Elucidation of Biological Networks Across Complex Diseases Using Single-Cell Omics

Abstract

Single-cell multimodal omics (scMulti-omics) technologies have made it possible to trace cellular lineages during differentiation and to identify new cell types in heterogeneous cell populations. The derived information is especially promising for computing cell-type-specific biological networks encoded in complex diseases and improving our understanding of the underlying gene regulatory mechanisms. The integration of these networks could, therefore, give rise to a heterogeneous regulatory landscape (HRL) in support of disease diagnosis and drug therapeutics. In this review, we provide an overview of this field and pay particular attention to how diverse biological networks can be inferred in a specific cell type based on integrative methods. Then we discuss how HRL can advance understanding of regulatory mechanisms underlying complex diseases and aid in the prediction of prognosis and therapeutic responses. Finally, we outline challenges and future trends that will be central to bringing the field of HRL in complex diseases forward.

SIGNIFICANCE IN ELUCIDATING GENE REGULATORY MECHANISM OF COMPLEX DISEASES

Fundamental biological questions include how individual cells differentiate into various tissues/cell types, how cellular activities are operated in a coordinated manner, and what gene regulatory mechanisms support these processes [1, 2]. Gene regulatory mechanisms involve diverse procedures carried out by cells to increase/decrease the production of specific cell products (metabolites or RNA). They mainly involve the regulation of the rate of transcription and translation, the processing of RNA molecules (e.g., alternative splicing), the control of the mRNA molecule stability, and the modulation of the pathway behavior and interactions [3].

Disorders in the above regulatory activities usually relate to the occurrence and development of complex diseases, which refer to diseases caused by a combination of genetic, environmental, and lifestyle factors [4–9]. Examples of complex diseases include asthma [10], epilepsy [11], hypertension [12], Alzheimer’s disease (AD) [13], manic depression [14], schizophrenia [15], cancer [16], diabetes [17], and heart diseases [18]. Early detection of these diseases relies on the identification of predictive biomarkers, which plays an important role in improving prognosis and identifying appropriate therapeutics. Although some predictive diagnostic tests, e.g., MammaPrint and Oncotype DX, have been identified as biomarkers and applied to clinical practice for breast cancer prognosis, it remains a challenge to reverse-engineer the gene regulatory mechanisms underlying these biomarkers [19]. Furthermore, gene regulatory mechanisms disrupted by genotype aberrations may cause altered expression of genes and lead to pathogenic phenotypes [20, 21]. For example, genetic aberrations that disrupt topologically associated domains (TAD) were found to rewire promoter-enhancer interactions and lead to limb malformations [20]. Accurately translating genotype to pathological phenotype requires effective approaches to unravel the gene regulatory mechanisms.

Meanwhile, discovering the pathways that are perturbed and dysregulated in each disease type is critical for elucidating the gene regulatory mechanisms relating genotypes to phenotypes and may direct efforts to develop treatments. Pathway perturbation encompasses functional alterations of a pathway induced by internal or external factors, and pathway dysregulation refers to metabolic disorders and cellular homeostasis disruption [22]. From a systems biology viewpoint, perturbed or dysregulated pathways foster the development or deterioration of complex diseases [5]. Therefore, interpretation of pathway perturbations and dysregulations could shed light on improving diagnosis and personalized treatments [5].

HETEROGENEOUS REGULATORY LANDSCAPE ENABLED BY SINGLE-CELL OMICS TECHNOLOGIES

Although bulk omics approaches for interrogating genetic variations that drive the regulation of pathways have determined genetic aetiologies for thousands of human diseases, they are typically limited to studies at the ecosystem or organism level [23, 24]. Practically, the diversity of individual cells within an ecosystem is far greater than we can measure by studying a mixture of cells of the same organism, and the genomes within the cells of an individual multicellular organism are not always the same [23]. In these settings, omics assay at the single-cell level represents a powerful, high-resolution tool for probing specific molecular mechanisms and pathways and uncovering the nature of cell heterogeneity [22, 25].

The fast-development of single-cell omics technologies has enabled comprehensive profiling of genetic, epigenetic, spatial, proteomic, and lineage information [23, 26, 27]. These technologies have provided exciting opportunities for systematical investigation of rare cell types, cellular heterogeneity [28], evolution [28], and cell-to-cell interactions [29] in a wide range of tissues and cell populations [30–32]. The generated multimodal information from individual cells has enabled the elucidation of the cellular reprogramming, developmental dynamics, communication networks in disease development, and identify unique malfunctions of individual cells [1, 2, 26, 30, 31, 33–35]. Interrogation of the contributions of individual cells to pathogenesis and disease progression [23], especially the immune cells [35], enables aetiology interpretation and therapeutics identification in complex diseases [30].

