Systematic evaluation of the heterogeneity of topologically associating domain boundaries in a large genomic context in humans Authors: Ying Sun, Niels Tommerup, Lars Juhl Jensen, Jan Gorodkin Published: 2026 ABSTRACT AbstractTopologically associating domains (TADs) are generally considered as a homogeneous basic units of genome folding, which is critical for transcriptional regulation. However, recent studies indicate that both the TAD domain structures and the boundaries between them are not as homogeneous as originally recognized. Here, we address the heterogeneity of the TAD boundaries in the human genome at a large scale, which varies between active and inactive chromatin and across cell lines and tissues. To address this, based on the well-annotated TAD boundaries extracted from multiple cell lines and tissues, we examine their nucleotide content, resulting in two main clusters, one GC-rich and one AT-rich, which are mainly distributed in active and inactive chromatin, respectively. Also, they contain different types of repetitive sequences and have different epigenetic patterns, with more CTCF binding motifs in the GC-rich cluster. Hence, our observations of the TAD boundary content provide novel insights into TAD genomic architecture. In addition, we find that cell- or tissue-specific boundaries are less evolutionarily conserved than other boundaries. We highlight the importance of TAD boundary diversity in different functional contexts and discuss the importance of the different types of repetitive sequences and epigenetic patterns in the two main types of boundaries. Introduction Topologically associating domains (TADs) are fundamental genomic units characterized by increased DNA–DNA interactions within a TAD than between TADs [1, 2]. TADs have been linked to several biological processes, such as transcriptional regulation and DNA replication [2, 3]. Many TADs are formed at the same genomic location in different cell lines and tissues and are even conserved between species [1, 4–6]. The boundaries of a TAD hinder the interactions beyond the TAD, forming a locally insulated environment [1]. The disruption of TAD boundaries may cause ectopic cis-regulation and dysregulate gene expression, which may lead to disease [7–9]. In recent years, the development of sequencing and image-based technologies has accelerated our understanding of the spatial genome organization [4, 6]. TADs are hierarchical in their structural organization, where smaller TADs, sub-TADs, are nested inside the larger TADs [2, 10, 11]. One important architectural protein is CTCF, which acts as an insulator that limits DNA interactions with the neighboring TADs [1]. However, some TAD boundaries have only a few or no CTCF binding sites at all [12, 13]. TAD boundaries also appear in different contexts corresponding to the different levels of TAD hierarchy [14]. The above evidence indicates their diversity. Several studies have discussed the diversity of TAD boundaries [15–22]. However, the diversity has not yet been systematically examined, and the functional impact of potential differences remains unaddressed as well. In light of the many enriched genomic elements surrounding TAD boundaries and evolutionary conservation of these boundaries [1, 23], the nucleotide content of TAD boundaries may differ from that within TAD regions and play essential roles in influencing genomic folding [24]. However, our understanding of the exact relationship between nucleotide composition and TAD boundaries remains limited. In this study, we systematically explored the heterogeneity of TAD boundaries at a large genomic context of humans between active and inactive chromatin, and the heterogeneity across cell lines and tissues. First, we compiled a set of well-annotated TAD boundaries from multiple human tissues and cell lines. Second, we found that the TAD boundaries were split into GC-rich and AT-rich ones that were mainly located in active and inactive chromatin, respectively. The GC-rich boundaries were enriched for CTCF binding sites and cohesin complex, while AT-rich boundaries were enriched with EZH2 (a functional enzymatic component of PRC2, which is reported to participate in histone modification and transcriptional repression [25]) binding sites. Additionally, GC-rich boundaries have higher boundary strength and are more likely to display hierarchical TAD structures compared with AT-rich boundaries. The different epigenetic patterns and the boundary properties suggest the potential of multiple mechanisms to insulate interactions between TADs. Third, by comparing the TAD boundaries across cell lines and tissues, we found that some boundaries were cell- or tissue-specific while others were not. Notably, we observed that the cell- or tissue-specific boundaries were less evolutionarily conserved than others. Materials and methods TAD and TAD-related data To explore the heterogeneity of TAD boundaries, we constructed two types of datasets. First, we collected TAD datasets identified by different TAD callers from different sources. Specifically we obtained 29 TAD datasets identified by DomainCaller [1] processed by 3D Genome Browser (version 1.0) [26] with 40-kb resolution (DomainCaller dataset), and 17 TAD datasets identified by InsulationScore [27] processed by 4DN Data Portal with a 5-kb resolution (InsulationScore dataset) (datasets collection until 27 February 2025) [28]. All the TAD datasets were mapped to human hg38 reference genome. However, the datasets in both databases have been made under different conditions, some of which are not readily comparable. For example, TAD datasets from flavopiridol-treated samples may have different TAD coordinates due to inhibited transcription elongation [29]. Hence, such types of samples were discarded. Similarly, other types of conditions were removed (Supplementary text “Selection of TAD samples”). Details of the selected and discarded TAD datasets are shown in Supplementary Fig. S1 and Supplementary Table S1. Second, we downloaded publicly available TAD boundary-related annotation data from the UCSC Genome Browser [30], GENCODE [31], 4DN Data Portal [28], and other published datasets [5, 14] either annotated in hg38 or otherwise lifted over to hg38 coordinates. The downloaded data covered the following genome annotations used in our analysis: transcription factor binding sites (TFBSs), CpG islands, PhastCons scores, repeat sequences, DNase I hypersensitive sites, lamina-associated domains (LADs), A/B compartments, cytogenetic bandings, genome annotation, insulation scores, and TAD hierarchical levels. Among these annotated datasets, repeats and CpG islands are identified on DNA sequences without cell or tissue information, and the TFBS and DNase I hypersensitive sites are clusters of peaks from ChIP-Seq and DNase-seq across multiple cell lines and tissues. A/B compartments are collected from multiple cell lines and tissues, while LAD is only based on fibroblasts. Further details, including dataset descriptions, cell line or tissue origins, and version information, are summarized in the Supplementary Table S2. Calling TAD boundaries from TAD locations Since different TAD callers more often agree on the boundary assignment than the TAD structures themselves [2, 32–34], we took outset in the called TAD boundaries, rather than the whole TAD structure. However, the precise TAD boundary coordinates may be slightly different across different samples (see example in Supplementary Fig. S2), leading us to discard some boundaries for selecting representative ones in a local region. We started by calling TAD boundaries on the DomainCaller dataset. Since the TADs from 3D Genome Browser have a 40-kb resolution [1, 26], we used this 40 kb as the resolution of the extracted TAD boundaries, where we defined a 40-kb region (±20 kb) surrounding TAD start and TAD end as the TAD boundary. Then, we extracted boundaries by a recursive process across different datasets. First, we collected boundaries across the 29 datasets to obtain a genome-wide map. Then, we