }y}$$

where $n$ represents the interaction frequency (PET counts) for each loop, $x$ denotes the piled-up read counts associated with loop anchor 1, and $y$ indicates the read counts associated with loop anchor 2.

Reproducibility assessment

The ‘multiBigwigSummary’ function in deepTools⁵⁷ was used to assess the reads correlation in ensemble ChAIR-RNA (bin size = 10 kb), and -ATAC (bin size = 10 kb) data. For ensemble ChAIR-PET data, we calculated the stratum-adjusted correlation coefficient (SCC) between datasets using HiCRep⁵⁸ (bin size = 50 kb, hsmooth = 5, dBPMax = 2500000).

Bulk RNA-seq, ATAC-seq, and ChIATAC data processing

Raw sequencing reads were evaluated by FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc) and low-quality reads and adapter sequence were removed by Trimmomatic⁵⁹. Reads after trimming were mapped by HISAT2⁶⁰ for RNA-seq data and BWA⁶¹ for ATAC-seq data. ChIATAC data were processed using ChIA-PIPE⁵². MACS2 was used to call peaks and the peak intensity from MACS2 output was normalized with ‘SPMR’ option (signal per million reads). Significant loops were first called by ChIASig⁶² with PET ≥ 3 (PET distance > 8 kb) and FDR < 0.05. The loops were further filtered by anchor support; only significant loops with both anchors supported by peaks were retained.

Benchmark analysis

To compare ChAIR with published multi-omic (scRNA-seq+scATAC-seq) methods, the datasets generated from the same cells (K562) measuring scRNA-seq and scATAC-seq were taken from the source data in ISSAAC-seq⁶. The dsciATAC²⁵ data was downloaded from GSM3507342. sci-Hi-C²¹ data was downloaded from GSE84920.

Aggregate analysis of ChAIR-ATAC signal at peaks and TSS loci

To assess ATAC signals at the open chromatin loci (±3kb) and TSS loci in ensemble ChAIR-ATAC data and related methods, ChAIR-ATAC peak sites and TSS regions were sorted in descending order based on intensity, and these loci were used as the reference to match with all other datasets also in descending order.

Aggregate compartment and TAD analysis

The eigenvector values (bin size = 100 kb) and insulation score (bin size = 25 kb) were calculated using cooltools⁶³ and HiCExplorer⁶⁴, respectively. Aggregate compartment analysis was done using the ‘saddle’ function in cooltools. The compartment strength was defined according to previous study⁶⁵ and normalized to a scale ranging from −1 to 1. The aggregated signal of TADs was quantified based on the ratio of the mean signal value from a central 4×4 pixel area, representing the aggregated TAD signal, to the mean signal value of a 15×15 pixel area located at the upper right corner of the contact heatmap as background noise.

Aggregate chromatin loop analysis

We identified intra-chromosomal chromatin loops (PET ≥ 3, VC_sqrt_score ≥ 0.001) associated with S and G2/M marker genes in Patski cells and cell group marker genes (log₂FC > 0.5) in mouse brain. For chromatin loops with one anchor intersecting a marker gene's promoter, we checked their aggregated signal in the VCSQRT normalized 2D contact heatmaps (bin = 5 kb), applying the 'aggregate' function in FAN-C⁶⁶. To assess the strength of TSS–Enhancer interactions within a 180×180 pixel matrix, we calculated the ratio of the mean signal within a central 4×4 pixel area (representative of TSS-Enhancer interactions) to the mean signal within a 57×57 pixel background area positioned in the upper right corner.

ChromHMM analysis

Chromatin states of human K562 and mouse brain cells were downloaded from GSM936088 and https://github.com/gireeshogu/chromatin_states_chromHMM_mm9, respectively. The coordinates of the chromatin states were lifted over to hg38 for K562 and mm10 for mouse brain cells using LiftOver. The chromatin state enrichment was calculated by the following equation:

$$Enrichment=\frac{n}{N}/\frac{l}{L}$$

where $n$ means number of peaks in a state, $N$ means total number of peaks, $I$ means total length of a chromatin state, and $L$ means total length of all chromatin states.