A central question that arises now is how we can discover the network models that control cellular differentiation and drive transitions from one cell type to another. Specifically, a co-expression network (GCN) is a comprehensive compilation of co-expression linkages among genes, which can be inferred from single-cell RNA-Sequencing (scRNA-Seq) [36–41]. From scRNA-Seq data, gene regulatory networks (GRN) can be constructed to model the regulatory interactions between TFs/non-coding-RNAs and target genes [25, 42–54]. Computed from single-cell ATAC-Sequencing (scATAC-Seq), cis-co-accessibility networks (CCAN) in single cells indicate how accessible cis-regulatory elements (CRE) orchestrate gene regulatory mechanisms dynamically [55–57]. Methylation-associated GRNs (MGRN) could be inferred from single-cell DNA methylation sequencing (scMethyl-Seq) to capture the impacts of epigenetic factors on gene regulatory mechanisms [58, 59]. To quantify the interplays between chromatin loci in a 3-D space, chromatin interaction networks (CIN) in individual cells can be identified from single-cell Hi-C (scHi-C) data [60]. In single cells, CRE-gene interaction networks (CGN) can be constructed through the integration of scRNA-Seq and scATAC-Seq, which provide a detailed view of how CREs influence gene expression in single cells [61–65]. Similarly, TF-CRE interaction networks (TCN) in individual cells are mainly inferred by integrating scRNA-Seq and scATAC-Seq [61–65]. Although each network provides a different glimpse into the entire cell heterogeneity [66], they can jointly build up a heterogeneous regulatory landscape (HRL) that characterizes cell-type-specific genetic and epigenetic regulatory relationships. All the above networks are thus called HRL-associated networks, and they are outlined in Figure 1, where their applications to complex diseases are also summarized.

Figure 1.

Graphical elucidations of different networks using single-cell data and applications to complex diseases. Abbreviations: gene co-expression network (GCN), gene regulatory network (GRN), cis-co-accessibility network (CCAN), methylation-associated gene regulatory network (MGRN), chromatin interaction network (CIN), CRE-gene interaction network (CGN), and TF-CRE interaction network (TCN).

Moreover, single-cell multimodal omics (scMulti-omics) opens up new frontiers by simultaneously measuring multiple modalities. In scMulti-omics, the information of one modality can be leveraged to improve the interpretation of another. Currently, at most four types of single-cell omics can be measured simultaneously, leading to 13 combinations [66]. Specifically, there are nine double-modality sequencing techniques (e.g., sci-CAR-Seq combines RNA expression and chromatin accessibility), three triple-modality sequencing techniques (e.g., scNMT-Seq combines RNA expression, chromatin accessibility, and DNA methylation), and one quad-modality sequencing technique (ECCITE-Seq combines RNA expression, protein expression, T cell receptor, and perturbation). Hence, scMulti-omics has brought about new resources for understanding HRL [66].

Overall, HRL underlies the multi-level genotype and phenotype observations across diverse cell types. Changes of HRL alongside the cell developmental trajectory reflect the pathological process in complex diseases. By comparing HRLs across different cell populations, the discovery of the causal genetic patterns driving the cell developmental trajectories can assist in the prediction of prognosis and treatment responses for complex diseases. Besides, HRL diversity between cell populations can be exploited to guide the identification of some rare cell populations associated with complex diseases [67].

HRL-ASSOCIATED NETWORK INFERENCE FROM SINGLE-CELL OMICS DATA IN COMPLEX DISEASES

All aspects of life activities or complex diseases in living cells are mediated by various regulatory mechanisms that are jointly characterized by an HRL. HRL-associated networks can then be built to delineate the interplays between molecules within the HRL. These networks, including GCN [68], GRN [42], CCAN [69], MGRN [70], CIN [71], CGN [72], and TCN [73], were originally defined on tissue level and then extended to individual cells in single-cell data. Some CCANs seek to exploit the co-accessibility between promoters, which is called promoter-promoter co-accessibility network [74]. CINs can be further classified into promoter-promoter interaction networks and promoter-enhancer interaction networks [75]. Rapid advances in single-cell sequencing technologies have provided exciting opportunities for constructions of HRL-associated networks and applications to complex diseases. Taking leukemia as an example, five HRL-associated networks have been applied including GCN, GRN, MGRN, CGN, and TCN (Figure 1). Specifically, the GCN of acute myeloid leukemia (AML) discovered aberrant co-expression between stemness and myeloid priming genes in primitive AML cells [76]. SINCERITIES was utilized to infer the GRN driving the differentiation of monocytic THP-1 human myeloid leukemia cell into macrophages [47]. The MGRN constructed in chronic lymphocytic leukemia (CLL) suggested that conserved CpG sites are protected from DNA methylation by TF binding through direct exclusion of methylases or negative selection caused by disrupted CLL regulatory code [58]. CGN was leveraged to detect putative cis-regulatory linkages of genes specific to mixed-phenotype acute leukemia (MPAL) [64]. Meanwhile, TCN was inferred to identify TFs that regulate the leukemia genes [64]. Besides, CGN was also applied to CLL and discovered several cell clusters enriched for enhancers specific to CLL cells [65], and meanwhile TCN was constructed to identify a consistent regulatory program in CLL cells [65]. In the following sub-sections, we provide an overview of the available tools for HRL-associated network inference. Around 21 examples were collected to elucidate seven HRL-associated networks in Table 1, covering the details such as networks, diseases, sequencing methods, number of cells, and accession IDs.

Table 1.

Examples of HRL-associated network inference and applications in complex diseases.

