Evolution of CTCF binding sites in the human genome

Abstract

The CCCTC-binding factor (CTCF) is a master regulator of topologically associating domains (TADs), which shape 3D genome architecture and gene regulation. While TAD evolution across species has been extensively studied, the evolution of CTCF binding sites (CBSs) in primates remains underexplored. Here, we employed a deep learning model to predict genome-wide CBSs in five primate species, including human, chimpanzee, gorilla, orangutan, and macaque. We found that approximately half of human CBSs are conserved across all five species. These primate-conserved CBSs cluster near TAD boundaries, exhibit stronger motif matches, and are subject to stronger selective constraints. In contrast, human-specific CBSs show signatures of positive selection. We identify a positively selected human-specific CBS that emerged de novo from ancestral noncoding sequence via three derived substitutions in the CTCF binding motif. This site anchors a TAD containing PDE9A, a gene involved in learning and memory, which exhibits strengthened enhancer–promoter interactions in humans relative to chimpanzees. Human-specific CBSs contribute to human-specific gene expression regulations, where nearby genes, such as PRDM16, ADCY1, and RIMS1, are differentially expressed and exhibit stronger or human-specific promoter–enhancer interactions in humans, compared to chimpanzees. We further demonstrate that distinct classes of transposable elements contribute differentially to conserved and human-specific CBSs. Moreover, genetic variants near human CBSs are more enriched for associations with complex traits and diseases. Our results highlight that natural selection has shaped CBS evolution in primates, driving both conservation and innovation in chromatin organization and contributing to human-specific regulatory evolution.

Graphical Abstract

Graphical Abstract.

Introduction

It has been proposed that the phenotypic divergence between humans and chimpanzees may stem more from changes in gene regulation than from alterations in protein-coding sequences (Kanton et al. 2019; Song et al. 2021). Regulatory evolution is increasingly recognized as a key driver of the unique traits that have emerged during human evolution (Kanton et al. 2019; Song et al. 2021; Starr et al. 2024). There is growing evidence that the evolution of 3D genome architecture can contribute to the emergence of regulatory and morphological novelties in vertebrates, such as the wing-like pectoral fins in the little skate, the adaptive intersexuality in the Iberian mole, and corticogenesis in primates (Real et al. 2020; Luo et al. 2021; Marlétaz et al. 2023). These examples illustrate how changes in genome architecture can reshape enhancer–promoter interactions and developmental programs, providing mechanistic links between genome structure and evolutionary innovation.

Three-dimensional genome organization operates at multiple architectural levels in vertebrates, such as compartments and topologically associating domains (TADs), and arises from the interplay of distinct mechanisms, including loop extrusion, phase separation, and nonlinear transcriptional activity (Rao et al. 2014; Dixon et al. 2015; Sabari et al. 2018; Zuin et al. 2022; Szalay et al. 2024; Rossi et al. 2025). Among these, TADs are genomic regions characterized by a high frequency of self-interactions that function primarily to insulate and regulate gene expression (Dixon et al. 2012; Bonev et al. 2017; Zhan et al. 2017). The CCCTC-binding factor (CTCF) is a key architectural protein involved in the formation and maintenance of TADs. The turnover of CTCF binding sites (CBSs) in the genome can influence the TAD structure, thereby modulating regulatory interactions across large genomic blocks. Such changes have the potential to profoundly reshape gene regulatory networks, potentially exerting a greater impact than other more localized regulatory changes. Therefore, investigation on the evolution of CBSs in humans and other primates could help understand their contribution to the evolution of TAD architecture and associated phenotypic differences.

Disruption of CTCF binding has been shown in cell lines to result in the abrogation of TAD insulation and the formation of ectopic cis-regulatory interactions between neighboring TADs, which can lead to gene misexpression and contribute to developmental disorders and cancer (Lupiáñez et al. 2015; Franke et al. 2016; Hnisz et al. 2016). Initial studies suggested that TADs and TAD boundaries are highly conserved across species, supported by comparable TAD structures at the Hox and Six loci (Woltering et al. 2014; Gómez-Marín et al. 2015); conserved TAD numbers, boundary positions, and gene ordering within TADs (Patalano et al. 2025); enrichment of genomic rearrangement breakpoints at TAD boundaries but depletion within TADs (Krefting et al. 2018; Lazar et al. 2018); and a genome-wide pattern of negative selection against structural variations that disrupt TAD boundaries (Fudenberg and Pollard 2019). However, most of the evidence supporting the conservation of TADs is indirect or limited to a small number of loci. Recent studies suggested that conservation of TADs may be more limited than previously thought, even between species as closely related as human and chimpanzee (Eres et al. 2019). Indeed, conserved TAD boundaries tend to exhibit stronger insulation, higher CBS enrichment, and greater enrichment of housekeeping genes (Luo et al. 2021). Despite the important role of CTCF in TAD formation, it remains unclear to what extent cross-species differences in TADs are driven by evolution of CTCF binding profiles.

CTCF is a major mediator of chromatin looping and TAD boundary insulation and is involved in diverse regulatory functions, including transcriptional activation and repression, facilitation of proper enhancer–promoter interactions, and establishment of euchromatin–heterochromatin boundaries (Phillips and Corces 2009). CTCF binding motifs are highly conserved across vertebrates (Kim et al. 2007; Schmidt et al. 2012), and CBSs located within TAD boundary regions have been shown to be under stronger evolutionary constraints in both humans and mice (Kentepozidou et al. 2020; Radke et al. 2021). In Drosophila genomes, more conserved CBSs are subject to stronger purifying selection, while diverged and younger CBSs exhibit signatures of positive selection (Malik et al. 2012). In the human genome, signals of positive selection in CBSs in different human cell lines have been identified, and brain-related cell types have been found to exhibit the highest proportion of positive selection (Liu and Robinson-Rechavi 2020). Rare mutations in human CBSs have been linked to the formation of novel, population-specific TADs and are associated with altered disease susceptibility (Wu et al. 2025). However, the evolution of CBSs across primate species and within human populations has not been systematically examined.

Transposable elements (TEs) have been proposed to contain and disseminate CBSs (Schwalie et al. 2013; Choudhary et al. 2020, 2023; Diehl et al. 2020; Luo et al. 2021). TE activation has been reported to drive species-specific expansions of genome-wide CBSs in rodents, dogs, and opossums, but not in humans (Schmidt et al. 2012; Schwalie et al. 2013). For CBSs located within TAD boundary regions, the contributions of TE types are also distinct among species (Choudhary et al. 2020; Diehl et al. 2020; Luo et al. 2021). For example, short interspersed nuclear elements (SINEs) contribute significantly to TAD boundary CBSs in mice, moderately in dogs and rhesus macaques, but only minimally in humans (Choudhary et al. 2023). Evolutionarily old TEs, including MIR, LINE-2, Charlie, and MaLR, were enriched in conserved TAD boundary regions in human, macaque, and mouse brains, while young TEs, such as AluY (SINE) and L1PA (LINE), were significantly enriched in human-specific TAD boundary regions, which may contribute to the origin of human-specific TADs in the brain (Luo et al. 2021). However, most studies have focused on the contribution of TEs to CBSs genome-wide or at TAD boundary regions, while their contributions to CBSs across different conservation levels—potentially reflecting the timing of TE activity—remain unexplored.

To investigate the evolutionary dynamics and functional consequences of CBS divergence in primates, we employed a deep learning framework to predict genome-wide CBSs across five species: human, chimpanzee, gorilla, orangutan, and macaque. Integrating these predictions with available TAD maps, we aimed to (i) characterize and compare the genomic features of conserved versus human-specific CBSs, (ii) quantify the action of natural selection on different types of CBSs, (iii) assess the contribution of TEs to the emergence and divergence of CBSs in humans, (iv) examine how CBS evolution influences higher-order chromatin organization and gene regulation, and (v) investigate whether variants in human CBSs are linked to human disease phenotypes. Through these analyses, we seek to elucidate how CBS evolves in humans and contributes to human-specific regulatory architecture and traits.

Results

Computationally predicted CTCF binding profiles in primate genomes

We first used BPNet (Avsec et al. 2021) to computationally predict CTCF binding profiles in the genomes of five primates, including human, chimpanzee, gorilla, macaque, and orangutan. To assess the reliability of BPNet-derived predictions, we performed a chromosome-level hold-out validation using chromosomes 1, 8, and 9 as an independent test set. Across all species, BPNet-predicted and observed CTCF ChIP-seq log-counts showed Pearson and Spearman rank correlation coefficients greater than 0.70 on the held-out chromosomes (Fig. S1), supporting predictive performance and generalizability of BPNet deep learning models.

CTCF recognizes highly conserved consensus DNA sequence motifs across these species (Fig. S2). In total, we identified between 21,535 and 42,669 CBSs for each species. To further evaluate prediction accuracy, we compared the predicted human CBSs with empirically identified CBSs from human cell lines. Among the 56 human cell lines analyzed by Li et al. (2013), 95.5% of our predicted CBSs were present in more than 20 cell lines. Substantial overlap was also observed between our predicted human CBSs and available ENCODE CTCF ChIP-seq peaks in human cell lines (Table S1, Fig. S3a), with induced pluripotent stem cells (iPSCs) and bipolar neurons showing some of the highest overlaps among all tested cell types (Fig. S3a).

We further analyzed the features of our predicted CBSs in the human genome and compared them with empirical CBSs reported by Nanni et al. (2020). Specifically, we grouped adjacent CBSs into mono-plets, di-plets, tri-plets, and tetra-plets and classified these clusters into four pattern classes based on their orientations (Fig. 1a). By examining the chromosomal distributions of these four pattern classes, we found that convergent patterns tend to have larger cluster size than divergent patterns, regardless of the number of CBSs involved (Fig. 1b). Furthermore, the convergent and divergent patterns exhibit opposite spatial distribution trends between 5 and 131 kb, with convergent patterns being depleted and divergent patterns enriched in windows up to about 131 kb (Fig. 1c). These spatial clustering patterns of our predicted human CBSs closely resemble the empirical data (Nanni et al. 2020), supporting the reliability of our prediction.

Figure 1.

Spatial distribution of predicted human CBS clusters. a) Classification of CBS clusters by their relative orientation. Adjacent CBSs in human genome were clustered into mono-plets, di-plets, tri-plets, and tetra-plets and categorized by patterns of their relative orientation. The imbalance between pattern classes within each multi-plets was assessed though Pearson chi-square tests. b) The size distribution of CBS clusters shows that convergent patterns tend to have significantly larger cluster size. c) Convergent CBS cluster patterns are depleted, and divergent patterns are enriched in windows up to about 131 kb.

Taken together, the held-out validation and empirical comparisons support the reliability of our predicted CBSs, indicating that they recapitulate key empirical features of CTCF binding and provide a robust basis for downstream analyses.

Clustering of conserved CBSs around TAD boundary

To investigate the evolution of human CBSs, we classified all CBSs according to their conservation across five primates using a human-centric approach. Specifically, we categorized CBSs as primate-conserved (five-way) if shared by all five primates, partially conserved between human and other primate species (four-way, three-way, and two-way), or species-specific if unique to a single primate species (Fig. 2a). We identified 11,543 five-way CBSs, accounting for approximately 50% of the predicted human CBSs (Fig. 2a), consistent with previous reports of high CBS conservation in mammals (Schmidt et al. 2012). To assess whether the classification of human-specific CBSs was affected by alignment failure in repetitive regions, we performed additional synteny-based and flanking-sequence mapping validation analyses (see Materials and methods section, Table S2). These analyses indicated that the majority of human-specific CBSs were supported by syntenic alignments and showed high flanking-sequence mappability across nonhuman primates.

Figure 2.

Distribution of CBSs with different conservation levels around TAD boundaries. a) Number of CBSs with different conservation levels. Arrow points to human-specific CBSs. b) Distance of human CBS with different conservation levels to its nearest human TAD boundary. c) Distribution of CBSs around TAD boundary in 5 kb windows. Colors denote conservation level same as in (b). d) Distance of human CBS with different conservation levels to its nearest CBS. CBSs located within the ± 50 kb TAD boundary regions (TB) and those located outside these regions (nonTB) are compared. e) Number of TAD boundary regions harboring clustered and singleton CBS of different conservation levels. Results are robust to definitions of CBS conservation and TAD boundary-window sizes (Figures S4 and S5).

We then examined the spatial distribution of CBSs with different conservation levels near TAD boundaries, utilizing published TAD coordinates information from consensus Hi-C map of human iPSCs (Eres et al. 2019). We found that highly conserved CBSs tend to localize closer to TAD boundaries (Fig. 2b), and more CBSs with higher conservation level clustered near the TAD boundaries (Fig. 2c). Based on this clustering pattern, we defined TAD boundary regions as the ±50 kb flanking sequences surrounding each boundary, and classified CBSs within these regions as TAD boundary-associated. We observed that for all CBSs, distance to the neighboring CBS decreases with increasing conservation level (Fig. 2d). For TAD boundary-associated CBSs, neighboring CBSs were significantly closer together than those outside TAD boundary regions, a pattern that was absent in human-specific CBSs (Fig. 2d). These results indicate that conserved CBSs tend to cluster more densely in TAD boundary regions. Finally, we clustered CBSs less than 10 kb apart, following Kentepozidou et al. (2020), and found that a large number of TAD boundary regions contain five-way CBSs, and these regions have a higher proportion of clustered CBSs (Fig. 2e). These results support the conclusion that conserved CBSs preferentially cluster near TAD boundaries, consistent with their important roles in maintaining conserved TAD structure.

Importantly, all these patterns we observed are robust to different parameter choices for the definitions of conserved CBSs (1 bp or stricter 50% reciprocal overlap, or motif-centered matching between orthologous CBSs; Fig. S4) and TAD boundary region sizes (±25, ±50, ±75 kb; Fig. S5). These robustness analyses confirm that our results are not sensitive to the specific parameter settings used but reflect genuine biological features in primate genome evolution.

Purifying selection and positive selection drive the evolution of five-way and human-specific CBSs, respectively

We observed that five-way CBSs have higher phastCons scores than human-specific CBSs, independent of their association with TAD boundaries (Fig. 3a). Notably, for both five-way and human-specific CBSs, the ones located within TAD boundary regions show significantly higher phastCons scores than those located outside (Fig. 3a, Table S3). These findings suggest that five-way CBSs within TAD boundary regions play important roles in maintaining conserved TAD structure and are subject to strong evolutionary constraints across species.

Figure 3.

Selection on CBSs of different conservation levels. a) PhastCon scores of different types of CBSs. hs, human-specific CBS; TB, within the ± 50 kb TAD boundary regions. For significance, see Table S3. b) Tajima's D scores of human-specific and five-way CBSs, in comparison with noncoding and coding variants in the nearest gene, and randomly selected 25 bp intergenic sequence in-between. Syn, synonymous sites; Nonsyn, nonsynonymous sites; Inter, intergenic region. (c) Proportion of adaptation (α) for human-specific and five-way CBSs, using long intron or synonymous sites of nearest gene as neutral reference, respectively. MAF10 and MAF25 are MAF cutoff of 10% and 25%, respectively. Error bars represent 95% confidence intervals obtained by jackknife resampling across CBS loci within each group. d) Sequence alignment of the positively selected CBS when using the synonymous sites as neutral reference and MAF cutoff of 10%.

We then examined signatures of selection on five-way and human-specific CBSs. Purifying selection tends to reduce polymorphism levels and skew the allele frequency distribution toward rare variants, which is indicated by a more negative Tajima's D value compared to neutral sites. Using human polymorphism data from the African population (Bergström et al. 2020), we found that Tajima's D value of five-way CBSs within the ±50 kb TAD boundary regions was significantly lower than that of nearby randomly sampled intergenic sequences, and slightly lower (n.s.) than that of synonymous sites (Fig. 3b). In contrast, for human-specific CBSs, particularly sites within TAD boundary regions, Tajima's D values were highly variable and slightly higher (n.s.) than intergenic neutral references (Fig. 3b), hinting potential involvement of positive selection.

To further delineate the selective signatures underlying the evolution of CBSs, we employed the McDonald and Kreitman (MK) test to distinguish neutrality, negative selection, and positive selection (McDonald and Kreitman 1991), by comparing the levels of polymorphism within species and divergence between species. Positive selection is indicated by higher divergence to polymorphism ratio at these sites, compared to neutral references, whereas negative selection has the opposite effect (Andolfatto 2005; Parsch et al. 2009). Using the nearest-gene-matched synonymous sites and long intron as locally matched neutral references, our analyses revealed that five-way CBSs showed significantly lower divergence relative to polymorphism, consistent with negative selection (Table 1). In contrast, human-specific CBSs exhibited significantly higher divergence relative to polymorphism, consistent with a greater contribution of positive selection.

Table 1.

Comparison of polymorphism and divergence in five-way and human-specific CBSs, using synonymous sites and long intron of the nearest gene of each CBS as neutral reference.

Category	Divergent	Polymorphic	Divergent/polymorphic ratio	Comparison	P-value
5-way	1,862	409	4.55	—
Synonymous	48,348	8,530	5.67	vs. 5-way	<0.01
Long intron	4,351,535	790,765	5.50	vs. 5-way	<0.01
5-way TB	1,421	309	4.60	—
Synonymous	38,833	6,863	5.66	vs. 5-way TB	<0.01
Long intron	3,454,152	629,762	5.48	vs. 5-way TB	<0.01
hs	487	60	8.12	—
Synonymous	5,152	907	5.68	vs. hs	0.01
Long intron	446,047	76,929	5.80	vs. hs	0.01
hs TB	261	36	7.25	—
Synonymous	2,966	532	5.58	vs. hs TB	0.17
Long intron	254,268	44,307	5.74	vs. hs TB	0.22

hs, human-specific CBSs; TB, CBSs within ±50 kb TAD boundary regions.

We further used an extended MK framework to estimate α, the fractions of divergent substitutions between species driven to fixation by positive selection, for five-way and human-specific CBSs using both locally matched neutral reference types and two minor allele frequency (MAF) thresholds. Although α estimates for human-specific CBSs showed greater variance, likely due to the smaller number of variants, they were consistently higher than those for five-way CBSs. Permutation tests further supported this pattern: under the 10% MAF threshold, human-specific CBSs showed significantly higher α than five-way CBSs using either long introns (P = 0.020) or synonymous sites (P = 0.013) as neutral references, whereas the difference was marginal or not significant under the 25% MAF threshold (Fig. 3c, Fig. S6), suggesting a higher proportion of adaptive substitutions fixed in human-specific CBSs. Together, these results suggest that a substantial proportion of substitutions in human-specific CBSs may have been driven to fixation by positive selection.

There is limited power to identify individual CBS showing signature of positive selection through α, due to limited polymorphism and divergence sites within its short length. Nonetheless, we found one positively selected human-specific CBS using the synonymous sites as neutral reference and MAF cutoff of 10%. This CBS harbors a human-specific G-to-A substitution at a highly conserved position in CTCF binding motif, along with three additional substitutions, together leading to the de novo origin of human-specific CBS from ancestral noncoding sequences (Fig. 3d). Since this CBS site is located within a TAD boundary region, we further examined the interaction frequency between enhancer and promoter pairs for genes situated within the TAD anchored by this CBS site. We found that the enhancer–promoter interaction frequency of cGMP-hydrolyzing phosphodiesterase-9a (PDE9A) gene was significantly higher in humans than in chimpanzees (Fig. S7). PDE9A was reported to express in cognition-relevant brain regions such as the cortex and the hippocampus (Lakics et al. 2010) and is involved in the induction of long-term potentiation and cognitive function (Hutson et al. 2011; Kroker et al. 2014; Alexander et al. 2016).

TEs contribute preferentially to human-specific CBSs and TAD boundaries

TEs have also been implicated in the evolution of chromatin architecture by introducing CBSs (Schmidt et al. 2012; Schwalie et al. 2013; Choudhary et al. 2020, 2023; Diehl et al. 2020; Kentepozidou et al. 2020). To assess the contribution of TEs to CBSs of varying evolutionary conservation, we analyzed the overlap between TE-derived sequences and human-specific versus five-way conserved CBSs. Using a 40% CBS-TE overlap threshold, we found that about 35.7% of human-specific CBSs overlap with TEs, significantly higher than the 10.8% observed for five-way CBSs (Fig. 4a). This enrichment was consistent both within and outside TAD boundary regions, suggesting a greater contribution of TEs to human-specific CBSs. Notably, for CBSs within human-specific TAD boundary regions, the proportion of TE-overlapping CBSs increased further to 46.0% and 13.4% for human-specific and five-way CBSs, respectively (Fig. 4a).

Figure 4.

Contribution of TEs to human CBSs. a) Contribution of TEs to CBSs of varying evolutionary conservation, at 40% overlap threshold (minimum TE coverage of CBS). Sample sizes are shown within bars. b) Fraction of CBSs associated with TEs using different overlap thresholds. The significance was estimated by 10,000 permutations from matched-superpool of ENCODE CTCF ChIP-seq peak-derived regions. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. c, d) Stacked bar plots showcasing the distribution of repeat elements derived CTCF across major repeat elements classes (c) and major repeat elements families (d). e, f) Volcano plots displaying enrichment of individual TE families for (e) human-specific CBSs and (f) five-way conserved CBSs. Fold enrichment and empirical P-values were calculated from matched-superpool permutations (Table S5 and S6).

To evaluate whether the observed TE association patterns were artifacts of specific overlap threshold or local genomic context, we performed two additional robustness analyses. First, the main conclusions remained consistent across a range of CBS-TE overlap thresholds from 20% to 70%, with human-specific CBSs showing higher TE association than five-way conserved CBSs and human-specific TAD boundary CBSs exhibiting the strongest TE association (Fig. 4b). Second, in a matched-superpool permutation framework that controlled for chromosome, length, guanine-cytosine (GC) content, mappability, distance to the nearest gene, and distance to the nearest TAD boundary, the TE enrichment patterns remained significant (Fig. 4b, Table S4). Together, these analyses indicate that the observed TE enrichment of human-specific CBSs, particularly near human-specific TAD boundaries, is robust to both parameter choice and local genomic confounding effects. These results suggest that TEs have played a disproportionately large role in establishing CBSs associated with newly evolved TAD boundary regions in humans.

We further dissected the contribution of specific TE lineages to the observed enrichment in human CBSs, for TE lineages at both the class and family levels. At the class level, long terminal repeat (LTR) elements and LINE elements were consistently enriched for all human-specific CBSs (hs), human-specific CBSs within TAD boundary regions (hsTB), and human-specific CBSs within human-specific TAD boundary regions (hsHSTB) (Fig. 4c, Table S5), evaluated by matched-superpool permutations. The family-level analyses revealed that this signal of LTRs was driven primarily by endogenous retrovirus (ERV) lineages, especially ERV1, ERVK, and ERVL-MaLR (Figs 4d, e). Within the hsHSTB category, ERVL also showed a significant enrichment, suggesting its potential role in shaping human-specific chromatin architecture at TAD boundary regions. The enrichment signal of LINE elements was largely attributable to the L1 family. In addition, the DNA transposon hAT-Blackjack was significantly enriched in hsTB context (Fig. 4e, Table S5).

In contrast to human-specific CBSs, five-way conserved CBSs show depletion for all major TE groups, including LTRs, LINEs, SINEs, and DNA transposons (Table S6), suggesting purifying selection acting on these conserved genomic architectures. However, amid this global exclusion, the hAT-Tip100 family of DNA transposon emerged as an exception: it was the only lineage to show significant enrichment in five-way conserved CBSs (Figs 4d, f). These class- and family-level enrichment patterns are robust across varying CBS-TE overlap thresholds (Tables S5 and S6).

Impacts of CBSs near TAD boundary on gene regulation

To explore the functional relevance of different human CBS categories, we first compared the CTCF binding as a function of CBS conservation level and TAD boundary association. We computationally calculated the matching probability of each CBS to CTCF binding motif of corresponding species (Fig. S2) and used it as a proxy for CTCF binding strength (hereafter referred to as predicted CTCF binding score). We found that CBSs associated with TAD boundaries have significantly higher predicted CTCF binding scores than those located elsewhere (Fig. 5a), and this pattern was primarily driven by five-way CBSs (Fig. 5b). Notably, five-way CBSs exhibit higher predicted CTCF binding scores than less conserved ones (Fig. 5b), supporting that purifying selection may act to maintain these high binding score CBSs around TAD boundaries in primates.

Figure 5.

Functional effects of CBSs on gene regulation. (a) Comparison of CTCF binding scores between TAD boundary-associated CBSs (TB, within the ±50 kb TAD boundary regions, n = 15,393) and nonboundary-associated CBSs (nonTB, n = 6,102). CTCF binding score is defined as the matching probability to the species-specific predicted CTCF binding motif (see the Materials and methods section). b) CTCF binding scores stratified by CBS conservation level and boundary association status, showing that five-way conserved CBSs exhibit significantly higher binding affinity, particularly those associated with TAD boundaries. Colors denote boundary association as in (a). P-values in (a and b) were estimated using two-sided Wilcoxon rank-sum tests, and significance levels are *P < 0.05; **P < 0.01; ***P < 0.001. c) Percentage of DE tandem gene pairs separated by CBSs. DE genes were defined as those with a FDR q-value < 0.01 and |log₂ fold change| > 1. Number within bars are the count of tandem gene pairs (n) in each category. hs, human-specific CBS; TB, within the ±50 kb TAD boundary regions; No CBS, tandem gene pairs not separated by any CBS. d) Comparison of contact frequencies for TADs anchored by human-specific CBSs within human-specific TAD boundary regions versus their homologous regions in chimpanzees. Dashed lines indicate median values. Contact frequencies were compared between humans and chimpanzees using a two-sided Wilcoxon rank-sum test.

TAD boundaries insulate gene expression between neighboring TADs. To evaluate the functional impacts of CBSs on gene regulation, we analyzed the expression patterns of tandem gene pairs flanking these sites using RNA-seq data from human iPSC lines (Pavlovic et al. 2018). Tandem genes transcribed in the same direction are known to display more coordinated expression than nontandem genes (Caron et al. 2001; Lercher et al. 2002). Compared to tandem gene pairs not separated by CBSs, the fraction of differentially expressed (DE) tandem genes was higher for all CBSs (Fig. 5c). For human-specific CBSs within the ±50 kb TAD boundary regions, the fraction of DE tandem genes was even higher (Fig. 5c), consistent with the insulating role of CBSs within TAD boundary regions.

Among the 762 human-specific CBSs, 63 sites are located near human-specific TAD boundaries, which anchor TADs or sub-TADs (smaller TADs nested within larger TADs) present only in humans. The contact frequencies within these human-specific TADs anchored by the 63 sites are significantly higher in humans compared to homologous regions in chimpanzees (Fig. 5d). Among the 63 genes that are nearest to these 63 CBSs, only 44 genes have available promoter–enhancer interaction data in both humans and chimpanzee, in which 19 of them showed no detectable contacts in either species. Among the remaining 26 genes with valid data, 18 exhibited human-specific or human-enhanced promoter–enhancer contacts, while 8 showed higher contact frequencies in chimpanzees (Table S7). Specifically, eight genes displayed contacts exclusively in humans, including ADCY1 (adenylyl cyclase 1) and RIMS1 (regulating synaptic membrane exocytosis 1). Another 10 genes showed elevated contact frequencies in humans, with PRDM16 (PR/SET domain 16) exhibiting the most pronounced fold change (Fig. S8, Table S7). To examine the functional impact of these 3D genome and promoter–enhancer interaction changes, we integrated human–chimpanzee comparative transcriptomic data from three independent sources (Khrameeva et al. 2020; Wang et al. 2025; Zhuo et al. 2025), mostly from comparisons of brain regions. Interestingly, 15 out of the 18 genes with human-specific or human-enhanced promoter–enhancer contact frequencies were reported to upregulate in humans across multiple brain regions in at least one of these datasets (Table S8), including PRDM16, ADCY1, and RIMS1, potentially linking human CBS evolution to gene expression evolution in brain regions via TAD structures, suggesting potential contributions to the cognitive and neuronal traits that distinguish humans from other primates.

In summary, we showed that five-way CBSs near TAD boundaries have higher predicted CTCF binding scores and are under purifying selection; that all CBSs, particularly human-specific CBSs near TAD boundaries, drive gene expression divergence; and that human-specific CBSs contribute to human-specific expression regulation, collectively underscoring the functional importance of CBSs in gene regulation.

Association of polymorphisms in human CBSs with GWAS hits and disease phenotypes

We identified 14,189 polymorphic single nucleotide polymorphisms (SNPs) present within 10,001 CBSs using polymorphism data in human populations from the 1000 Genomes Project (https://www.internationalgenome.org/) and gnomAD v4.1.0 (https://gnomad.broadinstitute.org/). This dataset included 589 SNPs in 375 human-specific CBSs and 7,292 SNPs in 5,241 primate-conserved five-way CBSs. Potential functional impacts inferred using Combined Annotation-Dependent Depletion (CADD v1.7, https://cadd.gs.washington.edu/) suggest that 39 and 2,235 of these variants located within human-specific and five-way CBSs may be deleterious (CADD scores > 15), respectively.

Many variants show population-specific allele frequencies, potentially contributing to population phenotypic diversity or disease risk. The African population contains nearly half of the population-specific human CBS variants, consistent with their higher genetic diversity overall (Ramachandran et al. 2005; Tishkoff et al. 2009). Among the CBS variants that is absent in the African population within the 1000 Genome dataset, we found a T-to-C mutation specific to Eurasian populations at a highly conserved position in CTCF binding motif (Fig. S9a). The ancestral T allele is likely not recognized by CTCF, and this derived C mutation probably created a segregating CBS site in Eurasian populations. This human CBS is located near a human-specific TAD, while the homologous TAD structure was absent in chimpanzees who carry the ancestral allele T, likely linked to the absence of this CBS. In humans, the gene C14orf132, located within this human-specific TAD, has higher enhancer–promoter interaction in human than in chimpanzee (Fig. S9b, c). Interestingly, C14orf132 is highly expressed in the human brain (UniProt accession: Q9NPU4) and is DE between human and chimpanzee brains (Khrameeva et al. 2020), suggesting potential contribution of this CBS to brain-related phenotypes or diseases.

To explore the functional relevance of human CBSs to disease and phenotypic variation, we performed heritability enrichment analysis of variants located within ±50 kb of human CBSs using stratified LD score regression (S-LDSC) and summary statistics from 176 genome-wide association studies (GWAS) datasets curated in public repositories (Finucane et al. 2015). We found significant enrichment of these CBS-associated variants across a number of human complex traits and diseases (Table S9). Among these, we followed the trait-grouping scheme of Wu et al. (2025) and displayed a subset of 42 traits representing a broad spectrum of cardiopulmonary, hematological, immunological, metabolic, and neuropsychiatric categories (Fig. 6). In each category, variants near human CBSs showed significant heritability enrichment in several traits including Alzheimer’s disease, type 2 diabetes, height, triglycerides, inflammatory bowel disease, and allergy eczema (Fig. 6, Table S10). Notably, Alzheimer's disease showed the highest fold enrichment, with 46.9% of SNPs accounting for 71.3% of SNP-based heritability (Table S10). These results are robust when using five-way conserved CBSs (Fig. S10, Table S10), probably driven by the fact that five-way conserved sites constitute most of the total CBS set, while the human-specific CBS contains relatively few polymorphic sites, limiting its power to detect significant enrichments across these traits (Fig. S11, Table S10). Moreover, these enrichment patterns for the 42 traits were consistent between the ±50 and ±25 kb windows (Table S11), confirming that our findings are robust to variations in annotation window size.

Figure 6.

Heritability enrichment for genetic variants located within the ±50 kb of human CBSs. A subset of 42 traits representing a broad spectrum of human complex traits and disease were displayed, following the trait-grouping scheme of Wu et al. (for full results of 176 datasets, see Table S9). Enrichment is the proportion of heritability explained by these SNPs to the proportion of SNPs in these regions. Value >1 indicates that SNPs near CBSs contribute disproportionately more to trait heritability relative to their genomic coverage. Error bars represent jackknife standard errors estimated by S-LDSC. Asterisks indicate significant enrichment after Bonferroni correction: *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001. Notably, these SNPs exhibit the highest fold enrichment of heritability for Alzheimer's disease (highlighted).

We further examined the overlap between CBSs and loci reported to be associated with human diseases and traits from GWAS (https://www.ebi.ac.uk/gwas/). We identified 92 CBSs harboring GWAS variants associated with traits such as intelligence, type 2 diabetes and metabolic level (Table S12). These variants are enriched at TAD boundary-associated CBSs, especially for five-way CBSs (P = 0.013, Fisher's exact test). Three SNP variants in human-specific CBSs are associated with GWAS traits, including rs806276 in mitogen-activated protein kinase kinase kinase 7 (MAP3K7) associated with attention deficit hyperactivity disorder, rs61502187 in ficolin 2 (FCN2) having carbohydrate binding and opsonic activities, and rs35682 in the ghrelin and obestatin prepropeptide (GHRL) gene associated with energy homeostasis. Notably, rs35682 is located within the core conserved site of the CTCF binding motif.

Collectively, these analyses highlight the potential contribution of both conserved and human-specific CBSs to human health and phenotypic variations. Future functional validation of candidate CBS variants at disease-associated loci will be essential to elucidate their precise mechanistic roles.

Discussion

In this study, we investigated the evolution of genome-wide CBSs in five primate species, including humans. Using a deep learning approach, we predicted genome-wide CBSs and showed that primate-conserved CBSs are enriched at TAD boundaries, exhibit higher predicted CTCF binding scores, and are subject to strong purifying selection, consistent with their essential roles in maintaining stable chromatin architecture and regulatory insulation. In contrast, human-specific CBSs display signatures of positive selection. In addition to de novo evolved CBSs from ancestral noncoding sequences, we found a substantial contribution from TEs, particularly in human-specific TAD boundary regions. Human CBSs are associated with altered enhancer–promoter interactions, gene expression divergence in brain-related genes, and genetic variants linked to complex traits and diseases, underscoring the role of CBS turnover in shaping human-specific regulatory organization and phenotypic evolution.

Both purifying and positive selection sculpt the evolution of CBSs in humans

Our results reveal distinct evolutionary forces shaping conserved and human-specific CBSs, with purifying selection acting on ancient, primate-conserved five-way CBSs, particularly those located at TAD boundary regions, and positive selection driving the fixation of human-specific CBSs. These results provide evidence that the evolution of 3D chromatin architecture is not only shaped by neutral processes or mutational biases but is also subject to both purifying and positive selection, reflecting the important role of CBS evolution in maintaining genome stability and enabling regulatory innovation.

The higher phastCons scores observed for five-way CBSs, especially those within TAD boundary regions, are consistent with their critical role in preserving chromatin domain organization. This pattern supports the notion that these sites are under strong evolutionary constraint, likely because disruptions of these CBSs can lead to aberrant enhance–-promoter interactions, gene misexpression, and disease phenotypes. The fact that even within the same conservation category, boundary-localized CBSs exhibit higher constraint than nonboundary CBSs further underscores the functional importance of spatial genome architecture in gene regulation. This result aligns with previous studies showing that TAD boundary regions have a high proportion of conserved bases in humans (McArthur and Capra 2021) and that CBSs at all conservation levels within TAD boundary regions have higher selective constraints in mice (Kentepozidou et al. 2020).

In contrast, human-specific CBSs display lower sequence conservation, and more importantly, signatures of positive selection, as evidenced by elevated Tajima's D values, increased divergence-to-polymorphism ratios in McDonald–Kreitman tests, and a significantly higher proportion of adaptive substitutions (α). This suggests that many human-specific CBSs have risen to fixation through natural selection favoring novel regulatory configurations, which may have contributed to lineage-specific phenotypic evolution. This is consistent with a previous study in Drosophila, which shows that newly evolved CBSs exhibit a significantly stronger signature of positive selection compared to older CBSs (Malik et al. 2012).

One particularly interesting example is a human-specific CBS that emerged de novo through a G-to-A substitution at a key position in the CTCF binding motif, accompanied by three additional mutations. This site lies within a TAD boundary region and is associated with increased enhancer–promoter interaction frequencies for the PDE9A gene in humans compared to chimpanzees. Given that PDE9A is expressed in brain regions critical for learning and memory, such as the hippocampus and cortex, and plays a role in long-term potentiation (Hutson et al. 2011; Kroker et al. 2014; Alexander et al. 2016), this locus may represent a case of adaptive rewiring of neuroregulatory circuits during human evolution. The fixation of this CBS could have fine-tuned the spatial regulation of PDE9A, thereby contributing to enhanced cognitive functions in humans, a hypothesis that aligns with growing evidence linking 3D genome reorganization to brain evolution (Luo et al. 2021; Winick-Ng et al. 2021). It is also worth noting that while we detected a genome-wide signal of positive selection on human-specific CBSs, individual site-level inference remains challenging due to the short length of CBSs and limited polymorphism data. Nevertheless, our results highlight the potential for future studies using larger population datasets to uncover additional adaptive events.

Distinct TEs contribution to human-specific and primate-conserved CBSs

Our results demonstrate a significant contribution of TEs to CBSs, particularly those associated with human-specific TAD boundaries. We observed a significantly higher overlap between TEs and human-specific CBSs compared to evolutionarily conserved five-way CBSs. This pattern was further accentuated at human-specific TAD boundaries, highlighting the role of TEs in CBS evolution and in human-specific 3D chromatin architecture. These findings are in general agreement with previous studies showing that waves of retrotransposon expansion have remodeled CTCF binding landscapes across mammalian evolution (Schmidt et al. 2012; Schwalie et al. 2013; Thybert et al. 2018; Choudhary et al. 2020, 2023; Diehl et al. 2020; Okhovat et al. 2023). TE activation has been reported to drive species-specific expansions of genome-wide CBSs in rodents, dogs, and opossums but not in humans (Schmidt et al. 2012; Schwalie et al. 2013). However, CBSs located within the human-specific TAD boundaries were enriched in LTR elements (Okhovat et al. 2023). Moreover, a previous study has reported increasing TE abundance at chromatin loops with decreasing conservation between human and mouse (Diehl et al. 2020). Our data extend this model by classifying human CBSs into different conservation levels, which not only reflect their functional importance but also reflect the evolutionary time of their origin. This allows us to dissect human CBSs temporally to examine TE activity. We show that LTRs, especially ERV1 and ERVL-MaLR, are overrepresented in human-specific CBSs, supporting their roles in human regulatory innovation since the divergence from chimpanzees. Notably, we observed a marked increase in the ERVL family in CBSs within human-specific TAD boundaries, which may reflect distinct sequence features or integration preferences that make this family more compatible with boundary activity. The enrichment of these elements aligns with reported bursts of ERV activity during primate evolution (Belshaw et al. 2005; Wildschutte et al. 2016).

In addition, our finding suggested that distinct lineages within the hAT superfamily of DNA transposons may have been co-opted at different evolutionary timepoints. The hAT-Blackjack family is specifically enriched in human-specific CBSs within TAD boundary regions, while the hAT-Tip100 family is significantly overrepresented in five-way conserved CBSs. hAT-Tip100 has been reported to broadly enriched at CBSs across all human TAD boundary regions in human–mouse comparison (Okhovat et al. 2023), and our results suggest this enrichment is likely driven by five-way conserved CBSs.

The consistent underrepresentation of Alu elements (SINE) across all CBS types, despite their high copy number in the human genome, further underscores that TE co-option is not random. As previous studies reported, only a subset of TEs carries sequences amenable to CTCF binding and proper orientation, and selective pressures likely filter out insertions that disrupt essential regulatory functions (Choudhary et al. 2020, 2023).

Collectively, our results reinforce the paradigm that TEs are not merely genomic parasites but key drivers of structural and regulatory evolution. By comparing TE contributions to CBSs across conservation levels and genomic contexts, we support the model that different TE classes may have shaped the 3D genome through distinct evolutionary trajectories.

Potential contribution of CBS evolution to human evolution

CBS turnover may have influenced gene regulatory networks linked to cognitive and neuronal traits that distinguish humans from other primates. We identified several brain- or neuron-related candidate genes associated with human-specific CBSs that may play roles, including the PDE9A gene discussed above whose associated CBS shows signatures of positive selection, and several genes exhibiting enhanced enhancer–promoter interactions in humans, such as RIMS1, PRDM16, and ADCY1. ADCY1 is a member of the adenylyl cyclase family, which plays critical roles in learning, memory, and movement (Devasani and Yao 2022). Transgenic mice overexpressing ADCY1 in the forebrain exhibit enhanced long-term potentiation, improved object recognition memory, and slower extinction of contextual memory (Wang et al. 2004). RIMS1 encodes a RIM protein, a key component of presynaptic active zones that regulates synaptic vesicle docking and the coupling of voltage-gated calcium channels. In the hippocampal CA1 region, postsynaptic RIM is essential for NMDA receptor–mediated synaptic transmission and contributes critically to synaptic plasticity and memory formation (Wang et al. 2018; Smaga et al. 2020). PRDM16 is a key transcriptional regulator in diverse cell types (Chuikov et al. 2010). In brain development and homeostasis, PRDM16 has been shown to regulate the maintenance and proliferation of neural stem cells (Chuikov et al. 2010; Shimada et al. 2017; Chui et al. 2020), and it can regulate a temporal transcriptional program to promote progression of cortical neural progenitors (He et al. 2021).

Considering the important functions of human CBSs, mutations at these sites, particularly conserved sites in CTCF binding motifs, could change CTCF binding ability and should have important functional consequences. This is supported by the finding that CBSs harbor several GWAS signals for human complex traits or diseases, highlighting the biomedical relevance. For example, Wu et al. 2025 demonstrated that in B-lymphoblastoid cells, two human-specific nucleotide substitutions occurring at the most evolutionarily conserved positions within the CTCF binding motif contribute to the formation of human-specific CTCF loops, thereby influencing disease susceptibility. Using S-LDSC, which accounts for linkage disequilibrium and MAF biases, we detected statistically significant heritability enrichments in functional annotations overlapping CBSs for multiple trait categories, including neuropsychiatric, metabolic, hematological, immunological, and cardiopulmonary traits. These findings suggest that genetic variants within regions near human CBSs contribute significantly to the heritability of diverse complex traits.

Genetic variations in CBS could lead to diversification of TAD structures within human populations (Wu et al. 2025). Interestingly, we also identified cases of polymorphisms at conserved CTCF motif that have different allele frequencies among human populations (Fig. S9). For example, a T-to-C mutation specific to Eurasian populations created a new human CBS, which is associated with a human-specific TAD and likely contributes to the increase in the enhancer–promoter interactions within this TAD. Within this TAD, we found C14orf132 gene with human-specific high expression in brain.

Advantage and limitation

We employed a deep learning model, BPNet (Avsec et al. 2021), to computationally predict CTCF binding profiles in the reference genomes of human, chimpanzee, gorilla, macaque, and orangutan. The use of computationally predicted CBSs offers several important advantages for our study. First, deep learning–based predictions enable systematic, genome-wide identification of CBSs across multiple primate species in a consistent framework, which would be difficult to obtain from experimental assays in matched cell types and conditions, given the inconsistency in cell lines used in previous TAD and CTCF studies in primates (Schwalie et al. 2013; Eres et al. 2019). By leveraging this unified prediction strategy, we were able to directly compare the genomic distribution, conservation patterns, and TAD boundary associations of conserved and human-specific CBSs, and then link these patterns to natural selection, TE contributions, and gene expression divergence. We validated our predictions quantitatively and qualitatively using three methods. First, we validated our BPNet models using a strict chromosome-level hold-out strategy and demonstrated high correlations between predicted and observed signals in all five species (Fig. S1). Second, we compared with published CBSs obtained from 56 human cell lines (Li et al. 2013) and confirmed that over 95% of our predicted CBSs were present in more than 20 cell lines, with iPSCs and bipolar neurons among the most highly overlapped cell types (Fig. S3). Third, the features of our predicted CBSs in the human genome are comparable to empirical CBSs reported by Nanni et al. 2020, with regard to their spatial distribution and clustering pattern.

At the same time, using predicted rather than experimentally mapped CBSs imposes limitations that need to be acknowledged. Predictions may include false positives or miss context-dependent binding events, and they do not capture cell-type specificity, chromatin state variation, or dynamic binding changes across developmental stages (Kim et al. 2007; Cuddapah et al. 2009; Li et al. 2013). Thus, while computational prediction provides the breadth and cross-species comparability necessary to uncover global evolutionary and regulatory trends in CBS turnover, targeted experimental validation in relevant tissues and developmental contexts will be essential to confirm the causal roles of specific CBSs and refine estimates of their functional and evolutionary significance.

Another limitation of this study is that the datasets used for regulatory and chromatin features are derived from iPSC-based datasets, which provide a useful but incomplete representation of human cells. Because transcription, regulatory interactions, and chromatin architecture are highly cell-type specific, our analyses should be interpreted as capturing a shared baseline of 3D genome organization across primates rather than cell-type-specific programs. This framework allowed us to make consistent cross-species comparisons while minimizing confounding effects from tissue composition and developmental stage, but it also means that the functional consequences of CBS evolution may vary across cellular contexts. Future studies using matched cell-type-specific multiomic datasets across primates will be needed to determine how these patterns differ among tissues, lineages, and developmental states, and to refine the contribution of CBS evolution to human-specific regulatory innovation.

Materials and methods

BPNet prediction of CBSs and their quantitative validation

To predict genome-wide CTCF binding profiles for five primates, we utilized the deep learning model BPNet, which predicts base-resolution binding profiles of transcription factors from DNA sequences with high performance (Avsec et al. 2021). The genomes of human (hg38), chimpanzee (panTro6), gorilla (gorGor6), orangutan (ponAbe3), and macaque (rheMac10) were obtained from UCSC Genome Browser (https://hgdownload.soe.ucsc.edu/goldenPath/currentGenomes/). ChIP-seq data from lymphoblastoid cell lines for each species was used to train BPNet (bpnet train) (Schwalie et al. 2013).

To assess predictive performance and generalizability, we performed a chromosome-level hold-out validation using chromosomes 1, 8, and 9 as an independent test set. For each species, BPNet-predicted log-counts were compared with observed ChIP-seq log-counts within empirically defined peak regions on these held-out chromosomes. Model performance was quantified using Pearson correlation and Spearman rank correlation between predicted and observed log-counts for each species. The consistently high correlations across species indicate that BPNet-predicted binding profiles are in good agreement with empirical CTCF binding data and support their use for downstream analyses.

To interpret the sequence determinants learned by the models, base-level contribution scores for input sequences were computed using the bpnet contrib command. These scores quantify the relative importance of individual nucleotides in predicting CTCF binding. Task-specific motif discovery was performed using TF-Modisco (bpnet modisco-run), which aggregates contribution scores to identify interpretable sequence motifs. For each motif, a contribution weight matrix (CWM) was generated by averaging the contribution scores of all four possible nucleotides at each position. The resulting motifs were compared to the consensus mammalian CTCF binding motif (Schmidt et al. 2012), and the m0_p0 pattern was identified as the best-matching CTCF motif in all five species (Fig. S2). We then used CWM-scan to calculate motif-matching probabilities for m0_p0, which were referred to as predicted CTCF binding score and used as proxies for CTCF binding strength. To avoid biasing our dataset toward only strong, conserved binders, we retained motifs spanning a broad range of motif-matching probabilities (0.2 to 1.0). This inclusive low threshold was chosen to capture low- to medium-scoring sites, which are often associated with species-specific and evolving regulatory elements.

Spatial clustering and orientation analyses of predicted CBSs

To assess the genomic organization of predicted CBSs, we analyzed their spatial clustering and orientation patterns as previously described by Nanni et al. (2020). Briefly, we grouped CBSs into spatial clusters based on their genomic proximity using the bedtools cluster function (v2.30.0). We defined d as the maximum allowable genomic distance (in base pairs) between adjacent CBSs for them to be assigned to the same cluster. To capture organizational features across different genomic scales, we performed the clustering analysis iteratively across a broad range of d values. Specifically, we varied d in logarithmically increasing steps ranging from 4 to 2,048 kb, following the protocol described by Nanni et al. (2020). This range was selected to span local regulatory interactions to chromosome-scale domains. Within each cluster, configurations were classified according to the number of sites (mono-plets, di-plets, tri-plets, or tetra-plets) and their relative orientations. Orientation patterns were categorized as follows: (1) Same, all sites oriented in the same direction; (2) Convergent, sites oriented toward each other; (3) Divergent, sites oriented away from each other; and (4) Convergent + Divergent, a mixed pattern observed in tri-plets and tetra-plets. Strand annotations (forward: >; reverse: <) were used as input for the ctcf_analysis.py script, which summarized all orientation configurations genome-wide.

Determination of CBS conservation levels across species

To investigate the conservation of CTCF binding profiles across five primates and compare their features around TAD boundaries, we projected predicted CBSs from four nonhuman primates (chimpanzee, gorilla, orangutan, and macaque) onto the human reference genome (hg38). We obtained the pairwise whole-genome alignment chain files and filtered synteny net files (syn.net) from the UCSC Genome Browser (https://hgdownload.soe.ucsc.edu/goldenPath/currentGenomes). These synteny net files retain only syntenic alignments and exclude noncollinear regions, segmental duplications, and repeat-rich clusters that are prone to mapping artifacts. We then extracted syntenic chains from the full alignment sets using netChainSubset with the -wholeChains option, generating species-specific syntenic chain files (syn.chain) for coordinate conversion. This step restricts our coordinate mapping to syntenic regions and minimizes false discoveries of human-specific CBSs due to missing synteny. These filtered chains were used with pslMap to map CBS coordinates across species, following the pipeline in Zhu et al. (2007) (https://hgdownload.soe.ucsc.edu/admin/exe/linux.x86_64). The resulting PSL files were converted to BED format using pslToBed from the UCSC Utilities toolkit (https://hgdownload.soe.ucsc.edu/admin/exe/linux.x86_64).

Conserved CBSs were defined as human CBSs that overlapped with orthologous CBSs in at least one other species by at least one base pair, whereas human-specific CBSs were defined as human CBSs with no overlap with CBSs in the four nonhuman primates.

To assess the robustness of our human-specific CBS calls and minimize potential misclassification due to repetitive or poorly mappable regions, we performed two additional validation analyses. First, we examined whether the 762 human-specific CBSs were represented in at least one of the filtered pairwise syntenic alignments. We found that 753 of 762 human-specific CBSs were present in at least one syntenic alignment file, indicating that the majority of these loci are not due to alignment failure. Second, for each human-specific CBS, we extracted a 10 kb window centered on the site (±5 kb flanking sequence) and mapped these windows to the four nonhuman primate genomes using pslMap and the corresponding filtered human-to-primate syntenic chain files. We quantified alignment coverage for each window by summing the aligned bases across PSL records and dividing by the total 10 kb window length. Most windows mapped with high coverage to the nonhuman primate genomes, with median alignment coverage exceeding 95% in each species. The proportions of loci with ≥80% alignment coverage were 91.7% in chimpanzee, 91.5% in gorilla, 87.0% in orangutan, and 76.6% in macaque (Table S2).

Examination of CBSs clustering pattern near TAD boundaries

To characterize the genomic features of CBSs relative to TAD boundaries in humans, TAD annotations were obtained from Eres et al. (2019). For each CBS, the distance to the nearest TAD boundary was computed using bedtools closest. Similarly, the distance to the closest CBS was calculated to assess local site spacing. To evaluate CBS clustering around TAD boundaries, we applied the approach described by Kentepozidou et al. (2020), where CBSs were classified as clustered or singleton using bedtools cluster, with clusters defined based on user-specified maximum intersite distances (10 kb). TAD boundary regions were then annotated according to the highest conservation level (conserved or species-specific) of overlapping CBSs.

To assess the robustness of our observed patterns to specific parameter choices, we repeated key analyses using different parameters. For CBS conservation definition, we reclassified conserved sites using two alternative criteria: ≥50% reciprocal overlap or motif-centered matching (summit overlap within ±5 bp) between human and nonhuman orthologous CBSs. Specifically, for motif-centered matching, we identified the precise motif summit (position of maximum contribution score) for each CBS using our base-resolution BPNet predictions and defined conservation as motif summits overlapping between species overlap within a stringent ±5 bp window. For TAD boundary association, we recomputed CBS enrichment and clustering patterns using TAD boundary windows of ±25 and ±75 kb in addition to the original ±50 kb window. These alternative definitions yielded qualitatively consistent results (Figs. S4 and S5), confirming that the observed clustering patterns do not depend critically on the exact parameter choices.

Natural selection analyses of CBS evolution

PhastCons scores of 17-way primate alignments were downloaded from the UCSC Genome Browser (https://hgdownload.soe.ucsc.edu/goldenPath/currentGenomes/Homo_sapiens/phastCons17way/). Human polymorphism data from 22 African Yoruba individuals were obtained from Bergström et al. (2020). Tajima's D was calculated using VCFtools (Tajima 1989; Danecek et al. 2011) for each CBS, as well as for synonymous and nonsynonymous sites of its nearest gene and a randomly sampled 25 bp intergenic region between the CBS and its nearest gene. The nearest gene for each CBS was identified using bedtools closest. Synonymous and nonsynonymous sites were annotated using ANNOVAR (Wang et al. 2010). Genomic positions with more than 10% missing data were excluded from analysis (Bailey et al. 2025).

For the MK test and extended MK test, the polymorphism dataset was filtered and downsampled to 40 chromosomes. Human–chimpanzee (hg38-panTro6) whole-genome alignments were obtained from UCSC (https://hgdownload.soe.ucsc.edu/goldenPath/hg38/). Callable sites shared between the postfiltered polymorphism data and the divergence alignment were retained for analysis, and only polymorphic and divergent sites within these shared callable regions were included in the calculation of the fraction of adaptive fixations (α).

To control for local variation in mutation rates and genomic context, we used locally matched neutral reference regions for each CBS. For each locus, we identified the nearest gene and extracted two neutral reference classes: synonymous sites and long introns. Long introns longer than the median intron length of 1,029 bp were selected, and regions within 150 bp of splice junctions were excluded (Gazave et al. 2007; Farlow et al. 2012). Because CBSs are much shorter than the reference regions, polymorphic and divergent counts in the reference regions were length-normalized by the ratio of CBS length to reference length before summation across loci. α was estimated using the extended MK framework:

$$\alpha =1-\frac{{\mathrm{D}}_{\mathrm{ref}}{\mathrm{P}}_{\mathrm{CBS}}}{{\mathrm{D}}_{\mathrm{CBS}}{\mathrm{P}}_{\mathrm{ref}}}$$

where P_CBS and D_CBS denote the numbers of polymorphic and divergent sites in CBSs, and P_ref and D_ref denote corresponding length-normalized counts of polymorphic and divergent sites in the locally matched neutral references. Synonymous total sequence lengths were estimated using multipliers from Huber et al. (2017) and Zhen et al. (2021). The significance of differences in divergence between CBSs and neutral controls was assessed using Fisher's exact test. To reduce the influence of slightly deleterious mutations, polymorphic sites with MAF below 10% or 25% threshold were excluded (Fay et al. 2001; Charlesworth and Eyre-Walker 2008; Booker et al. 2017). We also independently estimated α using synonymous sites and long introns as alternative neutral references. The 95% confidence intervals for α were obtained by jackknife resampling across CBS loci within each group. Moreover, to determine whether the human-specific CBSs showed a significantly higher proportion of adaptive substitutions than the five-way CBSs, we performed permutation tests based on the difference in mean α between the two groups $\mathrm{\Delta }\alpha ={\overline{\alpha }}_{\mathrm{hs}}-{\overline{\alpha }}_{5-\mathrm{way}}$. Specifically, the 21,535 human CBSs were randomly partitioned into two nonoverlapping groups, matched to the sizes of the human-specific and five-way CBS sets, and the difference in mean α was recalculated over 10,000 permutations to generate an empirical null distribution. The permutation P-value was defined as the proportion of randomized Δα at least as large as the observed Δα.

TE enrichment in CBSs and robustness analyses

RepeatMasker annotations for the human genome (hg38) were downloaded from the UCSC Genome Browser. TEs were classified into four major classes: LINEs, SINEs, LTRs, and DNA transposons. Additional repeat categories, including simple repeats and satellite DNA, were also annotated but represented only a minor fraction of the human genome (Ahmad et al. 2020).

To assess the contribution of repeat elements, particularly TEs, to different CBS categories, we overlapped CBSs with annotated repeats using bedtools intersect. A CBS was considered repeat-associated if at least 40% of its length overlapped a repeat element. For each CBS category defined by conservation status and/or TAD boundary association, we counted the number of repeat-overlapping sites and calculated the proportion contributed by each repeat class.

To evaluate the robustness of the TE enrichment results, we performed two additional analyses. First, we repeated the TE overlap analysis across a range of overlap thresholds from 20% to 70% in 10% increments. Second, to control for local genomic context, mappability, and chromatin environment, we constructed matched-superpool permutations of 25-bp genomic windows centered on CTCF ChIP-seq peak summits from ENCODE experiments across multiple human cell lines (Table S1). After excluding blacklisted regions, this superpool was restricted to regions that are sequenceable, mappable, and CTCF-permissive, and each candidate window was annotated by chromosome, GC content, mappability, distance to the nearest gene, and distance to the nearest TAD boundary. For each CBS, matched null regions were sampled from the same genomic stratum defined by these features. Specifically, GC content was binned into low (≤0.3), mid (0.3 to 0.5), and high (>0.5) categories, and mappability was quantified using ENCODE-style mappability tracks and binned into low (<0.7) or high (≥0.7) categories. Distance to the nearest TAD boundary was categorized as “near” (0 to 50 kb) or “far” (>50 kb), and distance to the nearest gene was categorized as “promoter proximal” (0 to 2 kb) or “intergenic” (>2 kb). For each CBS, one 25 bp matched region was drawn without replacement from the same genomic stratum, and stratified permutation tests were performed by shuffling CBS labels within these strata 10,000 times to generate an empirical null distribution for TE-associated fractions. The empirical P-value was calculated as the proportion of permuted samples yielding a TE fraction equal to or greater than the observed value, with a +1 correction applied to both numerator and denominator to avoid zero probabilities.

Differential gene expression analyses for tandem gene pairs

RNA-seq data from the same human iPSC lines used for TAD inference were downloaded from the Gene Expression Omnibus (GEO accession GSE110471) (Pavlovic et al. 2018). Raw sequencing reads were assessed for quality using FastQC (v0.11.9), and adapter sequences and low-quality bases were trimmed using Trimmomatic (v0.39) (Bolger et al. 2014). Cleaned reads were aligned to the human reference genome (hg38) using STAR (v2.7.9a) (Dobin et al. 2013), and gene-level counts were generated with HTSeq (v0.9.1) (Anders et al. 2015) based on Ensembl gene annotations. Read counts were normalized and differential expression analysis was performed using the DESeq2 R package (v1.20.0) (Love et al. 2014). Genes with a false discovery rate (FDR) q-value <0.01 and |log₂ fold change| > 1 were considered DE. Tandem gene pairs were identified using MCScanX, based on genomic proximity and shared synteny within the human genome (Wang et al. 2012).

Calculation of contact frequency within TAD regions

The contact frequency for whole TADs or enhancer–promoter pairs of candidate genes was calculated using Pgltools (www.github.com/billgreenwald/pgltools) (Greenwald et al. 2017). TADs, contacts, and reads mapping files for humans and chimpanzees were obtained from Eres et al. (2019). The enhancer–promoter pairs for candidate genes were derived from Li et al. (2023). Briefly, we converted the contacts and reads mapping files to paired-genomic-loci (PGL) files with the pgltools formatbedpe and samTopgl, respectively. We then employed coverage to quantify the number of reads from the mapping file that overlap each entry in the contact file. The contact files with mapping counts were intersected against enhancer–promoter pairs of candidate genes using the pgltools intersect.

Heritability enrichment analyses

We used S-LDSC (http://github.com/bulik/ldsc) (Finucane et al. 2015) to evaluate whether CBSs were enriched for the heritability of complex traits relative to their genomic proportion. CBS coordinates were lifted from hg38 to hg19 using liftOver, and annotation files were generated with make annot.py. LD scores were computed with ldsc.py using SNPs from the HapMap3 reference panel (https://data.broadinstitute.org/alkesgroup/LDSCORE/w_hm3.snplist.bz2).

Because the number of polymorphic sites in CBSs is limited, we aggregated all human CBSs and included SNPs within ±50 kb of each CBS coordinate. We obtained summary statistics of 176 GWAS studies curated in public repositories (https://zenodo.org/records/7768714, accessed 2025 March 25) (Finucane et al. 2015; Gazal et al. 2017; Hujoel et al. 2019). S-LDSC was performed for all 176 datasets, with Fig. 6 showing 42 traits representing cardiopulmonary, hematological, immunological, metabolic, and neuropsychiatric categories, following the trait-grouping scheme of Wu et al. (2025). When multiple datasets were available for the same trait, we selected the study with the largest effective sample size for visualization.

To assess whether heritability enrichment differed by evolutionary conservation, we repeated the analysis separately for human-specific CBSs and five-way conserved CBSs. This stratification enabled comparison of lineage-specific and ancestral regulatory contributions, although we note that the human-specific subset contains fewer polymorphic sites and therefore has lower statistical power, and five-way conserved CBSs represent a large proportion of human CBSs and thus may drive the pattern observed in all human CBSs. We also performed a sensitivity analysis using a narrower 25 kb flanking window to evaluate the robustness of our window choice.

Author contributions

Y.Z. and K.Z. designed the study. K.Z. and J.Z. performed bioinformatics analyses. K.Z. and Y.Z. wrote the manuscript.

Supplementary material

Supplementary material is available at Molecular Biology and Evolution online.

Funding

This work was supported by the Zhejiang Provincial Natural Science Foundation (LQ20C040001 to K.Z.), the National Natural Science Foundation of China (32100484 to K.Z. and 32370662 to Y.Z.), the “Pioneer” and “Leading Goose” R&D Program of Zhejiang (2024SSYS0032), and the Research Center for Industries of the Future (RCIF) at Westlake University.

Data availability

Public datasets were curated and used for this work (sources summarized in Table S13), and no additional sequencing data were produced. The code for α calculation and repeat elements analyses is available at https://github.com/kanglizhu/-calculation-and-repeat-elements-analyses.