Haplotype analysis

The SNP information of mouse strains C57BL/6J and SPRET/EiJ were taken from mouse genomes website https://ftp.ebi.ac.uk/pub/databases/mousegenomes/REL-1505-SNPs_Indels/strain_specific_vcfs/). All heterozygous SNPs were removed, and SNPs with different bases between two strains were defined as haplotype SNPs. ChAIR reads not overlapping with haplotype SNPs were removed. Reads containing more than 80% of the haplotype SNPs specific to either strain were assigned to their respective strain origin.

Cell-cycle phasing by chromatin contact

K562 and Patski cells were phased into different cell-cycle stages based on chromatin contact following previous method¹⁹. We followed metrics below to phase the cells:

Group	Cell-cycle phase	Sort order	Ordering by
1	Post-M	% mitotic ≥ 30 ∧ %near ≤ 50	-% near
2	G1	% near ≤ 63
3	Early to Mid-S	63 < % near ≤ 78.5
4	Mid-S to G2	% near > 78.5
5	Pre-M	% near > 50 ∧ %near + 1.8 X % mitotic > 100

Cell-cycle pseudotime analysis

To explore the cell cycle dynamics of K562 and Patski cells through their gene expression data, we initiated our study by assigning cell cycle scores to each cell, following the methodology outlined in the Seurat cell cycle vignette (https://satijalab.org/seurat/articles/cell_cycle_vignette.html). Then, we performed PCA and identified the centroids for each cell-cycle phase (G1, S, and G2/M) in the PCA space using the first two PCs. For our calculations, we defined the centroid of a cell-cycle phase as:

$${\{x,y\}}_{centroid}=\left\{\frac{{\sum }_{i=1}^n{x}_i}{n},\frac{{\sum }_{i=1}^n{y}_i}{n}\right\}$$

where $x_i$ means PC1 of cell $i$, $y_i$ means PC2 of cell $i$ and $n$ is the number of cells in the specific cell-cycle phase.

Further, we computed the Euclidean distance from a given cell (e.g., S cell) to the centroid of its previous cell-cycle phase (e.g., G1) and compared it with the distance to the centroid of the later cell-cycle phase (e.g., G2/M) to determine the differential distance (DD) as follows:

$$D{D}_i=\sqrt{{(x_i\text{-}{x}_{later})}^2+{(y_i\text{-}{y}_{later})}^2}\text{-}\sqrt{{(x_i\text{-}{x}_{pre})}^2+{(y_i\text{-}{y}_{pre})}^2}$$

where {$x_{\mathit{later}}$, $y_{\mathit{later}}$} means the centroid of the later cell-cycle phase, {$x_{\mathit{pre}}$, $y_{\mathit{pre}}$} means the centroid of the previous cell-cycle phase.

The cells were subsequently ranked according to their DD values, descending from the highest to the lowest within each phase. Cells with higher DD values are positioned closer to the previous phase, while those with lower DD values are nearer to later phase, indicating their progression through the cell cycle. For metacell analysis, we aggregated single cells from the same phase, starting from the lowest to the highest rank, to form metacells, each encompassing 1 million PETs.

3D genome modeling

We grouped cells from different cycle phases of Patski cells and various types of mouse brain cells into metacells, each containing one million PETs, for haplotype-revolved 3D modeling using dip-c package²². For mouse brain dataset, we reconstructed 3D genome structure using nuc_dynamics⁶⁷ version:1.3 (parameters: “-m 5 -f n3d -s 8.0 4.0 2.0 0.4 0.2 0.1”) with five replicate structures. The output was in “3D genome (3DG)” format (tab delimited: chromosome name, genomic coordinate (bp), x, y, z) because the original PDB format did not allow > 99,999 atoms. Each replicate started from different initial coordinates and involved repeated random sampling. Particles that harbored few contacts, such as centromeres and heterochromatic repeats, were removed from the final 3D structure. For each particle, the number of contacts within 0.5 Mb was recorded, the bottom 6% of all particles were removed. After removal of repetitive regions, shared genomic particles were extracted from the replicates and aligned with the Kabsch algorithm in a pairwise manner. We used dip-c package²² to convert 3DG to mmCIF for visualization in PyMol (run “set connect_mode, 4” before loading).