Network	Disease	Seq method	#Cell	Tool	Main contribution	Accession	Ref
Applications to complex diseases
GCN	Prostate cancer	scRNA-Seq	144	WGCNA	Identify key biological processes regulated by the androgen receptor, which is essential to the development of normal and cancer prostate glands.	GSE99795	[77]
	AML	scRNA-Seq	38,410	-	Discovered aberrant co-expression between stemness and myeloid priming genes in primitive AML cells, indicating dysregulated transcriptional programs in malignant cells.	GSE116256	[76]
	Melanoma	scRNA-Seq	4,645	COAC	Constructed a localized GCN to predict the survival rate of melanoma patients.	https://github.com/ChengF-Lab/COAC	[78]
GRN	RMS	scRNA-Seq	9,970	-	Identified driver TFs, e.g., SOX8, underlying childhood RMS by dissecting the GRN.	GSE116344	[79]
	T2D	scRNA-Seq	2,491	-	Discovered the altered GRN architecture in the pancreas from T2D patients.	http://sandberg.cmb.ki.se/pancreas/	[52]
	AD	scRNA-Seq	22,951	-	Identified rewiring of microglia gene regulation of AD by comparing the characteristics of GRNs in AD and controls.	GSE98971	[52]
	Melanoma	scRNA-Seq	4,645	ACTION	Use cell-type-specific GRN to identify several novel biomarkers (e.g., CCND1 and SNAI2) and TF-gene linkages (e.g., MITF-CTSK), which may be associated to melanoma.	Upon reasonable request	[80]
	Melanoma	scRNA-Seq	674	GENIE3	Infer GRN from melanoma scRNA-Seq data and identified TFs such as SOX10 and RXRG as central hubs governing the neural crest stem cell states.	GSE116237	[81]
	Luminal breast cancer	scRNA-Seq	9,979	PIDC	Leveraged PIDC to identify the regulatory relationships between genes from the scRNA-Seq profiles of CD44high and CD44low cells in luminal breast cancer.	GSE122743	[82]
	Uveal melanoma	scRNA-Seq	8,598	GRNBoost2	Utilized GRNBoost2 to infer GRN from scRNA-Seq profiles of uveal melanoma tumor cells, and identified gene modules together with their regulatory motifs.	GSE139829	[83]
	Myeloid leukemia	scRNA-Seq	960	SINCERITIES	Used SINCERITIES to infer the GRN driving the differentiation of monocytic THP-1 human myeloid leukemia cell into macrophages.	https://cabsel.ethz.ch/tools/sincerities.html	[47]
CCAN	T2D	scATAC-Seq	1,456	Cicero	utilized Cicero to nominate the cell-specific target genes at 104 non-coding T2D genome-wide association studies signals.	dbGap: phs001188.v2.p1	[84]
	AD	scATAC-Seq	∼100,000	Cusanovich 2018	Identify cell-type-specific enrichments of the heritability signal for hundreds of complex traits and diseases, such as heritability for AD in microglia.	GSE111586	[56]
	Heart diseases	scATAC-Seq	∼80,000	JRIM	Reported several mutations associated with tissue-related heart diseases such as myocardial hypoplasia, defective atrial, and congenital heart defects.	GSE111586	[57]
MGRN	CLL	scMethyl-Seq	1,821	-	Suggested that conserved CpG sites are protected from DNA methylation by TF binding through direct exclusion of methylases or negative selection caused by disrupted CLL regulatory code.	GSE109085	[58]
	Breast cancer	scMethyl-Seq	-	iRegulon	Found that stem- and proliferation-related TFs are transcriptionally active in circulating tumor cells, which correlates with a poor prognosis in breast cancer.	ENA: PRJEB25101	[59]
CIN	Breast cancer	scHi-C	-	-	utilized CIN to identify 110 putative target genes, e.g., PEX14 and DLX2, from interaction peaks within risk loci.	PRJEB23968	[85]
CGN/TCN	CLL	scRNA-Seq, ATAC-Seq	43,049	-	Identify a consistent regulatory program starting with a sharp decrease of NF-κB binding in CLL cells, which is followed by reduced activity of lineage-defining transcription factors, erosion of CLL cell identity, and acquisition of a quiescence-like gene signature.	GSE111015	[65]
	MPAL	CITE-Seq, scATAC-Seq	35,882	Seurat, Cicero	Identified 91,601 peak-to-gene links, and inferred RUNX1 as a potential oncogene of MPAL.	GSE139369	[64]
Other applications using single-cell data
GCN	-	scRNA-Seq	466	SCIMITAR	Revealed 92 progression-associated genes and uncovered three co-regulatory states from human fetal neurons.	GSE67835	[40]
GRN	-	scRNA-Seq	376	GRNVBEM	Identified 491 cell type marker genes for each of the four sample groups associated with low and high expression of CD41 in zebrafish, and detected two hub genes, slc2a6 and csf3r.	ArrayExpress: E-MTAB-3947	[48]
	-	scRNA-Seq	148–5,069, 49–8,522	IRIS3	Demonstrated superior performance compared to SCENIC on 19 scRNA-Seq datasets as well as reproducibility and robustness.	https://bmbl.bmi.osumc.edu/iris3/	[54]
CIN	-	scHi-C	10,696,180	scHiCluster	Identified TADs within single cells after imputation, and allowed single-cell clustering.	GSE84920, GSE80006	[60]

GCN inference in prostate cancer, melanoma and AML from scRNA-Seq

The inference of GCNs (Figure 2A) is advancing rapidly and is generating many new insights into the prediction of prognosis and treatment responses of prostate cancer [77], AML [76], and melanoma [78]. For example, Chen and colleagues used GCNs to identify key biological processes regulated by the androgen receptor, which is essential to the development of normal and cancer prostate gland [77]. Galen and collaborators discovered aberrant co-expression between stemness and myeloid priming genes in primitive AML cells, unveiling dysregulated transcriptional programs in malignant cells [76]. Peng and colleagues constructed a localized GCN using a subset of the cells from a melanoma patient-derived scRNA-Seq dataset and then used it to predict survival rates [78].

Figure 2.

