The hierarchical folding dynamics of topologically associating domains during early embryo development

Abstract

Background

Recent research has indicated a close connection between the three-dimensional (3D) structure of chromatin and early embryo development, with precise higher-order chromatin folding playing a significant role in mediating gene expression. However, the specific role of 3D genomic hierarchical structure and its dynamics in early embryo development remains largely unknown.

Results

In this study, we examined the hierarchical topological association domain (TAD) during early embryo development and its relationship with zygotic gene activation (ZGA), gene expression, and chromatin accessibility to gain a better understanding of the dynamics of TAD nesting levels during this developmental stage. Our findings show that ZGA precedes the establishment of hierarchical TAD, leading to widespread gene expression, an increase in the percentage of high-level TAD structures, and enhanced chromatin accessibility at higher hierarchical levels. Additionally, we utilized a deep neural network to investigate the formation of TAD boundaries and found that histone H3 lysine 4 trimethylation (H3K4me3) and histone H3 lysine trimethylation (H3K27me3) are key features in the establishment of TAD boundaries. Furthermore, we observed heterogeneous dynamics of hierarchical TAD among different species.

Conclusions

Overall, our study sheds light on the folding dynamics of hierarchical TADs during early embryo development and underscores their close relationship with transcriptional programs.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12915-025-02259-y.

Background

DNA is located in various regions of the cell nucleus and is tightly wound and folded within the 10-micron nucleus, resulting in a highly organized and hierarchical chromatin structure in three-dimensional space. At the million-base scale, chromatin is divided into A and B compartments, with the A compartment exhibiting higher transcriptional activity compared to the B compartment. On the scale of megabases, CCCTC binding factor (CTCF) and cohesin facilitate loop extrusion to form topologically associating domains (TADs) and chromatin loops, which regulate gene expression through long-range enhancer-promoter interactions [1]. The acquisition of cell identity or phenotype relies on the combination of genetic and epigenetic information within the cell. However, local microenvironmental changes, such as histone modifications and chromatin accessibility, may not fully elucidate the mechanism of promoter-enhancer interactions over long distances. The dynamics of intracellular chromatin conformation throughout development and its role in mediating long-distance interactions between promoters and enhancers to promote development are still unknown. This is crucial for a deeper understanding of cell differentiation, cell fate, and reprogramming. Adam Burton et al. proposed that chromatin structure may regulate cell fate at two levels: locally, by controlling the expression of specific genes (e.g., transcription factors, TFs), and globally, by creating a chromatin environment conducive to cell fate allocation through chromatin and nuclear organization dynamics [2].

Early embryo development is a series of highly ordered processes that are restricted by precise temporal and spatial control of gene expression. In recent years, low-input sequencing approaches have been developed [3, 4], leading to a gradual revelation of the spatial regulation of early embryo developmental stages. In mouse, TADs and loops are observable in the zygote, exhibiting weak insulation [3, 4]. TADs and compartments gradually consolidate after zygotic gene activation (ZGA), the pattern consistent with Drosophila. In human, pig, drosophila, zebrafish, medaka, and xenopus tropicalis, chromatin domains are consistently established at ZGA and strengthen over time [5–10]. However, ZGA is not essential for the establishment of TAD in mouse, which differs from human [6]. While the establishment of TAD coincides with ZGA in many species, other features exhibit more significant variation. For instance, the establishment of compartments and loops differs notably. Compartments are established at ZGA in mouse, drosophila, xenopus, and medaka [3, 4, 8–10]. However, in human, compartmentalization is established by the morula stage [6], whereas in zebrafish, it occurs after ZGA [9]. Flyamer et al. initially identified weakly enriched loops in differentiated mouse cells through single-cell high-throughput chromosome conformation capture (Hi-C) analysis [11]. Then, loops are detected in stage 11 in xenopus [5], during mid-gastrulation in medaka [10], and after mid-gastrulation in zebrafish [9]. Conversely, drosophila does not exhibit loops and instead preferentially establishes remote cis-regulatory modules (CRMs) through the pioneer transcription factor Zelda before the establishment of TADs to regulate the transcriptional program [12].

TAD boundaries are enriched with multiple factors, including histone modifications, transcription start sites, and housekeeping genes [13]. Proper higher-order chromatin folding can significantly impact gene expression [14]. Disruption of TADs and abnormal fusion of TAD boundaries can result in various developmental disorders and diseases [15–19]. It has been reported that the destruction of TAD boundaries can cause limb malformations in development [15, 16]. TAD, the basic structural unit of chromatin, changes in the number and size of TAD change during cell differentiation [20]. Hierarchically nested structures within TADs, termed “metaTAD” [21] or “subTAD” [22] have been identified. Currently, there are over 13 methods developed for the identification of hierarchical TAD structures [13, 22–42]. These innovative techniques play a crucial role in unraveling the complex organization of chromatin domains and providing deeper insights into genome architecture. Numerous researchers have shown that hierarchical TADs are associated with genetic, epigenomic, and gene expression activities [21, 26, 30, 32]. Fraser et al. found that rearrangement of metaTADs during cell differentiation will affect the transcriptional state [21]. Additionally, Bintu et al. demonstrated the existence of TAD and subTAD-like structures in single cells using multiplexed super-resolution fluorescence in situ hybridization imaging [36]. Ye et al. developed the HiCS algorithm to identify hierarchical chromatin domain structures from single-cell Hi-C maps and verified the existence of hierarchical TADs in embryonic stem cells, with increased CTCF, cohesin, chromatin accessibility, housekeeping genes, and gene expression activity at higher levels of hierarchical TAD boundaries [13]. However, the dynamic pattern and formation mechanism of hierarchical TADs during early embryo development are still poorly understood, and it is unclear whether hierarchical TADs differ between species.

