Tumor cell gene expression programs are typically driven by somatic mutations that alter the coding sequence or expression of proto-oncogenes (1) (Fig. 1A), and identifying such mutations in patient genomes is a major goal of cancer genomics (2, 3). Dysregulation of proto-oncogenes frequently involves mutations that bring transcriptional enhancers into proximity of these genes (4). Transcriptional enhancers normally interact with their target genes through the formation of DNA loops (5-7), which typically are constrained within larger CTCF-cohesin mediated loops called insulated neighborhoods (8-10), which in turn can form clusters that contribute to topologically associating domains (TADs) (11, 12) (Fig. S1A). This recent understanding of chromosome structure led us to hypothesize that silent proto-oncogenes located within insulated neighborhoods might be activated in cancer cells via loss of an insulated neighborhood boundary, with consequent aberrant activation by enhancers that are normally located outside the neighborhood (Fig. 1A, lowest panel).

Open in a separate window

Fig. 1

3D regulatory landscape of the T-ALL genome

(A) Mechanisms activating proto-oncogenes.

(B) Hi-C interaction map and TADs defined in hESC (H1), and cohesin ChIA-PET interactions (intensity of blue arc represents interaction significance), CTCF and H3K27Ac ChIP-Seq profiles and peaks, and RNA-Seq in Jurkat cells at the CD3D locus. ChIP-Seq peaks are denoted as bars above ChIP-Seq profiles.

(C) ChIA-PET interactions at the RUNX1 locus displayed above the ChIP-Seq profiles of CTCF, cohesin (SMC1) and H3K27Ac.

To test this hypothesis, we first mapped neighborhoods and other cis-regulatory interactions in a cancer cell genome using Chromatin Interaction Analysis by Paired-End Tag Sequencing (ChIA-PET) (Fig. 1B, Table S1). A T-cell acute lymphoblastic leukemia (T-ALL) cell (Jurkat) was selected for these studies because key T-ALL oncogenes and genetic alterations are well-known (13, 14). The ChIA-PET technique generates a high-resolution (~5kb) chromatin interaction map of sites in the genome bound by a specific protein factor (8, 15, 16). Cohesin was selected as the target protein because it is involved in both CTCF-CTCF interactions and enhancer-promoter interactions (5-7), and has proven useful for identifying insulated neighborhoods (8, 10) (Fig S1A-B). The cohesin ChIA-PET data were processed using multiple analytical approaches (Fig. S1-4, Table S2) and identified 9,757 high-confidence interactions, including 9,038 CTCF-CTCF interactions and 379 enhancer-promoter interactions (Fig. S4C). The CTCF-CTCF loops had a median length of 270 kb, contained on average 2-3 genes and covered ~52% of the genome (Table S2). Such CTCF-CTCF loops have been called insulated neighborhoods because disruption of either CTCF boundary causes dysregulation of local genes due to inappropriate enhancer-promoter interactions (8, 10). Consistent with this, the Jurkat chromosome structure data showed that the majority of cohesin-associated enhancer-promoter interactions had endpoints that occurred within the CTCF-CTCF loops (Fig. 1C, S2H). These results provide an initial map of the 3D regulatory landscape of a tumor cell genome.

We next investigated the relationship between genes that have been implicated in T-ALL pathogenesis and the insulated neighborhoods. The majority of genes (40/55) implicated in TALL pathogenesis (curated from the Cancer Gene Census and individual studies)(Table S3) were located within the insulated neighborhoods identified in Jurkat cells (Fig. 2A, S5); 27 of these genes were transcriptionally active and 13 were silent based on RNA-Seq data (Fig. 2A, Table S4). Active oncogenes are often associated with super-enhancers (17, 18), and we found that 13 of the 27 active T-ALL Pathogenesis Genes associated with super-enhancers (Fig. 2A-B, S5A). Silent genes have also been shown to be protected by insulated neighborhoods from active enhancers located outside the neighborhood, and we found multiple instances of silent proto-oncogenes located within CTCF-CTCF loop structures in the Jurkat genome (Fig. 2A, 2C, Fig. S5B). Thus, both active oncogenes and silent proto-oncogenes are located within insulated neighborhoods in these T-ALL cells.