Chromatin volume estimation was performed using consecutive fragments, adopting an ellipsoid of inertia fitting approach akin to prior work¹⁹. The square roots of the eigenvalues of the covariance matrix were utilized to estimate the semi-axes of the ellipsoid. Then we used the formula V=4/3πabc with a, b and c denoting the ellipsoid semi-axes lengths to compute the ellipsoid volume. Here we used the volumes of the fitted ellipsoids of inertia instead of the semi-axis ratios as our focus lay on evaluating internal compactness rather than assessing chromosome elongation. The final metric employed was volume estimation, achieved through Gaussian kernel density estimation. This method involved calculating the volume occupied by the chromosome by summing the Gaussian density over all model beads exceeding a specified cutoff value. These calculations were conducted using the Chimera software⁶⁸ via the ‘molmap’ command with parameters ‘resolution=3’ and ‘cutoffRange=5’. Chromosome density was calculated as the ratio of chromosome size in base pairs to the occupied volume, according to the formula d=n/V, where d represents the density estimate, n is the chromosome size in base pairs, and V is the volume derived from the multivariate kernel density estimator. This density measure is inversely proportional to volume estimates obtained through the ellipsoid approach and the radius of gyration, exhibiting high concordance with these methods. It quantifies chromatin concentration in terms of base pairs or the number of polymer beads within the volume approximately occupied by the structure.

To provide more details, we analyzed both the active (chrXa) and inactive copies (chrXi) of the X chromosomes in Patski cells, examining the correlation of genome folding density, volume, and radius of gyration between chrXa and chrXi. Previous studies have shown that the two X chromosomes (active vs. inactive) display distinct levels of compactness, and the chrXi had been shown to be more compact (about 1.2-fold) than its active counterpart^69,70. Our characterization revealed that chrXa has lower chromatin density, larger volume and a greater radius of gyration compared to chrXi, and these patterns remained constant throughout the cell cycle (Figure shown below), as anticipated. This result confirms the accuracy of our methods using genome folding density and volume to estimate chromatin compaction.

Diffusion-based C_TG enhancement

To enhance the cell-cycle analysis of Patski metacells, we utilized the C_TG method⁷¹, a diffusion-based method with default parameters at 1M resolution. C_TG converts Hi-C contact matrix to C_TG distance matrix, a short C_TG distance indicates proximal in physical space and vice versa.

Unsupervised cell clustering, annotation, and trajectory inference

The dimensionality of the ChAIR-RNA data was reduced by PCA (30 components) first and then with uniform manifold approximation and projection (UMAP), followed by Louvain clustering using Seurat. We applied Harmony⁷² to integrate ChAIR-RNA data from different age points of mouse brain cells to mitigate batch effects.

To annotate cells, we integrated ChAIR-RNA with a published scRNA-Seq mouse brain data (http://www.mousebrain.org/adolescent/downloads.html). Cells were then annotated by using the ‘FindTransferAnchors’ and ‘TransferData’ functions in Seurat v4. We then identified cell type marker genes using the ‘FindAllMarkers’ function in Seurat v4. Cell types with fewer than 100 cells were removed from downstream analysis.

The developmental trajectory was learned from the UMAP of specific cell clusters and constructed using the ‘learn_graph’ and ‘order_cells’ functions in Monocle3⁷³ according to the vignette available at: https://cole-trapnell-lab.github.io/monocle3/docs/trajectories/.

Dimensionality reduction analysis for gene-centric ChAIR-PET data

Dimensionality reduction analysis using all PETs was done using BandNorm⁷⁴. We then took into consideration of the gene-centric nature of ChAIR-PET data, and the dimensionality reduction analysis was performed based on the adjusted gene-centric PET count associated with promoter and gene body. The PET count for every active gene in each cell was adjusted by normalized gene expression as follow:

$$\text{adjusted}PETcount=\frac{RNA}{RN{A}_{min}}{}^{\ast }PET\_count$$

where $\mathit{RNA}$ is the normalized UMI count of the gene, and ${\mathit{RNA}}_{\mathit{min}}$ is the normalized UMI non-zero minimum in all cell. The $\mathit{PET}\_\mathit{count}$ is the PET count in promoter (TSS±3kb) and gene body.

The dimensionality of all PETs and the adjusted gene-centric PETs were reduced by PCA (30 components) first and then with UMAP using Seurat.