counted the frequency of each boundary across the datasets. All assignments with a count of 29 were kept, while all other boundaries within 400 kb (±40 kb * 5 bins) that have a count <29 were discarded. Next, the same was made for boundaries with a count of 28. The boundaries with a count of 28 were kept, while the others within 400 kb with a count <28 were discarded. This same filtering step was then repeated iteratively until the minimum of one boundary has been reached. In the rare cases of two counts within 400 kb, we arbitrarily kept the one with the lower coordinate according to the annotated sense strand. This reduced the initial number of 19 333 boundaries to 5664 in the DomainCaller dataset. Considering that TADs from different TAD callers may differ, we also applied this boundary extraction workflow on TAD boundary datasets identified by InsulationScore. Since the resolution of InsulationScore dataset is 5 kb rather than 40 kb for the DomainCaller dataset, we applied a 50-kb (±5 kb * 5 bins) genomic region for removing boundaries rather than 400 kb (±40 kb * 5 bins) in the pipeline for the DomainCaller dataset. It reduced the initial number of 76 411 boundaries to 28 361 on the InsulationScore dataset. Our boundary extraction workflow is detailed in Supplementary Figs S2 and S3A. Quality control of samples in the extracted TAD boundaries We employed a quality control step by identifying potential deviating or abnormal TAD samples, which could affect our analyses. Using a pragmatic cutoff, we defined consensus boundaries as those present in >50% of samples (15 out of the 29) in the DomainCaller dataset and discarded samples containing <60% of these consensus boundaries. This resulted in the removal of a single sample from thymus (Supplementary Figs S3B and S4A, D). Through trying different cutoffs, we obtained similar results (Supplementary Fig. S4B, C, and E). For the remaining 28 samples we re-called the TAD boundaries using the procedure described in the previous subsection (Supplementary Fig. S3C), resulting in 5561 boundaries in total across the 28 samples in the DomainCaller dataset. The number of samples assigned to each boundary ranged from 1 to 28. For the InsulationScore dataset, we applied the same procedure for abnormal TAD sample detection and found no abnormal TAD samples (Supplementary Fig. S4G). Testing different cutoffs for defining consensus TAD boundaries yielded similar results (Supplementary Fig. S4F). k-mer featurization of the TAD boundaries Various genomic elements, such as CTCF and transcription start sites (TSSs), have been reported to be enriched around TAD boundaries (±300 kb from the center) [23, 35], suggesting that these elements are involved in guiding the formation of the genome architecture [24]. In complement of exploring specific elements, others have used specific k-mers (e.g. 3-mers) in prediction schemes [36]. Here, we decoded TAD boundaries with a length of 40 kb by their k-mer counts for further exploration. Given that the TAD boundary is not a strand-specific structure, we considered the sequences from both strands on which we counted k-mers. Hence, the same count was assigned to a k-mer and its reverse complement, which were considered jointly. In addition, we filtered out TAD boundaries with at least one “N”, which removed 87 TAD boundaries in the DomainCaller dataset. We represented a TAD boundary by its k-mer counts (k ≤ 20) as follows: We used all k-mers for k ≤ 4 and the 250 most frequent k-mers among all TAD boundaries for 4 < k ≤ 20, as calculated using Jellyfish (version 2.2.6) [37]. Next, we chose the most informative k-mers only; that is, we required k-mers to have a count of at least 5 in >80% of the boundaries as a pragmatic cutoff. This resulted in a boundary to be represented by 1114 k-mer counts in the DomainCaller dataset. The same pipeline was applied to the InsulationScore dataset, using 40-kb (5 kb ± 17.5 kb) region for k-mer featurization. Thirty boundaries with at least one “N” in the sequence were removed. Applying this pipeline resulted in 1095 k-mers selected in the InsulationScore dataset for boundary representation, with 1002 k-mers overlapping those selected in the DomainCaller dataset. We simply chose 1114 k-mers from the DomainCaller dataset to represent TAD boundaries in the InsulationScore dataset, which also served as an independent test set for further TAD boundary prediction model [38]. For further details on the specific k-mers and statistics see Supplementary Fig. S5. The k-mer extraction section is illustrated in Supplementary Fig. S3D. Quality control of k-mers in the extracted TAD boundaries To optimize the downstream clustering of the boundaries (see next section), we first identified and removed the potential outliers. To do that, we employed a “pre-clustering” step in which we clustered the boundaries by their distance density to extract single boundaries far from any clusters, which then were considered as outliers. Since we focus on TAD boundaries across cell lines and tissues, the genomic context of TAD boundaries is a good resource for outlier detection. For this purpose, we used HDBSCAN, a density-based clustering algorithm for outlier detection [39] based on the 40-kb sequence composition represented by k-mer counts. Compared with other density-based clustering methods like DBSCAN, HDBSCAN uses varying densities in clustering assignment, guided by the local density of the data points, thereby producing more robust results. To do the clustering, the Euclidean distance between all pairs of TAD boundaries was calculated from the k-mer representation. Since the individual k-mers came on different scales, we rescaled by their Z-scores over all the boundaries using the StandardScaler normalization option in scikit-learn library (version: 1.1.2) [40]. Applying HDBSCAN (version: 0.8.27, min_cluster_size = 5) resulted in the detection of 202 outliers as defined by HDBSCAN on the DomainCaller dataset. We found min_cluster_size of 5 to be suitable after testing the range 2, …, 6. A value of 5 resulted in 202 outliers, a value of 4 and below resulted in very few (14 outliers), while a value of 6 resulted in 4479 (see details in Supplementary Fig. S6A). The detected outlier boundaries were discarded, resulting in 5272 boundaries in the DomainCaller dataset used for subsequent analyses. For the InsulationScore dataset, min_cluster_size values ranging from 4 to 9 resulted in similar outlier results with pairwise Jaccard index ranging from 0.91 to 0.98 (see details in Supplementary Fig. S6A). For consistency with the DomainCaller dataset, we chose min_cluster_size = 5, which resulted in 108 outliers detected and further filtered out. The outlier detection of TAD boundaries is illustrated in Supplementary Fig. S3E. Clustering of the TAD boundaries Whereas HDBSCAN readily pointed to outliers, it resulted in only a single large cluster after trying different parameters mentioned earlier. To further explore the clustering potential, we employed K-means clustering by scikit-learn library (version: 1.1.2) [40], which resulted in two clusters that, however, had an overlap (see Supplementary Fig. S7A and B for details). Therefore, we decided to use soft clustering and applied fuzzy c-means clustering (https://doi.org/10.5281/zenodo.5497844) on the pairwise Euclidean distance between all boundaries in their Z-score representation. To decide on the optimal number of clusters, we used two strategies. First, the fuzzy partition coefficient [41], which resulted in two clusters, and second, the silhouette score [42], which also resulted in essentially the same two clusters. Compared with hard clustering methods, fuzzy clustering method does not assign one boundary to a single cluster but gives a partial membership (a weight between zero and one) to each cluster. After clustering, we still saw two main clusters. However, pragmatically, we only assigned a boundary to a cluster only if its weight was at least 0.6; otherwise, we did not assign the boundary to any of the