Mini-review: Gene regulatory network benefits from three-dimensional chromatin conformation and structural biology

Abstract

Gene regulatory networks are now at the forefront of precision biology, which can help researchers better understand how genes and regulatory elements interact to control cellular gene expression, offering a more promising molecular mechanism in biological research. Interactions between the genes and regulatory elements involve different promoters, enhancers, transcription factors, silencers, insulators, and long-range regulatory elements, which occur at a ∼10 µm nucleus in a spatiotemporal manner. In this way, three-dimensional chromatin conformation and structural biology are critical for interpreting the biological effects and the gene regulatory networks. In the review, we have briefly summarized the latest processes in three-dimensional chromatin conformation, microscopic imaging, and bioinformatics, and we have presented the outlook and future directions for these three aspects.

1. Introduction

A basic problem in biology is how a fertilized egg develops into a creature that consists of hundreds of different cell types. Decades of research have revealed that oosperm development is a complex process that involves spatiotemporal specific gene expression as well as regulatory element modulation [1], [2]. Researchers have discovered that the genome structure is not one-dimensional from the initial central dogma to this post-genomics [3], [4]. Instead, it more likely has a complicated three-dimensional organization in the nucleus [5]. Concurrently, with the tremendous progress made in epigenetics, molecular mechanisms, and non-coding sequences of transcriptional regulation [6], [7], [8], [9], [10], massive research have found that biological phenotypes are regulated by the gene regulatory network (GRN), which is encompassed by multi-genes, regulatory elements, transcription factors (TFs) and co-factors. A gene is not only modulated by one gene, TF, or regulatory element, it tends to be regulated by multiple components of them. Moreover, recent studies found that the chromatin architecture could profoundly affect the GRN [11] (Fig. 1). For example, the folding of chromatin provides a distal regulatory element the chance to interact with its target genes and then regulate the expression of genes. Thus, proper analysis of the GRN is pivotal in studying the correlation between gene expression and biological phenotypes [12], [13], [14].

Fig. 1.

Diagram of the Gene Regulatory Network. (A) schematically, the gene regulatory network. (B) the diagram for the 3D architecture of the gene regulatory network. Note: Gene A is regulated by the gene B and regulatory element; the gene B is regulated by gene A, gene C, and TF2; the gene C is regulated by the gene B, gene A, and TF2; and the gene D is regulated by the gene B and TF1.

The focus of gene regulatory networks is to investigate the transcription factors and signaling components involved in developmental processes in which the chromatin architecture in the nucleus is established by transcription participators in essence. At present, some progress has been made in analyzing GRNs using a variety of methods. Transcriptome analysis is a prevalent technology for GRN analysis, which studies gene expression changes in a dynamic state. The network is inferred via a bioinformatics mode. There is room to improve its reliability and reproducibility. For instance, the noise in the transcriptome has a profound influence on classification accuracy. With the most recent advances in three-dimensional (3D) genomics and the relationship between GRNs and 3D structure, 3D and structural biological data have been integrated with the transcriptome and other -omics methods, greatly improving the interpretation of GRNs and their biological effects. In this review, we summarized the recent progress of these aspects as well as the advantages and limitations of the techniques used in analyzing the GRNs. We also illustrated some points for the future analysis of GRNs.