In this study, we investigated the hierarchical TAD structures at various stages of early embryo development, as illustrated in Fig. 1A. Initially, we utilized the OnTAD [26] to identify hierarchical TADs and determine the hierarchical levels of TAD boundaries. Our analysis revealed folding dynamics in the hierarchical TAD structure and transformations in the hierarchical level of TAD boundaries during development. Previous research has demonstrated that CTCF, cohesin, H3K4me3, histone H3 lysine 36 trimethylation (H3K36me3), transcription start sites (TSSs), and housekeeping genes, which are enriched in TAD boundaries, may contribute to the formation of mammalian chromatin domains [13]. Thus, we incorporated TAD boundary with different epigenetic features (chromatin accessibility, H3K4me3, histone H3lysine9 trimethylation (H3K9me3), and H3K27me3) as inputs to construct a deep neural network to explore the formation mechanism of hierarchical TAD. Additionally, we examined the temporal sequence of the ZGA program and TAD structure establishment. To determine the relationship between hierarchical TAD folding dynamics and gene expression during early embryo development, we employed a method developed in our previous research, the TAD hierarchical score (TH score), to score the TAD nesting level of the protein-coding gene. Furthermore, we explored the heterogeneity between early human and mouse embryos at the level of the hierarchical TAD.

Fig. 1.

Overview of the hierarchical TAD analysis process during early embryo development. A The schematic overview for analyzing TADs folding dynamics during early embryo development. B The pattern of hierarchical levels of TAD and TAD boundary for mESC. C The correlation of the hierarchical TADs between different datasets

Results

Optimal resolution of resolving hierarchical TADs during early embryo development

To examine the consistency of our two mouse early embryo datasets, we conducted a clustering analysis of the Hi-C matrices from the two datasets. Due to limitations in the high-throughput chromosome conformation capture (Hi-C) experiments, the results revealed a noticeable batch effect between the datasets from different laboratories (Additional File 1: Fig. S1 A). Consequently, we chose one dataset (GSE82185) for hierarchical TAD analysis during early embryonic development and cross-validated it with the other dataset (CRA000108).

Subsequently, we investigated the chromatin interactions of mouse embryonic stem cells (mESCs) and observed a typical TAD nesting pattern (Fig. 1B, Additional File 1: Fig. S1B). The hierarchical levels of the TADs are designated from outer to inner. As illustrated in Fig. 1B, the TAD enclosed by the blue dashed line is designated as level 1, the TAD enclosed by the purple dashed line is designated as level 2, and the TAD enclosed by the green dashed line is designated as level 3. The TAD boundary is defined as the maximum number of TAD boundaries shared on the left or right side of the TAD. As depicted in (Fig. 1B), if one side of the boundary is shared by two TADs, and the other side has two or fewer TADs, the level of this boundary is level 2. Among the 13 hierarchical TAD identification tools evaluated by Xu et al., OnTAD stands out for its commendable performance in terms of computational speed and robustness in hierarchical TAD recognition [43]. Therefore, we then utilized the OnTAD [26] to investigate the dynamics of the hierarchical TADs during early embryo development. It is worth noting that the resolution of the analysis can impact the hierarchical structure of TADs. Therefore, we categorized genes based on the hierarchical TAD in which they were located. Our findings indicated that the hierarchical TAD in which the genes are situated can be effectively distinguished at a resolution of 20 kb (Additional File 1: Fig. S1 C, D). Hence, the 20kb resolution appears to be suitable for investigating the mechanism of hierarchical TAD establishment in early embryo development. Additionally, to assess the similarity of hierarchical TADs, we employed the SSIM algorithm [44] to analyze the consistency of hierarchical TADs in the two mouse early embryo datasets. The results demonstrated a high correlation in the TAD hierarchies between the two datasets, and the TAD hierarchies of the same cell stages could be well-clustered (Fig. 1C). These findings validate the reliability of our identified results.