Open in a separate window

Fig. 2

Active oncogenes and silent proto-oncogenes occur in insulated neighborhoods

(A) T-ALL Pathogenesis Genes. Colored boxes indicate whether a gene is located within a neighborhood, expressed and associated with a super-enhancer.

(B) Insulated neighborhood at the active TAL1 locus. The cohesin ChIA-PET interactions are displayed above the ChIP-Seq profiles of CTCF, cohesin (SMC1) H3K27Ac, and RNA-Seq track. A model of the insulated neighborhood is shown on the right.

(C) Insulated neighborhood at the silent LMO2 locus.

If some insulated neighborhoods function to prevent proto-oncogene activation, some T-ALL tumor cells may have genetic alterations that perturb the CTCF boundaries of neighborhoods containing T-ALL oncogenes. To investigate this possibility, we identified recurrent deletions in T-ALL genomes that span insulated neighborhood boundaries using data from multiple studies (Table S5A) and filtered for relatively short deletions (<500 kb) in order to minimize collection of deletions that affect multiple genes (Fig. S6A). Among the 438 recurrent deletions identified with this approach, 113 overlapped at least one boundary of insulated neighborhoods identified in T-ALL, and 6 of these affected neighborhoods containing T-ALL Pathogenesis Genes (Fig. S6B, Table S5B). Examples of two such genes, TAL1 and LMO2, are shown in Fig. 3A and Fig. 3G.

Open in a separate window

Fig. 3

Disruption of insulated neighborhood boundaries is linked to proto-oncogene activation

(A) Cohesin ChIA-PET interactions, CTCF and cohesin (SMC1) binding profiles at the TAL1 locus in Jurkat cells. Patient deletions described in (22) are shown as bars below the gene models. The deletion on the bottom indicates the minimally deleted region identified in (26).

(B) ChIP-Seq profiles of CTCF, H3K27Ac, p300 and CBP, and RNA-Seq at the TAL1 locus in HEK-293T cells. The region deleted using a CRISPR/Cas9-based approach is highlighted in a grey box.

(C) qRT-PCR analysis of TAL1 expression in wild type HEK-293T cells (wt), and in cells where the neighborhood boundary highlighted on (B) was deleted.

(D) Model of the neighborhood and perturbation at the TAL1 locus.

(E) 5C contact matrices in wild type HEK-293T cells and TAL1 neighborhood boundary – deleted cells. The position of the region removed in the mutant cells is highlighted with an arrow.

(F) Distance adjusted z-score difference (5C) maps at the TAL1 locus (ΔCTCF - wild type HEK-293T). Note the increase in the 5C signal adjacent to the deleted region. CTCF and H3K27Ac binding profiles in wt cells are displayed for orientation.

(G) Cohesin ChIA-PET interactions, CTCF and cohesin (SMC1) binding profiles at the LMO2 locus. Patient deletions described in (22) are shown as bars below the gene models.

(H) ChIP-Seq binding profile of CTCF and H3K27Ac, p300 and CBP, and RNA-Seq at the LMO2 locus in HEK-293T cells. The region deleted by a CRISPR/Cas9-based approach is highlighted in a grey box.

(I) qRT-PCR analysis of LMO2 expression in wild type HEK-293T cells (wt), and in cells where the neighborhood boundary highlighted on (H) was deleted.

(J) Model of the neighborhood and perturbation at the LMO2 locus

(K) 5C contact matrices in wild type HEK-293T cells and LMO2 neighborhood boundary – deleted cells. The position of the region removed in the mutant cells is highlighted with an arrow.

(L) Distance adjusted z-score difference (5C) maps at the LMO2 locus (ΔCTCF - wild type HEK-293T). Note the increase in the 5C signal adjacent to the deleted region. CTCF and H3K27Ac binding profiles in wt cells are displayed for orientation.