2. Construction of a gene regulatory network supported by sequence-based technology

Studying the interaction between chromatin, regulatory elements, and transcription factors is an important part of the transcriptional regulatory network. Characterizing various components of transcriptional regulation, such as promoters, enhancers, and super-enhancers (SEs) and their 3D interactions is a standard method for interpreting transcriptional regulation [15], [16]. The specific expression of genes is hightly determined by regulatory elements, and enhancers are the important one of regulatory elements [17]. Genome-wide enhancers and SEs can be identified by sequence-based techniques (such as STARR-seq) [6], ChIP-seq (such as H3K27ac, H3K4me1, p300) [18], [19] and chromatin accessibility (ATAC-seq, DNase-seq) [20], [21]. Among them, SEs are specific regulators of transcriptional regulatory networks that heavily interact with multiple genes and other elements involved in cell identity. According to the report, SEs are considered large clusters of regulatory elements and possess stronger abilities to function in GRNs compared with typical enhancers [22]. Research in humans found that the super-enhancers in the retina are enriched for two TFs (CRX and NRL) involved in photoreceptors, and genes that overlap with SEs are retinal-specific and have a higher expression [23]. The binding of gene and tissue-specific transcription factors is a key component and driver of gene expression regulation [24]. Another important issue with complex regulatory networks is the superposition effect of multiple enhancers. In regulatory networks, each enhancer is independent and functions in a synergistic manner [25], [26]. The study of the Indian hedgehog (Ihh) gene showed that the Ihh gene was regulated by nine specific enhancers in a synergistic manner, and that gene expression was up-regulated with the increase in enhancer copy number [27]. In the study of the PIM1 gene in human cells, it was found that the inhibition of this gene must simultaneously knock out several enhancers rather than a single enhancer [28]. Therefore, an important factor in gene transcription regulation is composed of complex enhancer sets, which together achieve fine regulation of genes. Nevertheless, the distance between the enhancer and the promoter is not a major factor in activating gene regulation, and studies using ChIA-PET in human and mouse cells have found that many of these enhancers regulate the promoter distally [29], [30]. In the study of mouse Shh and SOX9 genes, it was also found that the enhancers interacting with the promoters of these genes were located outside the 1 Mb region [16], [31]. It is possible that the precise regulation of GRNs is partially because of the existence of these distant interactions that have shaped them into a complex regulatory conformation.

In 2009, Hi-C [32] and ChIA-PET [33] technologies were developed, which represented a milestone in the field. Hi-C [32] and ChIA-PET [33] technologies can uncover the spatial chromatin conformation, from which researchers can directly infer the interactions among genes and regulatory elements. Following that, various 3 C-based technologies, such as in situ Hi-C [34], single-cell Hi-C [35], Capture-Hi-C [36], and others were developed. These techniques have expanded the knowledge of GRN construction and illustrated the targeted relationships among them. Hi-C and other technologies have revealed the existence of TADs (Topologically Associating Domains). TADs are DNA fragments ranging from hundreds of kb to several Mb in the genome, containing one or more genes and regulatory elements, which can be regarded as a highly self-interacting genomic unit. Hi-C and other technologies have revealed the existence of TADs, which are thought to reflect the presence of physical interactions between DNA domains that organize chromosomes into distinct structural and functional units [37], [38]. TADs are highly conserved in different cell types and developmental stages; their positions are almost unchanged [39], [40], [41]. The relationship between TADs and gene transcription has always been a controversial? topic. During the differentiation of embryonic stem cells, the number of TADs is constantly changing [42]. The occurrence of some diseases or the body’s emergency response will also be accompanied by changes in TADs [43], [44], [45]. TAD-relative gene expression was also found to change when TAD borders were lost [46]. These examples demonstrate that TADs promote enhancer-promoter interaction in a cell type-independent manner in transcriptional regulation [38]. In embryonic development and cell differentiation, TADs become more structured, possibly because they instruct enhancer-promoter interaction and directly participate in transcriptional regulation [47].

In the matrix from Hi-C or other technical analysis, promoter-enhancer interactions were detected at high frequencies on the baseline of the domain formed by loop extrusion [48], [49]. These interactions form a certain chromatin structure through loop extrusion, which plays an important role in the interaction between promoters and enhancers in GRNs [1], [50]. Loop extrusion also promotes segregation, enhancer-promoter interactions, and other genomic functions [14]. It is important to dissect the constitution and biological effects of GRNs. The interaction between regulatory elements mediated by loop extrusion is generally more pronounced at the boundary of TADs and is generally limited within TADs. So loop extrusion regulates gene expression by promoting communication between enhancers and promoters, thereby affecting biological traits [14], [51], [52], [53], [54]. Additionally, insulators (boundary elements) are also critical to gene expression. Insulators are generally at the boundary of TADs, defined by CTCF and cohesin complexes [34], [37], [55], [56]. Because the insulator determines the boundary region, it determines which interactions in the genome can be formed [57], [58], [59]. For example, the deletion of the TAD boundary element of the Shh gene resulted in a weakened interaction between its promoter and enhancer, and its gene expression level also decreased by 50% [60]. The above series of studies have shown that in the genome, the regulatory elements are annotated by regulatory element identification technology, and the enhancer-promoter interactions are identified by chromatin conformation capture technology, represented by Hi-C. In each TAD, transcription factors mediate enhancer-promoter interactions to form spatial DNA loops, and there are also connections between multiple different DNA loops, forming a complex GRN to jointly regulate biological traits.