Cell Identity Score analysis

Current cell type annotation in single-cell data relies almost exclusively on scRNA-seq data. To fully delineate cell identity using multi-omic data, we characterized cells with additional genomic features. Based on marker genes (with log₂ fold change > 0.5), we extracted 3D epigenomic features associated with the marker genes in each of the ChAIR-RNA defined cell groups, including promoter-associated chromatin contact data (intra-chromosomal PETs ≥ 3, VC_sqrt_score ≥ 0.001, and distance < 2 Mb), ATAC signals overlapping with the gene body (coding regions), promoters, and distal enhancers connected to marker gene promoters. Thus, the features analyzed included following features:

1. Transcription (T)

2. Gene body ATAC signal (A)

3. Promoter ATAC signal (P)

4. Enhancer ATAC signal (E)

5. Marker gene associated promoter-enhancer loops (L)

We then focused on 23 out of 45 cell groups²⁴ that possess sufficient data for quantitative analysis of the matched features. To assess the power of each of the genomic features for specificity of cell identity, we devised Cell Identity Score (CIS) to assess the uniqueness of cellular features critical in distinguishing among different cell groups. First, we generated normalized cell group-specific feature signal matrices (NFM) for each mono-modal feature and multi-modal feature. The cell group-specific feature signal was normalized relative to the total signal of all cell group-specific features. We used following formula for generating mono-modal (e.g., T) NFM:

$$\text{mono}NF{M}_{mn}=\frac{T_{mn}}{{\sum }_{i=1}^m{T}_{in}}$$

For generating NFM for multi-modal feature (e.g., L*A*P*E), we used formula:

$$\text{multi}NF{M}_{mn}=\frac{L_{mn}{}^{\ast }A_{mn}{}^{\ast }P_{mn}{}^{\ast }E_{mn}}{{\sum }_{i=1}^m(L_{in}{}^{\ast }A_{in}{}^{\ast }P_{in}{}^{\ast }E_{in})}$$

where “$mn$”, $m$ represents the rows of the matrix, each corresponding to signal from one of the 23 cell groups; $n$ represents the columns of the matrix, each corresponding to a specific feature related to cell identity. For example, $T_{mn}$ represents the row $m$ and column $n$ of the marker gene expression matrix. $i$ represents row $i$ of the matrix, so $in$ represents row $i$ and column $n$ of the matrix.

To measure the distinguishing capacity for individual cell group-specific feature, we calculate the individual CIS using the NFM and the following formula:

$$\text{Individaul}CI{S}_i=NF{M}_{ij}(i=j)∕\frac{{\sum }_{j=1}^{n}NF{M}_{ij}}{(n\text{-}1)}(i\ne j)$$

To measure the distinguishing capacity for all cell group-specific features, we evaluated the distinction between unique signals (reflected by diagonal elements of the NFM) for each cell type and the general background (represented by the off-diagonal elements) using the formula:

$$\text{Group}CIS=\frac{{\sum }_{i=1}^m{\sum }_{j=1}^{n}NF{M}_{ij}}{n}(i=j)∕\frac{{\sum }_{i=1}^m{\sum }_{j=1}^{n}NF{M}_{ij}}{n(n\text{-}1)}(i\ne j)$$

where “$mn$”, $m$ represents the rows of the matrix, each corresponding to signal from one of the 23 cell groups; $n$ represents the columns of the matrix, each corresponding to a specific feature related to cell identity. For example, ${\mathit{NFM}}_{mn}$ represents the row $m$ and column $n$ of the NFM. $ij$ represents row $i$ and column $j$ of the matrix.

Each of mono-modal genomic features could have its own CIS value, and the mono-modal features could be integrated to generate composite 3D epigenomic feature by multiplying the signals from aforementioned features (i.e., L*A, L*P, L*E, L*A*E, L*P*E, L*A*P*E) as seen in figure below.