On (C) and (I) data from n=3 independent biological replicates are displayed as mean + SD; P<0.01 between wt and boundary-deleted cells (two-tailed t-test).

If deletions overlapping neighborhood boundaries can cause activation of proto-oncogenes within the loops, then site-specific deletion of a loop boundary CTCF site at the TAL1 locus should be sufficient to activate these proto-oncogenes in non-malignant cells. TAL1 encodes a transcription factor that is overexpressed in ~50% of T-ALL cases and is a key oncogenic driver of this cancer (19, 20). TAL1 can be activated by deletions that fuse a promoter-less TAL1 gene to the promoter of STIL (19) and this was observed in many patient deletions (Fig. 3A). Several patient deletions, however, retained the TAL1 promoter (endpoint >5kb from promoter) but overlapped the CTCF boundary site of the TAL1 neighborhood (Fig. 3A), and TAL1 was active in the samples harboring these deletions (Fig. S7A-B). This suggests disruption of the insulated neighborhood, allowing activation of TAL1 by regulatory elements outside of the loop. We tested this idea by CRISPR/Cas9 mediated deletion of the TAL1 neighborhood boundary in human embryonic kidney cells (HEK-293T) (Fig. 3B). In these cells, the TAL1 proto-oncogene is silent as evidenced by low H3K27Ac occupancy and RNA-Seq (Fig. 3B). However, at least one active regulatory element occurs ~60kb upstream of TAL1, adjacent to the CMPK1 promoter, as evidenced by high levels of H3K27Ac and p300/CBP (Fig. 3B) and enhancer reporter assays (Fig. S8A-B). Deletion of a ~400 bp segment encompassing the boundary CTCF site, which abolished CTCF binding (Fig. S8A), caused a 2.3-fold induction of the TAL1 transcript (Fig. 3C), suggesting that the integrity of the neighborhood contributes to the silent state of TAL1 (Fig. 3D). Supporting this model, contacts between DNA regions that are normally within and outside of the neighborhood were increased (Fig. 3E-F, S10). Furthermore, deletion of the CTCF site in primary human T-cells also caused a small but detectable activation of TAL1 (Fig. S8C-G). These results are consistent with the model that the silent state of the TAL1 proto-oncogene is dependent on the integrity of the insulated neighborhood (Fig. 3D).

We further tested the model that site-specific perturbation of a loop boundary is sufficient to activate a proto-oncogene at the LMO2 locus. The LMO2 gene encodes a transcription factor that is overexpressed and oncogenic in some forms of T-ALL (14, 20). The region upstream of the LMO2 promoter is recurrently deleted in T-ALL and these deletions are linked to LMO2 activation (Fig. 3G); a previous study proposed that deletion of cryptic repressors located in the deleted region enable activation of LMO2 (21). Analysis of a T-ALL patient cohort (22) revealed deletions that overlap the CTCF boundary site of the LMO2 neighborhood, and that patient cells harboring these deletions had generally high levels of LMO2 expression (Fig. S9A-B). CRISPR/Cas9-mediated deletion in HEK-293T cells of a ~25 kb segment encompassing the insulated neighborhood boundary CTCF site and two additional CTCF sites that could act as boundary elements, caused a 2-fold increase in the LMO2 transcript (Fig. 3H-J), and a large-scale rearrangement of interactions around LMO2 as evidenced by 5C analysis (Fig. 3K-L, S10). These results indicate that the deleted CTCF sites contribute to the silent state of the LMO2 proto-oncogene (Fig. 3J).