Overall, while the three-dimensional structure of chromatin is important for understanding gene regulatory networks, current research is mainly focused on describing the 3D genomic architecture, with less emphasis on its functional implications. However, the 3D structure provides important insights into gene regulatory activity and can serve as a starting point for further investigations into the function of GRNs. At the same time, most of these studies are carried out in population cells, so in the future, we should study the dynamics among different levels of chromatin structure and gene expression regulation in single cells and also combine epigenetics, phase separation, and other fields [14], [61], which will help us better understand the relationship between the 3D chromatin structure and GRN function.

3. Imaging technology supports the construction of gene regulatory networks

In recent years, rapid advances in imaging technology have provided an intuitive way to catalog gene regulatory networks. Visualizing the structure of chromatin and transcription complexes is a key step to understanding the transcription process and dissecting the biological effects of GRNs. The development of microscopy imaging technology allows us to observe the chromatin scaffold and the internal structure of chromatin [62], [63], [64], which is sometimes consistent with the results of Hi-C. For example, chromEM can integrate electron diffraction (ED) and electron tomography (ET) and visualize chromosome fibers and chromosome scaffolds [65] using modified soybean ascorbate peroxidase (APEX2) as a marker to observe the localization of chromatin in cells [66], [67]. Electron microscopy and confocal microscopy revealed highly and homogeneously condensed regions in chromatin and revealed that gene transcription occurred in the nuclear region [68], [69]. With increasing resolution, the proximity of loop anchors during gene transcription was observed under a high-resolution microscope [70], [71]. This series of studies shows that the electron microscope has been able to observe the real existence of transcriptional regulatory networks and reveal the impact of these spatial structures on gene expression. These have improved our understanding of the intuitive cooperation of GRNs and how they function during different biological processes.

Imaging technology is a powerful tool to investigate the process of transcriptional regulation. The physical distance of the site in the genome could be inferred to determine whether there is interaction. In the field of optical microscopy, DNA-FISH can judge the interaction by verifying the spatial distance between the two sites. Although this distance is related to the set threshold and resolution, it can be close to the real situation to some extent to study the interaction in gene transcription [72], [73], [74]. When DNA-FISH is combined with frozen sections, the genome can be observed at a higher resolution, such as in various cells where more than 20% of the chromosomal regions of the genome interact with other regions. This interaction is more pronounced in transcriptionally active regions [75], [76]. This phenomenon also suggests the universality and importance of regulatory networks throughout the genome. Chromatin folding and gene regulation are dynamic processes that require us to observe them in living cells. Genome editing techniques (such as the improved dCas9 system) can track the dynamic changes of the interaction between two genomic sites in living cells [77], [78], [79]. Although these techniques have shown the existence of GRN conformations, there has been no in-depth study of specific molecular mechanisms.

Using cryo-electron microscopy, the Xu Lab [80], [81] discovered the critical role of nucleosomes and the preinitiation complex (PIC) in the transcription process. For example, they obtained high-resolution nucleosomes and a human 26-subunit mediator cryomicroscope structure and revealed the mechanism by which the mediator and PIC are involved in the transcription process (Fig. 2). Previous studies have shown that nucleosomes play an important role in modifying proteins, processing RNA, and assembling transcription initiation complexes, but the mechanism has not been elucidated [82], [83]. In a recent study, Chen et al [84]. used cryo-electron microscopy to find that nucleosomes promote the organization of PIC-mediators on promoters by binding to the TFIIH subunit p52 and mediator subunits MED19 and MED26, which may facilitate transcriptional initiation. PIC-mediator exhibits multiple nucleosomes binding patterns, supporting the structural role of nucleosomes in PIC-mediator assembly coordination. This reveals the molecular mechanism of PIC-mediator organization on chromatin and emphasizes the importance of nucleosomes in regulating transcriptional initiation. A series of cryo-electron microscopy-based results advances our understanding of the mechanism of transcriptional regulation on the basis of the space network.

Fig. 2.

The pattern of mediator and PIC involvement in the transcription process.

Using microscopic imaging techniques, scientists have revealed the relationship between structure and gene expression regulation from another perspective. These results can be compared and integrated with the results of sequence-based techniques. From observing the interaction of regulatory elements with fluorescent groups to observing the high-resolution structure of transcriptional participants by cryo-electron microscopy, microscopic imaging technology has been called an indispensable part of the study of GRN. However, the quality and resolution of sample products can directly affect the results of observation. In addition, the throughput of these technologies is limited, and observations of transcription processes or participants are limited to a few critical biological processes. The improvement of sample preparation and resolution and the comprehensive tracking of transcription processes in living cells are the future development directions for understanding gene regulatory networks through imaging techniques.