Cell type-specific enhancer identification

To identify potential cell type-specific enhancers for transcription activation in mouse brain cells, we examined 23 cell groups with sufficient data. We first identified cell specific peaks and connected them to their target genes by leveraging intrachromosomal interactions within a 2 Mb range, requiring at least three PET-supported promoter-enhancer (P-E) interactions with ATAC peak support. As a result, we identified 562 high-quality enhancer loci that displayed high cell type specificity (Extended Data Fig. 5a). Overall, the majority of cell type-specific CREs in this set were located in non-coding regions of the genome: approximately 58% were in distal intergenic regions, 17% in introns, and 18% in promoter regions of other genes, leaving merely 5% in exons (Extended Data Fig. 5b). ChromHMM analysis further revealed that these loci were associated with transcription active states, with the highest enrichment for strong enhancers (Extended Data Fig. 5c). More details can be found in Extended Data Fig. 5d-j.

Signal enhancement by HICSR

As the number of useable reads for certain cell types in mouse brain dataset was relatively low for both accurate identification and visualization of key 3D genomic features, we generated an enhanced version of the original contact maps using HiCSR⁷⁵, a deep-learning-based framework designed to computationally enhance Hi-C contact signals when sequencing depths are limited.

Specifically, raw contact maps at various resolutions (10 kb, 25 kb, 50 kb, 100 kb, 250 kb, 500 kb, and 1 Mb) were extracted from the multi-resolution cooler (.mcool) files. These contact signals were then normalized to the range [−1, 1] for each chromosome using the following equations:

$$X^c={\log }_2(1+M^c)$$

$$T^c=\frac{2{}^{\ast }X^c}{{\max }_{i,j}\{X_{i,j}^c\}}\text{-}1$$

Where $M^c$ represents the raw contact matrix of chromosome $c$ and $T^c$ represents the transformed matrix.

Subsequently, the normalized contact maps ($T^c$) were fed into HiCSR. A pre-trained HiCSR model with default parameters (input_size=40, output_size=28) was employed to enhance the signal strength within the contact maps.

Since the HiCSR output ($T^{c{,}^{\prime }}$) also falls within the [−1, 1] range, it needed to be converted back to the scale of raw interaction frequencies typically observed in high-coverage Hi-C experiments. To achieve this, a reference contact matrix file from a DNase Hi-C library containing around 450 million usable reads of the mouse brain tissue was downloaded from the 4DN data portal⁷⁶ with the accession number 4DNFIM38DJNE. $M^{c,H}$ represents the raw contact matrix for chromosome $c$ in this reference library. The raw interaction signals for $T^{c{,}^{\prime }}$ were estimated using the following equation:

$$X^{c,H}={\log }_2(1+M^{c,H})$$

$$T^{c,H}=\frac{2{}^{\ast }X^{c,H}}{{\max }_{i,j}\{X_{i,j}^{c,H}\}}\text{-}1$$

$$V_{max}=\frac{2}{n}{\sum }_{\text{-}0.01<T_{i,j}^{c,H}<0.01}{X}_{i,j}^{c,H}$$

$$M^{c{,}^{\prime }}=2\frac{(T^{c{,}^{\prime }}+1){}^{\ast }V_{max}}{2}\text{-}1$$

Finally, the enhanced contact maps ($M^{c{,}^{\prime }}$) were written back to the .mcool format using the “toCooler” command from the HiCPreaks package available at https://pypi.org/project/hicpeaks/.

Integrate ChAIR with Stereo-seq

To examine the spatial distribution patterns of the identified cell types in P95 ChAIR data from mouse brain cells, we mapped ChAIR RNA data (excluding cell types from cerebellum, olfactory bulb, and hind brain) onto Stereo-seq data using ‘FindTransferAnchors’ and ‘TransferData’ functions in Seurat. As a result, each bin50 spot within the Stereo-seq data was annotated with a corresponding cell type from the ChAIR dataset. We then manually validated marker gene expression by comparing it with in situ hybridization data (ISH) from the Allen Brain Atlas (atlas.brain-map.org). Only the cell types that demonstrated the expected marker gene expression in the ISH data were retained for downstream analysis.

Megacontact pseudotime analysis

Mouse brain cells were sorted according to the proportion of chromatin contacts distances greater than 2Mb, arranging them from the lowest to the highest percentage. Contact decay profiles for each cell as previously described⁶⁵ were generated for visualization. Excitatory neurons with the highest proportion of ultra-long metacontacts were set as the starting point in the chromatin contact pseudotime analysis. Conversely, non-neurons with the lowest proportion of ultra-long megacontacts were placed at the beginning for their respective analyses.

