SATB2 organizes the 3D genome architecture of cognition in cortical neurons

Summary

The DNA-binding protein SATB2 is genetically linked to human intelligence. We studied its influence on 3D epigenome by mapping chromatin interactions and accessibility in control versus SATB2-deficient cortical neurons. We find that SATB2 affects the chromatin looping between enhancers and promoters of neuronal activity-regulated genes, thus influencing their expression. It also alters A/B compartments, Topologically Associating Domains and Frequently Interacting Regions. Genes linked to SATB2-dependent 3D genome changes are implicated in highly specialized neuronal functions and contribute to cognitive ability and risk for neuropsychiatric and neurodevelopmental disorders. Non-coding DNA regions with SATB2-dependent structure are enriched for common variants associated with educational attainment, intelligence and schizophrenia. Our data establish SATB2 as a cell-type specific 3D genome modulator, which operates both independently and in cooperation with CTCF to set up the chromatin landscape of pyramidal neurons for cognitive processes.

Graphical Abstract

Wahl et al. establish SATB2 as a cell-type specific 3D genome modulator in cortical neurons. It acts independently or in cooperation with CTCF to modify the 3D genome architecture. SATB2 affects specifically genetic loci involved in synaptic function, cognition, learning and memory.

Introduction

Emerging evidence suggests a role of spatial genome organization in both normal cognition and cognitive impairment¹. Recent studies on candidate genes indicate that organization of DNA in chromatin loops contributes to genetic risk architecture of cognitive diseases with early childhood/young adulthood onset, including autism spectrum disorder (ASD) and schizophrenia (SCZ)^2,3. Structural DNA variants associated with neuropsychiatric disorders are positioned in intergenic or intragenic non-coding sequences and influence neuronal gene expression by bypassing the linear genome to directly interact with target genes^4,5. Cell type-specific nuclear mechanisms affecting the higher-order chromatin landscape of gene regulatory networks in neurons are not well understood. We investigate this by studying the function of SATB2, a nuclear protein that in the CNS is selectively expressed in pyramidal neurons of cortex and hippocampus^6–8, two brain regions crucial for cognitive function.

A growing body of evidence has linked SATB2 to higher brain functions. Genome-wide association studies (GWAS) have identified SATB2 as a locus associated with human cognitive ability⁹. Common variation in the genes regulated by SATB2 or encoding SATB2-interacting proteins influences human cognitive ability and contributes to SCZ^10,11. Rare de novo mutations within the SATB2 locus cause SATB2-associated syndrome, an autosomal dominant disorder characterized by severe intellectual disability¹². Mouse mutants, in which SATB2 is deleted in adult forebrain pyramidal neurons, display cognitive defects^13,14. Potent effects of SATB2 on gene transcription include regulation of neuronal activity-dependent immediate early genes (IEGs)¹⁵. SATB2 interacts with proteins implicated in de novo chromatin loop formation via CCCTC-binding factor (CTCF)/cohesin-independent mechanism^11,16. As a nuclear matrix protein SATB2 binds to matrix/scaffold attachment regions within the genome^8,17. These findings suggest a potential function as a modulator of loop formation and three-dimensional (3D) chromatin architecture. While recent studies in non-neuronal cell types have provided evidence for alterations in chromatin accessibility and individual chromatin loops upon SATB2 loss^18–20, a genome-wide analysis of SATB2 effects on 3D chromatin structure is currently lacking.

Our study identifies epigenetic profiles and higher-order chromatin interactions that depend on SATB2 in cortical pyramidal neurons by integrating high-resolution, multidimensional datasets from Satb2 conditional knockout (cKO) and floxed mice. We observe highly coordinated alterations in chromatin accessibility and chromatin loop landscape upon SATB2 deletion. SATB2 effects on 3D genome organization extend beyond chromatin loops to large-scale architectural levels. With remarkable specificity, SATB2-mediated 3D epigenome modelling occurs at genomic loci that are functionally associated with cognition and contribute to human cognitive ability and risk for neuropsychiatric and neurodevelopmental disorders.

Results

SATB2 deletion in cortical neurons causes extensive alterations in chromatin accessibility and chromatin loop landscape

We mapped SATB2 binding sites, SATB2-dependent chromatin interactions and accessibility in cortical cultures derived from Satb2^flx/flx::Nes-Cre(Satb2 cKO) and Satb2^flx/flx (floxed) mice¹¹ (Fig. S1A). ATAC-seq, CUT&RUN, and Hi-C libraries were highly reproducible across replicates from each genotype²¹ (Fig. S1B-D). Hi-C libraries were also highly similar to an in vivo Hi-C dataset derived from NeuN+-sorted neurons of adult mouse cortex²² (Fig. S1D). Satb2 cKO and floxed cultures did not differ morphologically (Fig. S2A) and SATB2 loss-of-function was reversible since its strong effects on gene transcription¹⁵ were rescued by AAV-mediated acute reintroduction in cKO cultures (Fig. S2B, S2C). Analysis of SATB2 genome-wide occupancy by CUT&RUN identified a total of 14964 binding sites, of which 7819 overlapped with promoters (Fig. 1A, 1B). Chromatin state modeling by ChromHMM using neonatal mouse forebrain datasets^23,24 indicated SATB2 peaks to be predominantly associated with active chromatin states and enhancers (Fig. 1C). Gene Ontology (GO) analysis of the genes with SATB2-bound promoters showed an overrepresentation of terms related to synapse organization, regulation of membrane potential and ion transmembrane transport (Fig. 1D). Profiling accessible chromatin in cKO versus floxed neurons unveiled a robust effect of SATB2 on open chromatin landscape (Fig. S3A). Over 28% of all identified open chromatin regions (OCRs) exhibited differential accessibility between genotypes (differentially accessible OCRs, dOCRs, 25651 peaks). The remaining 64266 peaks were invariant across genotypes. While invariant OCRs exhibited similar enrichments for enhancer and promoter states, dOCRs exhibited a six-fold higher enrichment for enhancers in comparison to promoters (Fig. 1E). Of all dOCRs, 12280 displayed decreased accessibility in cKO compared to floxed neurons (lost dOCRs), while 13371 showed greater accessibility in cKO relative to floxed neurons (gained dOCRs). Both types of dOCRs were enriched for SATB2 binding sites (Table S1) and showed an increased SATB2 CUT&RUN signal compared to neighboring regions (Fig. S3B). Analysis of TF binding site enrichment revealed substantial differences between lost and gained dOCRs. Motifs for AP-1 and RFX families of TFs had the highest footprinting score at lost dOCRs, whereas motifs for homeobox TFs and ARID3B had elevated footprinting at gained dOCRs (Fig. 1F). Next, genes near invariant and dOCRs were functionally annotated using as custom background genes that map to any OCRs. For the genes mapped to invariant OCRs, this analysis revealed a prevalence of housekeeping functions such as metabolic processes (Fig. S3C). In contrast, genes near both lost and gained dOCRs were enriched in synapse- and development-related pathways, including regulation of membrane potential, synapse organization and neuron projection development (Fig. S3D, S3E). Thus, both SATB2 peaks and dOCRs were strongly and specifically associated with genes encoding neurotransmitter receptors, ion channels, and synaptic organization proteins. We next studied SATB2’s impact on chromatin looping in merged replicate Hi-C maps. Out of 33225 loops called in the floxed Hi-C sample, 11847 (36%), were present in floxed but weakened or absent in the cKO sample (lost loops). Conversely, 9230 out of 29656 identified loops (31%) were detected in cKO but weakened or absent in the floxed sample (gained loops) (Fig. 1G, 1H). Anchors of both lost and gained loops were significantly enriched for SATB2 binding sites and over 15% of them overlapped with at least one SATB2 peak (Table S1). To explore the relationship between differential loops and dOCRs, we intersected the differential loop anchors with dOCRs. Lost loop anchors were highly enriched for lost dOCRs, whereas gained loop anchors significantly overlapped with gained dOCRs (Fig. 1I, Table S1). Therefore, changes in chromatin accessibility and chromatin looping can be coordinated by a single protein, such as SATB2. Invariant loop anchors were not enriched for dOCRs, supporting the specificity of the correlation between differential looping and differential accessibility.

Figure 1: SATB2 deletion in cortical neurons causes 3D epigenome remodeling.

A. Genomic distribution of SATB2 peaks.

B. Enriched genomic annotations of SATB2 peaks, calculated by HOMER, ***P < 0.001

C. Enriched chromatin states of SATB2 peaks, inferred by P0 forebrain ChromHMM 18-state model, for all enrichments and depletions, P < 0.00001.

D. GO analysis for biological processes of gene promoters occupied by SATB2. The top ten GO terms are displayed.

E. Enrichment analysis of dOCRs and invariant OCRs over chromatin states, inferred by ChromHMM 18-state model, for all enrichments and depletions, P < 0.00001.

F. Occupancy prediction of TFs at OCRs. TF motifs with higher footprinting score in floxed (blue) and cKO (red) samples are colored.

G. Aggregate peak analysis centered at differential and invariant loops in cKO and floxed Hi-C contact matrices (right) and differential plots (left), illustrating the log2 fold change in loop strength in cKO vs floxed neurons.

H. Representative 5kb-resolution Hi-C matrix plots with gene (black) and SATB2 peak (blue) tracks. Lost differential loops (purple) and gained differential loops (green) are encircled.

I. Genome browser tracks of the same regions (as in Fig. 6H), illustrating the overlap between dOCRs and differential loop anchors (grey). ATAC-seq tracks from cKO neurons are colored in green, from floxed neurons in purple. dOCRs are shown as colored bars, differential loops are presented as arcs.

SATB2 links regulatory elements to promoters of activity-regulated and cognition-associated genes

Since SATB2 binds predominantly to promoters (Fig. 1A), we first examined promoter-based SATB2-driven alterations in chromatin accessibility and looping. Integrating differential ATAC-seq and Hi-C profiles unveiled two types of chromatin loops that connect distal elements to promoters in a SATB2-dependent manner: invariant loops linking dOCRs to gene promoters, and differential loops connecting distal elements to promoters (Fig. 2A, Fig. S4A-C, Table S2). By overlapping invariant loop anchors with dOCRs and gene promoters, 472 genes were linked to lost dOCRs and 646 genes to gained dOCRs (Table S3). Both gene sets were enriched for SATB2-dependent differentially expressed genes (DEGs), as previously identified in cKO vs. floxed cortical neurons¹⁵ (Table S1). A tight correlation was observed between SATB2-driven open chromatin and gene expression changes. DEGs assigned to lost dOCRs displayed decreased expression in cKO neurons, while DEGs assigned to gained dOCRs were more strongly expressed in cKO neurons (Fig. 2B). Among the genes connected to lost dOCRs, neuronal activity-regulated genes (ARGs)²⁵ were strongly overrepresented (Table S1), consistent with the previously observed decreased ARG expression and impaired IEG response in cKO neurons¹⁵. AP-1 TFs are key regulators of activity-driven transcription²⁶. AP-1 TF motifs were strongly enriched at promoter-interacting lost dOCRs (Table S4), suggesting that decreased accessibility of AP-1 sites at lost dOCRs is responsible for the deficient IEG response in cKO neurons. Motifs for MEF TFs, known to be important for synapse formation and elimination²⁷, were also enriched at promoter-interacting lost dOCRs (Table S4). Correspondingly, MEF2C-regulated genes²⁸, including Egr3, Npas4, and Kcna4²⁹ were overrepresented in the promoter-interacting lost dOCR gene set (Table S1), which was also strongly enriched for synapse- and memory-linked GO terms (Fig. 2C). Similar GO terms were nominally significant when genes connected to any OCRs via invariant loops were used as background (Table S5). Genes interacting with gained dOCRs were enriched for terms associated with forebrain development, regulation of neurogenesis, neuron migration and differentiation (Fig. 2C). Several genes expressed in neural progenitors but not in SATB2-positve mature neurons, such as Nrp2, Gdpd5, and Gja1^30–32, interacted with gained dOCRs and correspondingly were found to be upregulated upon SATB2 loss¹⁵. Thus, SATB2-mediated decreased accessibility of their regulatory elements is a potential mechanism for downregulation of these genes during late cortical development.