4. Bioinformatics approaches to profiling gene regulatory networks

In the past two decades, advances in high-throughput biological data collection have greatly advanced our understanding of GRNs, which is based on bioinformatics analysis via mathematical modeling. In addition to the experimental methods described above for detecting chromatin contacts and GRN, network modeling techniques, such as network topology, transcriptome data modeling, chromatin conformation modeling, and deep learning, have been used to analyze its corresponding function and predict its behavior. GRN is often decomposed into basic building blocks with assumed functions, which extend the concept of structure-function relationships to the network schema level [85]. These patterns include auto-regulation (one gene regulates its own expression) [86], Feed Forward Loop (FFL) (one regulator regulates another regulator and a ternary structure of genes, which occurs more frequently in yeast GRN than other types[87]), Feedback Loops (looping action between multiple regulators and genes) [88], Single-Input Module (single regulators regulate multiple genes) and Dense Overlapping Regulon (composed of a set of regulators that control a set of output genes) [89] (Fig. 3). For functional studies, modularization by clustering can effectively identify co-expressed genes. The DREAM project uses genome-wide association study (GWAS) data to test and compare a module identification approach to human protein-protein interaction (PPI) data [88]. It has the advantage of using the diffusion state distances of the kernel clustering instead of the shortest path metric typically used to use PPI structure prediction functions to more accurately identify biological functions (Fig. 4A).

Fig. 3.

The topology of the gene regulatory network. Note: The dark-green dots indicate regulators, and the light-green dots represent genes.

Fig. 4.

Different visualizations of gene regulatory networks. (A) The protein-protein-interaction occurs during fibroblast differentiation into myoblasts. (B) Pol II ChIA-PET cis and trans interactions in mouse ESCs.

(A) (data source: STRING); (B) (data source: GSM1084137).

Combined with RNA-seq, machine learning algorithms are also an important technology for studying GRNs. GeneXPress first completes the mapping of GRNs based on a regression algorithm [90]. Regression-based methods use linear and nonlinear regression to predict mRNA levels as a function of chromatin markers and/or TF occupancy and can infer mRNA prediction models for individual conditions. GENIE3 [91], GRNBoost [115], LiPLike [93], and TIGRESS [92] are all based on this algorithm (Table 1). TIGRESS performs a series of feature selections for each target gene and uses randomization-based techniques to score evidence of regulatory interactions [92]. S. Feizi et al. used network deconvolution to solve the problem of identifying the direct relationship between variables connected in the network, which is a frequent problem in GRN mapping. Based on Bayesian inference [94], LeMoNe uses a centroid clustering method to infer GRNs, assigning genes and conditions to modules, and then assigning regulatory programs to gene sets [114]. ARACNe-AP applied partial correlation coefficients and data processing inequalities to distinguish direct and indirect dependencies, relating changes in TF mRNA levels to sets of other genes’ mRNA levels [95]. These methods infer cellular networks mainly using regression, correlation analysis, and Bayesian networks as the main tools, which require the least input data and have wide applicability. However, due to the high experimental cost and the fact that the sample size is much smaller than the number of TFs, the results are susceptible to noise. Various machine learning approaches have performed poorly on computationally simulated data evaluations of E. coli and Saccharomyces cerevisiae [96], [97]. This type of algorithm is also highly dependent on the quality of the data and the uncertainty in determining the relevant parameters. Noise, modeling form, parameter estimation methods, and the amount of experimental data are all crucial factors that affect the results [98]. Meanwhile, different analysis strategies and model choices can lead to varying effects of network inference. It is also a complex problem to choose the appropriate scheme to ensure the biological correctness of the results [99]. Therefore, in the process of exploring the GRN, we need the integrative analysis of multiple types of omics data, such as 3D genomic data, which opens a novel path.

Table 1.

Summary of tools used for GRN analyze.

