Application of Hi-C and other omics data analysis in human cancer and cell differentiation research

Graphical abstract

In this paper, we mainly concluded the Hi-C data analysis methods from genome 3D structure, A/B compartment, TADs, and loops detection, how to apply the visualization technology into the identification of chromosome feature structure, and correlation analysis among multi-omics data about human cancer and cell differentiation.

Abstract

With the development of 3C (chromosome conformation capture) and its derivative technology Hi-C (High-throughput chromosome conformation capture) research, the study of the spatial structure of the genomic sequence in the nucleus helps researchers understand the functions of biological processes such as gene transcription, replication, repair, and regulation. In this paper, we first introduce the research background and purpose of Hi-C data visualization analysis. After that, we discuss the Hi-C data analysis methods from genome 3D structure, A/B compartment, TADs (topologically associated domain), and loop detection. We also discuss how to apply genome visualization technologies to the identification of chromosome feature structures. We continue with a review of correlation analysis differences among multi-omics data, and how to apply Hi-C and other omics data analysis into cancer and cell differentiation research. Finally, we summarize the various problems in joint analyses based on Hi-C and other multi-omics data. We believe this review can help researchers better understand the progress and applications of 3D genome technology.

1. Introduction

With the completion of the Human Genome Project and the progress of other model organisms' genome projects, how to deal with the massive amount of molecular biology information is a considerable challenge. The existing multi-omics research is divided into genomics, transcriptomics, proteomics, epigenomics, and other omics research.

Transcriptomics studies how the same genome can result in different cell types and how gene expression is regulated. Among the genome analysis “omics” technologies, RNA-Seq [1] can be used to identify genes in the genome or to identify which genes are active at a specific time point, and read counts can be used to simulate relative gene expression levels accurately.

Epigenes are the genome's supporting structure, including protein and RNA binders, alternative DNA structures, and chemical modifications of DNA. Among technologies used to study epigenes, MNase-seq [2], DNase-Seq [3], ATAC-Seq [4], FAIRE-Seq [5], [6] are all used to study open chromatin regions and determine chromatin accessibility by detecting transcription factor (TF) footprints. ChIP-seq [7] to study the interaction between proteins and DNA in nuclei. Hi-C [8], [9] is a technology to analyze the spatial structure of chromatin in cells, quantifying the number of interactions between genomic loci that are adjacent in 3D space, but may be separated by many nucleotides in the linear genome (in this paper, chromatin interactions may be just the product of the random ligation of two DNA fragments detected by the Hi-C experiment, may be an interaction of chromatin segments mediated by proteins, etc.).

In recent years, multi-omics data analyses about diseases and cell differentiation appeared as shown in Table 1 and Table 2. Gene structure changes can lead to different diseases, for example, Holoprosencephaly (a forebrain disease caused by mutations in the SBE2 enhancer element [10]); PPD2 (polydactyly of a triphalangeal thumb caused by mutations in the ZRS enhancer [11]) and adenocarcinoma of the lung (caused by duplication of MYC gene enhancer [12]). In addition to multi-omics data analysis, in 2015, Ya Guo et al., [13] used CRISPR technology to invert the CTCF site, which changed the genome topology and enhancer/promoter functions. In each of these diseases, the underlying genetic defect could not have been identified without the use of multi-omics technologies and analyses, so understanding the relationship between gene structure changes and gene expression, combined with gene-editing technology, is expected to treat various genetic diseases. Hi-C technology, as the basis for studying genome structure, is fundamental among these technologies. So, this paper will concentrate on the analysis of Hi-C with other omics data.

Table 1.

Multi-omics datasets for human cancer and disease.

Cancer/disease	Cell line	Sequencing method	Data ID/data link	Reference
liver, lung tumours, breast pancreas and lymphoma samples	GM12878	WGS, Hi-C	GSE63525	[35]
breast cancer, colorectal cancer, lung cancer	GM06990, K562	(SNP)-arrays, Hi-C	GSE19399,GSE18199, GSE18350	[36]
melanoma, prostate cancer, lung cancer, leukaemia	GM06990	SNVs, SNPs, Hi-C	GSE18199	[37]
cardiovascular disease	HCASMCs	ATAC-seq, RNA-Seq, Hi-C	GSE101498	[38]
the Crohn’s disease	T cells	ATAC-seq, RNA-Seq, Hi-C	GSE101498	[38]
the celiac disease	intestinal T cell	ATAC-seq, RNA-Seq, Hi-C	GSE101498	[38]
IDH mutant gliomas	IMR90, NHEK, KBM7, K562, HUVEC, HMEC, GM12878	ChIP-seq, RNA-seq, Hi-C, DNA methylation quantification	GSE70991	[39]
Pan-cancer analysis	3T3, HCC-15, CR, HCT116, IMR90 cell line	Hi-C	http://cancergenome.nih.gov/	[40]

^aEach column denotes the key properties of multi-omics data analysis with Hi-C technology for cancer and other diseases. ‘Cancer/disease column’ denotes the cancer or disease’s name, ‘cell line’ column denotes the cell line that analyzed, ‘sequencing method’ column denotes the sequencing methods that were used, ‘data ID/data link’ column denotes the availability of sequencing data, among them, data ID GSEXXX can be searched in the NCBI GEO database (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi), ‘reference’ column denotes the references where the data acquisition and analysis were described. Readers are referred to these papers for further information.

Table 2.

Multi-omics datasets for cell differentiation.

Species	Cell line	Sequencing method	Data ID	Reference
Homo sapiens	Embryonic stem cell	Hi-C, RNA-seq, Chip-seq	GSE116862	[41]
		Hi-C, RNA-seq	GSE105028	[42]
		Hi-C, RNA-seq	GSE86821	[43]
		ChIP-seq, Dnase-HiC	GSE90680	[44]
		Hi-C, HiChIP, ChIP-seq, ATAC-seq, eU-Seq	GSE105028	[42]
		Hi-C,RNA-seq and ATAC-seq	GSE106687	[45]
		Hi-C	GSE107148
		Hi-C	GSE86821	[43]
		Hi-C	GSE52457
Mus musculus	mESC	Capture Hi-C	GSE124698	[46]
		ATAC-seq, Hi-C,RNA-seq, ChIP-seq	GSE115933	[47]
		RNA-seq, ChIP-seq, Hi-C and Promoter Capture Hi-C	GSE100835	[48]
		Hi-C, RNA-seq	GSE89520
		ChIP-seq,Hi-C	GSE95533	[49]
		ChIP-seq,RNA-seq, ATAC-seq, WGBS and Hi-C	GSE138102	[50]
		Hi-C	GSE153884
		ChIP-seq,Hi-C and 5C	GSE156868	[51]
		In situ Hi-C	GSE118911	[52]
		ChIP-seq,ATAC-seq,Hi-ChIP,Hi-C	GSE113339
		RNA-seq,ChIP-seq, Xist CHART-seq, and in situ Hi-C	GSE116413	[53]
		Hi-C	GSE119805	[54]
		HiC,STEM-seq and RNA-seq	GSE109344	[55]
		ChIP-seq,RNA-seq, Hi-C	GSE119697	[56]
		Capture Hi-C	GSE114619	[46]
		Chip-Seq,Gro-Seq,Mnase-Seq, ATAC-seq and Hi-C	GSE82144	[57], [58]
		Hi-C	GSE110061
		Hi-C	GSE125656	[59]
		Hi-C	GSE146001	[60]
		Hi-C, RNA-seq	GSE118263	[61]
		Hi-C	GSE133246	[62]
		Hi-C	GSM4386021
		PLAC-seq. Hi-C, mRNA-seq and ChIP-seq	GSE146449	[63]
		Hi-C	GSE119347
		Hi-C	GSE124342	[64]
		Hi-C	GSE59027	[65]
		Hi-C, ChIP-Seq, RNA-Seq, DNase-Hypersensitivity	GSE72164	[66]
		DNA SPRITE, RNA-DNA SPRITE	GSE114242	[67]
		Hi-C	GSE130723
		Hi-C	GSE130725	[68]
		Hi-C, RNA-seq	GSE136307
		Hi-C	GSE152918	[69]

^aEach column denotes the key properties of multi-omics data analysis of cell differentiation using Hi-C technology. ‘Species’ denotes the species’ name, ‘cell line’ column denotes the cell line that was analyzed, ‘sequencing method’ column denotes the sequencing methods that were used, ‘data ID’ column denotes the availability of the sequencing data, data ID GSEXXX can be searched in the NCBI GEO database (https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi), ‘reference’ column denotes the references where the data were published. Readers are encouraged to seek out these papers for further information.

Many authors have written relevant reviews on Hi-C data in the past ten years, mainly divided into three types: Hi-C data fundamental technical analysis, Hi-C structure analysis method, and explanations of applications of Hi-C data. The first case mainly focused on developing 3C technology [8], [14], [15], [16], [17], [18], [19], [20], [21], [22] and fundamental analysis methods [23], [24], [25]. Some reviews summarize the various hierarchical analysis methods based on Hi-C [26], [27], [28], [29], [30], [31], while the others summarize the different hierarchical structures of the 3D genome applications to human disease development [31], [32], [33], [34]. This review will focus on Hi-C technology's basic principles and the multi-level chromosome structures that can be identified based on Hi-C technology: overall structure, A/B compartment, TAD, and loop. In addition, we will briefly introduce how to combine Hi-C data with other epigenomic data and transcriptome data to study the relationship with an understanding of human diseases. Finally, we will give examples to illustrate the application of multi-omics joint analyses to provide ideas for researchers who have just started 3D genome research.

2. Demand for Hi-C data visualization analysis

2.1. Principle theory of Hi-C

The most commonly used Hi-C experiment was proposed by Erez Lieberman Aiden as follows [9]: (1) formaldehyde cross-linking so that the spatially adjacent chromatin fragments are covalently connected; (2) restriction enzymes digestion to cut the genome and the use of biotin to label the cut ends; (3) use of DNA ligase to ligate the cut ends and create chimeric molecules; (4) purify and break DNA chimeric molecules, and isolate DNA fragments with the biotin tag; (5) sequence both ends of the fragments of the DNA library; (6) construct the chromatin interaction matrix by counting the number of chimeric molecules between any two regions of the genome.

Using Hi-C technology, we obtain raw sequency data, but if other visualization methods are needed, the following steps are required: 1) Perform linker trimming processing on raw data to obtain valid sequencing data; 2) Obtain comparison file (.sam format) by double-end sequence alignment to the reference genome; 3) Read alignment file and process it into a matrix, tuple or .hic format; 4) use normalization algorithms like KR [70], ICE [71] or others [72], [73], [74], [75], [76], [77], [78], [79] to normalize the data. After the above experiment and data processing steps, we obtain the Hi-C contact matrix. With the Hi-C data and other omics data, we can explore the chromosomes' architecture and study the relationship between chromosome structure and transcriptional expression.

2.2. Hi-C enhancement

In the past few years, the 3D genome analysis methods have rapidly improved, and a large amount of data appeared, but the current resolution of most Hi-C data ranges from 25 kb to 1 Mb. Some high-resolution Hi-C data (range from 1 kb to 10 kb) are only available in a few tissues or cell lines, which affects our analysis of structures at kilobase pair (kb) resolution. But the higher the data resolution is, the deeper the sequencing depth required and the greater the expense is. So, how to map existing low-resolution Hi-C data to high-resolution Hi-C data has become a hot spot in the past five years.

Many authors used the deep learning framework to enhance the resolution of Hi-C data in recent years. In 2018, Yan Z et al. [80] developed HiCPlus, a method based on a super-resolution convolutional neural network (SRCNN). This algorithm, which can infer from low-resolution Hi-C data, is highly similar to the original matrix, a high-resolution Hi-C matrix, using only 1/16 of the original sequence reads. In 2019, Tong L et al. [81], [82] developed two new calculation methods to enhance the resolution of Hi-C data: HiCNN, based on a 54-layer convolutional neural network, and HiCNN2, inclusive of three different deep convolutional neural network architectures. Liu Q et al. [83] proposed hicGAN to enhance low-resolution Hi-C data through Generative Adversarial Networks (GAN). Same as hicGAN [83], in 2020, Hong, Hao et al. [84] developed the DeepHiC method, which can reproduce high-resolution Hi-C data from down-sampled reads as low as 1%. Zhilan L et al. [85] developed SRHiC based on the ResNet and WDSR model. They improved the Res-block in ResNet to increase the network's nonlinearity and learning ability. Simultaneously, a small convolution kernel is used multiple times to reduce the contact matrix’s size instead of using a large convolution kernel at once. This method has a strong generalization ability.

2.3. Hi-C data analysis

Many methods can be used to analyze Hi-C data, such as principal component analysis, interaction network analysis, heat maps, etc., to analyze Hi-C data:

1). cis–trans analysis to determine the quality of the Hi-C library:

$\mathit{cis}-transinteractionratio=sumofcisregionfrequency/sumoftransregionfrequency$ Generally, the cis/trans ratio in high-quality Hi-C experiments is between 40 and 60 [86].

2). chromatin interactions visualization using heat maps: a) Whole-genome interactive heatmap; b) Interaction analysis between chromosomes; c) Interaction analysis within chromosomes as shown in Fig. 1.

Fig. 1.

A/B compartment visualization with annotation. From top to bottom, the following visualization is shown: RNA-seq, Dnasel, CTCF (Broad), H3K27ac annotation, eigenvector, subcompartments and Hi-C contact heatmap. For Hi-C contact heatmap, we choose to show observed/expected balanced Hi-C data to visualize A/B compartment with 500 kb resolution, blue means A compartment, red means B compartment, combined with RNA-seq, Dnasel and H3K27ac, we can see, the blue region had higher gene expression and higher signals of h3k27ac, dnasel and CTCF. (Note: this Figure is drawn using the juicer box tool [87], Data Source: Rao and Huntley et al. [18] Cell GM12878 Hi-C in situ chr1: 0 MB-120 MB). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3). Structural analysis: a) 3D structure visualization for the whole/local chromatin, available tools to reconstruct structure can be seen in Table 3; b) compartments analysis to find the open/closed regions as Fig. 1 shows, available tools to find A/B compartments can be seen in Table 3; c) TADs detection to find CTCF histone as shown in Fig. 2, available tools to detect TAD boundaries can be seen in Table 3; d) loop calling mediated by CTCF and other proteins, available tools to find loops can be seen in Table 3.

Table 3.

Tools to analyze Hi-C data at different structural levels.

Tool	Function/algorithm/download link	Resolution	Reference
loop Detection	loop detection	1 ~ 10 kb
HiCCUPS			[18]
HOMER			[115]
GOTHiC			[116]
Fit-Hi-C			[117]
HiC-DC			[118]
SIP			[119]
cLoops/cDBSCAN			[120]
Mustache			[121]
Chicago			[122]
PSYCHIC			[123]
diffHiC	differential analysis		[124]
FIND			[125]
HICcompare			[126]
TADs Detection		~40 kb
HMM	Directionality Index		[127]
DP	Dynamic programming		[128]
HicSeg	Two-dimensional segmentation		[129]
Arrowhead	Arrow matrix		[18]
insulation score	Insulation Square Analysis		[130]
DHDF	Cluster-based		[131]
TopDom	IdentifyTD, evaluate quality		[132]
TADtree	hierarchical TADs		[133]
TADs_Identification	Spectral identification		[134]
IC-Finder	Hierarchical clustering		[135]
MrTADFinder	network modularity based		[136]
3DNetMod	network modularity based		[137]
HiTAD	domain-based alignment		[138]
rGAMP	Gaussian Mixture model and Proportion test		[139]
HiCDB	local relative, insulation metric		[140]
deDoc	graph structural entropy		[141]
tadbit	breakpoint detection algorithm
TADBoundaryDectector	deepLearning-based		[142]
EAST	Haar-based algorithm		[143]
TADBD	Haar-based algorithm		[144]
TADCompare	Differential TADs		[145]
TADpole	hierarchy of TADs in intra-chromosomal interaction matrices		[146]
SpectralTAD	Spectral cluster		[147]
ClusterTAD	an unsupervised machine learning approach		[148]
Matryoshka	cluster		[149]
A/B compartment
PCA	A/B compartment	100 kb
HOMER			[115]
juicebox			[87]
CscoreTools	https://github.com/scoutzxb/CscoreTool
HiCPro	http://github.com/nservant/HiC-Pro		[151]
3D structure
contact-based
Gen3D	adaptation, simulated annealing, and genetic algorithm	200 kb	[152]
MOGEN	Gradient ascent	200 kb-1 Mb	[153]
GEM	manifold learning	1 Mb	[114]
GEM-FISH	polymer model	5 kb	[154]
SuperRec	multidimensional scaling	100 kb	[155]
distance-based
AutoChrom3D	considering the sequencing depth	8 kb	[73]
ChromSDE	semi-definite embedding approach	500 kb-1 Mb	[103]
ShRec3D	Short-path algorithm	3–150 kb	[156]
FisHiCal	SMACOF algorithm	1 Mb	[100]
MBO	manifold optimization	unknow	[107]
InfMod3DGen	Gradient ascent	unknow	[104]
3D-GNOME	Markov chain, Simulated annealing	1–2 Mb	[93]
Chromosome3D	Simulated annealing	500 kb-1 Mb	[97]
LorDG	lorentzian objective function	500 kb-1 Mb	[111]
HSA	Multi-track modeling, Markov chain, Simulated annealing	25 kb-1 Mb	[157]
miniMDS	Hierarchical modeling	10–100 kb	[108]
TADbit	https://github.com/3DGenomes/tadbit	unknow	[94]
mdsga	genetic algorithm	unknow	[95]
ShRec3+	two-step algorithm	1 Mb	[112]
3DMax	maximum likelihood algorithm	1 Mb	[88]
Hierarchical3DGenome	Hierarchical modeling	1–5 kb	[101]
EVR	Error-Vector Resultant	unknow	[99]
ShNeigh	Gaussian formula	unknow	[158]
Probability-Based
BAC, BACH-MIX	Bayesian Inference	40 kb	[159]
pastis	multidimensional scaling	100 kb-1 Mb	[90]
tRex	Monte Carlo sampling etc.	1 Mb	[92]
PGS	simulated annealing	50 kb-1 Mb	[159]
SIMBA3D	Bayesian Estimation		[160]
CHROMSTRUCT 4	Monte Carlo sampling		[161]
online tools
NDB	https://ndb.rice.edu/		[162]
Csynth	https://csynth.org		[163]
GSDB	sysbio.rnet.missouri.edu/3dgenome/GSDB		[164]
3D-GNOME 2.0	3dgnome.cent.uw.edu.pl/		[165]
3DGD	http://3dgd. biosino.org/		[166]
3DIV	http://3div.kr/		[167]
3DGB	http://3dgb.cs.mcgill.ca/		[168]

^aEach column denotes the key properties of available tools to analyze Hi-C data at different structural levels. ‘Tools’ denotes availability of open-source software for a method. ‘Function/algorithm/download link’ column denotes Function, algorithms used by a method or download link for access,’ Resolution’ column denotes the resolution of Hi-C data described in the published method’s, ‘Reference’ column denotes the references where the methods were published.

Fig. 2.

TADs and loops visualization with annotation. From top to bottom, the following visualization is shown: RNA-seq, Dnasel, CTCF (Broad), H3K27ac annotation and Hi-C contact heatmap. For the Hi-C contact heatmap, Squares of contact frequency along with the diagonal (yellow squares) indicate the TADs, peaks (black points) in the contact heatmap indicate the chromatin loops. (Note: this Figure is drawn using the juicer box tool [87], Data Source: Rao and Huntley et al. [18] Cell GM12878 Hi-C in situ chr1: 63,820,000–72,460,000). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

2.4. Analysis methods for multi-omics data

As shown in Table 4, researchers can perform the following analyses according to their actual research goals: chromatin feature structure identification, correlation analysis among different samples, and Hi-C multi-omics joint analysis. We can analyze RNA-seq, ChIP-Seq, and ATAC-seq data using visualization methods such as box plots, scatter plots, heat maps, and volcano plots to do the following studies: 1) distribution of sequencing reads on the whole genome; 2) statistical information on the enrichment area of sequencing data (Peak); 3) difference analysis of multiple samples; and 4) motif identification.

Table 4.

Common tools for the joint analysis of omics datasets.

Tools	Function/algorithm	Omics data	References
MACS	peak calling	ChIP-seq/RNA-seq/ATAC-seq	[169]
ChIPseeker	peak annotation	ChIP-seq	[170]
HOMER	peak calling and Motif analysis	ChIP-seq/ATAC-seq	[115]
	Calculate ChIP-Seq expression near TSS	ChIP-seq
BEDtools	Extracting promoter sequences	ChIP-seq	[171]
	RNA-seq coverage analysis	RNA-seq
SeqSite	detect transcription factor (TF)	ChIP-seq	[172]
EdgeR	Peak comparisons	ChIP-seq/ATAC-seq	[173]
DESeq2	Peak comparisons	ChIP-seq/RNA-seq/ATAC-seq	[174]
DiffBind	Peak comparisons	ChIP-seq/ATAC-seq	[175]
methyKit	DNA methylation analysis	WGBS	[176]
MethGo	genomic and epigenomic analyses	WGBS/RRBS	[177]
GATK	variant analysis	WGS	[178]
BGI Online	variant analysis	WGS/RNA-seq
SAMtools	variant analysis	WGS/RNA-seq	[179]
limma	differential expression	RNA-seq	[180]
Cufflinks	RNA-Seq analysis workflow	RNA-seq	[181]
RNA-Cocktail	RNA-Seq analysis workflow	RNA-seq	[182]
topGO	GO/KEGG enrichment	RNA-seq
DAVID	GO/KEGG enrichment	RNA-seq	[183], [184]
KOBAS	GO/KEGG enrichment	RNA-seq	[185]

^aEach column denotes the key properties of available tools to analyze other omics data. ‘Tools’ denotes availability of open-source software for a method. ‘Function/algorithm/’ column denotes the primary function, algorithms used by a method. ’Omics data’ column denotes the omics data that serves as the input for the tool. ‘Reference’ column denotes the references where the methods were published.

3. Identification of chromatin structure

3.1. 3D visualization of chromosomes

To better understand the relationship between chromosome structure and function, many researchers have reconstructed the 3D structures of chromosomes based on population Hi-C data at different resolutions, or on single-cell Hi-C data. Using the process shown in Fig. 3, before modeling, we need to obtain clean data by normalizing the raw Hi-C data [70], [71], [72], [73], [74], [75], [76], [77], [78], [79]. These methods can be classified as probability-based, distance-based or contact-based.

Fig. 3.

Process of simulating chromosomes’ 3D structure. First, data is preprocessed to get distance matrix or contact data, which is used as the second step’s input. Second, we could choose or not choose some prior conditions (such as biological characteristics, physical forces, and FISH data) to help us build chromosome reconstruction models. In the past ten years, most researchers chose to reconstruct chromosomes by probability-based inferential, distance-based or contact-based methods. At the end of the analysis, we will have a 3D coordinate set for visualization.

1) Probability-based: some researchers assumed the contact counts of Hi-C data follow a normal distribution [88], [89] or Poisson distribution [90], [91], [92] and designed a transfer function between the distribution intensity and spatial distance to infer a genome structure.

2) Distance-based: most reconstruction methods are based on the principle that the frequency of chromatin interaction is inversely proportional to the bin distance. Researchers [73], [88], [89], [93], [94], [95], [96], [97], [98], [99], [100], [101], [102], [103], [104], [105], [106], [107], [108], [109] determined the chromosomes structures based on optimization functions, such as semi-definite embedding [103], [113], manifold-based [107], [114], simulated annealing [97], Lorentzian objective function [111], maximum likelihood function [88], and shortest path [112].

3) Contact-based: some researchers set the biological characteristics and physical forces of chromosomes as a priori conditions [89], [94], [106], [109], [114], [152], [153], [154], [159], [186], [187], [188], [189], and directly use contact data to optimize the chromosome structure of contacts. Gen3d [152] reconstructs chromosomes with an adaptive, simulated annealing and genetic algorithm. MOGEN [153] uses the optimized scoring function to convert the Hi-C intrachromosomal and interchromosomal contact data into a 3D conformation set by satisfying as many contacts with high probability as possible. TADbit [94], GEM [114], and GEM-FISH [153] reconstruct spatial organization of chromosomes by combining their conformational energy.

The above methods can be used at different resolutions. At Mb-500 kb resolution, we can choose these methods [73], [89], [90], [91], [93], [94], [96], [98], [102], [103], [104], [105], [107], [109], [152], [153], [156], [187], [190], [191], [192], [193], [194] to observe the overall chromosome; at 5 kb resolution, we can choose the following methods [88], [95], [101], [108], [110], [112], [114], [154], [195] to observe the spatial structure of the whole chromosome or the local three-dimensional structure.

3.2. A/B compartment

In 2009, Erez Lieberman-Aiden, et al. [8] found the A/B compartments using principal component analysis. As shown in Fig. 1, the Hi-C heatmap region can be divided into A and B compartments, corresponding to the positive (the blue region) and negative (the read region) parts of the principal eigenvector. Through the study of gene expression levels, histone modifications, and DNase enzyme hypersensitivity sites corresponding to positive and negative regions, they found that in the regions with positive eigenvalue, there are more genes, and the corresponding gene expression levels are relatively high. The signal of the Dnase-sensitive DNA site is also relatively high. These characteristics indicate that these regions are more open and accessible, and the region of transcriptional activation is defined as the A compartment, which corresponds to the open chromatin region; on the contrary, the B compartment corresponds to the closed chromatin region.

In the next few years, most researchers verified the relationships between the structures of the A/B compartments and their functional characteristics by predicting chromosome A/B compartments [196], [197], [198], other researchers developed tools to analyze the A/B compartments, such as HiC-Pro [151] and CscoreTool[150].

3.3. TADs detection

In 2012, the concept of TADs was first proposed by E. P. Nora et al [200] and Dixon et al. [127], to explain the squares in the Hi-C matrix diagonal. They thought the interaction frequency within the TADs was significantly higher than the interaction frequency between two adjacent regions. In 2019, E. D. Wit [201] gave a definition of TADs that considers the mechanisms that shape them: TADs are an emergent property of an underlying biological mechanism, i.e. loop extrusion or compartmentalization and are dynamic genomic regions rather than a static structural feature of the genome. As shown in Fig. 2, a heat map is used to perform TADs boundaries identification analysis. The heat map indicates the chromatin interaction at 10 kb resolution. The genome interaction map is a symmetric matrix, so the information on both sides of the diagonal is equal in Fig. 2. As shown in this figure, let’s just see the upper right corner of heat map, the interaction intensity changes from weak to strong, which is indicated by the color of the cell changing from white to red. We can see some small triangular regions appear on the bottom edge repeatedly, which are depicted in red, indicating that the interaction frequency between chromatin fragments within these regions is high, and the frequency of interaction between adjacent triangular regions is lower. In this heatmap, these regions (yellow boxes in Fig. 2) are called TADs.

In the past ten years, most researchers identified the TADs by extracting one-dimensional features from the two-dimensional interaction matrix for segmentation, or by using the clustering algorithm. Regarding the first method: in 2012, to identify TADs in chromatin, DI (directionality index proposed by J. R. Dixon et al. [127]) was used to quantify the degree of bias in upstream or downstream interactions of genomic regions. By determining DI in the genome, we can determine the location of TAD boundaries in the genome. In the next ten years, many authors continued to improve the TADs recognition algorithm. Dynamic programming [128] has been used to reveal the TADs hierarchical structure. In addition two-dimensional segmentation [129], insulation score [130], laplacian graph clustering method [134], hierarchical clustering [137], unsupervised machine learning methods [148], the modular concept of network science [136], Gaussian Mixture model [139], local relative insulation metrics and multi-scale aggregation methods [140], have all been developed to detect contact domain boundaries. Based on structural information theory, a method called deDoc [141] proposed a solution to predict the structure of high-resolution TADs from low-resolution data.

3.4. Loop calling

People often refer to chromatin interactions as chromatin loops. But there are subtle differences between the two concepts. The chromatin loop is a circular structure formed by folding and wrapping chromatin due to protein and other mediation, the chromatin interaction may only be the product of the random connection of two DNA fragments detected by 3C-based experiments. Loops bring promoters and enhancers to closely together in space to regulate gene expression. With dynamic changes of the loop structures, such as new formation or disappearance, genes’ regulation will be affected to a certain extent [18] [202].

The chromatin loop can be identified by constructing Hi-C maps with a resolution of less than 5 kb. In 2014, S. S. Rao et al. [18] identified the positions of chromatin loops by using HiCCUPS (integrated into the developed juicer [87]) to search for pairs of loci, whose pixels with higher contact frequency than typical pixels in their neighborhood. (These pixels are defined as “peaks” in the Hi-C contact matrix and the corresponding pair of loci are called “peak loci”). As shown by the black mark in Fig. 2, we used the juicerbox software to mark the loops. Many other tools to find loops, such as HOMER [203] and GOTHiC [116], are available and are listed in Table 1.

4. Applications of integrated omics data analysis

As described in section 3 and shown in Fig. 4, we can combine Hi-C data of different resolutions with other omics data regarding epigenetic regulation at different structural levels and gene expression analysis.

Fig. 4.

Analysis of multi-omics data. The superscripts 0, 1, and 2 represent the structure, epigenetic regulation, and expression related sequencing technology and corresponding analysis methods. From the loop level, we can do KEGG pathway and CTCF analysis; from the TAD level, we can use ChIP-seq, ATAC-seq, and WGBS to detect transcription factor sites or do variant calling and histone modification analysis; from the A/B compartment level, we can use ChIP-seq, ATAC-seq, WGBS, and RNA-seq to detect promoter, enhancer, transcription factor sites or conduct gene expression analysis.

4.1. Multi-omics data analysis at the loop level

As discussed in subsection 3.4, we can identify whether the two ends of chromatin loops are regulatory elements and gene loci by determining the locations of loops, so as to obtain a list of genes specifically regulated by different loops [18],[202]. As can be seen in Fig. 4, analyses from a variety of data sets such as DNase-seq, ChIP-seq, RNAseq or ATAC-seq may be combined to draw a whole-genome loop model. Comparing loops of multiple samples, finding loops that have changed at the genome-wide level, and using RNA-Seq to count the expression of related genes, can helps explain the relationship between loops and the potential differences in transcription regulation among different samples.

As shown in Fig. 5, Fig. 5(A) represents the differential interaction map of the whole genome map of GM12878 and K562 cells, Fig. 5(B) represents the differential interaction map of the two cell lines GM12878 and K562 in the region of chromosome 1: 0:2,000,000 bp and the corresponding identification of differential loops, Fig. 5(C, D) uses arc diagrams to visualize the interaction loop’s location and corresponding differential loops of the GM12878 and K562 cell lines in the 0:2,000,000 bp of chromosome 1. After statistically analyzing the differential loops between samples, we find loops that have changed at the genome-wide level, and use RNA-Seq to do differential expression analysis, which helps explain the relationship between loops and gene transcription regulation among different samples.

Fig. 5.

Differential analysis of data from the GM12878 and K562 cell lines. A. Differential heatmap of all chromosomes between the GM12878 lymphocyte line and the K562 cell line. The red color in the figure indicates sites with stronger interactions in the GM12878 cell line than the K562 cell line, and the blued ones indicate weaker interactions in the GM12878 cell line than in the K562 cell line. B. Differential heatmap of chromosome 1 between the GM12878 lymphocyte line and the K562 cell line. The points on the diagonal lines are the identified loops. The dark purple points represent the loops of GM12878, and the dark blue points represent the loops of K562. C-D. The loops and differential loops’ location of the GM12878 lymphocyte line and the K562 cell line chromosome 1 from 0 to 20,000,000 bp. Each arc indicates chromatin interaction from the start site to the end site. E. ChIP-Seq data is used to visualize the H3K27me3 and H3K4me1 histone modification peaks in the GM12878 and K562 cell lines, chromosome 1 range of 0–20,000,000 bp. (note: GM12878 Hi-C in situ and K562 Hi-C in situ data is from reference [18], A-D are drawn with the juicer box tool [87], and the G is drawn with the IGV tool [204]). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Let’s give an example: Rao et al. [18] found 9448 loops in the GM12878 cell line, of which 2854 loops are related to known promoter-enhancer functions. The expression of the gene's promoter with a loop was significantly higher than without a loop. For example, there is a loop in the GM12878 cell line, which is connected to the SELL promoter and a distal enhancer SELP, where the gene transcription is turned on, and the expression is increased. However, there is no loop in the same location in the IMR90 cell line, and the gene is not expressed. Greenwald, et al. [202] used a Hi-C experiment to generate high-resolution chromatin loops of pancreatic islets in three samples, as well as ATAC-seq, and ChIP-seq data to identify the target genes of pancreatic islet enhancers. Finally, these loops were annotated with target genes of islet enhancer, which shows that enhancer looping is correlated with islet-specific gene expression.

4.2. Multi-omics data analysis on TADs level

As section 3.3 has described, although TADs are statistical constructs rather than structural components of the 3D genome, TADs are an emergent property of an underlying biological mechanism, i.e. Loop extrusion or compartmentalization [201], if TADs boundaries are destroyed, the loops structure may change. so, analyzing differential TADs boundaries along with many other omics data as described below can help to understand the relationship between changes in loops/compartments and their functions.

1) Differential TADs boundaries region gene expression can be analyzed by using RNA-seq and Hi-C data, as described in reference [205].

2) Differential TADs boundaies and CNV identification and analysis of gene expression in related regions by using RNA-seq, WGS and Hi-C data can be done as in references [206], [207].

3) Analysis of the distribution of transcription factors and binding sites, histone modifications in differential TADs’ boundaries region by using ChIP-seq and Hi-C data, can be performed as in references [208], [209].

For example, in 2017, Rubin et al. used Hi-C and ChIP-seq data [209] to jointly analyze the interaction patterns of enhancers and promoters throughout the genome during the differentiation of isolated and cultured human primary keratinocytes. They confirmed two types of enhancers-promoter interactions: one is a ‘gaining’ interaction, which is enhanced during differentiation and is consistent with the enhancer obtaining H3K27ac activation marker; the other is a ‘stable’ interaction with enhancers constitutively marked by H3K27ac. Furthermore, these two interactions were not detected in pluripotent cells, suggesting that this lineage-specific chromatin structure was established in precursor cells and remodeled during terminal differentiation.

In reference [207], the authors performed Hi-C, whole-genome sequencing (WGS), ChIP-seq, and RNA-seq on two multiple aneuploidy myeloma (MM) cell lines to study the 3D genome structure of multiple myeloma (MM) and its relationship with genomic variation and gene expression. The authors found that the average interaction count inside each CNV block was positively correlated with its copy number, which indicates that raw interaction counts in cancer Hi-C data are biased by CNVs. This suggests that we can detect CNV by inferring the interaction counts of Hi-C data. Similarly, combining Hi-C and WGS data can improve the detection of translocations. The CNV breakpoints and TAD boundaries significantly overlapped. Compared with normal B cells, the number of TADs in MM increased by 25%, the average size of TADs was smaller, and about 20% of the genomic regions switched their chromatin A/B compartments type.

4.3. Multi-omics data analysis on the A/B compartment level

On the A/B compartment level, RNA-seq, ChIP-seq, ATAC-seq, WGS, or WGBS data is analyzed with Hi-C data as follows. For example, biological changes in myocardial cells are a mainly cause of heart failure. M. Rosagarrido et al. [210] found that this problematic cell function failure results from gene expression changes and is affected by transcription factors and chromatin remodeling enzymes. In reference [210], a chromatin conformation study of myocardial cell lines induced by load stimulation and mouse myocardial cell lines lacking CTCF function was conducted with Hi-C and RNA-seq. The analysis explores the effect of the entire genome structure on heart failure—generally, changes in A/B compartments correlated with gene expression. A change from the A compartment to the B compartment correlated with the down-regulation of gene expression, while a change from the B compartment to the A compartment correlated with the up-regulation of gene expression. This study's transcriptome data showed that most genes in regions with changed A/B compartments occurred in diseased cell lines, and the regulation of the expression of the sera changed, including the activation of some pathogenic marker genes.

Reference [211] used mouse embryonic stem cells to comprehensively study the effects on the three-dimensional structure (Hi-C) and on the chromatin accessibility (ATAC-seq), caused by the knockout of the methyltransferase complex subunit MLL2. Authors also studied alterations in protein modification (ChIP-seq), and gene expression levels (RNA-seq) casued by the MLL2 knockout. They found that the deletion of MLL2 increased the Polycomb complex's occupation, reshaped long-distance gene interaction and histone modification, reduced the transcription levels of critical genes, and ultimately caused abnormal embryonic development.

5. Conclusion

Overall, the Hi-C and other omics data analysis methods have appeared in different ways to help researchers understand the relationship between function and genome structure. However, there are still many improvements that could be made to the Hi-C analysis methods.

For the 3D structure analysis, although we have many methods to simulate the 3D structure of the genome, we still face many obstacles: (1) How can we simulate the genome structure of at a resolution even higher than 1 kb? Maybe we can apply deep learning methods to 3D structure simulation; (2) How can we improve microscope resolution to see the genome structure at kb resolution, to verify the accuracy of the three-dimensional simulated structure? Maybe we can use image super resolution technology to enhance the images from microscope; (3) How can we verify the accuracy of the 3D reconstructed model, and not just compare with FISH (Fluorescence in situ hybridization) data? (4) How can we detect TADs or loops when obtaining 3D structures in order to understand the relationship between structure and function?

Many methods were proposed to detect TADs and loops, but how can we detect TADs or loops from low-resolution Hi-C data? One way is by detecting them from enhanced Hi-C data. How can we detect them more accurately, faster and with fewer parameters? These problems still need to be solved.

There are many problems to deal with in order to observe an accurate genome structure, we still have additional methods to help us understand the relationship between function and genome structure using multi-omics analysis, for example, CRISPR/CAS9 genome editing technology. Ultimately these approaches will help us to develop cancer treatments and accelerate drug development.

Author contributions

All authors contributed to the conception and writing of this manuscript.

Declaration of Competing Interest

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