Illustration of each type of HRL-associated networks in the context of evaluating cancer heterogeneity. (A) A GCN is an undirected graph where each node represents a gene, and a pair of genes are connected by an edge if there is a significant co-expression relationship between them based on available scRNA-Seq data. (B) In a GRN, each edge is directed from the regulator to the target gene corresponding to activation or inhabitation. (C) A CCAN is an undirected graph, where each node represents a CRE, and a pair of CREs are connected by an edge if there is a significant co-accessibility relationship between them based on available scATAC-Seq data. (D) An MGRN is a GRN where regulator-gene regulatory relationships are influenced by methylation states. (E) A CIN is an undirected graph, where each node represents a chromatin locus (CL), and a pair of loci are connected by an edge if a significant interaction exists between them based on scHi-C data. (F) A CGN is a directed graph with nodes representing CREs and genes and edges corresponding to the cis-regulatory relationships between the two kinds of nodes. (G) A TCN is a directed graph, where nodes denote TFs and CREs, and edges represent the binding activities of TFs to CREs.

Until now, a few tools have been developed to infer GCNs, among which LEAP, PPCOR, and SCIMITAR are three widely-used ones [39–41]. LEAP is an R package that utilizes the pseudo-time of the cells to find gene co-expression involving time delay from a scRNA-Seq dataset [39]. Harly and collaborators applied LEAP to identify the co-expressed genes alongside the development of innate lymphoid cells [86], an important immune cell in diseases such as asthma, autoimmune diseases, allergic rhinitis, and inflammatory bowel disease. PPCOR is an R package used for computing the partial/semi-partial correlations between gene expressions relative to the other genes [41]. SCIMITAR is a Python package to infer GCN dynamics throughout the biological progression from single-cell transcriptomes [40]. Although the application value of PPCOR and SCIMITAR to complex diseases remains unexplored, they possess such potential in the future.

GRN inference in rhabdomyosarcoma (RMS), luminal breast cancer, myeloid leukemia, melanoma, type 2 diabetes (T2D) and AD from scRNA-Seq

The applications of GRNs (Figure 2B) mainly aim at RMS [79], melanoma [80, 81], T2D [52], luminal breast cancer [82], myeloid leukemia [47], and AD [52]. For instance, by dissecting the GRN, Gryder and collaborators identified the driver TFs (e.g., SOX8) underlying childhood RMS [79]. Lacono et al discovered the altered GRN architecture in the pancreas from T2D patients [52]. In addition, they also identified the rewiring of microglia gene regulation of AD by comparing the characteristics of GRNs in AD and controls [52]. Mohammadi et al reconstructed a cell-type-specific GRN for the newly-discovered melanoma subtype from MITF-associated melanoma patients [80]. The GRN highlighted several novel biomarkers (e.g., CCND1 and SNAI2) and TF-gene linkages (e.g., MITF-CTSK), potentially associated with melanoma.

Presently, many GRN inference efforts have been performed via developing tools such as CSN [51], PIDC [50], GENIE3 [42], GRNBoost2 [83], Jump3 [44], SCODE [45], SINCERITIES [47], GRNVBEM [48], SCNS [49], and IRIS3 [54]. GENIE3 is a widely-used Bioconductor package for GRN inference [42], which has been integrated into several scRNA-Seq analysis pipelines, e.g., SCENIC and Arboreto [43, 46]. Rambow et al. used GENIE3 to infer GRN from melanoma scRNA-Seq data and identified TFs such as SOX10 and RXRG as central hubs governing the neural crest stem cell states [81]. PIDC is a Julia package developed based on multivariate information theory [50]. Hong et al leveraged PIDC to identify the regulatory relationships between genes from the scRNA-Seq profiles of CD44^high and CD44^low cells in luminal breast cancer [82]. GRNBoost2 is an efficient GRN inference tool integrated into the Python package, Arboreto [43]. Using GRNBoost2, Durante and colleagues inferred GRN from scRNA-Seq profiles of uveal melanoma tumor cells and identified gene modules together with their regulatory motifs [83]. SINCERITIES was designed based on linear regression and temporally dynamical changes in expression and implemented using R and MATLAB, respectively [47]. The efficacy of SINCERITIES was demonstrated by identifying GRN driving the differentiation of monocytic THP-1 human myeloid leukemia cells into macrophages [47]. IRIS3 is an integrated web server for cell-type-specific regulon (CTSR) prediction from human or mouse scRNA-Seq data [54]. Based on the reasoning that a CTSR can reliably characterize and distinguish the cell types, these CTSRs can aid in the elucidation of regulatory mechanisms. Albeit efforts to apply the other GRN inference tools, such as CSN, Jump3, SCODE, GRNVBEM, SCNS, and IRIS3, to complex diseases are still in their infancy, they will be essential for a deeper understanding of complex diseases.

CCAN inference in AD and T2D based on scATAC-Seq

Lately, applications of CCANs (Figure 2C) aim to provide insight into the regulatory mechanisms underlying AD [56] and T2D [84]. To date, several tools have been proposed to infer CCANs, including but not limited to, Cicero [55], Cusanovich2018 [56], and JRIM [57]. Cicero is an R package to predict cis-regulatory interactions from scATAC-Seq data [55]. Rai et al utilized Cicero [55] to nominate the cell-specific target genes at 104 non-coding T2D genome-wide association studies signals [84]. Cusanovich2018 focuses on identifying cell types, defining candidate tissue-specific enhancers, modeling TF regulatory grammar that specifies each cell type, and linking distal CREs to promoters of target genes [56]. Cusanovich2018 was used to identify cell-type-specific enrichments of the heritability signal for hundreds of complex traits and diseases, such as heritability for AD in microglia [56]. JRIM is an R package based on Cicero, aiming to investigate common and specific regulatory mechanisms across different cell types [57]. Based on the changes in promoter-associated interactions predicted by JRIM, several mutations were reported to be associated with tissue-related heart diseases such as myocardial hypoplasia, defective atrial, and congenital heart defects [57].