Tools	Description	Data	Performance	Restriction	Link	Ref
GeneXPress	regression algorithm	bulk transcriptomics	Accurately predict regulator function, targets and the experimental conditions under this regulation occurs	1. Fail to identify certain regulatory relationships and may occasionally predict regulatory relationships that do not hold； 2. A gene may belong only to a single module	(https://pypi.org/project/GeneXpress/#files)	[90]
TIGRESS	regression algorithm	bulk transcriptomics	Expresses the GRN inference problem as a feature selection problem, and solves it with the popular LARS feature selection method combined with stability selection	Heavy formatting requirements	(http://cbio.ensmp.fr/tigress)	[92]
LeMoNe	Bayesian inference	bulk transcriptomics	Learning process is considerably faster for larger data sets	--	(http://bioinformatics.psb.ugent.be/software)	[114]
Network Deconvolution	Network deconvolution	bulk transcriptomics	1. Distinguishing direct targets in gene expression regulatory networks; 2. Recognizing directly interacting amino-acid residues for protein structure prediction from sequence alignments; 3. And distinguishing strong collaborations in co-authorship social networks using connectivity information alone	Network deconvolution assumes that networks are linear time-invariant flow-preserving operators, which excludes nonlinear dynamic networks as well as hidden sinks or sources.	(http://compbio.mit.edu/nd/index.html)	[94]
ARACNe-AP	Information theory	bulk transcriptomics	A dramatic improvement in computational performance (200 × on average) over the previous methodology, while preserving the Mutual Information estimator and the Network inference accuracy of the original algorithm	--	(http://sourceforge.net/projects/aracne-ap)	[95]
GENIE3	regression algorithm	bulk transcriptomics	Work well in the presence of a large number of genes, are fast to compute, and are scalable	Prior knowledge about transcription factors can have a significant impact on the results	(https://bioconductor.org/packages/release/bioc/html/GENIE3.html)	[91]
GRNBoost	regression algorithm	bulk transcriptomics	Implementation drastically reduces the time needed to infer a GRN	Statically performs workload scheduling	(http://arboreto.readthedocs.io)	[115]
LiPLike	regression algorithm	bulk transcriptomics	Higher accuracy and removing false positive identifications from GRN predictions	--	(https://gitlab.com/Gustafsson-lab/liplike)	[93]
BART3D	genomic differential chromatin interaction	Hi-C	Can infer TRs associated with genome-wide differential chromatin interactions	Genomic distance needs to be set manually	(https://github.com/zanglab/bart3d)	[100]
3CPET	non-parametric Bayesian	ChIA-PET	Infer the most likely set of protein networks likely to be involved in maintaining chromatin interactions with reasonable performance and good stability	memory and time requirements for 3CPET gradually increase with the DNA–DNA corpus size	(http://www.bioconductor.org/packages/release/bioc/html/R3CPET.html)	[103]

By integrating 3D genome data with different ∼omics datasets, not only the identification of GRNs, but also the precise location of interactions between enhancer-promoter and other transcriptional regulatory elements can be achieved (Fig. 4B). For example, based on Hi-C and ChIP-seq data, BART3D designed a DCI scoring algorithm to infer TRs associated with genome-wide differential chromatin interactions and validated that genomic regions containing CTCF or RAD21 binding sites exhibited reduced levels of chromatin interactions after knockout [100]. Ma et al. [101] used single-cell Hi-C data to demonstrate that within differentiated and undifferentiated cells, the association of TFs in spatial networks is closely related to GRNs, and that lineage-specific TF binding and chromosomal subcompartment segregation are closely related potential causal relationships [101]. By modeling Hi-C data, Johan et al. found a characteristic of chromosomes in interphase during mitosis. Especially at anaphase, chromosomes fold into contiguous arrays of rings condensed around a central axis. The chromatin loops are nested within a ∼400-kb outer loop separated by an ∼80-kb inner loop [102]. This is an important complement to the observations made by electron microscopy. 3 CPET is based on a nonparametric Bayesian approach that infers the most likely protein complexes involved in maintaining chromatin interactions and the regions they may control [103]. Based on 3D genome technology combined with gene expression and chromatin state data, the GRN-loop model of induced early neural lineage commitment effectively identifies driving transcription factors and enhancers and then validates the impact of predicted genetic perturbations on lineage commitment [97]. We can more accurately interpret enhancer-promoter interactions by applying machine learning algorithms, epigenetic data, and three-dimensional genome technology for mutual verification. RIPPLE combines the published chromosome conformation capture (3 C) dataset with a minimal regulatory genome dataset to construct a minimal classifier by combining random forest and group lasso-based multi-task learning to predict enhancer-priming in cells and subinteractions in a circuit-specific manner [104]. IM-PET is based on identifying discriminative features of EP pairs, including enhancer and target promoter activity profile correlations, transcription factor and target promoter correlations, the co-evolution of enhancer and target promoter, and the interactions between enhancer and target promoter. The distance constraints were trained on genome sequence, three histone modification ChIP-Seq and RNA-Seq/microarray data, judged by the random forest method, and validated by ChIA-PET [105]. In addition, accurate identification of the target genes of regulatory elements is very important for the study of enhancer-promoter interactions. Sean Whalen et al. developed an algorithm that can detect single enhancer-promoter interactions in multiple cell lines with an error rate 15 times lower than simply using the nearest target gene [106]. HiC-Reg combines published Hi-C datasets with one-dimensional regulatory genome datasets (such as chromatin marks, structural and transcription factor proteins, and chromatin accessibility) to predict two interaction counts between genomic loci. This approach breaks through the traditional binary classification of interacting vs. non-interacting and can predict interactions for new chromosomes or cell lines of interest [107].