clusters (Supplementary Fig. S7C and D). A similar outcome was obtained for thresholds in the range of 0.5–0.7 in steps of 0.05, and hence the advancement of fuzzy c-means clustering over K-means clustering is limited. With a threshold of 0.6, 4600 (of 5272) boundaries in the DomainCaller dataset and 24 747 (of 28 223) boundaries in the InsulationScore dataset were assigned to clusters. We refer to the two clusters as classes, and they respectively contain 2868 and 1732 boundaries in the DomainCaller dataset and 8490 and 16 257 boundaries in the InsulationScore dataset, respectively. Analyzing the TAD boundary composition Comparing the k-mer content To compare the k-mer content between the two TAD boundary clusters, we searched for k-mers enriched in one cluster over all boundaries. We applied the one-sided Wilcoxon rank-sum test for each k-mer between one given cluster of TAD boundaries and all TAD boundaries [43]. The Bonferroni method was used for multiple test correction via the statsmodels library (version: 0.12.2) [44]. The k-mers with adjusted P-values of < .05 were regarded as enriched k-mers. To quantify the degree of enrichment, the fold changes of enriched k-mers were calculated as the mean k-mer counts in one class of TAD boundaries divided by the mean k-mer counts in all TAD boundaries. To interpret the biological relevance of the enriched k-mers, we further investigated their association with known motifs. Specifically, we derived de novo motifs from the enriched k-mers using MEME (version 5.5.7, parameters: -mod oops -minw 3 -maxw 10) [45]. Because the enriched k-mers were typically short (5–8 nt), we extended their sequences by extracting flanking regions of 40 nt (±20 nt) for the motif analysis. Motifs identified by MEME were subsequently compared with known motifs from the HOCOMOCO database (version: HOCOMOCOv11_full) [46] using Tomtom with default parameters, which is also from MEME Suite (version 5.5.7) [45]. Motifs showing significant similarity to known motifs (q-value < 0.05) were considered functionally associated, and the corresponding k-mers were regarded as relevant to these motifs. The remaining k-mers were regarded as associated with novel motifs. We selected the HOCOMOCO database for comparison because it contains high-quality, manually curated human transcription factor motifs downloaded via the MEME Suite website. Comparison of genomic elements between the TAD boundaries Genomic elements, including repeats, TSSs, CpG islands, TFBS, and DNase I hypersensitive sites [1, 47] were compiled as follows. We obtained genomic coordinates of repeats, TFBS, CpG islands (hg38) from the UCSC Genome Browser [30] and gene annotations (hg38) from GENCODE (version 39) [31]. To assign the elements to the TAD boundaries, we considered their overlap by the following criteria. For repeats, DNase I hypersensitive sites, and CpG islands, we kept those with at least 50% of size overlapping with the boundaries. For TSSs, we used those relative to the canonical transcripts [31]. Screening both strands of the TAD boundaries, we counted the number of different elements (see Supplementary Table S2). To compare the content of these genomic elements between the two classes of boundaries, we used the one-sided Mann–Whitney U test [43]. Comparison of domain-like structures between the TAD boundaries Domain-like structures, including LADs, cytogenetic banding, and A/B compartment, were also compared between the two classes of TAD boundaries. We downloaded the LAD annotations (hg19) from UCSC Genome Browser, and the coordinates were lifted over to hg38 [30]. TAD boundaries with at least 50% of size overlapping with LADs or cytogenetic bandings were extracted, and their number of occurrences in each class was counted (see Supplementary Table S2). These frequencies were compared using Fisher’s exact test since each TAD boundary either overlapped with a single LAD/cytogenetic banding or not [43] (LAD: one-sided Fisher’s exact test; cytogenetic banding: two-sided Fisher’s exact test). The A/B compartment annotations for DomainCaller dataset were collected from Schmitt et al., 2016 [5], which covered 21 of the 28 cell lines of the DomainCaller dataset (Supplementary Table S2). These annotations were processed by “prcomp” function in R based on the hg19 reference genome, and we lifted over their coordinates to hg38. In each sample, each 1-Mb bin region in the chromatin was annotated as A compartment or B compartment, and few were undetermined. For the InsulationScore dataset, their corresponding compartment annotations were collected from 4DN Data Portal using cooltools at a resolution of 250 kb, based on the hg38 reference genome (Supplementary Table S1). To understand the boundaries in A/B compartment states across cell lines and tissues, we overlapped their coordinates in each cell line or tissue with the genomic coordinates of the boundaries. We assigned a boundary to A compartment, based on the following situations: (i) the boundary was located entirely in one bin marked as A compartment; (ii) the boundary spanned two bins, and both of which were marked as A compartment; and (iii) the boundary spanned two bins, one marked as A compartment and the other as undetermined. We used similar criteria to assign a boundary in B compartment. A few boundaries spanning bins with different A/B compartments were excluded. Genomic elements in TAD boundaries Enrichment analyses of repeat elements, CpG islands, DNase I hypersensitive sites, TFBS, and TSSs for two classes of TAD boundaries were performed. To do the enrichment analysis, we first chose an equal number of 40-kb genomic segments from TAD regions with the same GC content distribution as TAD boundaries served as the background. We first extracted the background regions for the DomainCaller dataset. Based on the definition of DomainCaller, the chromatin regions can be divided into three types: TADs, TAD boundaries, and unorganized chromatin (the regions between neighboring TADs with a distance >400 kb) [1]. To obtain a consensus TAD region dataset, we applied a strict workflow to extract the TAD regions across cell lines and tissues: First, given that the length of TAD boundaries can be up to 400 kb [1], we extended our annotated TAD boundaries to 400 kb (±200 kb); second, we merged unorganized chromatin regions annotated across cell lines and tissues. Third, we extracted the remaining regions in the chromatin (not TAD boundaries or unorganized chromatin in any TAD samples) and called them consensus TAD regions. Fourth, we randomly chose the same number of 40-kb regions from these TAD regions with the same distribution of GC content as TAD boundaries, which were extracted as the background for the enrichment analysis in the DomainCaller dataset. For the InsulationScore dataset, we applied a slightly modified procedure for background selection because “unorganized chromatin” was not defined by InsulationScore. We first extended the merged TAD boundaries to 55 kb (5 kb ± 25 kb) and excluded these regions from background selection. We then randomly selected the same number of genomic regions with a GC content distribution identical to that of the TAD boundaries, which served as the background for enrichment analysis. Then, we performed enrichment analyses of repeats, CpG islands, DNase I hypersensitive sites, TFBS, and TSSs based on two classes of TAD boundaries against their corresponding background. Based on the criteria from [23, 48], we extracted the repeats, CpG islands, DNase I hypersensitive sites, and TFBS for which at least 50% of regions overlapped with the boundaries. CpG islands, DNase I hypersensitive sites, and TFBS are not strand-specific, so all the elements are included. Repeats and TSSs from both strands were also included in our analysis. Then, we calculated the number of the relevant genomic elements for each boundary and each background region, respectively. We used Mann–Whitney U test for the comparisons between different classes of TAD boundaries and their corresponding background