The landscape of hierarchical TAD dynamics in early embryo development

In this section, we examined the dynamics of hierarchical TAD in various stages of early embryo development. Initially, we observed a decrease in the number of TADs at the lowest level (level 1), accompanied by a gradual increase in the number of TADs at higher levels (level ≥ 2) (Fig. 2A; Additional File 1: Fig. S2 A, B, C; *P < 0.05, **P < 0.01, ***P < 0.001, chi-square test calculates significance). This suggests a shift in the folding dynamics of TAD during early embryo development. Subsequently, we computed the hierarchical level of TAD boundaries and analyzed the dynamics of TAD boundary levels in adjacent developmental stages (Fig. 2B; Additional File 1: Fig. S2D, E, F). Our analysis revealed that, in mouse embryo development, high-level TADs (level ≥ 3) were derived from low-level TADs in previous cell stages, as opposed to newly gained TADs (S2D, F; *P < 0.05, **P < 0.01, ***P < 0.001, chi-square test calculates significance). Then, to investigate the relationship between gene expression and the hierarchical level of TAD boundaries, we calculated the expression distribution of enriched genes on TAD boundaries. It was found that as the hierarchical level of TAD boundaries increases, gene expression levels also increase (Additional File 1: Fig. S2G; Wilcoxon-test calculates significance). It indicates a positive correlation between TAD boundary hierarchy and gene expression. Notably, the Zscan4d gene exhibited specific expression at the early 2cell stage and played a significant role in early embryo development (Additional File 1: Fig. S2H). To provide a comprehensive understanding of the changes in TAD and hierarchical TAD boundaries during early embryo development, we depicted the hierarchical TAD dynamics of the Zscan4d gene (Fig. 2C). Our findings showed that the hierarchical levels of TADs and TAD boundaries of the Zscan4d gene progressively increased from the zygote to mESC. However, Zscan4d was silenced after the late 2cell stage, possibly due to the gradual closure of chromatin accessibility and high level H3K27me3 signal (Fig. 2C; Additional File 1: Fig. S2H).

Fig. 2.

The landscape of the hierarchical TADs in early embryo development. A The dynamics of the hierarchical TADs dynamics during early embryo development (*P < 0.05, **P < 0.01, ***P < 0.001, chi-square test calculates significance). B The transformation of TAD boundaries in adjacent cell stages. C The landscape of hierarchical TADs of Zscan4d. D The deep neural network for predicting TAD boundaries. E The ROC curve of the deep neural network, the area under the curve is about 0.91. Individual data values are presented in Additional File 2

The mechanism of TAD boundary establishment during early embryo development is a topic of interest. Previous studies have suggested that Transcription factor (TFs), chromatin regulators, and histone modifications are significantly either enriched or absent in domain boundaries [13]. This suggests that histone modifications and regulatory factors may contribute to the formation of TAD boundaries. We analyzed the enrichment levels of epigenetic signals on TAD boundaries and observed that as the hierarchical level of TAD boundaries increased, chromatin accessibility and H3K4me3 signals also rose, while the signals of H3K9me3 and H3K27me3 gradually declined (Additional File 1: Fig. S2I). This may suggest that chromatin accessibility, H3K4me3, H3K9me3, and H3K27me3 make distinct contributions to the state of TAD boundaries. To explore this further, we integrated TAD boundary, transposase accessible chromatin with high-throughput sequencing (ATAC-seq), and chromatin immunoprecipitation sequencing (ChIP-seq) (H3K4me3, H3K9me3, and H3K27me3) features and constructed a deep neural network (Fig. 2D). We then used chromatin accessibility and histone modification signals enriched at TAD boundary sites to predict hierarchical TAD levels at the next cell stage. Specifically, we predict the hierarchical levels of TAD boundaries in the subsequent cell stage by characterizing the conserved TAD boundaries that exhibit raised and reduced hierarchical levels, along with signals of chromatin accessibility and histone modifications in the current cell stage. Model performance illustrated by receiver operating characteristic (ROC) (Fig. 2E). The model achieved an an AUC (area under ROC curve) of 0.91, an accuracy (ACC) of 0.88, and an F1-score of approximately 0.86 (Fig. 2E). When we only input the TAD boundary, there was a decrease in ACC, AUC, and F1-score (ACC: 0.79; AUC: 0.86; F1-score: 0.74) (Additional File 1: Fig. S2 J). It indicates that epigenetic signals contribute to the establishment of TAD boundaries. In order to investigate the epigenetic factors that play a pivotal role in the formation of TAD boundaries, we introduce individual epigenetic factors into the model to assess their impact on predictive outcomes. This approach enables us to gain insights into how specific epigenetic markers influence the delineation of TAD boundaries and enhance our understanding of the regulatory mechanisms governing chromatin organization. The results indicate a significant increase in the model’s AUC value when incorporating H3K4me3 (ACC: 0.84; AUC: 0.89; F1-score: 0.84) and H3K27me3 (ACC: 0.87; AUC: 0.90; F1-score: 0.85) (Additional File 1: Fig. S2 K). This elevation suggests that the inclusion of these epigenetic markers substantially enhances the predictive performance of the model. Such findings underscore the importance of these specific histone modifications in shaping the accuracy and effectiveness of the predictive model in delineating TAD boundaries. Additionally, we performed an ablation study to identify the key features in the establishment of TAD boundaries (Additional File 1: Fig. S2L). When H3 K4me3 or H3K27me3 is removed, the model’s accuracy only slightly decreases or remains unchanged (Additional File 1: Fig. S2L, M). Validation on the dataset of Ke et al. yielded consistent results (Additional File 1: Fig. S2 N, O, P, Q). We speculate that this minimal effect of ablating these histone signals could be attributed to potential information redundancy in the training process, minimizing the impact of the ablation experiment. Therefore, we further assessed the correlation between input features and predicted labels. Spearman correlation analysis showed a significant correlation between H3K4me3, H3K27me3, and the subsequent status of TAD boundaries (Additional File 1: Fig. S2S, t-test, p < 0.05). Additionally, we constructed logistic plots to quantify the contributions of H3K4me3 and H3K27me3 in the model’s predictive process, with results aligning as anticipated (Additional File 1: Fig. S2 T). The hierarchical levels of TAD and TAD boundaries undergo significant changes during early embryo development. Our findings indicate that not all epigenetic signals enriched at the TAD boundary are crucial for its formation and that H3K4me3 and H3K27me3 are pivotal in the formation and alteration of hierarchical TAD boundaries. In addition, we utilized SHAP to quantify the contribution of each variable in the prediction, and the SHAP results showed that H3K4me3, ATAC signal, and H3K27me3 contributed significantly to the raise and reduce of TAD boundary (Additional File 1: Fig. S2U). This aligns with results from our ablation and single-variable experiments, as these three approaches evaluate model performance from distinct perspectives.