MGRN inference in breast cancer and CLL based on scMethyl-Seq

DNA Methylation is involved in various cellular processes, including X chromosome inactivation, genomic imprinting, and silencing of transposable elements [87], whereas aberrant DNA methylation relates to several diseases, e.g., breast cancer [59] and CLL [58]. MGRN (Figure 2D) has been applied to the prediction of prognosis and treatment responses of breast cancer and CLL [58, 59]. For example, Gaiti et al. utilized MGRN to explore the impact of methylation on transcriptional activity and cellular phenotypes in CLL [58]. Experimental results suggested that conserved CpG sites are protected from DNA methylation by TF binding through the direct exclusion of methylases or negative selection caused by disrupted CLL regulatory code. Using MGRN, Gkountela and collaborators found that stem- and proliferation-related TFs are transcriptionally active in circulating breast cancer cells [59]. Though many single-cell scMethyl-Seq techniques are developed, they all suffer from low sequence coverage and robustness, as well as high costs and handling time [87]. Hence these MGRN inference tools have not been applied to complex diseases, yet it is possible that advancing of sequencing technologies will bring wide application space for these tools.

CIN inference in breast cancer from single-cell Hi-C

Clinical applications of CINs (Figure 2E) focus mainly on breast cancer [85]. For example, Baxter and colleagues utilized CIN to identify 110 putative target genes, e.g., PEX14 and DLX2, from interaction peaks within risk loci [85]. Presently, several tools have been proposed to infer CINs, e.g., scHiCluster [60]. scHiCluster is an R package for clustering on Hi-C contact matrices based on imputations using convolution and random walk. After imputation using scHiCluser, TAD can be identified in single-cell [60]. Although there is no study that applies scHiCluster to complex diseases, scHiCluster lays a basis for future advances.

Meanwhile, two HRL-associated networks, i.e. CGNs (Figure 2F) and TCNs (Figure 2G), are inferred from multiple single-cell modalities rather than from single modality solely. Applications of CGNs and TCNs mainly concentrate on leukemia (Table 1) [64, 65]. CGN was used to identify 91,601 putative cis-regulatory linkages of the leukemia-specific genes, such as marker gene CD69, in MPAL [64]. Besides, several TFs (e.g., RUNX1) that regulate the leukemia genes were also discovered through the inference of TCN [64]. Additionally, CGN identified four clusters in CLL, which are enriched for enhancers specific to CLL cells and/or B cells, B cells and T cells, NF-κB binding sites, and transcribed regions marked by H3K36me3 in hematopoietic cells, respectively [65]. Meanwhile, TCN discovered a consistent regulatory program starting with a sharp decrease of NF-κB binding in CLL cells, which is followed by reduced activity of lineage-defining transcription factors, erosion of CLL cell identity, and acquisition of a quiescence-like gene signature [65]. More examples of network inference from scMulti-omics will be discussed in the next section and Table 2.

Table 2.

Examples of scMulti-omics integration tools and applications in complex diseases.

Seq methods	Inputa	Disease	Cell #	Tool	Main contribution	Accession	Ref
Applications to complex diseases
HDST	RE+SI	Breast cancer	50,000	-	Distinguished cell types and niches from a breast cancer resection.	Single Cell Portal	[88]
CITE-Seq	RE+ES	Salivary gland squamous cell carcinoma	∼23,000	-	Identified two subpopulations of cancer stem and basal cells, and four macrophage subpopulations.	GSE124425	[89]
Other applications using single-cell data
scRNA-Seq, STARmap	RE+SI	-	71,000+2,500	LIGER	Spatially located fine subtypes of cells present in the mouse frontal cortex and predicted even complex spatial expression patterns across many individual genes.	GSE126836	[63]
scNMT-Seq	RE+CA+DM	-	1,828	MOFA+	Suggested that independent cell lineages are characterized by different epigenetic variations, such as global changes in ExE endoderm, and local patterns in embryonic endoderm and mesoderm.	GSE121708	[91]
scRNA-Seq, scATAC-Seq	RE+CA	-	464+415	CoupledNMF	Demonstrated that peaks specific to various mESC cell clusters are enriched in different tissues such as heart, forebrain, and midbrain.	GSE115968, GSE115970	[62]
scRNA-Seq, scATAC-Seq, HiChIP	RE+CA+CI	-	464+96	DC3	Simultaneously identify cell subpopulations and deconvolve the bulk data into subpopulations-specific data. Moreover, the accessibility, expression, and loop profiles inferred by DC3 lay a basis for further analyses of the regulatory systems, such as constructing subpopulation-specific gene regulatory networks.	GSE115968, GSE107651, GSE127807	[61]

^aInput data-type combination examples using the corresponding tools: CA, chromatin accessibility; CI, chromatin interactions; DM, DNA methylation; ES, epitope sequencing; PE, protein expression; RE, RNA expression; SI, spatial information.