regions, respectively (repeats: two-sided Mann–Whitney U test; CpG islands, DNase I hypersensitive sites, TFBS, and TSSs: one-sided Mann–Whitney U test) [43]. To quantify the degree of enrichment, we also calculated the fold changes of these genomic elements by the mean counts in one class of TAD boundaries divided by the mean counts in background regions. To ensure robust comparisons, we examined genomic contexts of multiple lengths: 40 kb, 100 kb (40 kb ± 30 kb), and 400 kb (40 kb ± 180 kb) surrounding each boundary. Since the InsulationScore dataset was derived from a 5-kb Hi-C interaction frequency matrix, which enables finer resolution analyses, we also performed comparisons at a 5-kb scale, corresponding to the central 5 kb within the 40-kb genomic region. Comparison of CTCF orientation between the TAD boundaries To know how CTCF orientation affects the function of TAD boundaries, we extracted CTCF orientations by Find Individual Motif Occurrences (FIMO, version 5.5.7) [49]. The position weight matrix of CTCF motif (ID: MA0139.1) was downloaded from JASPAR (release version: 2024) [50]. Based on the CTCF binding sites available from the UCSC Genome Browser [30], the CTCF orientations were extracted by FIMO on both strands with the default P-value of .0001. Then, based on the CTCF motif detected in which strands in the reference genome, boundaries were further classified into several types: the complex orientation means that there is more than one motif in this boundary and the CTCF motifs in this boundary are on both strands in the reference genome; the tandem orientation means that there is more than one motif in this boundary and all of them are on the same strand in the reference genome; single motif means that one motif in this boundary, no matter which strand it is; and no motif means no CTCF motif detected in this boundary. Then, the number of boundaries classified by CTCF motif orientations was compared between different classes of TAD boundaries by two-sided Fisher’s exact test. Gene Ontology enrichment analysis To further investigate the functional characteristics of the two classes of TAD boundaries, we extracted genes whose TSSs were located within the 40-kb TAD boundary regions for Gene Ontology (GO) enrichment analysis. GO enrichment for biological process terms was performed using the clusterProfiler package (version 4.12.0) [51]. Multiple testing correction was performed using the Benjamini–Hochberg method, and GO terms with adjusted P-values < .05 were considered significantly enriched. Boundary strength and hierarchical structure of TAD boundary classes Boundary strength and hierarchical structure, two properties of TAD boundaries, were also compared between the two classes of TAD boundaries in our study. Given the limited processed datasets with these properties, we chose the boundaries from GM12878 and HUVEC cell lines for comparison analysis on the DomainCaller dataset, whose interaction frequency matrices came from Rao et al., 2014 [10]. The boundary strength, calculated from Hi-C interaction matrix, reflects the degree of insulating effect between neighboring TADs. These values were calculated on hg38 reference genome by cooltools library [52] at a 5-kb resolution. See [27] for detailed boundary strength calculation. These processed boundaries were downloaded from 4DN Data Portal [28]. Considering that TAD boundary coordinates from different TAD callers might be different, we chose the boundaries overlapping with our merged boundaries (≥50% region overlapped with merged boundaries from GM12878 or HUVEC cell line) for the comparison analysis. We used the one-sided Mann–Whitney U test for the comparison of boundary strength [43]. We also compared TAD hierarchical structure between two classes of TAD boundaries, which describes the TAD hierarchical levels and indicates the complexity of DNA interaction in the local region. This hierarchical structure was called by OnTAD [14] with a 10-kb resolution, which were provided by OnTAD developers. The TADs genomic coordinates from OnTAD were based on hg19, and we lifted these over to hg38 reference genome. We chose the TAD boundaries processed by OnTAD in which at least 50% of boundary regions overlapped with our merged boundaries for the comparison analysis. We used the two-sided Fisher’s exact test for the comparison of TAD hierarchical levels [43]. For the InsulationScore dataset, we also performed a similar comparison of boundary strength and TAD hierarchical levels between the two classes of TAD boundaries. When comparing boundary strength between GC-rich and AT-rich boundaries, the median boundary strength across all TAD samples was used as the representative value for each boundary in the analysis. TAD boundary similarities across cell lines and tissues To know the similarity of TAD boundaries across cell lines and tissues, we also performed clustering analysis based on the existence of TAD boundaries from different cell lines and tissues. We represented a TAD boundary with a vector of length corresponding to the number of TAD samples. If the boundary is present in the given sample, the corresponding position is marked as 1 and otherwise 0. The vectors from all the TAD boundaries made up one matrix. We then applied the hierarchical clustering analysis by the Euclidean distance and the Ward method between all pairs of TAD boundaries [43]. At the top level of clustering, we obtained three clusters of TAD boundaries. Among three clusters of TAD boundaries, the occurrence of TAD boundaries among TAD samples varied a lot. The cluster with boundaries detected in a few TAD samples is here referred to sample-specific boundaries; the cluster with boundaries detected in most of the TAD samples is referred to sample-unspecific boundaries; the remaining cluster is simply referred to other boundaries. In light of the different stability of TAD boundaries across cell lines and tissues [53], we further compared the evolutionary conservation between different classes of boundaries. We chose PhastCons score for the comparison analysis, which quantifies the probability that a given genomic position is evolutionarily conserved based on sequence composition. The greater PhastCons scores in specific genomic regions indicate their functional importance during evolution. To minimize potential bias from functional elements with intrinsically high PhastCons scores, we excluded all protein-coding regions annotated as “CDS” in the genomic annotation (GENCODE V39). We calculated the mean PhastCons score for each boundary by UCSC bigWigSummary tool [54]. Then, based on the mean PhastCons score of each boundary, we compared the distributions of mean PhastCons scores between different classes of TAD boundaries by one-sided Mann–Whitney U test [43]. Results Annotating TAD boundaries To study TAD boundaries, we first need to annotate them based on the location of the TADs themselves. Since some TAD boundaries are cell- /tissue-specific [5, 10], we collected TAD samples from multiple cell lines and tissues to obtain a more complete annotation of TAD boundaries. Considering that the TADs from different TAD callers may vary, we collected TAD datasets identified by different TAD callers, including DomainCaller and InsulationScore. We first downloaded TAD coordinates (hg38) from the 3D Genome Browser [26], covering 29 samples representing different cell lines and tissues (Supplementary Fig. S1A and Supplementary Table S1). We started with defining the boundaries from 29 cell lines and tissues (see “Materials and methods” section). However, TAD annotated structures can differ between different cells/tissues and can even be absent, e.g. in sperm. We therefore focused on samples where there is an overall consensus in the location of TADs and the corresponding TAD boundaries. After assigning boundaries to each sample, we removed samples which deviate from the bulk part of samples (see “Materials and methods” section). Here, we considered the consensus of TAD boundaries across all the samples from 3D Genome