Zygote genome activation precedes the establishment of the TAD structure

ZGA is a critical event during early embryo development. It has been reported that the TAD establishment of the mouse does not depend on ZGA [4]. However, the precise temporal relationship between ZGA and the hierarchical organization of TADs remains to be elucidated. In our study, we examined the distribution of hierarchical TADs containing mouse ZGA and non-ZGA genes (Fig. 3A; Additional File 1: Fig. S3 A). When a gene was present in multiple TADs, we considered only the highest-level TAD for our analysis. We observed that the number of ZGA genes in high-level TADs decreases until the late 2cell stage, after which there is a gradual increase in high-level TAD formation (Fig. 3A; Additional File 1: Fig. S3 A, B, C; P < 0.05, chi-square test calculates significance). This trend experienced a transient dip at the 8-cell stage before continuing to rise in the inner cell mass (ICM) (P < 0.05, chi-square test calculates significance) ((Fig. 3A). Conversely, the number of high-level TADs among non-ZGA genes consistently increased prior to implantation (Additional File 1: Fig. S3B, C; P < 0.05, chi-square test calculates significance), suggesting that ZGA precedes hierarchical TAD establishment.

Fig. 3.

Temporal sequence of ZGA and TAD establishment. A The number of mouse ZGA genes enriched in hierarchical TADs. B The landscape of hierarchical TADs dynamics of the Sox2 gene. C The pattern of ZGA gene expression in early mouse embryo development. D The number of up-regulated ZGA genes enriched in hierarchical TAD. E The number of down-regulated ZGA genes enriched in hierarchical TAD. Individual data values are presented in Additional File 2

To further explore the dynamics of TAD organization, we mapped the landscape of hierarchical TAD changes at ZGA gene loci. For example, the TAD nesting degree of Sox2 (Additional File 1: Fig. S3D), a crucial ZGA gene implicated in mouse embryonic and limb development, significantly increased at the late 2cell stage (Fig. 3B). Additionally, the ZGA process involves both upregulated and downregulated genes. To discern the differences in TAD structure associated with these gene categories, we analyzed RNA-sequencing (RNA-seq) data from preimplantation embryonic development using fuzzy c-means clustering (FCM) [45, 46] (Fig. 3C). We classified the expressed genes into eight clusters and noted the expression patterns after early 2cell. Genes in cluster 4 were categorized as up-regulated, while those in cluster 2 were deemed down-regulated (Fig. 3C). Subsequent analysis of TAD levels in these genes revealed that up-regulated genes possess a higher number of hierarchical TADs compared to down-regulated genes (Fig. 3D, E; Additional File 1: Fig. S3D, E, F, G, H). Collectively, these findings indicate that TAD establishment in early mouse embryos occurs subsequent to ZGA. While TAD formation is not contingent upon ZGA, there is a clear association between high gene activity and TAD establishment. Our research contributes additional insights into the necessity of ZGA for the development of TAD structure.

The effect of hierarchical TAD on gene expression during early embryo development

We aimed to investigate the impact of hierarchical TAD on gene expression in early embryo development. To assess the TAD nesting level of each protein-coding gene in a simplified manner, we utilized the TH score method, as previously described [47] (Fig. 4A; Additional File 1: Fig. S4 A). Based on the changes in TH score from zygote to mouse embryo stem cells (mESC), we identified four gene patterns: fluctuate, raised, reduced, and unchanged (Fig. 4B; Additional File 1: Fig. S4B). Through Gene Ontology (GO) enrichment analysis, we discovered that genes with fluctuating TH scores (the pattern of “fluctuate”) were enriched in pathways such as protein localization and transcription factor binding during early embryo stages (Fig. 4C; Additional File 1: Fig. S4 C). This suggests the involvement of genes related to chromatin structure in the establishment of the hierarchical TAD. Additionally, we observed uniformly elevated expression of genes encoding structural proteins associated with TAD formation at late 2cell (Additional File 1: Fig. S4D–J). Subsequently, we obtained ATAC-seq data from early embryos and used HOMER [48] to scan the motif of CTCF. Our analysis revealed a rise in the number of ATAC-seq peaks where CTCF is located in late 2cell (Fig. 4D). Comparing the proportion of TAD levels of ZGA genes, we observed a significant increase in high-level (level ≥ 3) TADs compared to the stage before ZGA (Fig. 4E; *P < 0.05, **P < 0.01, ***P < 0.001, Wilcoxon-test calculates significance). TAD boundaries have been reported to be enriched with active epigenetic signals [26]. To explore changes in chromatin accessibility at different levels of TAD boundaries, we analyzed the chromatin accessibility signal per 5kb bin located at 0.5 Mb upstream and downstream at the center of the TAD boundary. Our analysis indicated that mid and high-level boundaries (level ≥ 2) were more enriched with chromatin accessibility signal compared to low-level boundaries (level 1), demonstrating differential epigenetic activity during early embryo development, particularly before and after ZGA (Fig. 4F, Additional File 1: Fig. S4I).

