High-throughput in silico screen uncovers key regulators of 3D genome architecture

Abstract

The vertebrate genome is spatially organized into topologically associating domains (TADs), primarily via cohesin-mediated loop extrusion which typically halts at convergent CTCF binding sites to establish domain boundaries. However, despite the essential roles of CTCF and cohesin in establishing TADs, a long-standing paradox persists: CTCF and cohesin binding sites dramatically outnumber observed TAD boundaries, suggesting the existence of undiscovered architectural factors. To identify such missing factors, we conducted high-resolution in silico screens using C.Origami, a multi-modal AI model for predicting chromatin interactions. Remarkably, we identified ZNF654 and JMJD6 as novel factors uniquely defining TAD boundaries. Experimental validation confirmed that ZNF654, an uncharacterized vertebrate-specific zinc-finger protein, interacts with CTCF to form an architectural protein complex that demarcates chromatin domains. Genetic knockout of ZNF654 weakens TAD boundary strength without influencing other CTCF or cohesin binding sites. JMJD6, a deeply conserved jmjC-family dioxygenase, marks the anchors of the strongest chromatin stripes at both TAD boundaries and enhancer-promoter sites, while deleting JMJD6 weakens or diminishes such interaction signature. These results revealed the long-sought factors that uniquely mark TAD boundary and chromatin interaction anchors which, together with CTCF and cohesin, demarcate chromatin domains during 3D genome organization. Last, the evolutionary trajectory of ZNF654 and JMJD6 offers key insight into the evolutionary origins of 3D genome organization across metazoan species.

Figure 1.. In silico screen identified ZNF654 and JMJD6 as top candidates for 3D genome regulation.

a, Schematic of the high-throughput in silico deletion screen workflow using C.Origami. b, Ranking of chromatin proteins by their enriched binding at structurally high-impact regions, with the top 10 factors highlighted. c, UMAP visualization of ChIP-seq signal intensities for CTCF, RAD21, ZNF654, and JMJD6 across ~100,000 key CREs. d, Genomic track view at the HOXB gene cluster locus in HEK293T cells, showing ChIP-seq signals for CTCF, RAD21, ZNF654, and JMJD6. e, Venn diagram showing overlap of CTCF, RAD21, ZNF654, and JMJD6 occupied sites. f, Heatmaps of ChIP-seq signal centered on CTCF sites grouped by co-occupancy with RAD21, ZNF654, and JMJD6.

Figure 2.. ZNF654 and JMJD6 mark the strong TAD boundaries and loop anchors

a, Micro-C pile-up visualization of chromatin interaction intensities flanking the loci belonging to the four CTCF-occupied groups. For genomic loci in each group, the Micro-C contact map is centered at the factor binding site, and visualized at 5Kb/pixel resolution. b, Heatmap visualization of insulation scores at each genomic locus (5-kb resolution). Loci in each category was sorted by CTCF strength. Average Insulation score profile was displayed on top of each heatmap. c, Aggregate peak analysis (APA) visualization of chromatin loops centered on peak pairs within each binding category. d, Aligned visualization of Micro-C contact maps and ChIP-seq tracks at a representative genomic region (chr1:240.7–241.9 Mb) in HEK293T and HCT116 cells. Genomic loci marking cell-type-specific and conserved structural sites are highlighted in yellow and purple, respectively. Other CTCF binding sites are marked in grey.

Figure 3.. ZNF654 strengthens TAD boundary through interacting with CTCF

a-b, CRISPR-Cas9 strategy for generating ZNF654 knockout (KO) alleles by dual sgRNAs targeting exon 1 of the gene, generating an 85bp deletion that leads to frameshift and early stop codon that disrupt the protein. c, Micro-C contact maps in WT (top) and ZNF654-KO (middle) HEK293 cells at 10Kb resolution. Arrows indicate weakened loops in ZNF654-KO cells. Three loops are highlighted and shown at 5 kb resolution in zoomed-in views. Green dashed box indicates a weakened boundary. Difference map (bottom) of chromatin interaction changes (KO relative to WT) shows increased inter-TAD interactions upon knockout of ZNF654. d, Pile-up visualization of chromatin interaction changes across the four CTCF-occupied groups, highlighting the dramatic increase of inter-TAD interactions spanning the ZNF654-occupied loci upon knockout of ZNF654. The three corner numbers in each heatmap indicate the total interaction changes in the dash-squared area of each corner. e, Scatter plot comparing loop strength between wild-type and ZNF654-KO cells at ZNF654–CTCF co-occupied loop sites, with the number of significantly different loops annotated. f, Aggregate peak analysis (APA) visualization of significantly weakened or strengthened ZNF654–CTCF loops. g-h, Same as (e-f) for ZNF654-anchored enhancer–promoter loops. i, Anti-FLAG co-immunoprecipitation pull-down followed by CTCF, RAD21, SMC1, SMC3, STAG1 and STAG2 western blot in HCT116 3xFLAG-ZNF654 cells. All proteins are expressed at their endogenous level. j, Reciprocal IP pull-down using antibodies against CTCF or cohesin proteins followed by blotting FLAG tag (3xFLAG-ZNF654).

Figure 4.. JMJD6 strengthens stripe and loop interactions.

a, High-resolution Micro-C maps (500bp resolution) comparison between HEK293T and HCT116 cells. Loops and stripes are annotated, with associated ChIP-seq tracks and calculated insulation and stripe scores shown below. Orange highlights mark loop extrusion anchors; blue and yellow highlight conserved and cell-type-specific stripe anchors, respectively. b, UMAP projection of ChIP-seq signals showing the co-occupancy by JMJD6 (newly validated binding), H3K27ac, H3K4me3, and RNA Pol II binding profiles. Note that JMJD6 also binds at TAD boundary regions, as presented in Fig. 1c. c-d, UMAP visualization of the distance to the nearest transcription start site (TSS, c) and stripe scores (d) for each CRE in HEK293 cells. e, Heatmaps of stripe scores (1 kb resolution, calculated from normalized Micro-C matrix) centered across the five CRE categories. JMJD6–H3K27ac group show the strongest stripe scores, followed by the C–R–Z–J co-occupying group. f, High-resolution (1Kb) aggregate peak analysis (APA) visualization of chromatin loops centered on peak pairs across the five CRE categories. g, Micro-C contact maps at the AC005307.3 locus in HEK293T cells showing loss of cohesin loops upon JMJD6 knockout. h-i, Knockout of JMJD6 leads to decreased strength of stripe signals in JMJD6-occupied regions (JMJD6–H3K27ac group and C–R–Z–J co-occupying group). j-m, Scatter plot and APA visualization comparing loop strength between wild-type and JMJD6-KO cells at JMJD6–CTCF co-occupied loop sites (j-k) and JMJD6-anchored E–P/P–P loop sites (JMJD6–H3K27ac, l-m).