Browser [26], which were defined by the occurrences of boundaries across samples. We found that the TADs from the thymus had a lower percentage of consensus boundaries (Supplementary Fig. S4C; see “Materials and methods” section for details). Employing a range of cutoffs to define the consensus boundaries, we found that the percentage of consensus boundaries in thymus was always lower than the percentage in other samples (Supplementary Fig. S4A–D). To answer why TADs in thymus are different, we checked the expression of architectural proteins for TADs (CTCF, SMC3, RAD21) in the thymus based on transcript expression levels in the Human Protein Atlas database (https://www.proteinatlas.org/) [55]. We found that the expression values of architectural proteins for TADs in thymus are almost two times higher than the expression values in other tissues (Supplementary Fig. S8). The higher expression of architectural proteins may lead to different DNA folding and even different TAD structures. Therefore, we regarded the TAD sample from thymus as an outlier and removed it, resulting in a total of 5561 TAD boundaries defined from the remaining 28 TAD samples (see “Materials and methods” section for details). We also collected 17 TAD boundary datasets across multiple cell lines and tissues identified by InsulationScore downloaded from the 4DN Data Portal (Supplementary Fig. S1B), resulting in 28 361 TAD boundaries. Next, we searched for outliers among TAD boundaries at the sequence content level. First, we found 87 TAD boundaries in the DomainCaller dataset and 30 TAD boundaries in the InsulationScore dataset with at least one “N” in their genomic sequences (gap regions in the genome) and for that reason discarded. Then, we used the vectors with k-mer counts to represent the nucleotide content. After several selection steps, 1114 unique informative k-mers were selected and counted for each TAD boundary. Based on these selected k-mers, the HDBSCAN clustering method was used for outliers detection, which is an effective outlier detection method with varying density [39] (see “Materials and methods” section). This resulted in 202 TAD boundaries in the DomainCaller dataset and 108 boundaries in the InsulationScore dataset with moderate GC content (31%–63%). These boundaries were not extreme in GC bias but remained distinct from other genomic regions and were therefore excluded from subsequent clustering analyses (see details in Supplementary Fig. S6). There are two possible reasons why their nucleotide contents differ. (i) The incomplete interaction frequency matrices may be caused by the difficulty of accurately calling the TADs and their corresponding boundaries (Supplementary Fig. S6D and E). By examining the genomic locations of these outliers, we found that ∼17% (36/202 in the DomainCaller dataset and 17/108 in the InsulationScore dataset) overlapped (covering over 50% of outlier regions) the ENCODE blacklist regions [56], which represent poorly assembled regions such as centromeric/telomeric. (ii) Although some of the regions are covered by sufficient reads (compared with those repetitive regions without many unique-mapping reads), these boundaries might only be a statistical artifact within a cell population but not exist in individual cells [57], explaining why their nucleotide contents are different. Hence, after employing filtering and merging steps (Supplementary Fig. S3), we obtained well-annotated TAD boundary datasets comprising 5272 boundaries in the DomainCaller dataset and 28 223 boundaries in the InsulationScore dataset, which were merged from multiple cell lines and tissues for subsequent heterogeneity analysis (Fig. 1 and Supplementary Table S3). Overview of the heterogeneity of TAD boundaries. A schematic illustration showing the workflow for merging and analyzing TAD boundaries across multiple cell lines and tissues. TAD datasets identified by different TAD callers, including DomainCaller [1] and InsulationScore [27], were collected and integrated to generate a comprehensive set of TAD boundaries. The merged boundaries were subsequently used to investigate the heterogeneity of TAD boundaries between active and inactive chromatin, as well as across cell lines and tissues. Clustering analysis reveals two main classes of TAD boundaries of different nucleotide composition To address the question of similarities and differences among the domain boundaries, we clustered our annotated boundaries. Representing each boundary as a vector of 1 114 k-mer counts (see “Materials and methods” section) and using the Euclidean distance of all pairs of boundaries as input to fuzzy c-means (see “Materials and methods” section), we obtained two main clusters strongly differing in their GC content. Given the nature of the fuzzy clustering, it came with a range of “intermediate” boundaries, which we did not wish to assign to any of the two clusters. Adjusting the maximum membership threshold provided a robust outcome for values (0.50, 0.55, 0.65, and 0.70) in the neighborhood of 0.60 (see “Materials and methods” section), comprised 2868 AT-rich boundaries and 1732 GC-rich boundaries in the DomainCaller dataset (Fig. 2A and Supplementary Fig. S9A) and 16 257 AT-rich boundaries and 8490 GC-rich boundaries in the InsulationScore dataset (Supplementary Fig. S10A). Consequently, the enriched k-mers for each class of TAD boundaries differ. Enriched k-mers in GC-rich boundaries are abundant in “G” and “C” (Fig. 2B and Supplementary Figs S9B and S10B). To explore the relationship between enriched k-mers and known motifs, we performed de novo motif discovery using the enriched k-mers and their flanking sequences with the MEME suite and compared with known motifs using Tomtom (version 5.5.7) [45]. We found that the motifs derived from GC-rich boundaries were primarily associated with zinc finger binding sites, for example, “GCGG” is a subset of SP2 motif. In contrast, motifs discovered from AT-rich boundaries did not match any known motifs in the HOCOMOCO (version 11) database. Functional comparison of two classes of TAD boundaries on DomainCaller dataset. (A) GC content distribution of two classes of TAD boundaries. (B) Top 15 enriched k-mers in two classes of TAD boundaries ranked by fold change. (C) Enrichment of CTCF binding sites for two classes of TAD boundaries. (D) Comparison of orientation of CTCF motifs between two classes of TAD boundaries. Enrichment of representative TFBSs and repeats for two classes of TAD boundaries, including EZH2 (E), SMC3 (F), RAD21 (G), and Alu (H). (I) Comparison of insulation scores from two classes of TAD boundaries from GM12878 cell line. (J) Distribution of TAD hierarchical levels from two classes of TAD boundaries from GM12878 cell line. The TAD hierarchical levels are defined by OnTAD [14]. To compile the enrichment analysis (C and E–H), the backgrounds were randomly selected from TAD regions with the same GC content as TAD boundaries. The mean values in the distribution of boundaries and random regions were applied for the calculation of fold change. A one-sided Mann–Whitney U test was used for the comparison of TFBSs, and insulation scores. A two-sided Mann–Whitney U test was used for the comparison of repeats. A two-sided Fisher’s exact test was used for the comparison of CTCF motif orientation and TAD hierarchical levels. * P-value ≤ .05; ** P-value ≤ .01; *** P-value ≤ .001. Considering the significant GC content differences between two classes of TAD boundaries, we hypothesized that they correspond to active (A compartment, gene-rich) and inactive (B compartment, gene-poor) chromatin regions. To test this hypothesis, we performed a series of analyses on the DomainCaller dataset. We first observed that two classes of boundaries showed significantly different distributions along the chromosomes (Supplementary Fig. S9C). For example, there were more GC-rich boundaries on chromosome 19 (two-sided Fisher’s exact test, P-value = 1.21e-29), where more genes are annotated. In contrast, there were more AT-rich boundaries on