The boundaries of chromosome neighborhoods may be disrupted in other cancers. A recent study noted that mutations in CTCF binding sites occur frequently in cancers (23), but it is unclear if mutations in boundaries are common as only a subset of CTCF sites form insulated neighborhoods (8, 10, 24). CTCF-cohesin bound loops are largely preserved across cell types (8, 9, 24), and a set of ~10,000 constitutive CTCF-CTCF loops shared by GM12878 lymphoblastoid, Jurkat cells and K562 CML cells (24) were identified for comparison (Fig. 4A, S11, Table S8). The boundaries of these neighborhoods were examined for somatic point mutations found in cancer genomes using the ICGC database containing data for ~50 cancer types, ~2300 WGS samples, and ~13 million unique somatic mutations (Table S9). We found a striking enrichment of mutations at the CTCF boundaries of constitutive neighborhoods (Fig. 4B, S12A, Table S10) compared to regions flanking the boundary CTCF sites (+/−1kb of the CTCF binding motif; P<10⁻⁴, permutation test (Fig. S12B), and in many instances these created a significant change in the consensus CTCF binding motif (Fig. S12C). Non-boundary CTCF sites did not show such enrichment (Fig. 4B, S12D, S14). The genomes of esophageal and liver carcinoma samples were particularly enriched for boundary CTCF site mutations (Fig. 4C-D, S12D-E, S13, Table S10), and there was no similar enrichment of mutations at the binding sites of other transcription factors (Fig. S15). In these cancers, a considerable fraction of the mutated neighborhood boundary CTCF sites were affected by multiple mutations (≥3 mutations per site) [280/1826 (15%) in esophageal carcinoma, and 54/1030 (5%) in liver carcinoma](Table S10), and recurrent mutations occurred more frequently in neighborhood boundary CTCF sites compared to non-boundary CTCF sites (Fig. S16A-C). The genes located within the most frequently mutated neighborhoods included known cellular proto-oncogenes annotated in the Cancer Gene Census and other genes that have not been associated with these cancers (Fig 4E-F, Table S11-S12). Two examples of proto-oncogene -containing neighborhoods where the activation of the gene located in the neighborhood has been observed in the respective cancer type are shown in Fig 4G-H. These results suggest that somatic mutations of insulated neighborhood boundaries occur in the genomes of many different cancers.

Open in a separate window

Fig. 4

Somatic mutations of neighborhood boundaries occur in many cancers

(A) “Constitutive neighborhood” at the NOTCH1 locus. CTCF ChIP-Seq and cohesin ChIA-PET interactions in Jurkat (T-ALL), GM12878 (lymphoblastoid) and K562 (CML) cells are displayed.

(B) Frequency of somatic mutations in the ICGC database at CTCF sites that form constitutive neighborhood boundaries (left), and CTCF sites that do not form neighborhood boundaries (right).

(C-D) Somatic mutations in (C) esophageal adenocarcinoma (ESAD-UK) and (D) hepatocellular carcinoma (LIRI-JP) at constitutive neighborhood boundary CTCF sites.

(E-F) Genes in constitutive neighborhoods whose boundary is recurrently mutated in (E) esophageal adenocarcinoma and in (F) hepatocellular carcinoma. The bars depict the number of mutations in the neighborhood boundary site. Proto-oncogenes annotated in the Cancer Gene Census are highlighted in red.

(G-H) Mutations in the boundary sites of the neighborhood containing (G) the LMO1 proto-oncogene in esophageal adenocarcinoma, and (H) the FGFR1 proto-oncogene in hepatocellular carcinoma.

The enrichment of mutations at the constitutive neighborhood boundary sites (+/−5bp of the motif) displayed on (B-D) compared to regions flanking the binding sites has a P-value <10⁻⁴ (permutation test).

In summary, disruption of insulated neighborhood boundaries can cause oncogene activation in cancer cells. With maps of 3D chromosome structure such as those described here, cancer genome analysis can consider how recurrent perturbations of boundary elements may impact expression of genes with roles in tumor biology. Our understanding of 3D chromosome structure and its control is rapidly advancing and should be considered for potential diagnostic and therapeutic purposes. Because control of 3D chromosome structure involves binding of specific sites by CTCF and cohesin, which is affected by protein cofactors, DNA methylation and local RNA synthesis (25), future advances in our understanding of these regulatory processes may provide new approaches to therapeutics that impact aberrant chromosome structures.