Transcriptomic age analysis

To assess the biological age of cells, we employed the SCALE using mouse ageing-associated genes from AgeingMap⁴⁵. SCALE assigns each gene a weight based on its correlation with chronological age and the proportion of cells expressing it. The transcriptomic age of each cell is then calculated by taking the dot product of gene expression z-scores and their respective weights.

Extended Data

Extended Data Fig. 1. Benchmark ChAIR data derived from K562 and Patski cells with related methods.

(a-d) Violin plots comparing ChAIR with other single-cell methods for the numbers of RNA unique molecular identifiers (UMIs) (a), counts of identified genes (b), numbers of captured chromatin fragments in ATAC-seq data (c), and numbers of ATAC-seq peaks (d). (e-f) Violin plots comparing ChAIR with Hi-C based single-cell methods in detecting chromatin contacts (e) and the ratio of contacts associated with ATAC peak sites (f). Medians for each data category and the total number of cells analyzed by different methods were provided. Center line indicates median, asterisk indicates average, box represents interquartile range (25th to 75th percentiles), and whiskers extend to 1.5× interquartile range.

Extended Data Fig. 2. Characterization of ChAIR data.

(a) Percentage of ChAIR-RNA reads that were mapped to exon, intron, and intergenic regions and metagene profiles (reads per kilobase per million mapped reads, RPKM) in K562 (top) and Patski cells (bottom) in reference with bulk RNA-seq data. (b) Signal enrichment of ChAIR-ATAC data, 10x ATAC, ChIATAC, bulk ATAC-seq, and sci-Hi-C data in K562 (top) and Patski (bottom) cells, at ATAC peak (left) and TSS sites (right). (c) Categories of ChromHMM defined chromatin states for open chromatin loci identified by bulk ATAC-seq data (n = 176,891) (left) and ChAIR-ATAC data (n = 201,540) (right) in K562. (d) Categories of ChromHMM defined chromatin states for chromatin loops in ChIATAC (n = 37,579) (left) and ChAIR-PET data (n = 232,988) (right). (e) Chromatin contact distance distribution of chromatin loops with different interaction frequencies in ChIATAC and ChAIR-PET data in K562 and Patski cells.

Extended Data Fig. 3. Cell cycle-specific gene expression and chromatin interactions.

(a) Scatter plots showing gene expression of marker genes specific to S and G2/M phases from ChAIR-RNA data in K562 and Patski cells. (b) PCA plots showing the distribution of individual K562 cells based on their cell-cycle marker genes (left) and further grouped into metacells based on cell-cycle pseudotime (right). (c) 2D contact heatmaps of ChAIR-PET data in K562 and Patski cells at G1, S, and G2/M stages. (d) Aggregate compartment analysis of ChAIR-PET data in K562 (top) and Patski (bottom). (e) Aggregate TAD analysis of ChAIR-PET data in K562 (top) and Patski (bottom). (f) Aggregate plots of chromatin contacts between promoters (TSS) and putative distal enhancer elements (E) for phase-specific marker genes. The intensity values of phase-specific TSS-E interactions were given for comparison.

Extended Data Fig. 4. Single-cell analysis of ChAIR data derived from mouse brain cells.

UMAP plots of ChAIR (P365) data from the following modalities: (a) Mono-modal ChAIR-RNA data, (b) Mono-modal ChAIR-ATAC data, (c) ChAIR-PET (all interactions) data, (d) ChAIR-PET (gene-centric interactions) data, and (e-g) the combination of different modalities. (h) Cell type-specific ChAIR-RNA profiles in matrix plot showing canonical marker gene expression profiles in 7 major brain cellular classes and in 121 cell types. (i) Cell type-specific ChAIR-ATAC profiles in matrix plot showing cell type-specific ATAC peak signals in 7 classes and in 121 cell types.

Extended Data Fig. 5. Characterization of cell type-specific enhancers identified by ChAIR.