chromosome 13, which is gene-poor (two-sided Fisher’s exact test, P-value = 9.26e-10). Moreover, this differential distribution correlates with the cytogenetic banding pattern, with the AT-rich TAD boundaries corresponding to dark Giemsa-positive bands and the GC-rich TAD boundaries corresponding to light Giemsa-negative bands (Supplementary Fig. S9D). In addition, we conducted a series of analyses, including comparing the content of repetitive sequences, TSSs, and epigenetic datasets between two classes of TAD boundaries. Since repetitive sequences have been proposed to participate in the genomic organization [58–60], we assessed their content in the two classes of TAD boundaries (see details in “Materials and methods” section). The annotations used in the following analysis were obtained from UCSC Genome Browser [30] and GENCODE [31]. As shown in Supplementary Fig. S9E, we can see different distributions of genomic elements between the two classes of TAD boundaries. We found more SINEs (one-sided Mann–Whitney U test, P-value = 0.00e + 00) overlapped with GC-rich boundaries, while we found more LINEs (one-sided Mann–Whitney U test, P-value = 2.34e-24) and LTR elements (one-sided Mann–Whitney U test, P-value = 3.07e-15) overlapped with AT-rich boundaries. Upon assessing the TSSs, more were found in the GC-rich boundaries (one-sided Mann–Whitney U test, P-value = 3.12e-183). Consistently, there was a larger number of CpG islands overlapped with GC-rich boundaries than with AT-rich boundaries (one-sided Mann–Whitney U test, P-value = 5.09e-271). We assessed the chromatin accessibility and transcriptional activity [61] by comparing the number of DNase I hypersensitive sites overlapped with the two classes of boundaries and found that more DNase I hypersensitive sites were found to be overlapped with GC-rich boundaries than AT-rich boundaries (one-sided Mann–Whitney U test, P-value = 6.53e-280). We also explored the relationship of two classes of TAD boundaries with other domain-like structures. LADs have a strong interaction with the lamina and have been linked to inactive chromatin and gene repression [62]. By comparing genomic coordinates between TAD boundaries and LADs (see “Materials and methods” section for overlap criteria), we found that 1399 AT-rich boundaries overlapped LADs, while only 221 GC-rich boundaries overlapped LADs (one-sided Fisher’s exact test, P-value = 1.68e-79) (Supplementary Fig. S9E). These above-mentioned characteristics are critical indicators of active and inactive chromatin [47, 63]. Reports in the literature generally regard the A-compartment as active chromatin while the B-compartment is generally regarded as inactive chromatin [64]. To further verify our conclusion, we collected A/B compartment datasets from 21 cell lines and tissues from Schmitt et al. [5], among which are subset of our 28 samples. We observed that 62%–83% of GC-rich boundaries are located in A-compartments (B-compartment: 17%–37%), while 67%–76% of AT-rich boundaries are located in B-compartments (A-compartment: 23%–32%). All the results indicate the existence of two classes of TAD boundaries associated with active and inactive chromatin, corresponding to GC-rich and AT-rich boundaries, respectively (Supplementary Figs S9–S14). Notably, boundaries with low membership scores in the fuzzy c-means clustering, which are located at the edges between clusters, tend to shift between A and B compartments during development (Supplementary Fig. S15). Associating function to the two classes of TAD boundaries Many studies indicate the vital roles of genomic elements for TAD boundaries such as transcription factors, repeats [23, 58–60]. To address potential roles of the boundaries, a series of comparative analyses of different types of repeats, TFBSs, CpG islands, DNase I hypersensitive sites, and TSSs were performed for the two classes of TAD boundaries (Fig. 2A–F, Supplementary Figs S16–S20, and Supplementary Table S4). For DomainCaller dataset, we randomly chose an equal number of 40-kb regions in TADs across 28 TAD samples with the same GC content distributions as the background (see “Materials and methods” section for details). Next, we counted the genomic elements in the background as we did for TAD boundaries. For each class of TAD boundaries, we compared the distribution of the number of genomic elements between TAD boundaries and background and calculated the fold changes. For CTCF binding sites, we found CTCF binding sites were enriched at both classes of TAD boundaries (one-sided Mann–Whitney U test, AT-rich versus random: P-value = 4.80e-41; GC-rich versus random: P-value = 3.97e-12), with more CTCF binding sites detected in GC-rich boundaries (one-sided Mann–Whitney U test, GC-rich versus AT-rich: P-value = 1.79e-273) (Fig. 2C). Considering that the boundaries with different orientations of the CTCF motifs are relevant to different replication timing stages [17], we also compared the CTCF orientations between the two classes of TAD boundaries. More GC-rich boundaries were covered with complex orientation of CTCF motifs (more than one motif detected in different strands of reference genome in the boundaries) (two-sided Fisher’s exact test, P-value = 2.90e-50) (Fig. 2D). It has previously been reported that the number of CTCF binding sites and their motif orientations affect the insulation capacity [65, 66]. So far, it is not clear how to create an insulated environment for those boundaries with few or no CTCF binding sites. It is reported that polycomb repressive complexs (PRCs) also participate in the TAD formation in fruit flies [67]. Our analysis revealed that EZH2 (a functional enzymatic component of PRC2, which is reported to participate in histone modification and transcriptional repression [25]) binding sites were enriched at AT-rich boundaries but lacking in GC-rich boundaries (one-sided Mann–Whitney U test, AT-rich versus random: P-value = 3.12e-12; GC-richversus random: P-value = 1.68e-03), despite the fact that there were only a few AT-rich boundaries (13.9%) with EZH2 binding sites (Fig. 2E). It indicates that other genomic elements (such as EZH2) may also help with the stability of genomic architecture for a subset of AT-rich regions. We further examined the distribution of cohesin complex (SMC3 and RAD21) across two classes of TAD boundaries, which is the architectural protein for TAD formation. They showed a similar distribution as CTCF that their binding sites (SMC3 and RAD21) were enriched over genomic background for both classes of boundaries (one-sided Mann–Whitney U test, SMC3 from AT-rich versus random: P-value = 5.52e-82; SMC3 from GC-rich versus random: P-value = 1.75e-61; RAD21 from AT-rich versus random: P-value = 6.46e-15; RAD21 from GC-rich versus random: P-value = 5.09e-04), with more binding sites in GC-rich boundaries (one-sided Mann–Whitney U test, GC-rich versus AT-rich for SMC3: P-value = 3.68e-156; GC-rich versus AT-rich for RAD21: P-value = 8.68e-180) (Fig. 2F and G). Previous studies have shown that the SINE element is highly enriched around TAD boundaries (±500 kb) in humans, especially Alu [1, 68]. Likewise, we also found that Alu was highly enriched at our obtained two classes of boundaries (two-sided Mann–Whitney U test, AT-rich versus random: P-value = 1.18e-25; GC-rich versus random: P-value = 8.69e-160). And more Alu elements in GC-rich boundaries were detected than in AT-rich boundaries (two-sided Mann–Whitney U test, GC-rich versus AT-rich: P-value = 0.00e + 00) (Fig. 2H). Expanding the genomic context to 100 and 400 kb for broader regional comparisons yielded consistent results (Supplementary Table S4). Given that the InsulationScore dataset was defined at a 5-kb Hi-C interaction resolution, this enabled enrichment analyses at a finer scale, which also produced consistent results (Supplementary Figs S18–S20 and Supplementary Table S4). To further explore the differences in boundary strength and TAD hierarchical levels between the two classes of TAD boundaries, we chose the TAD boundaries from GM12878 cell line among the DomainCaller dataset, one of the high-sequencing depth samples, as an example for the comparison analysis. The lack of annotation of TAD hierarchical levels in chromosome X prompted us to focus on the TAD boundaries in autosomes. In our merged boundary set of the 5026 boundaries from the DomainCaller dataset, 3133 boundaries are from the GM12878 cell line (excluding chromosome X) as part of our DomainCaller-based pipeline. We first compared the boundary strength between two classes of TAD boundaries to examine whether they show different insulation abilities, as measured by InsulationScore. Here, the insulation score was obtained from 4DN Data Portal and calculated on a 5-kb Hi-C interaction frequency matrix by cooltools [28]. Given this dataset was processed by a different resolution, the 1640 ones overlapping with our DomainCaller-based boundaries (see “Materials and methods” section for overlap details) were used for the comparison of boundary strength between two classes of TAD boundaries. Our analysis revealed that the GC-rich boundaries had higher boundary strength than AT-rich boundaries (one-sided Mann–Whitney U test, P-value = 6.52e-11) (Fig. 2I), which may be explained by their different distributions of CTCF binding sites and Alu elements, where more CTCF and more Alu elements at TAD boundaries give higher insulation scores [65, 69]. Then, we compared the TAD hierarchical levels between two classes of TAD boundaries. The TAD hierarchical levels were collected from OnTAD developers, which were processed by a 10-kb resolution. 2029 boundaries that were overlapped our DomainCaller-based boundaries (see “Materials and methods” section for overlap details) were used for the comparison of TAD hierarchical levels between two classes of TAD boundaries. According to OnTAD definition, the TAD hierarchical level increased from external to internal, and the outside TAD was defined as level 1, while the TAD hierarchical level ended at level 5 [14]. Those TADs without a hierarchical TAD structure were called singletons (Fig. 2J). There were more AT-rich boundaries from singletons (two-sided Fisher’s exact test, P-value = 1.14e-45), while GC-rich boundaries preferred to come from hierarchical TADs, especially from higher levels of TADs (two-sided Fisher’s exact test, level 3: P-value = 4.44e-03; level 4: P-value = 3.83e-07; level 5: P-value = 3.05e-09) (Fig. 2J). This further suggests the different genome folding of TADs for two classes of TAD boundaries and complex regulation in GC-rich regions. A similar pattern can also be detected in the boundaries of the HUVEC cell line (see details in “Materials and methods” section, Supplementary Fig. S21). Consistent conclusions were drawn from the InsulationScore dataset (Supplementary Figs. S18G–I). To assess the robustness of the enrichment results across multiple genomic elements, we expanded the genomic context of TAD boundaries from 40 to 100 kb (40 kb ± 30 kb) and 400 kb (40 kb ± 180 kb). Enrichment analyses performed within these extended regions yielded consistent patterns (Supplementary Table S4). We next performed the same analyses using the InsulationScore dataset at different genomic scales (40, 100, and 400 kb) and observed similar enrichment patterns, except that EZH2 was enriched in both boundary classes (Supplementary Fig. S18C and Supplementary Table S4). This discrepancy likely reflects differences in TAD definitions between the two callers. As the InsulationScore dataset was derived from a 5-kb Hi-C interaction matrix, it enabled enrichment analyses at a finer resolution (5 kb, centered within the 40-kb boundary regions), which also produced consistent results. However, when the TAD boundary was extended to 400 kb (40 kb ± 180 kb) on InsulationScore dataset, these genomic elements tend to be depleted at TAD boundaries, especially for AT-rich boundaries. This further indicates the differences of TADs and TAD boundaries between different TAD callers (Supplementary Table S4). In addition, we explored the potential biological functions of the two classes of TAD boundaries through GO enrichment analysis. Based on genes whose TSSs were located within the 40-kb boundary regions, enriched GO terms were identified using clusterProfiler [51]. For GC-rich boundaries, the significantly enriched GO terms were predominantly associated with protein synthesis, folding, transport, and RNA processing-related biological processes in both the DomainCaller and InsulationScore datasets. These findings suggest the involvement of GC-rich boundaries in dynamic chromatin loops that facilitate interactions between non-coding regulatory elements and genes. In contrast, AT-rich boundaries were enriched for GO terms related to cell cycle regulation and cilium organization in the DomainCaller dataset, whereas those associated with neural and sensory processes were enriched in the InsulationScore dataset (Supplementary Fig. S22). Although the connection between the enriched GO terms of AT-rich boundaries and chromatin organization is less direct, the distinct functional profiles of genes near AT-rich and GC-rich boundaries further support the functional diversity of TAD boundaries across active and inactive chromatin contexts. A large subset of cell lines and tissues have the same TAD boundaries To explore to what extent the individual samples contain the same TAD boundaries across cell lines and tissues, we first considered the fraction of boundaries present in a given fraction of the samples. We found that >31% (2134/5272 in the DomainCaller dataset and 8914/28 223 in the InsulationScore dataset) of the TAD boundaries were found in over 50% of samples (Fig. 3A and Supplementary Fig. S23A). To further explore the similarity of TAD boundaries, the TAD boundaries were divided into three clusters by hierarchical clustering method based on their distributions among samples (Fig. 3B and Supplementary Fig. S23B). The three clusters of TAD boundaries were presented in different numbers of TAD samples: the cluster of boundaries detected in only a few samples were called sample-specific boundaries; the cluster of boundaries detected in most samples were called sample-unspecific boundaries; the remaining cluster of boundaries were called other boundaries. This resulted in 1882 sample-specific boundaries, 1472 sample-unspecific boundaries, and 1918 other boundaries in the DomainCaller dataset and 17 523 sample-specific boundaries, 4396 sample-unspecific and 6304 other boundaries in the InsulationScore dataset. To explore whether these classes of boundaries were evolutionarily conserved at the sequence level, the mean PhastCons score for each boundary was calculated, which estimates the likelihood of purifying selection by DNA sequence [70]. In case the functional elements with higher selection pressure than other genomic regions which might impact on the downstream result, all the coding regions in TADs were removed before comparison (see “Materials and methods” section for details). Sample-unspecific boundaries with higher PhastCons scores were found to be more evolutionarily conserved than sample-specific boundaries (Fig. 3C and Supplementary Figs S23C and S24). Distribution of TAD boundaries across cell lines and tissues on DomainCaller dataset. (A) Statistics of the distribution of TAD boundaries across samples. (B) Clusters of TAD boundaries based on their distribution among samples, including sample-unspecific, sample-specific, and other boundaries. The dark blue marked with 1 means this boundary is detected in this TAD sample, while the white marked with 0 means this boundary is not annotated in this TAD sample. TAD number and mean TAD size (Mb), which indicate the sequencing depth of TAD sample [33], were also colored with blue and green, respectively. As the TAD number or mean TAD size increased, the corresponding color increased. (C) Comparison of PhastCons scores among different classes of boundaries. The color of each violin reflected the median value in that class of boundary. One-sided Mann–Whitney U test was applied for statistics. (D) Comparison of sequence content among different classes of TAD boundaries. To understand these sample-specific boundaries, we first compared their nucleotide content, referring to AT-rich or GC-rich obtained from above clustering analysis. The comparison results show that sample-specific boundaries tend to exhibit higher AT-rich content (one-sided Fisher’s exact test, P-value = 3.96e-18) (Fig. 3D and Supplementary Fig. S23D). There are three reasons that may contribute to the presence of sample-specific boundaries: (i) Sequencing depth and experimental technique for dataset generation influence TAD calling resulting in a subset of sample-specific boundaries. The TAD boundaries from samples with higher sequencing depth (such as TADs identified from Rao 2014 [10] dataset) or data from Micro-C [71] may result in more sample-specific boundaries (Supplementary Fig. S25). (ii) A subset of TAD boundaries is cell-/ tissue-specific. Although our current pipeline cannot fully distinguish technical artifacts from true biological specificity, boundaries identified from Hi-C datasets of comparable sequencing depth and generated by same experiment type are more likely to represent true sample-specific boundaries. These findings highlight candidates for further experimental validation. (iii) It has been reported that TADs in active chromatin have more distinct boundaries at the population level through the visualization of interaction frequency matrix [15], which may make them easier to be captured by the current TAD calling method. These boundaries without a sharp and abrupt ending at the interaction frequency matrix might, to some degree, result in sample-specific boundaries. To evaluate whether sample-specific TAD boundaries influenced the functional comparisons between boundaries in active and inactive chromatin, we excluded these sample-specific boundaries and repeated all analyses. The results remained highly consistent with those obtained from the full dataset, confirming that our findings were not driven by sample-specific boundaries (Supplementary Figs. S26 and S27; Supplementary Table S4). Discussion Here, we systematically analyzed the heterogeneity of TAD boundaries by their sequence content. We first obtained two main classes of TAD boundaries with different GC contents, GC-rich and AT-rich. Upon a deep exploration of these two classes of TAD boundaries, we found that they were mainly within active and inactive chromatin, respectively. These boundaries were different not only in the sequence content, but also in epigenetic patterns and TAD related features, including boundary strength and TAD hierarchical levels. For instance, our results showed a higher prevalence of CTCF binding sites in GC-rich boundaries compared to AT-rich boundaries, whereas EZH2 binding sites were enriched at AT-rich boundaries but depleted in GC-rich boundaries. It is currently accepted that TADs build the insulation environment by loop extrusion model involving CTCF and cohesin complex (RAD21 and SMC3) [72, 73]. The insulation potency of TAD boundaries is also influenced by the number of CTCFs surrounding them [65, 66]. The lack of CTCF binding sites in AT-rich boundaries indicates an alternative mechanism for TAD formation in inactive chromatin. Our findings suggest that some repressive elements may also help with the TAD formations, such as EZH2, which has been proved by experimental evidence at several genomic loci in human [25, 74]. In addition, the comparison of insulation score and TAD hierarchical structures between two classes of TAD boundaries showed GC-rich boundaries presented higher insulation capabilities and preferred to form hierarchical structures with sub-TADs. These results indicate that TADs in active chromatin with more complex insulated environments participate in a more complex regulatory network compared with the TADs in inactive chromatin. This observation is also supported in other studies by molecular dynamics simulations and imaging-based methods in selected genomic regions [15, 16]. By applying different resolutions ranging from 5 to 400 kb in this analysis, we observed consistent conclusions across resolutions that TAD boundaries are consistently classified into two major groups, with similar distributions of repetitive elements and epigenetic features. We also explored the heterogeneity of TAD boundaries across cell lines and tissues by the existence of TAD boundaries across different TAD samples. According to our defined consensus TAD boundaries across all the samples, we observed that TAD boundaries in thymus showed a lower concurrence than the boundaries in other cell line/tissues. Our hypothesis is that higher gene expression of CTCF and cohesin complex may create different insulated environments and form different TADs. Through clustering TAD boundaries by their distribution across different TAD samples, we obtained the sample-unspecific boundaries, which exhibited greater evolutionary conservation than sample-specific boundaries. All the differences indicate the heterogeneity of TAD boundaries. Nevertheless, there are still some limitations in our current study. First, while we discussed the difference between TAD boundaries in active and inactive chromatin (or A/B compartment), we also suspected the presence of sub-clusters within these TAD boundaries. This hypothesis is supported by several independent studies. The chromatin compartments can be subdivided into sub-compartments which exhibit distinct properties [10], implying potential variations in TAD boundaries in these regions. In addition, Emerson et al. have proposed to categorize TAD boundaries into six classes according to the corner-dots of chromatin loops and the orientation of CTCF motifs, linking them to different replication initiation zones [4]. These can be discussed in the future as more evidence comes to light. Second, our nucleotide content analysis was conducted using a restricted set of k-mer counts. Employing language-based deep learning models may give a better view of sequence pattern at the boundaries. But our current TAD dataset was called based on the 40 kb resolution interaction frequency matrix. Using the 40 kb sequences as the input brings computational and time complexity, which prevents us from utilizing this approach in the current study. Thirdly, experimental validation of the enriched genomic elements is still required. Currently we observed different distributions of these factors among two classes of TAD boundaries. Although our enrichment analyses provide insights into large-scale genomic contexts, and some of these factors (e.g. CTCF, cohesin complex, EZH2) have been experimentally shown to alter TAD structures upon deletion, it remains unclear whether these elements act as causal boundary-forming factors or merely reflect the underlying genomic environment. Future functional experiments (such as perturbation experiments targeting these elements) will be essential to clarify their mechanistic roles in TAD formation [75]. Fourthly, our analysis primarily focused on TAD boundaries merged from multiple cell lines and tissues. Developing a high-resolution, cell-/ tissue-specific TAD boundary atlas will greatly enhance our understanding of chromatin architecture. Such an atlas would allow integration of cell-/tissue- specific histone modification and epigenomic data, thereby providing deeper insights into the regulatory mechanisms of TAD formation and gene expression within specific chromatin domains. Overall, we examined the heterogeneity of TAD boundaries and highlight the importance of TAD boundaries. Our findings indicate the potential of multiple mechanisms for TAD formation. Moreover, the observed heterogeneity of TAD boundaries emphasizes the necessity of accommodating this variety when developing TAD callers or prediction software [38]. Recognizing and integrating the heterogeneity of TAD boundaries can enrich our comprehension of both TADs and their boundaries. Supplementary Material