Fig. 4.

The effect of hierarchical TAD on gene expression in early mouse embryo development. A The dynamics of the TH score of protein-coding genes during early mouse embryo development. B The bar plot shows the number of four gene patterns (fluctuate, raised, reduced, and unchanged) based on the TH score from zygote to mESC, respectively. C GO enrichment analysis of gene pattern (fluctuate) from zygote to mESC. D The number of ATAC-seq peaks where CTCF is located. E Percentage of different TAD levels in ZGA genes before and after the ZGA program. F Chromatin accessibility of TAD boundaries at different levels during early mouse embryo development. Individual data values are presented in Additional File 2

In summary, our analysis suggests that the dynamics of hierarchical TAD folding are closely linked to gene expression and chromatin accessibility during early embryo development, consistent with previous studies [13].

Species heterogeneity of hierarchical TAD dynamics in the early embryo

Previous studies have reported significant differences in TAD establishment between humans and mouse, with human TAD establishment relying on ZGA [6]. Despite previous evidence of TAD conservation across species [49], the existence of species-specific variations in hierarchical TAD structure formation remains unknown. Accordingly, we identified hierarchical TADs in early human embryos, which exhibited a similar pattern to early mouse embryos: a decrease in the number of lower-level TADs and an increase in higher-level TADs (Additional File 1: Fig. S5 A,B; P < 0.05, chi-square test calculates significance). Next, we investigated the transfer of high-level TAD boundaries (level ≥ 4) during the early human embryo stage. Our findings revealed that, in contrast to mouse embryos, the majority of high-level TAD boundaries were newly acquired (Additional File 1: Fig. S5 C, D, E, F; P < 0.05, chi-square test calculates significance). Additionally, we determined the homologous genes shared by humans and mouse during ZGA, including the Nanog/NANOG gene, which plays a significant role in early embryo development. Consequently, we present the hierarchical TAD structure of Nanog/NANOG gene in early human and mouse embryos, respectively (Fig. 5A, B). The results demonstrated a decrease in the TAD level of the Nanog in mouse following ZGA (Fig. 5A), whereas in humans (NANOG), the TAD level increased (Fig. 5B). To ensure the reliability of our findings, we additionally present the hierarchical TAD dynamics of the Nanog/NANOG gene at 10-kb (Additional File 1: Fig. S5G, H) and 40-kb resolution (Additional File 1: Fig. S5I, J). Significantly, both analyses exhibited a consistent pattern. Furthermore, we assessed the proportion of hierarchical TADs occupied by human ZGA and non-ZGA genes. In order to elucidate nuanced disparities, we classified the TAD hierarchy in the human and murine genomes into four distinct levels (Additional File 1: Fig. S5 K). Interestingly, unlike early mouse embryos, the ZGA genes occupied more of the mid and high levels of TAD (level ≥ 2) during ZGA (Fig. 5C; Additional File 1: Fig. S5L). Additionally, after ZGA, the proportion of high-level TADs (level ≥ 4) decreased and rose again prior to implantation in mouse and human early embryo (Fig. 5C, Additional File 1: Fig. S5L).

Fig. 5.

The differences in the early embryo between human and mouse in hierarchical TADs. A The landscape of hierarchical TADs of the Nanog genes during early mouse embryo development. B The landscape of hierarchical TADs of the NANOG genes during early human embryo development. C The number of human ZGA genes located in different levels of TAD. D The number of human up-regulated ZGA genes located in hierarchical TAD. E The number of human down-regulated ZGA genes enriched in hierarchical TAD. Individual data values are presented in Additional File 2