(a) The matrix of ATAC signals at cell type-specific enhancer loci (n = 562) across 9 cell groups. (b) Genomic annotations. (c) Chromatin states annotated by ChromHMM. (d-e) Number of enhancers in different cell groups (d) and number of associated cell type-specific marker genes (e). (f) Distances from enhancers to their nearest genes and to the targeted marker genes. The P value was calculated using the two-sided Wilcoxon test. (g-h) Numbers of enhancers represented in each cell group (g) and associated with each target gene (h). (i-j) Mean distance of enhancer to target genes in different cell groups (i) and for individual genes (j). Center line in violin plot indicates median, asterisk indicates average, box represents interquartile range (25th to 75th percentiles), and whiskers extend to 1.5× interquartile range.

Extended Data Fig. 6. Validation of the correlation between megacontact, key cellular features, and nuclear volume using image-based measurements.

(a-d) The correlation between ChAIR data derived measurements (Y-axis) and nuclear volumes (X-axis), inferred from DAPI-stained imaging data. The ChAIR data derived measurements included: (a) megacontact ratio, (b) genome folding density inferred from 3D models, (c) global transcriptional activity, and (d) global chromatin accessibility. The fitted lines were smoothed by linear models with shading indicating the confidence interval. (e-g) P95 ChAIR data mapped megacontact ratio (e), global transcriptional activity (f), and chromatin accessibility (g) in 17 mouse brain cells types. Notably, megacontact was negatively correlated with global transcriptional activity (measured by RNA-MERFISH40), chromatin accessibility, and nuclear volume (inferred by DNA-MERFISH40). Center line denotes median, box represents interquartile range (25th–75th percentiles), and whiskers extend to 1.5× interquartile range.

Extended Data Fig. 7. Validation of integrating ChAIR and Stereo-seq data.

(a-c) Canonical marker gene expression profile in ChAIR-RNA (a) and Stereo-seq data (b-c). Stereo-seq data with cell type annotations (b) and anatomic region annotations (c) were shown. (d) Spatial visualization of cells from various cerebrum cortex layers and the gene expression of marker genes in Stereo-seq and Allen Brain Atlas in situ hybridization (ABA ISH) data (atlas.brain-map.org). (e) Cell type-specific feature (that is, transcription, chromatin loop, promoter ATAC, enhancer ATAC, gene activity, and the combination of these information) signal matrices calculated by different features across region-resolved cell types. Normalized signal intensities were provided. Group CIS was provided to measure the overall specificity of the feature examined across all cell types.

Extended Data Fig. 8. Examples of chromatin rewiring in neurons during differentiation.

(a) Contact distance spectrum of cerebellar ExNs (CBNBL2 and CBGRC) at P2 stage. Cells were arranged from the least extent of megacontact to the highest. Contact frequencies were normalized to 0-1. (b) UMAP plots of ChAIR data with RNA and chromatin megacontact pseudotime information in P2 ChAIR data. (c) 3D models of genome folding architectures reconstructed from ChAIR-PET data for CBNBL2 and CBGRC at P2 stage. (d) Gene expression profiles of CBNBL2-specific gene (Igfbpl1) and CBGRC-specific gene (Cbln1) along the megacontact pseudotime. The fitted curves were smoothed by generalized additive models with shading indicating the confidence interval. The analyses done in (a-d) were also applied to InhNs from OB in (e-g).

Extended Data Fig. 9. Nuclear volume dynamics of ExNs during ageing.

(a) Example views of 3D models of genome folding architectures in TEGLU and CBGRC cells across five age points. (b-d) Boxplots of megacontact percentage (b), genome folding density inferred from the 3D models of genome folding architectures (c), and the global transcriptional activity measured in UMI (d) for TEGLU (red) and CBGRC (blue) cells, based on ChAIR data. Center line in boxplot denotes median, box represents interquartile range (25th–75th percentiles), and whiskers extend to 1.5× interquartile range.

Data availability

All datasets generated in this study including ChAIR and ChIATAC data have been submitted to Genome Sequence Archive in National Genomics Data Center with accession number PRJCA024774. Histological staining results used in this study are available from the Allen Brain Atlas (atlas.brain-map.org).

Code availability

Pipeline for processing ChAIR data (ChAIR-PIPE) is available at https://github.com/fengchuiguo1994/ChAIR-PIPE. ChAIR data visualization tool ChAIR-Viewer is available at https://github.com/fengchuiguo1994/ChAIR-Viewer.