Deep learning is another cutting-edge weapon for studying GRNs. It integrates multi-omics data for analysis and prediction. Simultaneously, biological function analysis is also carried out based on identifying GRNs. These tools identify transcriptional regulatory elements by analyzing sequence features or further differentiate network differences in tissue characteristics and cell types based on the identification of transcriptional regulatory networks. DeepSEA was developed as a fully sequence-based algorithmic framework for noncoding-variant effect prediction. It learns regulatory sequence codes from genomic sequences by simultaneously predicting large-scale chromatin-profiling data, including TF binding, DHS, and histone mark profiles [108]. BiRen uses a deep learning hybrid architecture to predict enhancers based on DNA sequence alone. BiRen exhibited superior accuracy, robustness, and generalizability in enhancer prediction relative to other state-of-the-art enhancer predictors based on sequence characteristics [109]. DGRNS adopts a hybrid deep learning framework. With the help of recurrent neural networks and convolutional neural networks, not only statistical features but also time-related features can be extracted. Based on single-cell RNA-Seq technology, the regulatory network was classified and analyzed to explore network differences between biological tissues [110]. At the same time, deep learning is also applied to the prediction of three-dimensional genome structures. It is also helpful for the identification of GRNs. Akita converts input DNA sequences into predicted locus-specific genome folds based on convolutional neural network, Hi-C, and Micro-C data and then estimates entire contact maps [111]. Similar to Akita, DeepC also predicts genome folding through the convolutional neural network, but differs in data preprocessing and training methods [112]. DeepC employs a non-linear normalization objective that requires pre-training models on a large number of epigenomic profiles and transferring weights to their full models. The calculated results of the differences between the two methods merit further discussion of their advantages and disadvantages in future comparative studies.

Other techniques include methods such as protein interactomes, eQTL data, the hierarchy of tissues, etc., which are jointly applied to the prediction of GRNs. Recent advances in genomic technology and computational modeling have revolutionized our ability to model GRNs. Researchers have applied reverse engineering methods to construct GRNs in cancer cells to improve the accuracy of the comparison of the differences in GRN between normal and cancer cells [113]. In recent years, the interpretation of multilayer networks has been a key issue for a comprehensive understanding of gene regulatory systems. Although several statistical and machine-learning approaches have been developed and applied to uncover mechanisms, a comprehensive understanding of large-scale multilayer networks remains a challenge. Data-driven methods based on artificial intelligence and multi-omics data, especially 3D genomics, provide biologically reliable results and are the direction of bioinformatics efforts in this field.

5. Summary and outlook

Research on GRNs is an important subject in the field of precision biology. Although the continuous development of chromatin conformation capture technology provides good technical support for the interpretation of GRNs, there is still much room for improvement. Experimentally, the quality of the data produced by different laboratories is uneven, with poor repeatability, and the data also has low resolution, a poor signal-to-noise ratio, and other issues. In terms of data processing, there are differences in the evaluation criteria of different software, resulting in differences in the results of even the same data, such as the number and overlap of loops obtained by Homer, HiCCUPs and FitHic [116] when loop calling. Homer and FitHic directly model individual interacting elements, considering the effects of systematic bias or linear distance across the genome. Among them, FitHic estimates that the interaction of random polymers is affected by distance by modeling the relationship between interaction and space twice and using the binomial distribution to model the interaction, giving the significance of the interaction. While HiCCUPS uses the Poisson process to consider the effects of neighbor interactions or other hierarchies, the first type of algorithm can identify more significant interactions from the results, but the false discovery rate is relatively high. While the latter is computationally more rigorous, it will lose some potentially practical information. More importantly, the conclusions of different studies vary greatly. For example, there are many inconsistent results in TADs and more studies are needed on the effects of chromatin structure on biological functions. Changes in TAD were found to cause changes in gene expression in a study of the mouse α-globin gene [117]. In mouse cells, silencing of the male X chromosome also leads to the rearrangement of the associated TAD [118]. These all illustrate the important influence of TADs on gene expression and traits. However, in other studies, it was found that the destruction of TADs structure had little effect on gene expression [16]. Alternatively, when it comes to the presence or absence of TAD, ChromEMT produces inconsistent results [65] when compared to the traditional concept. In addition, although MERFISH imaging technology observes TAD-like domains at the single-cell level, its high heterogeneity contradicts the high conservation of TAD in the traditional concept [119]. At the same time, the need for a unified TAD identification standard also makes the credibility of the results questionable. The inferred results of different TAD identification methods are also different. The resolution size is another factor that affects the results [120]. Current criteria for evaluating TAD identification tools include robustness concerning bin size and normalization strategy, the minimum required data volume, reproducibility, biological relevance, and computational efficiency [121]. As biological researchers, however, we should pay more attention to interpreting TAD for its biological meaning itself than for the statistical results of computational observations. This needs to be verified by a more suitable data set and other corresponding experimental observation methods. These results make us think about whether this is a defect of electron microscopy or whether TAD is only a statistical model; so far, there is no reasonable explanation.

At present, the research on gene regulatory interactions has focused on how enhancers interact with promoters under the mediation of TF. In the human genome, less than 15% of genes and regulatory elements have been annotated [122], and the function of the remaining 85% of non-coding sequences is unknown. Recently, the function of silencing regulatory elements in gene regulation has been reported [123], [124], [125]. However, because these reports only covered a portion of the silencer's genome region, whole-genome silencer identification is an important supplement to GRN. Additionally, with the dynamic changes in organisms, the corresponding GRNs will also change. Most of the current technologies show static chromatin conformation and regulatory networks. In more cases, we only describe that changes in the spatial conformation will affect gene transcriptional regulation or that changes in gene expression will cause changes in chromatin conformation. Can we change chromatin conformation, gene expression, or even phenotype through these regularities? A recent study reported that cell identity can be changed by introducing exogenous enhancer elements into cells [126], which makes us realize the substantial regulation of gene expression by spatial conformation. The development of various fields has promoted the completion of GRNs.

We briefly summarize the relationship between chromatin spatial structure and gene expression regulation. The chromatin spatial structure and epigenetic program play pivotal roles in the GRN. The chromatin spatial structure embraces loops and TADs. The TADs include vast cis-loops. The loop extrusion underlies the TAD formation [127]. The loops formed by the proximity of enhancer-promoter or silencer-promoter, which are mediated by TFs. These will then get involved in gene expression regulation. Furthermore, the boundary elements are typically enriched in the TAD boundaries, which frequently bind the CTCF. And repression of the boundary elements could alter the structure of TADs, which would have a further impact on gene expression (Fig. 5).

Fig. 5.

The relationship between chromatin spatial structure and gene regulatory network.

Lastly, we realize the high-throughput, multi-scale, and multi-modal imaging technology will provide an integrated view of chromatin organization and structural biology in its native genome context. By integrating multiple-omics datasets, a good bioinformatics algorithm can identify relationships among genes and regulatory elements, which would greatly facilitate the unveiling of the comprehensive map of individual cell GRNs, thus allowing a novel and fundamental understanding of how a cell operates and initiates life. In the near future, through the cooperation of imaging technology, genomics technology, computer science, physics, and other fields, a complete, dynamic, and precise GNR map will be achieved, which will facilitate the gene function analysis for the purpose.

CRediT authorship contribution statement

Xiusheng Zhu: Conceptualization, original draft, Writing – review & editing, Methodology, Formal analysis. Qitong Huang: Writing – original draft, Conceptualization, Formal analysis, Writing – review & editing. Jing Luo: Writing – original draft, Conceptualization, Formal analysis, Writing – review & editing. Dashuai Kong: Writing – review & editing, Conceptualization. Yubo Zhang: Conceptualization, Writing – original draft, Writing – review & editing, Supervision.

Declaration of Competing Interest

The authors declare that there are no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.