Generally, network inference methods relying on correlation calculation (e.g., semi-partial correlation and covariance) act as the main forces in networks including GCNs, CCANs, and MGRNs. The GCN inference methods, WGCNA [77], COAC [78], SCIMITAR [40], LEAP [39], and PPCOR [41], are all developed based on correlations. Similarly, the three CCAN inference methods, Cicero [55], JRIM [57], and Cusanovich2018 [56], also depend on covariance, a kind of correlation metric. GRN inference encompasses a variety of methods, such as regression models (GENIE3, GRNBoost2, and SINCERITIES) [42, 43, 47], information theory (PIDC) [50], Bayesian models (GRNVBEM) [48]. Furthermore, SCENIC and IRIS3 reconstruct GRNs from scRNA-Seq data by integrating TFs to co-expressed gene modules [46, 54]. Likewise, iRegulon infers MGRN via associating TFs to a gene list [59]. Chromatin interactions in CINs can be captured directly by scHi-C [60].

INTEGRATION OF SCMULTI-OMICS DATA FOR HRL CONSTRUCTION

Although these HRL-associated networks yield informative insights into cellular diversity and disease development, they measure only particular aspects of cellular identity, motivating the need to jointly build an overall HRL based on these HRL-associated networks (Figure 3).

Figure 3, Key Figure.

Towards HRL construction from the integration of single-cell multimodal omics.

Indeed, scMulti-omics technologies encompass the diverse characterization of a single-cell’s DNA methylation, chromatin accessibility, RNA expression, protein abundance, gene perturbation, and even spatial information [66]. Correspondingly, a few studies were conducted focused on the integration of different single-cell technologies, allowing for a deeper understanding of cellular reprogramming in complex diseases. For example, Vickovic et al developed a new single-cell technology by combining transcriptome and spatial information and demonstrated its clinical potential by distinguishing cell types and niches in a breast cancer resection [88]. Praktiknjo and collaborators utilized single-cell transcriptome and epitope profiling to identify two subpopulations of cancer stem and basal cells, and four macrophage subpopulations [89]. By combining flow cytometry, scRNA-seq, and ATAC-Seq, Rendeiro and colleagues identified a consistent regulatory program starting with a sharp decrease of NF-κB binding in CLL cells, which is followed by reduced activity of lineage-defining transcription factors, erosion of CLL cell identity, and acquisition of a quiescence-like gene signature [65].

Several integration tools, such as Seurat [90], MOFA+ [91], LIGER [63], CoupledNMF [62], and DC3 [61], have been developed and even applied to scMulti-omics data of complex diseases. Seurat is a multifunctional R package for scRNA-Seq and scMulti-omics analysis based on canonical correlation analysis [90]. Granja and colleagues integrated scRNA-Seq and scATAC-Seq data using Seurat and inferred RUNX1 as a potential oncogene of MPAL [64]. LIGER is an R platform allowing for the integration of gene expression, epigenetic regulation, and spatial relationships across single-cell datasets [63]. Experiment on unmatched scRNA-Seq and DNA methylation dataset showcased its ability to identify methylation regions anticorrelated with Arx expression [66]. MOFA+ is an R/Python package for comprehensive and scalable integration of single-cell multimodal data [91]. MOFA+ was applied to a scNMT-Seq dataset in which RNA expression, chromatin accessibility, and DNA methylation were assayed [91]. Experimental results demonstrated that independent cell fate commitment events undergo different modes of epigenetic variations, such as global changes in ExE endoderm and local patterns in embryonic endoderm and mesoderm. CoupledNMF is a Python tool to integrate gene expression and chromatin accessibility that are not measured on the same cell [62]. CoupledNMF enables a systematic mapping of CREs to genes and TFs to CREs, informative for downstream analyses such as inferring cell-cluster-specific regulatory networks at the single-cell level. The cell-cluster-specific gene expression profiles and chromatin accessibility profiles computed by CoupledNMF derived useful insights into the constituent subpopulations of a mouse embryonic stem cells (mESC) [62]. The results showed that peaks specific to various subpopulations are enriched in different tissues such as heart, forebrain, and midbrain. DC3 is a Python package for deconvolution and coupled clustering from bulk and single-cell genomics data (e.g., RNA-Seq, ATAC-Seq, and HiChIP) [61]. Similar to CoupledNMF, DC3 can also construct cell-cluster-specific regulatory networks based on the identified TF-CRE-gene triplets [61]. The experiment on a mixture of scRNA-Seq, scATAC-Seq, and bulk HiChIP from mESC demonstrated that DC3 can deconvolve bulk profiles into subpopulation-specific profiles, and meanwhile the subpopulation-specific profiles, in turn, leads to improved coupled clustering results of single-cell data [61]. Details of applications of scMulti-omics to complex diseases can be found in Table 2, including sequencing methods, diseases, number of cells, contributions, and accession IDs.

REMAINING CHALLENGES

Development of single-cell sequencing technologies and the applications in complex diseases have been astonishing in the past decade, but many challenges still exist and much remains to be explored. As new techniques arise to measure distinct cellular modalities, the paramount challenge is to integrate these datasets to better understand cellular identity and regulatory mechanisms. Because each technology measures only partial aspects of cellular identity, and noise may be introduced during cell isolation, most analyses merely focus on the clustering and subtyping of cells. Therefore, it is pressing need to leverage information in one dataset to improve the interpretation of another as examples in bulk data [72, 92, 93]. PECA is a statistical approach to infer transcriptional regulatory networks from paired expression and chromatin accessibility data across diverse cellular contexts [92]. The networks inferred by PECA provide a detailed view of how trans- and cis-regulatory elements work together to affect gene expression in a context-specific manner. TimeReg is a framework for time-course regulatory analysis based on paired expression and chromatin accessibility data from the time course, which can be used to prioritize CREs, to extract core regulatory modules, to identify key regulators driving changes of cellular states, and to casually connect the modules across different time points [72]. One solution to this challenge is to transfer the pre-trained models on tissue level to single-cell level. Another solution is to resort to artificial intelligence (AI), e.g., multimodal learning models [94].