Figure 2: SATB2-dependent regulatory element-promoter interactions involve synapse- and cognition-associated genes.

A. Schematic representation of the two types of SATB2-dependent chromatin interactions, connecting distal elements to promoters.

B. Fold change distributions of the DEGs linked to lost and gained dOCRs via invariant loops. P value is calculated using non-parametric Mann-Whitney test.

C. GO analysis of genes interacting with lost (left) and gained dOCRs (right).

D. Representative 5kb-resolution HiC map of a 500kb region of Chr17, showing ICED-normalized contacts in floxed (upper right triangle) and cKO (lower left triangle) neurons. Lost loops are encircled in purple. Gene track (black) with Dusp1 locus highlighted in red as a representative gene mapped to promoter-based lost loops. Peak track (blue) shows SATB2 binding sites.

E. GO analysis of genes assigned to promoter-based lost loops.

F. Fold change distributions of the DEGs assigned to promoter-based lost and gained loops. P value is calculated using non-parametric Mann-Whitney test.

G. Enrichment analysis of distal elements using ChromHMM-infered states and histone marks (P0 forebrain). Color bar shows the z-score value, for all enrichments and depletions, P < 0.001.

H. Bar chart illustrating the percentage of enhancer-promoter or promoter-promoter interactions within the two types of SATB2-dependent chromatin loops, connecting distal elements to promoters.

I. Occupancy prediction of TFs across OCRs at promoter-interacting distal elements. TF motifs with higher footprinting score in floxed are shown in blue and for cKO samples in red.

By intersecting differential loop anchors with promoters, 2978 genes were assigned to lost loops and 2133 genes to gained loops (Fig. 2D, Table S3). The lost loop gene set was specifically and highly enriched for terms linked to axons, synapses, cognition and forebrain development (Fig. 2E). By contrast, the gained loop gene set did not exhibit any significant GO enrichment and thus appeared to be mostly random. Intriguingly, the gene set assigned to lost loops and the gene set mapped to invariant loops connecting lost dOCRs to promoters showed little overlap in their composition, however they shared striking functional similarities. Both overlapped significantly with SATB2-dependent DEGs¹⁵, ARGs ²⁵ and MEF2C-regulated genes²⁸ (Table S1). Furthermore, the DEGs within both gene sets were expressed at lower levels in floxed vs cKO neurons (Fig. 2B, Fig. 2F) and the ATAC-seq peaks within the distal targets of the two types of loops exhibited enrichment for AP-1 and MEF TF motifs (Table S4).

In a subset of lost and gained promoter-based loops (43% and 52%, respectively), the same promoter was engaged in both gained as well as lost interactions (Fig. S4D, Table S2). Genes with both lost and gained loops at their promoter (Table S3), including Rgs7, Cacnalc, and Srgap2, overlapped significantly with the DEGs in cKO vs floxed cortical neurons¹⁵ (Table S1), and were highly significantly enriched for neuron-specific and synaptic GO terms (Fig. S4E). In another subset of SATB2-dependent promoter-based interactions, both the loop connecting the distal element to the promoter and the regulatory element itself (dOCR) were differential. Out of 318 genes mapped to this type of promoter interactions, 123 genes overlapped with SATB2-dependent DEGs, of which 105 were down-regulated in the cKO. This is in agreement with decreased expression caused by loss of interactions with active distal elements. Strong enrichment of neuronal GO terms, such as glutamatergic synapse, modulation of chemical synaptic transmission, and behavior was observed also for this gene set (Fig. S4F).

Finally, we characterized the epigenetic features of the distal targets connected to promoters in a SATB2-dependent manner (Fig. 2A), by using ChromHMM-defined chromatin states, neonatal forebrain H3K27Ac and H3K4me3 ChIP-seq data^23,33,34 as well as the SATB2 binding sites from this study. Distal targets were enriched for SATB2 peaks (Table S1), and for enhancer and promoter states (Fig. 2G, 2H), suggesting that SATB2-dependent chromatin loops are primarily of enhancer-promoter or promoter-promoter type, potentially forming neural enhancer-promoter aggregates³⁵. Notably, there were significant differences in TF binding site enrichment at the distal targets of both types of loops. Motifs for AP-1 and RFX proteins had the highest footprinting scores at lost dOCRs, whereas homeobox TF motifs and ARID3B showed elevated footprinting at gained dOCRs within these loop targets (Fig. 2I). In summary, by integrating Hi-C and ATAC-seq data in cKO vs floxed cortical neurons, we found that SATB2 orchestrates distal element-promoter interactions by: 1) modulating enhancer accessibility within stable, invariant chromatin loops and 2) influencing chromatin looping between promoters and distant enhancers. These SATB2-driven 3D epigenome changes occurred at genes responsible for specific synaptic functions and were closely aligned with gene expression.

Non-promoter-based differential loops occur at genomic regions occupied by CTCF

Next, we analyzed the characteristics of non-promoter (non-P)-based differential loops, which represented 22.4% and 19.6% of all chromatin loops called in Flx and cKO neurons. With a median length of ~355 kb these loops belong to the sub-megabase architectural level, for which CTCF is established as a major organizer of 3D topology³⁶. Therefore, we investigated potential links between CTCF and SATB2. In total, 37% of the SATB2 peaks identified in this study overlapped with CTCF peaks defined in neonatal mouse forebrain²³. We found an enrichment for CTCF but not SATB2 peaks at the anchors of both gained and lost non-P-based loops (Table S1, Fig. 3A). Since CTCF-dependent loops are regulated by multi-component protein complexes^37,38, we tested the potential of SATB2 and CTCF to cooperate directly or indirectly, via common co-interactors²³. The comparison of published SATB2 protein interactors in cortical neurons ⁿ, CTCF protein interactors^39,40 and CTCF loop participants/co-localizing factors³⁸ yielded a highly significant overlap (Fig. 3B). Co-immunoprecipitation of SATB2 with CTCF was observed in HeLa cells overexpressing both proteins (Fig. 3C). A similar co-IP with CTCF has previously been described for SATB1⁴¹. Thus, the effects of SATB2 on non-P-based differential loops might occur via modulation of CTCF effects on loops by interaction of both proteins in multi-protein complexes.

Figure 3: Non-P-based differential loops indicate SATB2 interactions with CTCF.

A. Average CTCF ChIP-seq signal over anchors of non-P-based SATB2-dependent differential loops.

B. SATB2 protein interactors¹¹ overlap significantly with a set of CTCF co-binding factors³⁸/CTCF protein interactors^39,40. Fisher’s exact test was used for statistical analysis, OR, odds ratio.

C. Co-immunoprecipitation of SATB2 and CTCF in HeLa cells overexpressing SATB2-V5, FLAG-CTCF or both. Co-IP was carried out using SATB2 antibody. Precipitates were analyzed by Western blotting, using antibodies against SATB2, CTCF and GAPDH. Representative example of 4 independent experiments. I= input, F= flow through, B= eluted beads.

SATB2 influences the 3D genome architecture of cortical neurons at multiple hierarchical levels

To gain insights into the global 3D genome folding pattern upon SATB2 loss, we explored distance-dependent chromatin interaction frequencies. Genome-wide average contacts decayed stronger with distance in cKO vs floxed Hi-C contact matrices at all distances (Fig. 4A). Accordingly, we observed weakened compartmentalization in cKO neurons of both inactive (B) and active (A) compartments (Fig. 4B, 4C). At the level of Topologically Associating Domains (TADs), the intra-TAD interaction frequency was decreased in the cKO compared to floxed Hi-C matrix (Fig. 4D). Furthermore, interactions between neighboring TADs were diminished in cKO compared to floxed matrices (Fig. 4E).

Figure 4: SATB2 deletion modifies global 3D chromatin architecture.

A. Chromosome-averaged distance-decay plots, illustrating intra-chromosomal contact frequency as a function of genomic distance in cKO and floxed Hi-C matrices.

B. Saddle-plots showing compartmentalization of the genome in cKO (left) and floxed (middle) samples. Shown are the average observed/expected (O/E) values of contact enrichment between pairs of 100 kb bins arranged by their compartment signal strength (EV1) from highest (most A-like) to lowest (most B-like). Differential saddle-plot (right), with green denoting less interactions in the cKO compared to floxed contact matrices.

C. Scatter plot (left) and boxplot (right) of the compartmentalization-strength scores per chromosome in cKO and floxed sample. P value is calculated using Wilcoxon matched-pairs signed rank test.

D. Aggregate TAD analysis (ATA) centered at TADs called in the floxed and cKO samples, using floxed (left) and cKO (middle) Hi-C contact matrices. Differential ATA plots (right), with blue denoting loss of interactions in cKO.

E. TAD-neighbour analysis: interactions between TADs, stratified on the number of intervening TADs (n), compared to floxed sample.

We next identified specific genomic regions that undergo A/B compartment and TAD rearrangements upon SATB2 deletion. At 100-kb resolution, pairwise differential analysis revealed 1451 significant compartment switches and strength changes between cKO and floxed neurons, affecting 5.6% of the genome (Fig. 5A, Table S2). The changes in compartment score (ΔΕV1) were positively correlated (Spearman r2 = 0.34, P <0.001) with chromatin accessibility changes (mean log2FC of all OCRs within a differential compartment). Likewise, the genes within differential compartments (631 genes, Table S3) showed a positive correlation between compartment score shifts (ΔΕV1) and expression changes (Fig. 5B) and were enriched in postsynaptic, dendritic, and memory-related GO terms (Fig. 5C). Genes linked to synaptic plasticity, such as Bdnf, Grm7, Fmr1, and Kcnb2, as well as SCZ risk-prioritized genes⁴², including Csmd1, Lrrc4b, Ptprd, and Epn2 (the later also associated with intelligence⁴³), were among those affected by compartment changes (Fig. 5D). TAD calling ⁴⁴ identified a similar number and average width of TADs and subTADs in floxed vs cKO Hi-C matrices (Fig. S5A, S5B). Of all TAD boundaries 12.7% were identified as differential (Fig. 5E, Table S2), displaying various types of changes (complex, split, merge, shift, strength, Fig. S5C)⁴⁵. In total 10.1% of the differential TAD boundaries overlapped with SATB2 peaks. In cKO neurons, the average insulation score decreased across 692 TAD boundaries (lost TAD boundaries), increased at 511 TAD boundaries (gained TAD boundaries), but did not change at invariant boundaries (Fig. 5F), corroborating the results of the differential TAD analysis. Genes nearby lost TAD boundaries (504 genes, Table S3) were expressed at lower levels in cKO neurons (Fig. 5G) and were strongly enriched for GO terms related to glutamate receptor signalling and neuron projection development (Fig. 4H). In contrast, genes nearby gained TAD boundaries (310 genes, Table S3) were expressed at higher levels in cKO neurons but lacked significant GO term enrichment. Lost but not gained TAD boundaries were enriched for genomic annotations known to be associated with TAD boundaries, e.g., promoters and CTCF binding sites ^46–48 (Fig. 5I).