By analyzing the timing of gene expression, we identified ZGA genes that were upregulated and downregulated during the early development of human embryos, based on their expression patterns (Additional File 1: Fig. S4M). In line with observations in mouse, the proportion of high-level TADs (level ≥ 4) in human upregulated ZGA genes also decreased during the ZGA period (Fig. 5D, Additional File 1: Fig. S5 N, O, P, Q, R; P < 0.05; chi-square test calculates significance). It suggests that the establishment of high-level TADs of upregulated ZGA genes does not occur immediately after ZGA but is re-established during the blastocyst stage before implantation. However, for ZGA down-regulated genes, the dynamics of hierarchical TAD levels in human and mouse showed differences (Fig. 5E, Additional File 1: Fig. S5S, T, U, V, W; P < 0.05; chi-square test calculates significance). Specifically, human high-level TADs (level ≥ 4) remained unchanged after ZGA, while the proportion of high-level TADs (level ≥ 4) increased by about 11.1% after ZGA in mouse (P < 0.05; chi-square test calculates significance) (Fig. 5E, Additional File 1: Fig. S5 T, U, V, W; P < 0.05; chi-square test calculates significance). A recent study has shown that chromatin accessibility, enrichment of housekeeping genes, and gene expression exhibit a progressive increase with TAD levels [13]. This suggests a positive correlation between high transcriptional activity and the likelihood of TAD boundary formation. We noticed that high-level TADs of up-regulated genes increased at 8cell while down-regulated genes decreased in human early embryos (Fig. 5E, Additional File 1: Fig. S5 T, U, V, W; P < 0.05; chi-square test calculates significance), which indicates that the high-level hierarchical TAD establishment relies on high transcriptional activity. These results indicate a dependence of high-level hierarchical TAD establishment on high transcriptional activity. However, this was not observed in early mouse embryos. Thus, these findings indicate that the hierarchical folding dynamics of TADs exhibit specificity at both species and cellular levels.

TAD boundaries are mostly conserved across species. To evaluate the hierarchical TAD boundaries whether conseved in human and mouse, we analyzed conserved TAD boundaries across cell stages in both species using Tcbf tools [50]. We found that the majority of TAD boundaries are shared between human and mouse, while a subset is species-specific (Additional File 1: Fig. S5X). Additionally, we mapped TAD boundaries in the human Chr6:7.0–8.5 Mb region (containing the NANOG gene) and the corresponding mouse genomic region. Consistently, most boundaries remained conserved across developmental stages between the two species (Additional File 1: Fig. S5Y).

As a conserved feature in vertebrates, the core mechanism of TAD boundaries-CTCF-dependent insulation-has been evolutionarily retained. However, the longer gestation period and complex developmental programs in human embryos likely drove the evolution of more species-specific TADs in human versus mouse, enabling specialized functional regulation unique to each species. Post-ZGA human embryos deploy transcriptionally active L1 retrotransposons to recruit RNA Pol II at TAD boundaries-a mechanism minimally used in mouse [51]. Primate-specific innovations include 499 human brain TADs with mutation-enriched boundaries directing subplate layer expansion [52], and high-altitude zokors’ chromosomal inversions that fuse TADs to cluster hypoxia-response genes for coordinated expression [53]. Additionally, CTCF and cohesin mediate TAD boundary formation across species, but developmental timelines diverge between mice and humans. Mouse sperm retain TAD structures, and embryos assemble functional TADs by the early 2cell using maternal CTCF/RNAs without requiring ZGA [3, 4]. In contrast, human TAD formation is ZGA-dependent, initiating at the 8-cell stage. Human sperm lack TADs and CTCF, which surges ZGA to establish boundaries [6, 51] (Additional File 1: Fig. S5X). Consequently, pre-ZGA human embryos exhibit ambiguous TAD organization compared to mouse. Thus, the conservation of TAD boundaries across species requires further investigation.

Discussion

In this study, we explored the hierarchical structure of TADs and calculated changes in TAD boundaries during early embryo development. A deep neural network was constructed to predict the formation of TAD boundaries, demonstrating that H3K4me3 and H3K27me3 play crucial roles in establishing TAD boundaries. Furthermore, the temporal order of ZGA and the formation of hierarchical TADs were also explored. The findings indicate that ZGA precedes the establishment of hierarchical TADs. TH scores were used to distinguish between cancer and normal samples, as they have been demonstrated to effectively capture cellular heterogeneity. The analysis using TH scores indicates a correlation between genes exhibiting altered levels of hierarchical TADs and chromatin structure. Our study suggests that hierarchical TAD dynamics are widespread during early embryo development, resulting in substantial changes in gene expression that potentially contribute to cell differentiation and development. Additionally, we provide evidence of the close association between hierarchical TAD folding, gene expression, and chromatin status through the examination of developmental gene localization and chromatin accessibility across genomic regions.

After fertilization, the totipotent zygote inherits two sets of genomes containing distinct epigenetic information from both the oocyte and sperm, and then undergoes significant epigenetic reprogramming. Although high-resolution sequencing technology has motivated scientists to explore the transcriptional regulatory networks in early human embryo development, several unresolved questions persist, necessitating further research to unveil the underlying molecular mechanisms. Specifically, what factors govern the reprogramming of chromatin structure during early embryo development? Our preliminary investigation, using a deep neural network, examined the mechanism underlying TAD boundary formation and revealed its association with H3K4me3 and H3K27me3. However, we acknowledge the limitations of our study, as we were unable to explore the precise formation process. The lack of explainability in artificial intelligence poses a significant challenge when applied to the field of biology. Going forward, greater efforts are necessary to unravel the enigmatic nature of this complex phenomenon.