Moreover, single-cell omics data typically suffer from technical issues including sparsity, biased cell subpopulations, and false-positive variants [23, 34, 95]. Sparsity mainly originates from dropout, low-throughput, and low coverage, and cell subpopulation biases are introduced during single-cell isolation, and false-positive variants result from the amplification or sequencing. These issues may lead to missing nodes or edges in constructed networks. One solution to deal with sparsity is to use bulk omics data to impute single-cell omics data. Typical tools for scRNA-Seq imputation include MAGIC [96], SCRABBLE [97], and SAVER-X [98]. Comparing the variant alleles discovered in the single-cells to the bulk population could ensure there were no biases in cell subpopulations. Besides, using the bulk sample as a reference can reduce false-positive variants (e.g., SAVER-X).

Furthermore, batch effects exist when integrating single-cell omics data across samples, experiments, and modalities. These effects result from varying resolution levels, the uncertainty of any measurements, and scaling of single-cell methodology to more cells and more features measured at once. Batch effects may thus influence the stability of inferred networks. One way is to develop techniques to measure different modalities within the same cells. Another way is to design tools to harmonize the modalities measured in different cells into a single reference, e.g., Seurat [90].

Finally, the construction of networks depends on many factors such as data type, network complexity & topology, and tool settings. Evaluating the reliability of the predicted networks remains a challenge. One common solution is to cross-validate the linkages in the network with known interactions from literature. However, this method may be inaccurate in evaluating negative interactions and global interactions, as their information is difficult to obtain. Perturb-Seq, a sequencing method combining CRISPR-mediated genes with scRNA-Seq, can be performed to curate TF-gene relationships, allowing for validation of the stability and plasticity of networks [99]. In addition, the functional homogeneity of network modules could be estimated by enrichment analyses against gene ontology terms, pathways, and motifs [40, 46, 48, 54, 56]. However, both Perturb-seq and scMulti-omics cross-validation require special sequencing techniques that not suitable for networks predicted from a single dataset.

CONCLUDING REMARKS

Exploring single-cell gene regulatory mechanisms provides the opportunity to uncover the unique malfunctions associated with complex diseases in individual cells. To elucidate the roles of TFs, genes, CREs, and chromatin loci in gene regulation, HRL-associated networks (GCN, GRN, CCAN, MGRN, CIN, CGN, and TCN) are constructed, which jointly build an HRL to characterize the multi-level genotype and phenotype observations across diverse cell types. We raise several questions to be answered regarding HRL in complex diseases using single-cell data (see Outstanding Questions). Meanwhile, several future trends in HRL in complex diseases are discussed below.

OUTSTANDING QUESTIONS.

• How to bridge different HRL-associated networks and cell-cell networks through data integration and modeling approaches?

• Will it be feasible to bridge basic cell/molecular biology models to clinically predictive models in complex diseases?

• Are there any data-driven ways to guide sample collection to model dynamic time-series networks?

• How to build networks based on all the different modalities?

• How to extend systems biology concepts to the tissue level?

• What other information can be used to interpret the networks in addition to single-cell omics data?

• How can high throughput perturbation experiments (e.g. CRISPR, single-cell) be used to inform and validate network and system information?

One clear trend is the increasing significance of HRL in complex diseases. HRL is informative in identifying the TFs, genes, CREs, DNA methylation states, and chromatin architectures that possibly link to complex diseases. First, HRL construction will be further refined by integrating other information, such as spatial, time-series, and treatment, leading to dynamic HRL alongside time-series. Besides, HRL reconstruction could promote the delineation of cell-cell differences via building cell-cell interaction networks. Uncovering the HRL by inferring cell-type-specific networks will provide biological insights into the development, evolution, and pathologies of organisms, and contribute to the prognosis and treatment of diseases. Lastly, the integration of bulk omics and single-cell omics data will be an important strategy in support of HRL elucidation, e.g., imputing gene expression from scRNA-Seq using bulk RNA-Seq data [66]. This topic will attract considerable attention to the advances in multimodal omics and the fast-development of AI.