Figure 5: SATB2 loss triggers alterations in A/B compartments and TAD boundary changes.

A Heatmaps illustrating changes in compartment state in cKO vs floxed neurons. Compartment switches from B to A and vice versa (left). Shifts in the compartment score within the same compartment (right).

B Correlation between changes in compartment score (ΔΕV1) and expression (log2FC) for DEGs located within differential compartments (permutation relationship test, n = 10000 permutations).

C GO analysis of genes mapped to differential compartments.

D Representative genome browser tracks illustrating a compartment switch (left) and a compartment strength change (right) in cKO vs floxed neurons. Shown are EV1 values and −log₁₀ P adj values of differential compartment calls at 100 kb-resolution.

E Representative Hi-C contact matrix of ICED-normalized contacts in cKO and floxed neurons at 10kb-resolution (chr2:108–111 Mb). Shown are TADs, called in floxed (red) and cKO (green) matrix. Gene (black) and SATB2 peak (blue) tracks are displayed.

F Tornado plots and average profiles showing insulation scores centred at lost, gained and invariant TAD boundaries in cKO and floxed Hi-C matrices. Scores are shown relative to the center of the boundary up to 500 kb in both directions.

G Fold change distributions of DEGs near differential TAD boundaries. P value is calculated using non-parametric Mann-Whitney test.

H GO analysis for genes nearby lost TAD boundaries as defined by GREAT⁸⁶ (association rule: basal + extension: 5 kb upstream, 1 kb downstream, and 100 kb max extension).

I Enrichment analysis of differential TAD boundaries over SATB2 binding sites, CTCF binding sites, H3K4me3 peaks, and promoter states, inferred by P0 forebrain ChromHMM. Color bar shows the z-score, for all enrichments and depletions, P < 0.001.

In summary, our data establish SATB2 as a cell type-specific modifier of 3D chromatin structure in pyramidal neurons, acting not only on chromatin loops but also on multiple larger scale hierarchical levels.

SATB2-dependent FIREs and super-FIREs harbor cognition-associated genes

Frequently Interacting Regions (FIREs) are cell type-specific local interaction hotspots, harboring active cis-regulatory elements^49,50. Contiguous FIREs are known to form clusters, called super-FIREs, which have the most significant local interaction frequency^50,51. We tested if SATB2 exerts effects on FIREs/super-FIREs in cortical neurons. A total of 6669 and 5428 FIREs were detected in floxed and cKO Hi-C contact matrices. Based on differences in FIRE score⁴⁹, 99 floxed- and 15 cKO-specific FIREs were identified. By filtering out super-FIREs common to both genotypes, 180 floxed- and 93 cKO-specific super-FIREs were defined (Fig. 5A, Table S2). Floxed- but not cKO-specific FIREs were enriched for enhancer states and for lost dOCRs (Table S1). Genes mapped to floxed- but not to cKO-specific FIREs/super-FIREs (Table S3) showed highly significant overrepresentation of GO terms associated with synapses, behavior, learning or memory and cognition (Fig. 5B). Genes residing in floxed-specific FIREs/super-FIREs showed higher expression levels in floxed compared to cKO neurons (Fig. 5C) and were strongly enriched for DEGs (Table S1). All DEGs mapped to floxed FIREs and 48/49 DEGs mapped to floxed super-FIREs were upregulated in floxed compared to cKO neurons.

We next tested if genes with the highest known correlation with human cognitive ability are present in floxed-specific FIRE/super-FIRE gene sets. These genes are contained within “cognitive task performance” (CTP) and “performance on a non-cognitive factor” (NCF) meta-loci in humans⁵². The human orthologues of approximately one third of all genes mapped to floxed-specific FIREs (13/45) and super-FIREs (29/114) belonged to CTP “driver” genes, including ANKS1B, GPHN, RERE, and PLCB1 (Table S1). Likewise, the human orthologues of 12/45 and 14/114 genes mapped to floxed-specific FIREs and super-FIREs were part of NCF “driver” genes, including DLG2, GRIN2B, and NRXN1 (Table S1). Among the floxed-specific FIRE/super-FIRE genes with human orthologue in CTP/NCF meta-loci several were also associated with autism spectrum disorder (ASD) (13/37, CTP; 11/21, NCF)⁵³, developmental brain disorders (10/37, CTP; 8/21, NCF)⁵⁴ and intellectual disability (ID) (10/37, CTP; 5/21, NCF)⁵⁵.

Genes harbored in NCF and CTP meta-loci are co-regulated⁵². Intra-nuclear proximity and increased connectivity of regulatory elements have been suggested as mechanisms for gene co-regulation⁵⁶. Considering the remarkable overrepresentation of NCF and CTP driver genes in floxed-specific FIRE/super-FIRE gene sets, we reasoned that SATB2 might regulate their spatial distance in the 3D nucleus. To test this, we modelled chromatin conformation in 3D using a consensus set of TAD domains and Hi-C data from floxed vs cKO neurons⁵⁷. 3D models consisted of 3478 chromatin bead domains, called for the diploid genome, averaging 1.42 Mb in length. In total, 43 domains were mapped to genes residing in floxed-specific FIRE/super-FIREs and having a human orthologue in NCF meta-loci (referred to as FIRE/super-FIRE/NCF genes) and 69 domains contained genes residing in floxed-specific FIRE/super-FIREs and having a human orthologue in CTP meta-loci (referred to as FIRE/super-FIRE/CTP genes). The pairwise Euclidean distances (EDs) between these domains were significantly shorter in floxed versus cKO 3D models (Fig. 6D, 6E, Fig. S6A). Similar results occurred when all NCF and CTP driver genes were considered, beyond those in floxed-specific FIREs/super-FIREs (P = 2.2e-16, Wilcoxon signed-rank tests). Next, we hierarchically clustered the 43 and the 69 domains, based on the pairwise EDs between them in floxed vs cKO 3D models. Cutting the trees at the 0.90 quantile led to clusters of domains closely positioned within the 3D nucleus called Euclidean hot spots⁵⁸ (Fig. 6F, Fig. S6B). Although the cKO and floxed dendrograms showed the same total number of Euclidean hot spots, their size and composition were very different between the genotypes, indicating differential position and clustering of these domains within the 3D sphere of cKO vs floxed nuclei (Fig. 6F, 6G, Fig. S6B, S6C). The dissimilarity between the two dendrograms was further supported by the low values of Cophenetic and Baker correlation coefficients, which measure similarity between tree topologies (Fig. S6D). Similarly, a weak correlation was observed when dendogrames were constructed by using all NCF and CTP driver genes (Fig. S6D). To validate the 3D genome modelling results, we performed 3D DNA FISH experiments. We measured the interallelic distances between Grin2b and Pde4b, two representative FIRE/super-FIRE/NCF genes, in floxed, cKO and AAV-SATB2-transduced cKO (“rescued”) nuclei. In agreement with our in silico Chrom3D models⁵⁷, the 3D DNA FISH data showed significantly shorter pair-wise distances between these genes in floxed vs cKO nuclei (Fig. 5I). After acute re-introduction of SATB2 in cKO nuclei the average interallelic distances between these two loci were not significantly different in floxed vs “rescued” nuclei.

Figure 6: SATB2-dependent FIREs and super-FIREs overlap with cognition-related genes.

A. Representative genome browser tracks illustrating floxed-specific FIREs and super-FIREs. The significance of FIRE scores, as calculated by one-sided Z-test, is depicted for floxed (green) and cKO (purple) neurons. Floxed-specific FIREs and super-FIREs are presented as blue bars under the significance tracks.

B. Genes located in floxed-specific FIREs/super-FIREs were more highly expressed in floxed compared to cKO neurons. P value is calculated using non-parametric Mann-Whitney test.

C. GO analysis of genes mapped to floxed-specific FIREs/super-FIREs.

D. In silico Chrom3D models of floxed (left) and cKO (right) cortical neuron spatial genomes. Colored beads represent the chromatin bead domains, which harbor the genes mapped to floxed-specific FIREs/super-FIREs and having a human orthologue in NCF meta-loci (FIRE/super-FIRE/NCF genes). The domains harboring the alleles of Grin2b (Chr6) and Pde4b (Chr13), representative FIRE/super-FIRE/NCF genes, are colored in khaki and grey, respectively. Pairwise Euclidean distances (EDs) are reported as numbers above the lines connecting khaki and grey beads.

E. Pairwise EDs between the domains harboring FIRE/super-FIRE/NCF genes (P = 3.27e-13, cKO vs floxed, Wilcoxon signed rank test) in cKO vs floxed 3D genome models. Shown are also the pairwise EDs between randomly permuted sets of domains of the same size in both models (n = 1000 permutations, cKO 3D model, P = 3.06e-09 for domains harboring FIRE/super-FIRE/NCF genes vs random domains; floxed 3D model, P = 0.016 for domains harboring FIRE/super-FIRE/NCF genes vs random domains, Wilcoxon rank-sum tests).

F. Hierarchical clustering of the domains illustrated as colored beads in Fig. 6D in floxed vs cKO 3D genome models. Red squares depict Euclidean hot spots. White squares show the domains harboring Grin2b and Pde4b alleles.

G. Comparison between dendograms resulting from hierarchical clustering in floxed vs cKO 3D genome models. “Unique” nodes, with a combination of labels/items not present in the other tree, are highlighted with dashed lines. Domains that belong to different subtrees are connected by green lines, domains within the same subtrees are connected by purple lines.

H. Representative z-projected (maximum intensity) confocal image of 3D FISH experiment using FISH probes against Grin2b (red) and Pde4b (green). DNA is stained with DAPI (blue). Scale bar, 5 μm.

I. Quantification of average interallelic distances between Grin2b (on Chr6) and Pde4b (on Chr13) 3D FISH loci in cKO, floxed, and AAV-SATB2-transduced cKO (“rescued”) neurons. 75–80 nuclei per genotype were quantified, P = 0.0056, Kruskal-Wallis one-way ANOVA, floxed vs cKO, P = 0.0028; floxed vs “rescued”, P = 0.378, Dunn’s multiple comparisons test.

Thus, SATB2 influences distances between NCF and CTP driver genes tied to cognitive dimensions and psychopathology within and beyond local interaction hotspots (FIREs/super-FIREs) in pyramidal neurons.

SATB2-dependent 3D epigenome changes contribute to cognitive ability and risk for neurodevelopmental and neuropsychiatric disorders

To explore the functions of the genes that are influenced by SATB2 3D epigenome effects, we combined all genes identified at individual 3D hierarchical levels into one composite set (referred to as SATB2 3D genome set). It consisted of genes mapped to: differential compartments, lost TAD boundaries, invariant loops connecting dOCRs to promoters, promoter-based lost loops, differential loops at shared promoter and floxed-specific FIREs/super-FIREs (Table S3). Due to the lack of significant GO term enrichment, the genes mapped to gained TAD boundaries, promoter-based gained loops and cKO-specific FIREs/super-FIREs were excluded, as they are likely to represent randomly acquired effects upon SATB2 loss. We then examined whether the SATB2 3D genome set as a whole is enriched for genes associated with neuron- and cognition-specific GO categories. The results showed a highly significant enrichment, with approximately 46%−50% of all genes within these categories being linked to SATB2-dependent 3D genome architecture. In contrast, control non-neuronal GO terms were not overrepresented or even displayed significant depletion (Fig. 7A). Intriguingly, the individual gene sets, determined by SATB2 effects at distinct 3D chromatin hierarchical levels, exhibited limited overlap between each other. More than 70% of the genes within the SATB2 3D genome set were exclusively present in only one of the hierarchical level-specific gene sets (Fig. 7A). Thus, SATB2 exhibits selective influence over the chromatin structure of a substantial portion of all cognition-associated genes, by operating across multiple 3D architectural levels while impacting distinct gene sets on each level.

Figure 7: Genes and non-coding regions mapped to SATB2-driven 3D chromatin changes contribute to cognitive ability and risk for neuropsychiatric disorders.

A. Neuron- and cognition-specific but not control GO categories are enriched for genes affected by SATB2 3D epigenome modeling. Bar charts showing the proportion of each GO category, occupied by individual gene sets. P values were calculated using Fisher’s exact test, significance level after Bonferoni correction of multiple testing = 0.00625, OR, odds ratio.

B. MAGMA-GSA of SATB2 3D genome sets in EA, IQ, SZ, BP, and MDD. Beta values (effect sizes) are plotted on the x-axis with P values shown beside each bar, *P < 0.05, **P < 01 and ***P < 0.001. Horizontal bars indicate standard error. All data associated with this graph are included in Table S6.

C. Analysis of SATB2 3D genome sets using data on de novo mutations (mis = missense; lof = loss-of-function), identified in SZ, ASD, ID, and DD, shown are Bonferoni-corrected P values, *P < 0.05, **P < 0.01, ***P < 0.001. Data are supplied in Table S7. Shown are total number of genes (per set) with a human orthologue (HUGO Gene Nomenclature Committee).

D. Enrichment of common risk variants for brain-related and non-psychiatric/non-brain related GWAS traits in the human orthologues of dOCRs and non-P-based differential loop anchors. Enrichment values are plotted as heatmap. #, P values significant after Bonferoni correction of multiple testing across all tests, *, P values nominally significant. Data are supplied in Table S8.

We next investigated if the individual gene sets within SATB2 3D genome set are implicated in human psychiatric disorders and cognitive traits. We assessed whether common genetic variation in the corresponding human orthologues is linked to major psychiatric traits by gene set analysis (GSA). For SNP-to-gene assignment, we employed MAGMA⁵⁹ and H-MAGMA ⁶⁰, the latter using chromatin interaction profiles from human brain tissue to refine gene mapping. MAGMA-GSA unveiled significant enrichments for genes associated with educational attainment (EA), cognitive ability/human intelligence (IQ), SCZ, bipolar disorder (BPD) and major depressive disorder (MDD) in several individual gene sets (Fig. 7B, Table S6). H-MAGMA-GSA, using gene-SNP pairs based on cortical neuronal Hi-C⁴⁹ and fetal brain Hi-C⁶¹ yielded similar enrichments (Table S6). To exclude that the enrichments detected for EA, IQ, SCZ, BPD, and MDD are a property of polygenic phenotypes in general, we included control phenotypes encompassing both brain- and non-brain-related diseases, such as Alzheimer’s disease, stroke, cardiovascular disease, Crohn’s disease, and type 2 diabetes. No enrichment was observed for any of the control phenotypes (Table S6). Genes mapped to gained TAD boundaries, promoter-based gained loops, and cKO-specific FIREs/super-FIREs, which likely represent non-physiological effects arising from SATB2 loss, did not show enrichment for genes associated with psychiatric disorders and/or cognitive ability. Only the gained TAD boundary gene set exhibited marginal enrichment for genes associated with SCZ (Table S6).

Next, we examined whether the individual gene sets exhibit enrichments for de novo mutations, reported in trio-based studies of specific disorders, including SCZ^62,63, ASD⁶⁴, intellectual disability⁶⁵, and developmental disorders (DD)^66,67. We focused on missense (Mis) and loss-of-function (LoF) mutations, while synonymous (Syn) mutations served as a control, as they are unlikely to be pathogenic. Control trios and trios consisting of unaffected siblings were analyzed as additional controls. We observed significant enrichment for LoF and/or Mis mutations, reported in ASD and DD patients, in the genes mapped to promoter-based lost loops, differential loops at shared promoters, and floxed-specific FIREs/super-FIREs (Fig. 7C, Table S7). Genes mapped to lost dOCRs interacting with promoters were enriched for LoF and Mis mutations found within the DD trios. The analysis of Syn mutations across all four disorders did not reveal enrichments in any of the individual gene sets (Table S7). Similarly, neither the control trios, nor the trios composed of unaffected siblings exhibited de novo mutation enrichment within any of the SATB2 3D genome subsets. The genes mapped to 3D epigenome changes gained in cKO did not exhibit enrichment for de novo mutations reported in any of the four trio-based studies.

Lastly, we investigated whether SATB2-driven changes in chromatin accessibility and non-P-based chromatin looping have relevance for non-coding genetic variation associated with neuropsychiatric diseases. For this purpose, we identified syntenic regions in the human genome for dOCRs and non-P-based differential loops through liftover analysis (Table S8). Remarkably, the human orthologues of mouse dOCRs exhibited enrichment for enhancer states identified in the human frontal cortex⁶⁸, suggesting potential functional conservation (Table S8). Subsequently, we conducted stratified linkage disequilibrium score regression (sLDSC) analysis to assess whether common genetic variants, associated with cognitive ability or risk for neuropsychiatric disorders, are enriched in the human syntenic regions. Out of five cognitive and psychiatric traits tested, we found a significant enrichment of SNP-based heritability for EA, IQ, and SCZ within the human orthologous areas of mouse dOCRs and non-P-based differential loops (Fig. 7D, Table S8). By contrast, no heritability enrichment was observed for control traits and disorders unrelated to psychiatry or the brain (Fig. 7D, Table S8). The specificity of the observed enrichments for EA, IQ, and SCZ, as opposed to the other two neuropsychiatric disorders tested, is in line with existing molecular genetic evidence of shared genetic factors between cognitive performance and schizophrenia^52,69. In summary, our findings suggest that SATB2-dependent 3D epigenome changes, particularly in chromatin accessibility, loops, and FIREs/super-FIREs, contribute to the genetic mechanisms, underlying human cognitive function, neuropsychiatric (SCZ and BPD) and neurodevelopmental (ASD and DD) disorders.

Discussion

Our work reveals 3D epigenome remodeling by SATB2 within the highly specialized nuclei of postmitotic pyramidal neurons. SATB2-linked chromatin alterations occur at all 3D architectural levels, closely correlate with gene expression and the affect loci are involved in specific neuronal processes, including cognition, learning and memory. SATB2 effects on 3D genome likely depend on multiple mechanisms, including direct binding to DNA as well as interaction with other proteins including CTCF, in multiprotein complexes of the nuclear matrix.

We find that a large fraction of ARGs/cognition-associated genes are engaged in SATB2-dependent chromatin loops and that the promoter-interacting distal targets of these loops are enriched for AP1-binding sites. Altered occupancy of AP1-binding sites at ARG distal regulatory regions is part of the epigenomic signature of neuronal activation^70,71. Thus, the impaired induction of ARGs in SATB2-deficient neurons¹⁵ is likely caused by reduced accessibility of AP1 sites and/or lost chromatin loops connecting AP1-bound distal regulatory elements to ARG promoters. Beyond AP-1 sites, we uncover extensive effects of SATB2 on chromatin accessibility genome-wide. Some of the identified dOCRs were proximal to promoters or interacted with promoters via invariant loops. The presence of these promoter-associated dOCRs closely correlated with changes in gene expression in cKO neurons, for example we observed increased accessibility and up-regulation of neural progenitor-specific genes^30–32 and decreased accessibility and down-regulation of ARGs.

A very abundant type of dOCRs was found at non-P-based differential loop anchors. Since chromatin state is primarily responsible for chromatin interactions within compartmental domains^72,73, this finding indicates a structural role of SATB2. Such an architectural role is also supported by our observation that SATB2 loss affects chromatin structure on multiple hierarchical levels, including compartments, TAD boundaries and FIREs. Ubiquitously expressed chromatin conformation regulators, such as CTCF and cohesin establish or maintain general aspects of 3D genome structure in all cell types⁷⁴. In highly specialized cells additional layers of 3D genome adjustments must co-exist with CTCF-dependent mechanisms that enable specific, adaptive transcriptional responses, tailored towards cell type-specific biological processes^75,76. Cell type-specific 3D genome organization likely depends on proteins with highly restricted expression pattern, such as SATB2, which within the adult CNS is almost exclusively expressed in pyramidal neurons that form the cellular foundation of cognitive processes. Our data indicate that ubiquitous and cell type-specific layers of 3D genome organization cooperate. The enrichment of CTCF peaks at anchors of non-P-based differential loops and the direct or indirect interaction between recombinant SATB2 and CTCF imply the existence of multi-protein complexes, containing both proteins. This is consistent with recent studies demonstrating that CTCF co-interactors³⁷ and CTCF peak co-localizing factors³⁸ can modulate the strength of CTCF-mediated loops⁷⁷. Prominent examples of CTCF co-localizing factors that are also SATB2 interactors are BCLAF1, NR2F2, and HDAC1, all highly expressed in pyramidal neurons⁷⁸.

The recent discoveries of 3D genome effects exerted by matrix proteins, including HnRNPU⁷⁹, SAFB⁸⁰ and Matrin-3⁸¹ demonstrate that a protein does not necessarily have to be directly bound to DNA in order to affect 3D genome structure. SATB2, a well-established constituent of the nuclear matrix^8,17, may also modulate 3D chromatin interactions without direct DNA binding, accomplished through its involvement in multimeric nuclear matrix complexes. An adaptor function as a mediator of CTCF interactions with nuclear matrix has been demonstrated for SATB1⁴¹.

We identify a highly selective subset of FIREs/super-FIREs that is SATB2-dependent. A hallmark of this subset is its enormous specificity since it harbors genes that are highly enriched (odds ratios over 10) for the genes that are most tightly associated with human cognition, i.e. the mouse orthologues of the genes within CTP and NCF meta-loci in humans⁵². Our modelling of the spatial genome organization provides a potential SATB2-dependent mechanism for the observed co-regulation of these genes, i.e. higher intra-nuclear proximity increasing connectivity between cis-regulatory elements. Thus, in cortical neurons SATB2 might influence the formation of higher-order transcription hubs, which have emerged as a central mechanism for transcriptional bursting/co-activation of genes within a functional gene set^82,83.

SATB2 has been associated with general cognitive ability⁹ and risk for SCZ^84,85. Yet, a thorough understanding of its contribution to these traits is currently missing. By leveraging common and rare genetic variation data, we uncover SATB2-dependent 3D epigenome remodeling as a candidate mechanism, contributing specifically to human cognitive function and risk for SCZ. Lifting-over mouse to human genome, we show that human orthologues of SATB2-dependent 3D genome sets are enriched for genes associated IQ and SCZ. Even more surprisingly, human orthologues of non-coding regions undergoing SATB2-dependent chromatin modifications are enriched for SNP-based heritability for EA, IQ, and SCZ. This indicates that the function of SATB2 in orchestrating cognition-related gene regulatory networks via coordinated effects at multiple 3D epigenome levels, has remained conserved during evolution of complex vertebrate brains.

Limitations of the study

Given the strong connection between cognition and neuroplasticity, one limitation of this study is that the dynamics of SATB2-dependent 3D genome alterations upon neuronal activation has not been investigated. Another limitation is that the impact of SATB2 on the 3D epigenome of human pyramidal neurons has not been addressed. When suitable human SATB2 loss-of-function models become available, these studies can then be expanded to explore the role of SATB2 in neuropsychiatric disorders, including SATB2-associated syndrome. Third, to unravel the exact molecular mechanisms underlying SATB2-dependent 3D epigenome effects, it will be necessary to delve into the specific roles played by SATB2 homomeric and heteromeric protein complexes, employing structural mutants of SATB2.

STAR Methods

I. RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to the lead contact, Galina Apostolova (galina.apostolova@i-med.ac.at).

Materials availability

All mouse lines and viral constructs used in this study are available from the lead contact with a completed Materials Transfer Agreement.

Data and code availability

• Accession information of published data sets analyzed in the manuscript are included in the Key Resources Table along with information for original datasets. Original data generated in this study have been deposited at GEO and are available as GEO record GSE222609. Original Western blot and microscopy images used for figures have been deposited at Mendeley Data. The DOI is listed in the Key Resources Table. Additional microscopy data reported in this paper will be shared by the lead contact upon request.

• Study analysis pipelines and code are available at the following Github repository: https://github.com/sespesogil/SATB23DGenome. A version of record has been logged in Zenodo DOI:10.5281/zenodo.10373960. Information for all other code used can be found in the Key Resources Table.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-SATB2	Abcam	#ab92446
Anti-CTCF	Abcam	#ab128873
Anti-GAPDH	Sigma	#MAB374
Anti-mouse IgG HRP-conjugated	Cell Signaling	#7076S
Anti-rabbit IgG HRP-conjugated	Cell Signaling	#7074S
CUTANA™ Rabbit IgG	Epicypher	SKU: 13–0042
Chemicals, peptides, and recombinant proteins
(−)-Bicuculline methochloride	Tocris	#0131
NBQX disodium salt	Tocris	#1044
AMPure XP beads	Beckman	#A63881
Pefabloc® SC	Roche	#11429868001
Sodiumbutyrat	Sigma	#B5887
cOmplete EDTA-free protease inhibitor	Roche	#05892791001
Tn5 Transposase	Illumina	#15027865
jetPRIME Transfection reagent	Polyplus	#101000015
Concavalin-A conjugated beads activation	Epicypher	#21–1401-EPC
AG-MNAse	Epicypher	#15–1016-EPC
Critical commercial assays
Arima-HiC+ Kit	Arima Genomics	#A51008-ARI
Accel-NGS® 2S Plus DNA Library Kit for Illumina	Swift Bioscience	#21024-SWI
2S Indexing Kit for Illumina	Swift Bioscience	#26148-SWI
QIAGEN MinElute Kit	QIAGEN	#28004
NEBNext® Ultra Library prep kit	New England Biolabs	#E7645S
NEBNext® Multiplex Oligos for Illumina	New England Biolabs	#E7335S
Purelink RNA Micro Scale Kit	Invitrogen	#12183016
Dynabeads Protein G Co-Immunoprecipitation Kit	Invitrogen	#14321D
DNA high sensitivity D-5000 screen tapes	Agilent	#5067–5592
Deposited data
CUT&RUN, ATAC-seq, RNA-seq and Hi-C data	This paper; GEO	GSE222609
RNA-seq data	Feurle et al.15	GSE157375
ChromHMM 18-state model of postnatal day 0 mouse forebrain	ENCODE	ENCSR301UGN
ChromHMM 18-state model of human adult frontal cortex	ENCODE	ENCFF382TUC
CTCF ChIP-seq	ENCODE	ENCFF430PPJ
H3K27ac ChIP-seq	ENCODE	ENCFF676TSV
H3K4me3 ChIP-seq	ENCODE	ENCFF160SCR
Hi-C data adult mouse cortex	Espeso-Gil et al.133	SRP154319
Hi-C data B cells	Vian et al.134	4DNESE1VMAMD
Hi-C data olfactory neurons	Monahan et al.135	4DNESEPDL6KY
Original Western blot images	This paper; Mendeley Data	DOI: 10.17632/g873yxyxdn.1
Original Imaging data	This paper; Mendeley Data	DOI: 10.17632/g873yxyxdn.1
Experimental models: Organisms/strains
Satb2 flx/flx : :Nes-Cre	This paper	available upon request
Satb2 flx/flx	This paper	available upon request
Hela cells	ATCC	#CCL-2.2
Recombinant DNA
pKS004-pCAGGS-3XFLAG-CTCF-eGFP	Addgene	#156438
PEF-DEST51-V5-SATB2	This paper	available upon request
AAV-hSyn-EGFP virus serotype 8	Addgene	#50465-AAV8
AAV-hSyn-V5-mSatb2	This paper	available upon request
RP23–310P11-OR BAC mouse FISH probe	Empire Genomics	#RP23–310P11-OR
RP23–182N11-GR BAC mouse FISH probe	Empire Genomics	#RP23–182N11-GR
Software and algorithms
HiC-Pro	Servant, N. et al.87	https://github.com/nservant/HiC-Pro
HiCRep	Yang, T. et al.21	https://github.com/TaoYang-dev/hicrep
GENOVA	Van Der Weide, R.H. et al.90	https://github.com/robinweide/GENOVA
Genrich	N/A	https://github.com/jsh58/Genrich
dcHiC	Chakraborty, A. et al.136	https://github.com/ay-lab/dcHiC/tree/master
annotater	Cavalcante, R.G. et al.92	https://github.com/rcavalcante/annotatr
SpectralTAD	Cresswell, K.G. et al.44	https://github.com/dozmorovlab/SpectralTAD
HiCcompare	Stansfield, J.C. et al.93	https://github.com/dozmorovlab/HiCcompare
TADCompare	Cresswell, K.G. et al.45	https://github.com/dozmorovlab/TADCompare
GREAT	McLean, C.Y. et al.86	https://great.stanford.edu/great/public/html/
Mustache	Roayaei Ardakany, A. et al.94	https://github.com/ay-lab/mustache
FIREcaller	Crowley, C. et al.51	https://github.com/yycunc/FIREcaller
Arrowhead	Durand, N. et al. 95	https://github.com/aidenlab/juicer/wiki/Arrowhead
Chrom3D	Paulsen, J. et al.57	https://github.com/Chrom3D/Chrom3D
pheatmap	Kolde, R. et al.89	https://github.com/raivokolde/pheatmap
dendextend	Galili, T. et al.96	https://github.com/talgalili/dendextend
ChimeraX	Meng, E. et al. 137	https://www.cgl.ucsf.edu/chimerax/
BWA-MEM	Li, H. et al. 99	https://github.com/lh3/bwa
SAMtools	Danecek, P. et al.100	https://github.com/samtools/samtools
deepTools2	Ramírez, F. et al. 101	https://deeptools.readthedocs.io/en/develop/
featureCounts	Liao, Y. et al. 102	https://subread.sourceforge.net/
edgeR	Robinson, M.D. et al. 103	https://bioconductor.org/packages/release/bioc/html/edgeR.html
RUVseq	Risso, D. et al.104	https://bioconductor.org/packages/release/bioc/html/RUVSeq.html
HOMER	Heinz, S. et al. 105	http://homer.ucsd.edu/homer/index.html
ChiPseeker	Yu, G. et al. 106	https://bioconductor.org/packages/release/bioc/html/ChIPseeker.html
g:Profiler	Kolberg, L. et al. 107	https://biit.cs.ut.ee/gprofiler/gost
TOBIAS	Bentsen, M. et al. 108	https://github.com/loosolab/TOBIAS
Bowtie2	Langmead, B. et al. 138	https://github.com/BenLangmead/bowtie2
SEACR	Meers, M.P. et al. 111	https://github.com/FredHutch/SEACR
STAR	Dobin, A. et al. 113	https://github.com/alexdobin/STAR
DESeq2	Love, M.I. et al. 115	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
Metascape	Zhou, Y. et al. 117	https://metascape.org/gp/index.html#/main/step1
clusterProfiler	Wu, T. et al. 112	https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html
regioneR	Gel, B. et al. 118	https://bioconductor.org/packages/release/bioc/html/regioneR.html
MAGMA	de Leeuw, C.A. et al. 59	https://ctg.cncr.nl/software/magma
H-MAGMA	Sey, N.Y.A. et al. 60	https://github.com/thewonlab/H-MAGMA
sLDSC	Finucane, H.K. et al. 130	https://github.com/bulik/ldsc
UCSC liftOver	N/A	https://genome.ucsc.edu/cgibin/hgLiftOver
denovolyzeR	Ware, J.S. et al. 132	https://github.com/jamesware/denovolyzeR
Other
Bioinformatics pipeline and code	This paper; Github; Zendo	https://github.com/sespesogil/SATB23DGenome/tree/main DOI:10.5281/zenodo.10373960

II. EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Primary cortical cultures

Dissociated cortical neurons from neonatal Satb2 ^flx/flx::Nes-Cre and Satb2^flx/flx mice were plated at a density of 0.125×10⁶ cells/cm² in 60-mm culture dishes, 6-well plates or 12-well plates, previously coated with poly-L-ornithine (Sigma, 0.5 mg/ml, overnight at 4°C) and laminin (Sigma, 0.005 mg/ml, 1 h at 37°C). For 3D DNA FISH, glass coverslips, coated following the same protocol, were used. Neurons were plated in minimal essential medium (MEM, Gibco), supplemented with 10% horse serum, 10% penicillin/streptomycin and sodium pyruvate. After pre-culturing for 2 h, medium was replaced with Neurobasal A medium (NBM-A, Gibco), supplemented with GlutaMAX, Penicillin/Streptomycin, B27 (Thermo Fisher Scientific). Neurons were cultured for 14 days at 5 % CO2 and 37 °C. Glial cell proliferation was inhibited by adding cytosine arabinoside (5 μΜ, Sigma) to the culture medium at DIV3. On DIV4, half of the medium was replaced with fresh medium. On DIV 14, cortical neurons were first pre-treated with NBQX (20 μΜ, Tocris) for 16 h, and then treated with bicuculline (50 μΜ, Tocris) for 1 h.

All experiments, in which animals were used, were approved by the “Austrian Animal Experimentation Ethics Board”.

HeLa cell culture and transfections

HeLa cells were grown in DMEM medium supplemented with 10 % FBS and 1 % penicillin/streptomycin (Thermo Fisher Scientific). Transfection was performed using jetPRIME Transfection Reagent (Polypus), according to manufacturer’s instructions. Forty-eight hours after transfection total protein lysates were prepared and used for co-immunoprecipitation.

III. METHOD DETAILS

Co-Immunoprecipitation

HeLa cells were lysed on the culture dish in IP lysis buffer (Pierce). The lysates were incubated for 10 min on ice while shaking, followed by a brief centrifugation at 13 000 × g. The supernatant was used in immunoprecipitation reactions using the Immunoprecipitation Kit-Dynabeads Protein G (Thermo Fisher Scientific) according to the manufacturer’s instructions. Briefly, 50 μl of protein G Dynabeads were covalently linked to 5 μg of anti-SATB2 antibody (ab92446, Abcam), beads were mixed with cell lysate (2 mg total protein) and incubated overnight at 4 °C. On the next day, the beads-antibody-protein complexes were washed 3 times with washing buffer, re-suspended in 2 x Roti-Load sample buffer (Roth) for elution and incubated at 95 °C for 5 min. The eluates were separated on a 10 % SDS-PAGE used for further immunoblotting analysis.

Western blotting

PVDF membranes were blocked with 5% milk in TBST (0.1% Tween 20 in TBS) for 1 hour and then incubated overnight at 4°C with the primary antibodies diluted in blocking solution. After incubation with HRP-coupled secondary antibodies, blots were incubated with ECL reagent (Bio-Rad) and developed using BioRad’s Chemiluminescence Detection System (ChemiDoc^™).

Hi-C

For in situ Hi-C, Arima-HiC kit was used, according to the manufacturer’s instructions. Briefly, 2.5×10⁶ primary cortical neurons were fixed directly on the dish by addition of formaldehyde to a final concentration of 1 % for 10 min at room temperature. After washing twice with 1X PBS cells were detached using a cell lifter and collected in 1X PBS, containing protease inhibitors. Cells were pelleted at 500 g for 10 min at 4 °C and pellets were directly snap-frozen in liquid nitrogen.

After nuclei isolation and chromatin digestion, biotin was added and filled ends were ligated. After reverse crosslinking and purification, DNA was shared using a M220 Covaris sonicator at duty factor 12 %, PIP 75, CPB 200 for 60 s to achieve input fragment size of 400 bp. Additionally, DNA was size-selected using AMPure XP beads, according to Arima-HiC kit manual. Hi-C libraries were prepared using Swift Biosciences Accel-NGS 2S Plus DNA Library Kit according to Arima-HiC Kit User Guide (November 2018).

Libraries were sequenced at New York Genome Centre on Illumina HiSeq2500 (125 bp paired-end, sequencing depth of 250 million reads, n=2) and at Novogene UK on Illumina Novaseq (150 bp paired-end, sequencing depth of 650 million reads, n=4).

Mapping, filtering and normalization

Sequencing reads were mapped to the mouse reference genome assembly (mm10), artifacts were filtered and libraries were ICED normalized using HiC-Pro default parameters and guidelines (v.2.11.4)⁸⁷. For some downstream analyses, KR (Knight-Ruiz) normalization was used⁸⁸. Data reproducibility was assessed using HiCRep method ²¹ with a wrapper function renamed as ‘HiCReproducibility.R’ (https://github.com/sespesogil/SATB23DGenome/blob/main/HiCReproducibility). Briefly, Juicer Hi-C files were transformed into compatible Hi-C matrices at 250 kb-resolution and observed frequencies per chromosome were extracted. Next, pairwise reproducibility values (SCC scores) were computed using the function ‘get.scc’ with the following parameters: h = 1, lbr = 0, ubr = 5000000. Results were transformed into a quadratic matrix as input for the pheatmap R package⁸⁹ for visualization purposes.

Contact distance decay plots were generated by using the function ‘DecayFreq.R’ (https://github.com/sespesogil/SATB23DGenome/blob/main/Decayplots/DecayFreq.R) that computes and aggregates ICED normalized frequencies at 100 kb-resolution.

Hi-C matrix visualization

Hi-C matrices were displayed at different resolutions using the function ‘hic_matrixplot’ from GENOVA package⁹⁰ (https://github.com/robinweide/GENOVA). To provide full accessibility to Hi-C visualizations, we developed and deployed a dedicated Shiny app to visualize samples at different resolutions: 5, 10, 25, 100 and 250 kb.

Compartment analysis

To visualize genome compartmentalization, Saddle plots were generated using ‘saddle’ function followed by ‘visualize’ from the GENOVA R package ⁹⁰ (https://github.com/robinweide/GENOVA). To distinguish A versus B compartments, instead of H3K27ac active mark suggested by GENOVA we used a consensus ATAC-seq peak BED file derived from cKO and floxed samples by using Genrich tool (v0.6.1). The output results from the function ‘saddle’ were subjected to the function ‘quantify’ to compute the corresponding compartment strengths for pooled replicate samples. To compare the compartment strength scores between cKO and floxed sample, a scatter plot was generated using compartment strength score values of all chromosomes.

Differential compartment analysis was carried out by using dcHiC⁹¹ (v2.1). Input files were prepared following dcHiC guidelines to obtain observed vs expected (O/E) correlation matrices (https://github.com/ay-lab/dcHiC/tree/master). Briefly, Hi-C files at 100 kb-resolution were first preprocessed using ‘preprocess.py’ running in python v3.7.10. Output was used as input for ‘dchic.py’ script to compute compartment differences between samples.

To annotate genes residing in differential compartments, we used annotatr package⁹². For gene ontology analysis, we used only genes, which promoter (defined as sequences < 1 kb upstream of the TSS and 1–5 kb upstream of the TSS) overlapped with differential compartments and had expression value base mean > 100 counts¹⁵.

Topologically associating domains (TADs)

TADs were computed at 25 kb-resolution using SpectralTAD⁴⁵. Briefly, ICED normalized HiC-Pro-derived matrices were converted into ‘bedpe’ format using the ‘hicpro2bedpe’ function from HiCcompare⁹³. Then, pooled samples were used as input for SpectralTAD, which was run iteratively for each chromosome.

Differential TAD boundaries between cKO and floxed Hi-C contact matrices were identified by TADCompare⁴⁵. First, SpectralTAD TAD caller was used to pre-define TAD boundaries and then they were compared between the genotypes. Regions with absolute differential boundary score > 2 (P value smaller than 0.05) were considered as differential boundaries. To annotate genes nearby differential boundaries, GREAT tool was used⁸⁶, with the following association rule: basal + extension: 5 kb upstream, 1 kb downstream and 100 kb max extension.

In aggregate TAD analyses (ATA), TADs were first rescaled to a uniform size and then the result was averaged across the genome, as described in the GENOVA R package⁹⁰. To compute the interaction density within TADs and their 1,2,. . . , N neighbors, “TAD + n” analysis was used. Briefly, ‘TAD + n’ plots were generated using ‘intra_inter_TAD’ followed by ‘visualize’ functions of GENOVA R package⁹⁰.

Chromatin loops

Chromatin loops were called by Mustache (v1.2.0), a local enrichment-based loop calling method⁹⁴. Replicate Hi-C matrices were pooled to be able to perform the analysis at 5 kb resolution. The following parameters for differential loop detection were used: P value (-p) of 0.05 and sparsity (-st) of 0.8, together with KR normalization (“-norm KR” parameter). To assign differential chromatin loops to genes, we intersected loop anchors with gene promoters (defined as sequences < 1 kb upstream of the TSS and 1–5 kb upstream of the TSS) using annotatr package⁹². Loops annotated to genes, which were common between lost and gained promoter-based loops were considered as shared promoter differential loops.

Frequently interacting regions (FIREs)

Frequently interacting regions (FIREs) were called using the FIREcaller tool⁵¹. Briefly, ‘sparseToDense.py’ HiC-Pro utility was used to create quadratic matrices at a resolution of 10 kb. To be able to run FIREcaller, a mappability file corresponding to ArimaHiC fragmentation was required. We provide this file for mm10 here: https://github.com/sespesogil/SATB2_3D_Genome/tree/main/FIREs/resources. We provide code examples for FIRE calling (see “Code Availability”). Differential FIREs were identified as follows: genomic regions that had floxed FIRE scores greater than qnorm (0.975) and cKO FIRE scores lower than qnorm (0.9) were defines as floxed-specific FIREs; conversely, cKO-specific FIREs were defined as genomic regions with cKO FIRE score greater qnorm (0.975) and floxed FIRE score lower than qnorm (0.9)⁴⁹. Floxed- and cKO-specific super-FIREs were identified by filtering out the super-FIREs common to both samples on a genome coordinate level⁴⁹. To link differential FIREs and super-FIREs to genes, annotatr package⁹² was used to intersect FIREs and super-FIREs with all genic annotations (1–5 kb upstream of the TSS, promoter (< 1 kb upstream of the TSS), 5UTR, first exons, exons, introns, CDS, 3UTR).

3D genome modelling

Hi-C data from floxed and cKO cortical neurons were used to model chromatin conformation in 3D. TADs were called for each of the pooled Hi-C libraries (floxed and cKO) at 50 kb-resolution and KR normalized (“-k KR” parameter) using Arrowhead^95,96. Next, we created a consensus TAD file for both conditions by merging the two bed files using ‘bedtools merge’ and used it as an input file for Chrom3D⁵⁷. We generated gtracks using https://github.com/sespesogil/automat_chrom3D. The analysis was restricted to diploid autosomal interactions at 50 kb and 1 Mb for intra and inter-chromosomal interactions, respectively. Chrom3D was run with the nucleus parameter set, with a radius of 3 μm for 1 million iterations: -r 3.0 -n 1000000 -l 5000 --nucleus. Mapping regions to floxed vs cKO 3D models was performed using: https://github.com/sespesogil/automat_chrom3D_colors. The coordinates of the promoters of the mouse orthologues of NCF and CTP genes were obtained by using “AnnotationHub” package (v3.4.0). Pairwise Euclidean distances between regions of interest were computed using: https://github.com/sespesogil/automat_euclidean. The same utility was applied to compute hierarchical clustering using the complete linkage method (‘hclust’ function), integrated in pheatmap R package⁸⁹. The resulting cluster dendograms were compared using dendextend package⁹⁶ and the function ‘CutTheTree.R’ (https://github.com/sespesogil/automat_euclidean). The following parameters were computed: entanglement (measures the quality of the alignment of the two trees, varies between 1 (full entanglement) and 0 (no entanglement)), “Baker’s Gamma Index”, and “Cophenetic” correlation coefficients (both measuring similarity between tree topologies. Statistical significance of Baker’s Gamma Index was calculated by a permutation test with 1000 iterations. Models were visualized using ChimeraX (v1.0)⁹⁷.

ATAC-seq

For each experiment, we used 1×10⁶ primary cortical neurons plated on a single well of a 6-well plate. The neurons were washed twice with 1X PBS before addition of nuclei isolation buffer (1X PBS, 0.1 % Triton X-100, 5 mM AEBSF Pefabloc (Sigma), 5 mM Sodium butyrate, 1 tablet cOmplete EDTA-free protease inhibitor (Roche)). The dish was incubated on a shaker at 4°C for 7 min before detaching the cells with a cell lifter and gently resuspending them with a P1000 pipet. Cell suspension was transferred to an Eppendorf tube and spun down at 500 g at 4 °C for 10 min. Supernatant was removed and pellet was gently resuspended in 1X PBS with protease inhibitors. Intact nuclei were counted in a Neubauer chamber. A total of 25 000 nuclei were used for tagmentation in 4X TD buffer (132 mM Tris-acetate pH 7.8, 264 mM Potassium-acetate, 40 mM Magnesium acetate, 64 % Dimethylformamide, 0.005 % Digitonin and 1.5 μl Tn5 Transposase, Illumina)). Reaction was incubated at 37 °C for 30 min and purified using the MinElute Kit (QIAGEN) according to the manufacturer’s protocol. Libraries were prepared as previously published⁹⁸. Each library was given a unique barcode⁹⁸ and amplified for 11 cycles. Afterwards, the PCR reaction was purified using 1X AMPURE XP Beads (Beckman) and eluted in 33 μl of 10 mM Tris HCl pH 8.0. Libraries were analyzed with Qubit (Thermo Fisher Scientific) and fragment size was accessed using the high sensitivity D-5000 screen tapes on a Tapestation (Agilent). ATAC-seq libraries were sequenced at Novogene UK on a NovaSeq at 150 bp paired-end with a sequencing depth of 150 million reads per replicate (n=3).

Pair-end data were mapped by using BWA-MEM (v0.7.17)⁹⁹ and the resulting SAM files were converted into sorted and indexed BAM files (SAMtools v1.1)¹⁰⁰. The reproducibility of replicate ATAC-seq libraries was assessed by calculating Pearson correlation between BAM files by using deepTools2¹⁰¹. Peaks were identified per experimental condition by employing Genrich tool (v0.6.1). Additionally, a consensus peak file across all conditions was produced by using bedtools ‘merge’. The number of read counts per region was calculated by using featureCounts (subread v.2.0.1)¹⁰². Next, differential analysis was done by using edgeR¹⁰³. Principal component analysis showed a technical bias and samples were batch-normalized using RUVr method of RUVseq¹⁰⁴. Homer (v4.11) was used for TF motif analysis ¹⁰⁵. Genomic annotation and GO enrichment analysis was done by using ChiPseeker¹⁰⁶ and g:Profiler¹⁰⁷. Footprinting analysis and occupancy prediction was performed by using TOBIAS (v0.12.10)¹⁰⁸.

CUT&RUN

On DIV14, primary cortical neurons were washed twice with room temperature 1X PBS. Nuclei were isolated by addition of Nuclei Extraction Buffer (NEB, 20 mM Hepes (KOH) pH 7.9, 10 mM KCl, 0.5 mM Spermidine, 0.1 % Triton X-100, 20 % Glycerol, 1 tablet cOmplete EDTA-free Protease inhibitor (Roche)). Cells were incubated on a shaker at 4 °C for 15 min, lifted from the dish using a cell lifter and gently resuspended using a P1000 pipet. The lysate was transferred to a fresh Eppendorf tube and spun down at 600 g at 4 °C for 3 min. The supernatant was carefully discarded and the pellet was resuspended in NEB. Intact nuclei were stained with Trypan Blue stain and counted in a Neubauer chamber. A total of 100 000 nuclei were used for each CUT&RUN reaction.

CUT&RUN protocol was adapted from the previously published protocol¹⁰⁹. Per reaction, 150μl Concavalin-A conjugated beads (Epicypher) were activated by washing twice in 1ml cold Binding Buffer (20mM HEPES-KOH, pH 7.9, 10mM KCl, 1mM CaCl2, 1mM MnCh) on a magnetic stand and then resuspended in 300 μl Binding Buffer. Freshly extracted nuclei were added to the activated beads and incubated at RT for 10 min. Beads-bound nuclei were then blocked in Blocking Buffer (20 mM Hepes pH 7.5, 150 mM NaCl, 0.5 mM Spermidine, 0.1 % BSA, 2 mM EDTA, 1 tablet cOmplete EDTA-free Protease inhibitor) at RT for 5 min. After washing in 1ml cold Washing Buffer (WB, 20 mM Hepes pH 7.5, 150 mM NaCl, 0.5 mM Spermidine, 0.1 % BSA, 1 tablet cOmplete EDTA-free Protease inhibitor), nuclei were resuspended in 250 μl cold Washing Buffer. Primary antibodies (anti-SATB2, ab92446, Abcam or rabbit IgG, CUTANA^™ Rabbit IgG, Epicypher) were diluted 1:50 in 250 μl cold WB and added to the beads (final dilution of 1:100 in 500 μl volume). Beads were incubated with rotation at 4 °C for 2 h. Beads were washed twice in 1ml cold Washing Buffer and resuspended in 250μl cold Washing Buffer. 12μl of pAG-MNAse (15–1016-EPC Epicypher, 1:20) was added to the mix and incubated with rotation for 1h at 4°C. Next, beads were washed twice in 1ml cold Washing Buffer and resuspended in 150μl cold Washing Buffer. The samples were equilibrated to 0 °C in an ice-water bath and 3μl of 100 mM CaCl2 were added to the reaction. After an incubation of 30 min at 0 °C on a shaker, 150μl 2X STOP buffer (200 mM NaCl, 20 mM EDTA, 4 mM EGTA, 50 μg/ml RNAseA, 40 μg/ml Glycogen) was added and beads were incubated for 20 min at 37 °C. DNA was purified by phenol-chloroform extraction and precipitated in 95 % ethanol at −80 °C. Pellet was dissolved in TE buffer pH 8.0 and stored at −20 °C.

Libraries were prepared using the NEBNext Ultra Library prep Kit (E7645S, New England Biolabs) according to the manufacturer’s protocol. Adapters were used at a working concentration of 1.5 μM. Libraries were PCR amplified with unique indexes (New England Biolabs) for 14 cycles. After clean-up, fragment size and concentration was analyzed by high sensitivity D-5000 screen tapes on a Tapestation (Agilent) and Qubit (Thermo Fisher Scientific) before sequencing. When needed, libraries were size-selected with SPRIselect beads (B23317, Beckman Coulter). All pre-made libraries were sequenced at Novogene UK on Illumina Novaseq platform (150 bp paired-end sequencing). Each replicate (n=2) was sequenced at a sequencing depth of 30 million reads.

After running a quality control on raw reads with FastQC, Illumina Universal adapters were trimmed with Trim Galore! using default parameters (--paired --illumina). Next, the trimmed reads were aligned to the mouse genome (mm10) with Bowtie2 (v2.4.5)¹¹⁰ with the following settings: bowtie2 --dovetail --local --very-sensitive-local --no-unal --no-mixed --no-discordant -q --phred33 -I 10 -X 700 -p 48 -x mm10. Subsequently, the mapped reads were filtered with SAMtools view. Reads with a mapping quality score (MAPQ) < 2 as well as optical and PCR duplicates are excluded. The reproducibility of replicate CUT&RUN libraries was assessed by calculating Pearson correlation between BAM files by using deepTools2¹⁰¹. Genome coverage bedgraph files were computed with bamCoverage (bin size 50 bp) and normalized to reads per kilobase per million (RPKM) to minimize bias arising from different sequencing depths and batches. Genome-wide average signal was computed and subtracted from the RPKM- normalised bedgraph files to further exclude residual background noise. Peaks were called with SEACR v1.4¹¹¹ under relaxed conditions and normalized to the IgG control bedgraph file. Replicates were handled as follows: the analysis steps from generation of genome coverage bedgraph file to peak calling were performed in parallel for each of replicate, merged replicates and two pseudoreplicates of the merged file (each generated by random down sampling to 50% of the total reads with Picard DownsampleSAM/BAM). Peaks were called using the corresponding IgG replicate, merged replicates or two pseudoreplicates. To generate a final list of consensus peaks, the peaks overlapping between the two replicates and the merged file (generated with Bedtools Intersect) were retained and added to the peaks overlapping between the two pseudoreplicates. Peaks were called with SEACR v1.4¹¹¹ under relaxed conditions and normalized to the IgG control bedgraph file. Genomic annotation and GO enrichment analysis was done by using ChIPseeker¹⁰⁶ and clusterProfiler¹¹². Peaks were annotated using annotatr package⁹² and assigned to genes, if overlapping with gene promoters (1 kb upstream of TSS). Genomic enrichment analysis was performed by Homer (v4.11)¹⁰⁵.

RNA-seq

Primary cortical neurons (1×10⁶) were transduced with AAVs (AAV8-hSyn-EGFP or AAV8-hSyn-V5-SATB2) at multiplicity of infection 1.5 χ 10⁵ at DIV4 and used for RNA-seq at DIV14.

Cells were washed twice with 1X PBS and collected in 1 ml of Trizol (Invitrogen, #15596026). RNA was extracted using the PureLink^™ RNA Micro Scale Kit (Invitrogen, 12183016) according to the manufacturer’s protocol. RNA quality was analyzed using an Agilent Tapestation. Samples with RIN score > 8.4 were used for library preparation and sequencing. PolyA-enriched libraries were generated by Novogene UK and sequenced at Illumina Novaseq platform at 150 bp PE, with a sequencing depth of 30 million reads per replicate.

Pair-end FastQ files were mapped to mm10 using STAR aligner (v2.7.6a)¹¹³. Resulting BAM files were sorted and indexed using samtools (v1.11). Next, BAM files were used to generate a count matrix using featureCounts (Rsubread v2.0.1)¹¹⁴. Differential gene expression analysis was performed in DESeq2¹¹⁵, accounting for hidden batch effects by the removal of unwanted variation (RUVg, k =5) method¹⁰⁴. For differential gene expression, a threshold cutoff of adjusted (Benjamini-Hochberg) P value < 0.05 was applied. Rld-normalized expression values (rlog, DESeq2) were used for clustering and heatmap analysis. The gene expression profiles of cKO vs floxed cultures (both AAV-EGFP-transduced) and cKO vs “rescued” cultures (AAV-EGFP-transduced and AAV-SATB2-transduced, respectively) were compared by means of a rank-rank hypergeometric overlap (RRHO) analysis¹¹⁶. RRHO heat maps that graphically visualize correlations between two expression profiles were generated at http://systems.crump.ucla.edu/rankrank/.

Fluorescence in situ hybridization (FISH)

DIV 14 primary cortical neurons (floxed, cKO, and AAV8-hSyn-V5-SATB2-transduced cKO), cultured on glass coverslips, were washed twice in 1X PBS and fixed in 4 % PFA in PBS for 10 min at room temperature (RT). After washing in 1X PBS cells were stored at 4°C until used for 3D DNA FISH.

Coverslips were transferred to a 12-well plate for permeabilization, blocking and washing steps. In detail, cells were permeabilized in 0.5 % Triton-X in PBS for 10 min at RT. After washing in 1X PBS for three times, cells were blocked in PBS+ blocking solution (0.1 % Casein, 1 % BSA, 0.2 % fish skin gelatin in 1X PBS, pH 7.9). Blocking was performed at RT for 1h while mild shaking. After additional three washing steps in 1X PBS, cells were permeabilized on ice in permabilization solution (0.1 M HCl, 0.7 % Triton-X, in ddH20) for 10 min. Cells were then washed 3 times with 2X SSC for 3 min each. Cellular DNA was denatured using 70 % formamide in 2X SSC for 30 min on 80 °C. After denaturation, coverslips were rinsed in ice cold 2X SSC and inverted over a 10 μl drop of each DNA probe in diluted in probe buffer according to the Empire Genomics protocols. Humidity was generated in a sealed chamber using 70 % formamide in 2X SSC and FISH reaction was incubated in this chamber overnight at 37°C for 16h up to 24 h. After hybridization, coverslips were carefully rinsed three times in preheated 70 % formamide in 2X SSC at 42°C for 3 min each, washed another three times in 2X SSC at RT and quickly rinsed with 1X PBS. Before mounting, coverslips were rinsed with ddH₂0 and mounted on a microscopy slide in Fluoromount mounting medium with DAPI (Roth).

Z-stacks were acquired at 0.130 μm intervals using Zeiss LSM700 microscope (Zeiss Plan-Apochromat 63x/1.40 oil M27 objective) at 1024× 1024 pixel resolution and voxel size of 0.050 μm x 0.050 μm x 0.13 μm. Z-stacks were analyzed using Imaris (v.10.0.1) image analysis software. First, the volume of each nucleus was measured by using the “cells” tool. Only nuclei that were separated enough to be identified as individual nuclei by the mask tool were considered in the downstream analysis. Next, the “spots” tool was used to identify the FISH signal in each of the channels. The quality threshold was applied to limit the FISH signal to two distinct spots per nucleus in each channel. The x, y and z coordinates of the center of the mass of each spot were determined and the interalleleic pairwise distances were calculated using the following formula:

$$Interallelicpairwisedistance=\sqrt{{(x1+x2)}^2+{(y1+y2)}^2+{(z1+z2)}^2}$$

Per nucleus, the average of four pairwise interallelic distances was calculated and normalized to the volume of the nucleus. Per condition, 75–80 nuclei were scored.

Gene Ontology enrichment analysis

Metascape¹¹⁷ was used to identify overrepresented GO terms in the genes, affected by SATB2-mediated 3D epigenome alterations. The following gene-sets were analyzed: genes residing in differential compartments, genes located nearby differential TAD boundaries, genes mapped to promoter-based differential loops, genes interacting with dOCRs via invariant loops, genes assigned to floxed-specific FIREs/super-FIREs. As ontology sources, GO Molecular Functions, GO Biological Processes, and GO Cellular components were used. Terms with a P value < 0.01, a minimum count of 3, and an enrichment factor > 1.5 (the ratio between the observed counts and the counts expected by chance) were collected and grouped into clusters based on their membership similarities. Representative GO terms per cluster are displayed in the figures. All data associated with these graphs are included in Table S5. P values were calculated based on a cumulative hypergeometric distribution, and q-values were calculated using the Banjamini-Hochberg procedure for multiple testing. All genes in the genome were used as enrichment background. For the GO analysis of genes interacting with dOCRs via invariant loops custom background was used, consisting of genes connected to any OCRs via invariant loops.

GO enrichment analysis of genes with SATB2 bound promoters was performed by using ChIPseeker¹⁰⁶ and clusterProfiler¹¹². Peaks were annotated using annotatr package⁹² and assigned to genes, if overlapping with gene promoters (1 kb upstream of TSS).

Genomic annotation and GO enrichment analysis of dOCRs and invariant OCRs was done by ChiPseeker¹⁰⁶ and g:Profiler¹⁰⁷. The genes assigned to all identified OCRs were used as a custom background.

Overrepresentation analysis

We employed Fisher’s exact tests to test for significant overlaps between the genes influenced by SATB2 3D epigenome effects and different gene lists (Table S1).

The list of DEGs between floxed and cKO neurons was obtained from Feurle et al.¹⁵. In this study, the expression of 1372 and 1639 genes was found to be decreased and increased, respectively (adjusted p-value < 0.05, log2FC threshold = 0.3). Functional GO annotation of this DEG set has revealed several overrepresented GO categories, among them “cellular response to hormone stimulus”, “head development” and “trans-synaptic signaling”¹⁵.

The list of ARGs was sourced from Tyssowski et al.²⁵, MEF2C-regulated genes from Harrington et al.²⁸, and the ‘driver’ genes, mapped to cognitive task performance (CTP) meta-loci and performance on a non-cognitive factor (NCF) meta-loci form Lam et al.⁵². Gene products associated with the GO terms “axon development”, “postsynapse”, “neuron projection morphogenesis”, “glutamatergic synapse”, “cognition”, “T-cell activation”, “immune response” were downloaded from http://amigo.geneontology.org/amigo/landing. GREATv4 genes (GREATv4.genes.mm10.tsv) that were based on extremely high-confidence gene predictions were used as background reference set (n = 21395).

Association analysis of genomic regions based on permutation tests

A dedicated wrapper named ‘Enriched permutate’ was created to perform permutation analysis of genomic regions (see Code Availability). Briefly, it computes the probability of genomic features to overlap using regioneR package¹¹⁸. Permutations were performed using bins of 5 kb (for the analysis of loop anchors) and 1 kb (for the analysis of SATB2 peaks and OCRs) over mm10 genome, and of 500 bp (for the analysis of lifted-over dOCRs) over hg38 genome, with a total number of 100 000 iterations. For permutations of mouse regulatory elements, ChromHMM 18-state model of postnatal day 0 mouse forebrain was used (ENCSR301UGN)³³. Before permutations, multiple enhancer states (EnhG, Enh, EnhLo, EnhPois, EnhPr), Transcription Start Site states (Tss, TssFlnk, TssBiv), repressive states (ReprPC and ReprPCWk), quiescent states (QuiesG, Quies, Quies2, Quies3, and Quies4), and the two states associated with actively transcribed genes (Tx and TxWk) were concatenated into a representative single state (Enh, Tss, ReprPC, Quies, Tx). For association analysis of human regulatory elements, ChromHMM 18-state model of frontal cortex female adult (67 years) and female adult (80 years) from donor(s) ENCDO311AAA, ENCDO312AAA was used (ENCFF382TUC)^119,120. Enhancer states (Enh A1, EnhA2, EnhBiv, EnhG1, EnhG2, and EnhWk) were concatenated as described above.

For enrichment analysis of histone marks and CTCF peaks, the following postnatal day 0 mouse forebrain datasets were used: CTCF (ENCFF430PPJ); H3K27ac (ENCFF676TSV); H3K4me3 (ENCFF160SCR)^119,120.

Conventional MAGMA and Hi-C-coupled MAGMA (H-MAGMA) gene-set analyses

Gene-set analysis (GSA) was used to analyze the effect of multiple SNPs and their joint effect toward a phenotype within defined genes of interest. GSA was performed using MAGMA⁵⁹ using test and control GWAS summary statistics. First, SNPs from European 1000 Genomes phase 3 cohort (https://ctg.cncr.nl/software/MAGMA/ref_data/g1000_eur.zip) were annotated in relation to their location within genes on the GRCh37/hg19 human build using start/stop coordinates (https://ctg.cncr.nl/software/MAGMA/aux_files/NCBI37.3.zip) and a 20 kb window. Next, P values for each genes association with the relevant phenotypes were calculated, using a linear principal components regression model accounting for linkage disequilibrium (LD) between SNPs and employing 1000 Genomes European cohort data as a reference panel (gene-based analysis). Finally, a competitive GSA was performed based on P values from the gene-based analysis, testing if the gene-set has a stronger association with the phenotype of interest than other genes in the genome. MAGMA-GSA also accounted for gene size and gene density¹²¹.

While MAGMA uses locational proximity to annotate SNPs to genes, H-MAGMA incorporates tissue-specific Hi-C chromatin interactions to assign SNPs to genes⁶⁰. The previously generated annotation files (“Cortical_Neurons.genes.annot” and “Fetal_brain.genes.annot”), based on cortical neuronal and fetal brain Hi-C data were used to provide SNP-to-gene relationships⁴⁹. Gene-based analysis and competitive GSA were performed as in MAGMA above.

GWAS Data

For both MAGMA and H-MAGMA analyses, gene sets were tested for enrichment of genes associated with five neurodevelopmental test phenotypes using genome-wide association study (GWAS) summary statistics for schizophrenia (SCZ)⁹, Intelligence (IQ)⁹, Educational Attainment (EA)¹²², Bipolar Disorder (BPD)¹²³, and Major Depressive Disorder (MDD)¹²⁴. In addition to the test phenotypes, five control phenotypes were also used: Alzheimer’s disease (AD)¹²⁵, Stroke¹²⁶, Coronary Artery Disease (CAD)¹²⁷, Crohn’s Disease(CD)¹²⁸, and Type 2 Diabetes (T2D)¹²⁹.

Stratified Linkage Disequilibrium Score Regression (sLDSC)

To investigate if genomic regions were enriched for heritability contributing to neuropsychiatric disorders and cognition-associated phenotypes, stratified Linkage Disequilibrium Score Regression (sLDSC) (https://github.com/bulik/ldsc)^130,131 was performed. Regions on mm10 mouse genome were lifted over to the human hg19 genome using UCSC liftOver (https://genome.ucsc.edu/cgi-bin/hgLiftOver) with default settings. To avoid significantly larger human regions inflating heritability enrichment values, lifted over regions more than 2 SD larger than the mean were removed. HapMap Project phase 3 SNPs with a MAF > 0.05 in the lifted over regions were considered in this analysis. LD scores between SNPs within a 1 centimorgan (cM) window were estimated using the 1000 Genomes Phase 3 European reference panel. SNP heritability for each phenotype was stratified within each set of regions using a model accounting for heritability associated with 53 functional genomic annotations found in the baseline model¹³⁰. Enrichment for heritability compared to the baseline model was calculated with a corresponding P value. Enrichments surviving a Bonferroni-corrected P < 0.00125 were considered as significantly enriched.

Enrichment analysis for genes containing de novo mutations

Lists of de novo mutations (DNMs) identified in patients with ASD (n = 6,430), ID (n = 192) and in unaffected siblings (n = 1995) and controls (n = 54) based on exome sequencing of trios were sourced from⁶⁴ and⁶⁵. Genes containing DNMs reported in SZ patients (n = 3394) were taken from⁶² and⁶³. DNMs identified in Developmental Disorder (DD) patients were sourced from⁸¹ but were subject to additional filtering based on posterior probability of de novo mutations, as described in⁶⁶. DNMs were categorized as synonymous, missense and loss-of-function (includes nonsense, frameshift and splice site mutations). Each gene-set was tested for enrichments for genes harboring DNMs in each phenotype using the R package denovolyzeR¹³². To do this, denovolyzeR counts the numbers of observed DNMs and derives the expected number of DNMs in a given population based on the mutability of the gene and the number of trios sequenced¹³². Enrichment of DNMs in a test gene-set was investigated using a two-sample Poisson rate ratio test, using the ratio of observed to expected DNMs in genes outside of the gene-set as a background model.

IV. QUANTIFICATION AND STATISTICAL ANALYSIS

Data were obtained from independent biological replicates (cultured cortical neurons isolated from independent mouse litters, as described in the method details). Descriptions of data collection, data quantification, and statistical methods used to analyze each experiment are provided in the respective sections in the method details or in figure legends. Exact P values and number of replicates (n) can be found in the main text and/or figure legends.

V. ADDITIONAL RESOURCES

“SATB2 3D Genome Viewer”: a ShinyR App, based on GENOVA package, for visualization of the Hi-C data, generated in this study, available at: https://shiny.i-med.ac.at/satb2/. Chromatin loops and TADs in floxed and Satb2 cKO Hi-C matrices can be visualized at different resolutions and user-selected genomic coordinates. The App will be publicly available as of the date of publication.

Supplementary Material

2

Table S1. Overrepresentation analysis and enrichment analysis for SATB2 peaks and CTCF binding sites. Related to Figures 1–3, 6.

3

Table S2. Genomic coordinates of SATB2-dependent regulatory element-promoter interactions, differential A/B compartments, TAD boundaries, and FIRES/super-FIREs. Related to Figures 2, 5–6.

4

Table S3. Gene sets, representing SATB2 effects at individual 3D genome hierarchical levels. Related to Figures 2, 5–7.

5

Table S4. Homer TF motif enrichment analysis. Related to Figure 2.

6

Table S5. Gene Ontology enrichment analysis of gene sets, using Metascape¹¹⁷. Related to Figures 2, 5-6.

7

Table S6. MAGMA-GSA and H-MAGMA-GSA analysis of SATB2 3D genome sets. Related to Figure 7.

8

Table S7. Enrichment analysis for genes containing de novo mutations, reported in trio-based studies of SCZ, ASD, intellectual disability, and developmental disorders. Related to Figure 7.

9

Table S8. Stratified Linkage Disequilibrium Score Regression (sLDSC) analysis. Related to Figure 7.

Highlights.

• SATB2 loss in cortical neurons affects 46 % of cognition-related genes

• SATB2 modifies 3D genome architecture at multiple hierarchical levels

• SATB2 loss triggers extensive changes in chromatin accessibility

• SATB2-dependent chromatin modifications contribute to neuropsychiatric risk