According to Marcelo Nollmann et al., remote cis-regulatory modules (CRMs) were established in Drosophila through the pioneer transcription factor Zelda to regulate gene expression before the establishment of TAD [12]. Our study demonstrates a decrease in the proportion of high-level TADs during ZGA, indicating a weakening of the typical hierarchical TAD during this process. Are there sub-TAD-mediated long-range enhancer-promoter interactions in early mammalian embryos to ensure proper activation of the transcriptional program? Wang et al. discovered that chromatin transcription was active in spermatocytes during the pachytene stage in rhesus monkeys, despite the absence of TAD and A/B compartment, but a refined A/B compartment was present. This suggests that local interactions between active regions might regulate the transcriptional program during early embryo development [54]. However, it is unclear whether localized interaction between adjacent active sites is similar to CRM centers in Drosophila. Our study is limited to highlighting the species heterogeneity of hierarchical TAD folding dynamics. Further investigation is needed to determine if the chromatin spatial interaction patterns differ between early human or mouse embryos and the localized interactions observed in rhesus monkeys.

Conclusions

In our study, we investigated the dynamics of TADs during early embryonic development and examined the mechanisms underlying the formation of hierarchical TAD boundaries. Our results revealed that the presence of H3K4me3 and H3K27me3 plays a crucial role in the establishment and modification of hierarchical TAD boundaries. We observed that the activation of ZGA precedes the formation of hierarchical TADs. Furthermore, we found that hierarchical TAD structures dynamically regulate gene expression and determine cell function. Additionally, our research demonstrated the presence of species heterogeneity in hierarchical TADs.

Methods

Data source

Early human and mouse embryo data were manually collected from the European Nucleotide Archive (ENA; https://www.ebi.ac.uk/ena/browser/search) and the NCBI Gene Expression Omnibus (GEO; https://www.ncbi.nlm.nih.gov/geo/) (Barrett et al. 2013). Hi-C data of 2 early mouse embryos and 1 human early embryo were obtained from Du et al., Ke et al., and Chen et al. (details in Additional File 3: Table S1) [3, 4, 6]. The assay for ATAC-seq for chromatin accessibility was obtained from Wu et al. [55]. The processed RNA-seq and ChIP-seq (H3K4me3, H3K9me3, H3K27me3) data were downloaded from the GEO database [55, 56]. Our study was based on the mm10 or hg19 genome reference. The data sets aligned to mm9 were converted to mm10 genome reference by CrossMap [57].

Data processing

Hi-C data processing

The paired-end reads were processed using HiC-Pro (v3.1.0) (https://github.com/nservant/HiC-Pro) and the contact matrices were generated at resolutions of 10-kb, 20-kb, 40-kb, and 100-kb and normalized using the iterative correction and eigenvector decomposition (ICE). The command is “HiC-Pro -i fastqfile -o hic_result -c config.txt”. Then, we converted the format of contact matrices using the HiCExplorer (version 3.5.1) command “hicConvertFormat -matrices –inputFormat hicpro -bedFileHicpro -outFileName –outputFormat h5” for calculating and visualizing [58].

ATAC-seq data processing

The quality of the raw sequencing data is evaluated by FastQC (http://www.bioinformatics.babraham.ac.uk/projects/fastqc), and low-quality read ends and adapter sequences are filtered and removed using Trim Galore! (https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/). ATAC-seq reads were aligned to the mm10 reference genome after cutting adaptors using bowtie2 with default parameters. The alignment files were processed and merged using SAMtools. PCR duplicates were removed using sambamba tools (https://github.com/biod/sambamba). Peaks were called by MACS2 [59] with “- -shift 100- -extsize 200- -nomodel” options. For downstream analysis, we normalized the read counts by computing the numbers of reads per kilobase of bin per million of reads (RPKM).

Motif analyses for ATAC-seq peaks

To find the sequence motif enriched in distal ATAC-seq peaks, findMotifsGenome.pl from the HOMER [48] program was used. For (Fig. 4D, we performed motif-enrichment analysis for ATAC-seq peak regions using HOMER [48] for CTCF motif with the option: -size 200. Then, the proportion of CTCF motif located in ATAC-seq peaks at each stage of the early embryo was calculated.

Identification of hierarchical TAD

We use the OnTAD algorithm to identify hierarchical TADs from Hi-C matrix data [26]. The first step of OnTAD calculates candidate TAD boundaries by an adaptive local minimum search algorithm. In the second step, the recognized candidate boundaries are used to assemble the TAD structure by a recursive algorithm. We take the 20-kb resolution Hi-C interactions matrix as input and use sparseToDense scripts to convert the sparse matrix into a dense matrix, and then calculate the hierarchical TAD structure with the following parameters: OnTAD matrix - penalty 0.1 -minsz 2 -maxsz 100. The output files were used for the next step of analysis and visualization.

The calculation of the TAD boundary level and status switch

We used the method developed by An et al. for calculating TAD boundary levels [26]. TAD boundaries were defined as the maximum number of shared TAD boundaries on the left or right sides. Specifically, if the boundary is shared by no more than one TAD on each side, the boundary level is 1, and so on. If there are two TADs at the left of the boundary and three TADs to the right, the boundary level should be defined as level 3.

To investigate the dynamics of hierarchical TAD boundary, the TAD boundaries at one stage that do not overlap with TAD boundaries in the previous stage were defined as the gained TAD boundary at this stage. The TAD boundaries at one stage that do not overlap with TAD boundaries at the next stage were defined as the lost TAD boundary at this stage. The TAD boundaries that overlap in the two stages but the level are elevated are defined as raised, and the level decreases are defined as reduced.

Domain boundary ATAC-seq signal enrichment analysis

Domain boundary profiles were identified using the boundaries’ centers as anchors with binning the ATAC-seq signals in 5-kb bins ± 0.5 Mb up and downstream of the anchor. The parameters are “computeMatrix reference-point –referencePoint center -S file.bw -R boundary file –beforeRegionStartLength 500000 –afterRegionStartLength 500000 –binSize 5000 –skipZeros -o matrix.mat.gz”. Results were plotted with deepTools2 “plotProfile” command [60].

Deep neural network construction

We chose the dataset shared by Hi-C and epigenetic data for TAD boundary prediction, which mainly includes late 2cell, 8cell, ICM, and mESC. We used ten-fold cross-validation to divide the training set and the test set. We calculated the switch of TAD boundaries in adjacent cell stages and identified conserved TAD boundaries with the level of raised and reduced. In addition, chromatin accessibility and histone modification signals that completely overlap with the TAD boundary sites as feature inputs to construct the neural network model using the Pytorch function library. Before the data passes the fully connected neural network to start the forward propagation computation, we first encode the high-dimensional vectors for the current cell stage and the corresponding layer information into trainable hidden vectors of 16 dimensions using torch.nn.Embedding(). At the same time, we concat the H3K4me3, H3K27me3, H3K9me3, and the next cell stage category, and construct 3 fully connected hidden layers through torch.nn.Linear(). The activation function of each layer is Tanh, and the outputs are the hidden vectors of 8 dimensions. Finally, the output implied vectors are predicted using a Sigmoid classifier, which outputs the probability of TAD boundary layer change for the next cell stage. For model optimization, we use the CrossEntropy function (CrossEntropy) as the predictive loss function, combined with Adam optimizer for backpropagation model parameter tuning to make it fit the training set data.

Identification of ZGA genes

ZGA-only genes were analyzed as previously described [55], which were defined as those not expressed in oocytes (FPKM ≤ 0.5) but that are activated (FPKM > 1) after ZGA (in either the 2cell, 4cell, 8cell embryos, or ICM). To analyze the heterogeneity of hierarchical TAD patterns between human and mouse, the human and mouse ortholog gene groups were identified using the homologene package based on the NCBI HomoloGene database (https://www.rdocumentation.org/packages/homologene/).

Gene ontology analysis

The clusterProfiler tools were used to identify the GO terms [61].

Abbreviations

TAD

Topological association domain

3D structure

Three-dimensional Structure

ZGA

Zygotic gene activation

CTCF

CCCTC binding factor

CRMs

Cis-regulatory modules

TSS

Transcription start site

ICM

Inner cell mass

ESC

Embryonic stem cell

Hi-C

High-throughput chromosome conformation capture

RNA-seq

RNA-sequencing

ATAC-seq

Transposase accessible chromatin with high-throughput sequencing

ChIP-seq

Chromatin immuno-precipitation sequencing

GO

Gene Ontology

TF

Transcription factor

FCM

Fuzzy c-means clustering

ROC

Receiver operating characteristic

H3K4me3

Histone H3 lysine 4 trimethylation

H3K27me3

Histone H3 lysine trimethylation

H3K9me3

Histone H3 lysine 9 trimethylation

H3K36me3

Histone H3 lysine 36 trimethylation

Authors’ contributions

M.L., H.C., and H.L. conceived and designed the study. X.B1. and X.T. conducted data processing and wrote the manuscript. Y.W. and S.Y. created the models of the manuscript. X.X., P.H., Y.L., and J.X. contributed to the design of methodology of the data analysis. J.W., H.T., and Y.Z. contributed to the discussion of the manuscript. B.C. and M.T. contributed to sample collection. L.L. and R.W. provided technical support. Y.S., C.R., and X.B2. participated in reviewing and providing suggestions for the study. All authors read and approved the final manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (No.62422318 and No.62173338 to H.C., No.31900488 to H.L.), National Key Research and Development Program of China (No.2024YFA1307700 to X.B. and No. 2023YFF0725500 to H.C.), the National Youth Science Fund Project of the National Natural Science Foundation of China(No.82001520 to X.T.), and the Beijing Nova Program of Science and Technology Z191100001119064 (awards to H.C.).

Data availability

All data generated or analysed during this study are included in this published article, its supplementary information files and publicly available repositories. The raw FASTQ data for Hi-C of early human and mouse embryo are available in the the european molecular biology laboratory’s european bioinformatics institute (EMBL-EBI) under bioproject PRJNA326112 and the Genome Sequence Archive with the accession number (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE82185), CRA000852 (https://ngdc.cncb.ac.cn/gsa/browse/CRA000852) and CRA000108 (https://ngdc.cncb.ac.cn/gsa/browse/CRA000108) [3, 4, 6]. The raw datasets of ATAC-seq were obtained from the GEO database with the accession number GSE66581 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE66581) [55]. The processed RNA-seq and ChIP-seq (H3K4me3, H3K9me3, H3K27me3) data were downloaded from the GEO database with the accession number GSE66582 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE66582), GSE71434 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE71434), GSE97778 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE97778) and GSE73952 (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE73952) [55, 56]. The source code is available at https://github.com/XuemeiBai-296/hierarchical_TAD [62].

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.