Another trend is the increasing application of AI and machine learning (ML) methods in single-cell omics analyses (Figure 4), especially deep learning (DL) models [94]. The capability to extract new insights from the exponentially increasing volume of genomics data requires more expressive ML models. By effectively leveraging large datasets, DL can capture dependencies in sequential and graph-structured data using various models such as deep neural network (DNN), recurrent neural network (RNN), convolutional neural network (CNN), and graph-convolutional neural network [94]. DNNs are mainly used to model complex dependencies, which include predicting the percentage of spliced exons, discovering TF binding motifs, and prioritizing potential disease-causing genetic variants [94]. Besides, DNN can be easily extended to take multiple modalities as inputs in order to leverage complementary information from between them [94]. Applications of CNNs include discovering local patterns in sequential data, e.g., predicting molecular phenotypes, classifying TF binding sites, denoising ChIP-Seq, and enhancing Hi-C data resolution [94]. The CNN model, CNNC, has been proposed for supervised gene relationship inference, which improves upon prior methods in tasks ranging from predicting transcription factor targets to identifying disease-related genes to causality inference [100]. RNNs have been used to model long-range dependencies in sequences, e.g., predicting TF binding sites or RNA binding proteins, detecting methylation states, and discovering miRNA targets [94]. Tasks that graph-convolutional neural networks can be trained for include modeling dependencies in graph-structured data, such as deriving features from protein-protein interaction networks, modeling polypharmacy side effects, and predicting binarized gene expressions [94]. In the scenario where data are scarce, training a model from scratch might be infeasible. Transfer learning (TL) can be used to initiate the model with the majority of parameters from another model trained on a similar task [94]. The utility of TL has been demonstrated for sequence-based predictive models of chromatin accessibility by training models in other cell types [94]. In federated learning (FL), model instances are deployed on distinct sites and trained on local data to optimize a global model [94]. By avoiding data transfer among devices, FL could attain high efficiency and facilitate the respect of medical data privacy. FL has achieved advantageous performance in predicting hospitalizations due to heart diseases within a calendar year based on patients’ electronic health records prior to that year [101]. Moreover, generative models (e.g., generative adversarial network) could be used to simulate clinical data that can be analyzed by others without privacy violation [94]. Finally, to make the inner working of ML models understandable to humans, three strategies for interpretable ML have been proposed [102]: 1) measuring how changes in the input data impact model predictions (perturbing strategy), 2) training an interpretable model using the same data to approximate the predictions of the to-be-interpreted model (surrogate strategy), 3) inspecting the structure and parameters in a trained model (probing strategy).

Figure 4.

Showcase the power of AI and ML, especially graph-convolutional neural network, in the prediction of network models.

Although DL models have showcased outstanding performances in this field, potential limitations remain. For example, the performance of DL models strongly depends on the choice of models and hyperparameters [94]. The increase in the number of parameters might pose overfitting challenge. One solution is to develop an effective method for hyperparameter optimization. Besides, some models, e.g., RNN, are time-consuming and cannot be easily parallelized [94]. This limitation might be overcome by designing more efficient DL models.

HIGHLIGHTS.

• Advances in single-cell sequencing technologies open a window to understanding the Heterogeneous regulatory landscape (HRL) encoded in complex diseases, by inferring various biological networks.

• The development of scMulti-omics technologies combined with scMulti-omics integration tools provides multimodal measurements of HRL.

• Application of HRL to complex diseases poses opportunities and challenges.

• Among the remaining challenges, establishing a robust benchmarking pipeline is paramount.

• In support of the integration of diverse single-cell modalities, bulk and single-cell omics, the deep learning and AI omics become the major trends.

GLOSSARY

Artificial intelligence (AI)

a.k.a. machine intelligence, is intelligence demonstrated by machines simulating natural intelligence

cis-co-accessibility network (CCAN)

an undirected graph, where each node represents a CRE and a pair of CREs is connected by an edge if there is a significant co-accessibility relationship between them based on available scATAC-Seq data

Chromatin interaction network (CIN)

an undirected graph, where each node represents a chromatin locus, and a pair of regions are connected by an edge if there is a significant interaction between them based on available scHi-C data

CRE-gene interaction network (CGN)

a directed graph with nodes representing CREs and genes, and edges corresponding to the cis-regulatory relationships between the two kinds of nodes

Convolutional neural network (CNN)

a kind of neural network that performs convolution operations in at least one of their layers

Deep learning (DL)

a type of ML algorithms that uses multiple layers to progressively extract higher-level features from the raw input

Deep neural network (DNN)

a type of artificial neural network with multiple layers between the input and output layers

Federated learning (FL)

a.k.a. collaborative learning, is a ML technique that trains an algorithm across multiple decentralized edge devices or servers holding local data samples, without exchanging their data samples

Gene co-expression network (GCN)

an undirected graph, where each node corresponds to a gene, and a pair of nodes is connected with an edge if there is a significant co-expression relationship between them

Graph-convolutional neural network

a neural network designed to work directly on graphs and leverage their structural information

Gene regulatory network (GRN)

a directed graph with nodes representing transcriptional regulators and genes, and edges corresponding to the regulatory relationships between these two kinds of nodes

Heterogeneous regulatory landscape (HRL)

a large genomic region of heterogeneous cell types containing several long-range-acting regulatory sequences that control target genes in a coordinated manner

Methylation associated gene regulatory network (MGRN)

a GRN where regulator-gene regulatory relationships are influenced by methylation status

Machine learning (ML)

an application of AI that improves automatically through experience

Perturb-Seq

a.k.a. Crisp-Seq or CROP-Seq, is a high-throughput method of performing scRNA-Seq on pooled genetic perturbation screens

Perturbing strategy

a category of ML interpretation strategies that measure how changes in the input data impact model predictions

Probing strategy

a category of ML interpretation strategies that involve directly inspecting the structure and parameters in a trained model

Recurrent neural network (RNN)

a type of artificial neural network where connections between nodes form a directed graph along a temporal sequence

Single-cell multimodal omics (scMulti-omics)

a category of technologies measuring multiple types of molecules from the same individual cell

Surrogate strategy

a category of ML interpretation strategies that involve training an inherently interpretable model (e.g., a linear model) to approximate the black-box model using the same data

TF-CRE interaction network (TCN)

a directed graph, where nodes denote TFs and CREs, and edges represent the binding activity of TFs to CREs

Transfer learning (TL)

an ML technique that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem