Integrative approaches in genome structure analysis

Abstract

New technological advances in integrated imaging, sequencing based assays, and computational analysis have revolutionized our view of genomes in terms of their structure and dynamics in space and time. These advances promise a deeper understanding of genome functions and mechanistic insights into how the nucleus is spatially organized and functions. These wide arrays of complementary data provide an opportunity to produce quantitative integrative models of nuclear organization. In this article, we highlight recent key developments and discuss the outlook for these fields.

Graphical Abstract

eTOC Blurb

Boninsegna et al. review recent advances in the use of integrated imaging studies and multi-modal data-driven computational modeling to study the structure and dynamics of genomes. They discuss how these approaches enable a deeper understanding of gnome functions and nuclear organization.

Introduction

The three-dimensional (3D) folding of genomes and their nuclear organization greatly affect and control gene transcription and other nuclear functions (Dekker et al., 2017). Aberrant chromatin folding is linked to disease, such as cancer, developmental defects and premature aging disorders (Misteli, 2010, 2020). Disease-associated mutations are located, in the majority of cases, at noncoding regulatory regions, mostly distal in sequence from affected genes. Thus, to better understand disease related mechanisms, we need to understand not only 1D sequence relationships of genes along chromosomes, but decipher the effects of mutations on the 3D chromatin structure and its dynamic changes over time. Besides local chromatin folding also the nuclear microenvironment, the nuclear locations of genes relative to nuclear bodies and compartments, affect the permissiveness of gene expression. Thus, it is of great interest to gain deeper knowledge about the 3D nuclear organization of genomes along with the molecular factors and biophysical principles that determine it. However, deciphering the higher order nuclear organization of whole genomes is a challenging task. The human genome contains more than 3 billion DNA basepairs and undergoes dramatic compaction, regulated by a multitude of processes. The last decade has seen unprecedented progress, mainly due to the development of sequencing-based assays, microscopy techniques, and computational modeling (Kempfer and Pombo, 2020). Together these techniques revealed new levels of chromatin organization and uncovered some of the fundamental mechanisms shaping chromatin organization from the scale of nucleosomes to the entire nucleus. Here, we provide an overview of some of these methods and discuss the uncovered organizational principles that shape spatial genome organization and define the structural biology of the nucleus.

We first discuss details of genome structural organization, both at the level of local chromatin folding and global nuclear scale. We then highlight a selection of recent state-of-the-art high throughput imaging approaches that provide an integrated view of genome structure and function. Finally, we discuss data-driven modeling approaches as an indispensable tool for multi-modal integration of genomics and imaging information for achieving more predictive quantitative models of nuclear genomes.

Overview of spatial genome organization

To give an overview about the spatial organization of genomes, we first divide processes controlling genome structure into those acting on the local chromatin fiber topology and those defining the global localization of chromatin in the nucleus.

Local chromatin topology.

To initiate transcriptional activation, promoter-enhancer interactions often act over considerable sequence distances. These interactions are coordinated by the spatial accessibility of regulatory elements, controlled by the chromatin fiber topology through the formation of chromatin loops, often located at the borders of so-called topologically associating domains (TADs) (Figure 1, left). These chromatin domains show preferred interactions within and depleted interactions between TAD chromatin regions (Dixon et al., 2012). Several technologies identify chromatin interactions on a genome-wide scale (Kempfer and Pombo, 2020) and revealed the ubiquitous nature of chromatin loops (Rao et al., 2014), TADs, and sub-domains (Phillips-Cremins et al., 2013). Among those are ligation-based chromosome conformation capture (3C) methods (in situ Hi-C, micro-C) (Kempfer and Pombo, 2020; Lieberman-Aiden et al., 2009; Rao et al., 2014), Chromatin Interaction Analysis by Paired-End Tag Sequencing (Chia-PET) (Fullwood et al., 2009), and ligation-free Genome Architecture Mapping (GAM) (Beagrie et al., 2017), Split-Pool Recognition of Interactions by Tag Extension (SPRITE) (Quinodoz et al., 2018) and Chia-Drop (Zheng et al., 2019). In ensemble Hi-C data, chromatin loops emerge as locally enriched contact frequency peaks (Rao et al., 2014), while TADs are revealed as squared blocks of enriched contact frequencies along the diagonal of the Hi-C contact frequency map (Dixon et al., 2012; Nora et al., 2012; Sexton et al., 2012).

Fig. 1.

(Top panel) Chromatin organization at different scales. Chromatin forms loops and topologically associating domains (TADs), the boundaries of which are enriched with CTCF and cohesin and other protein factors (left). Active and inactive chromatin are further segregated into nuclear compartments (middle) and organized around nuclear bodies such as nuclear speckles, nucleoli, and nuclear lamina (right). (Bottom panel) Tools for detecting chromatin organization at different scales. See text for references.

Loop extrusion is one of the primary mechanisms for chromatin loop/TAD formation and subsequent chromatin condensation (Fudenberg et al., 2016; Sanborn et al., 2015). The process is initiated when a loop-extruding factor such as cohesin or condensin selectively attaches to the chromatin fiber and actively reels in the chromatin fiber in both directions, thus leading to a progressively growing loop until the loop-extruder encounters an extrusion barrier at specific boundary elements. The insulator CCCTC-binding factor protein (CTCF) acts as an extrusion barrier and stabilizes a loop if its binding motifs are oriented in a convergent sequent orientations at the loop anchors (Rao et al., 2014). Chromatin loops are inherently dynamic structures with a life-time in the tens of minutes, resulting in an overall irregular polymorphic chromatin fiber in interphase genomes, confirmed by recent super-resolution microscopy and electron microscopy studies (Bintu et al., 2018; Ou et al., 2017).

Global chromatin organization.

Besides local topology, also the global locations of genes within the nucleus play a critical role in gene function. Processes shaping the global nuclear architecture are driven by two main components, chromatin compartmentalization and spatial associations to nuclear landmarks (Misteli, 2020).

Nuclear compartmentalization

Nuclear compartmentalization is a major factor in the spatial segregation of chromatin in different functional states (Hildebrand and Dekker, 2020). Chromosome conformation mapping shows characteristic checkerboard patterns for low frequency long range and inter chromosomal interactions. These indicate preferential, but dynamic, associations of chromatin with similar functional profiles, leading to a spatial segregation of chromatin into at least two compartments, transcriptionally active (A) and inactive (B) compartments (Lieberman-Aiden et al., 2009), (Figure 1, center). These compartments were subsequently refined, at high sequencing depth, into 5 primary Hi-C subcompartments (Rao et al., 2014). Spatial segregation of eu- and heterochromatin is likely driven by phase separation (Hildebrand and Dekker, 2020; Solovei et al., 2016), which creates distinct phases from a single homogeneous mixture through weak but multivalent interactions of binding factors (either proteins or nucleic acids) that bridge chromatin of a given type and form condensates once a critical concentration of the binding factor is reached (Sabari et al., 2020). For instance, heterochromatin protein 1 (HP1) binds methylated H3K9me3 histone tails and oligomerizes with HP1 at other chromatin regions, which then condenses into a heterochromatic phase (Larson et al., 2017; Strom et al., 2017). Also, some active regions are likely to form condensates during transcription initiation, which increases the local concentration of RNA polymerase machineries and mediator complexes. These condensates are driven by the activator transcription factors, like MYC, which are able to form local condensates (Sabari et al., 2020). Processes driving global chromatin compartmentalization are distinct and act antagonistically to those shaping the local chromatin folding (i.e., chromatin looping and TAD formation). Loop extrusion, being a non-equilibrium process, overrides locally the fine-scale compartment patterns that would otherwise be visible at longer-range interactions. Thus, a common hierarchical organization with TADs as building blocks of compartments does not exist. TADs can exist without compartments and vice versa (Mirny et al., 2019). For instance, depletion of loop extrusion factor cohesin reduces TADs, while revealing finer compartment patterns. Also the depletion of extrusion barrier CTCF weakens TADs, while leaving compartments and chromatin compaction unaffected (Nuebler et al., 2018; Wutz et al., 2017).

Spatial association with landmarks.

The second major component shaping global genome organization are preferential associations of chromatin to nuclear bodies and nuclear compartments (Misteli, 2010, 2020). The nuclear locations of genes relative to nuclear bodies affect the permissiveness of gene expression (Chen et al., 2018). Transcriptional active regions coalesce at nuclear speckles, but also nuclear pore complexes and PML bodies, while regions of transcriptional repression are often associated with the nuclear lamina, perinucleolar chromatin as part of the heterochromatin compartment. Some genes appear to move towards nuclear speckles upon transcriptional activation (Khanna et al., 2014). Moreover, the nascent transcript levels of some genes are also directly correlated with their mean distances to nuclear speckles (Kim et al., 2019). Thus, stochastic gene expression could be dependent on stochastic positioning of genes relative to nuclear speckles and subsequent “gene expression amplification” if a gene is found in a preferential nuclear microenvironment”. Thus, a gene’s nuclear microenvironment, defined by their relative locations within the nucleus, can play key role in gene function (Figure 1, right).

Several experimental technologies probe quantitatively the mean distances (TSA-seq) or association frequencies of genes to nuclear speckles (Chen et al., 2018), lamina associated domains (Guelen et al., 2008), and nucleoli (Vertii et al., 2019), including recent super-resolution large-scale imaging approaches (Su et al., 2020; Takei et al., 2021a). These studies revealed the distinct preferences of genes to being reproducibly localized at specific nuclear bodies, on a genome-wide scale.

However, collecting information about the spatial locations of all genes with respect to all nuclear bodies simultaneously within the same cell, at the same time, is challenging, especially when considering cell-to-cell variability of a gene’s microenvironment within a population of cells (Finn et al., 2019). Recent high-throughput integrative imaging and integrative modeling approaches (Boninsegna et al., 2021; Yildirim et al., 2021a) could provide a promising solution to this challenging problem, by offering the opportunity to describe the nuclear microenvironment of genes on a genome-wide scale. In the following sections, we will focus on these two technologies.

Integrated imaging methods for genome structure analysis

Significant advances in our understanding of genome organization were made by fluorescence microscopy. 3D Fluorescence In Situ Hybridization (FISH) (Cremer et al., 2012) uses DNA sequences tagged with fluorescent dyes as probes to hybridize to complementary target regions in the genome of fixed cells and enables simultaneous visualization of a small number of loci in single cells (Kempfer and Pombo, 2020). Recent advances in combinatorial labeling of color schemes, combined with superresolution imaging, have boosted the field tremendously and enabled genome-scale multiplexed imaging at unprecedented scale (Payne et al., 2021; Su et al., 2020; Takei et al., 2021a).

A catalyst for a variety of super resolution techniques were advances in Oligopaint technology (Beliveau et al., 2012), i.e. massively parallel synthesis of DNA oligonucleotide (oligo) libraries, which are highly programmable in design to rapidly diversify FISH probes to a multitude of target genomic regions. The technology generates hundreds of thousands of unique oligos for labeling kilo- to mega base genomic regions, which can then be imaged by super-resolution microscopy to visualize different levels of genome organization at nanometer-scale resolution (Beliveau et al., 2015). The high adaptability of oligopaint probes has allowed extensive genome organization studies.

Global volume and shape of TADs in different chromatin states.

Direct visualization of DNA with oligoPAINT probes revealed previously unseen structural details of chromatin domains in different epigenetic states (Boettiger et al., 2016). The study revealed dramatic differences in packaging, volume occupancy and chromatin intermixing of chromatin within domains in functional different states, pointing to distinct mechanisms of chromatin folding.

Interactions between TADs.

Another study (Wang et al., 2016) tailored oligopaints for multiplexed sequential FISH imaging that allows 3D tracing of individual chromosomes in the nucleus. This was achieved by two rounds of oligo hybridizations, in which a primary oligo probe recognizes the targeted chromosome, while a secondary oligo probe recognizes a readout sequence in the primary oligo, which was specifically designed to target each individual TAD. This process provided insights on how TADs are arranged to form compartments in chromosomes. It was also shown that the mean TAD distances highly correlate with contact frequencies in Hi-C data.

Internal structure of TADs.

A similar approach was used to study the internal structure of TADs in thousands of individual cells at super-resolution (Bintu et al., 2018). The TAD genomic regions (about 1–2Mb long) were divided into segments of 30kb, each individually imaged by sequential rounds of FISH imaging. Imaging revealed an abundance of TAD-like globular domain structures in single cells; however, domain boundaries varied largely between cells with a statistical preference to be located at CTCF/cohesion binding sites. Cohesin depletion led to a loss of preferential domain boundary locations, while the abundance of globular domains overall was not affected, indicating that cohesin was not required for formation and maintenance of single cell globular domains.

Combining Hi-C with super-resolution imaging.

A recent study combined super-resolution microscopy with Hi-C data modeling. OligoDNA-PAINT microscopy with sequential hybridization traced the structures of a ~8 Mb region of human chromosome 19 (Nir et al., 2018) for both homologous copies, revealing that maternal and paternal homologous can be differentially organized. The genomic region was divided into segments, each sequentially imaged by a large number of oligos; the resulting cloud of oligo locations was translated into a chromatin localization density for the targeted segment, and 3D structures of the chromatin fiber were generated that best fitted into such density map while recapitulating Hi-C contact frequencies. It was shown that consecutive compartment segments showed only minimal entanglement, suggesting they were distinct physical entities, which interacted with others to form two spatial clusters corresponding to A and B compartments.

Another promising development is in situ sequencing of DNA in the nucleus to jointly identify DNA sequences directly at their in situ nuclear locations. The approach has been applied to either targeted (STARmap) or untargeted chromosomal regions (FISSEQ (Lee et al., 2014), oligoFISSEQ (Nguyen et al., 2020)).

Integrated imaging to jointly probe structure and transcriptome.

New approaches were recently introduced that used multimodal imaging to trace chromosome structures together with mRNA transcriptomics, epigenetic chromatin states and locations of nuclear bodies simultaneously in the same intact cell, thus producing an integrated view of genomes in their functional context (Zhuang, 2021).

One study was performed with SeqFISH+ (Takei et al., 2021a), an expansion of the seqFISH protocol (Lubeck et al., 2014), which used multiplexed DNA FISH to image the nuclear locations of more than 3600 loci in individual mouse ES cells. The subnuclear localizations of 17 histone modifications and nuclear bodies were imaged in the same cells by sequential immunofluorescence (IF), while nascent RNA transcripts of 70 genes were detected by RNA seqFISH, a multiplexed RNA FISH approach that uses a sequential color code to distinguish nascent RNA of different genes (Figure 2A).

Fig. 2.

(A) [Adapted from (Takei et al., 2021a)] (Top panel) 3D images of immunofluorescence markers for heterochromatin (H3K9me3, DAPI, MinSat), active chromatin (H3K9ac,H3K27ac), RNAPIISer5-P, SINEB1 and SF3A66 in a single cell. (Bottom panel) 3D images of H3K9me3, SF3a66, and Lamin-B1 together in the same single cell (left), pixels of LINE1 by DNA FISH and X chromosome (middle), representative 3D images of loci in Chr 12, 16, 18 and 19 together with the fibrillarin marker in the same cell (right). (B) [Adapted from (Payne et al., 2021)] Representative mouse zygote with >7,000 spatially localized IGS reads colored by chromosome (left); distance to nuclear lamina (middle) and closest nucleolus precursor body (right) as from laminB1 immunofluorescence and DAPI respectively. Chromosome territories are showing within clearly resolved maternal and paternal pronuclei. (C) [Adapted from (Su et al., 2020)] 3D rendering of all detected chromatin loci color-coded by chromosomes in a single cell (top); detected foci of nascent transcripts shown as spheres in the same cell, color-coded by identities of the imaged genes, and the transcription burst size is represented by the sphere size (middle); volume-filling representations of detected nuclear speckles (yellow), nucleoli (blue), and lamina (gray) in the same cell (bottom).

The study uncovered several interesting structure function correlations. Chromosomes displayed territories with high structural variability between cells. However, a subset of loci was invariably associated with specific immunofluorescence marks in a large proportion of cells. These marks relate to nuclear bodies and globules of histone modifications. Subsequently, these loci do not show variations in their physical microenvironment between cells. Such loci with “fixed” structural preferences naturally constrain a chromosome’s spatial organization to features of the nuclear architecture. This extends previous observations that suggested lamina, speckles and the nucleolus to be deterministic scaffolds of chromosome organization.

The nuclear volume was further divided into a dozen zones based on observed locations of histone marks and nuclear bodies. These zones, containing chromatin with shared functional profiles, formed physically distinct regions in the nucleus. Mapping locations of genes to nuclear zones revealed characteristic zone associations for some loci. Loci were often localized at interphases between two zones, such as surfaces of nuclear bodies. Clustering of immunofluorescence data showed several architectural cell states, similar to states detected in mESC cells by differential gene expression analysis. The same imaging pipeline was subsequently applied to study also mouse brain cortex to differentiate cell types and confirm invariant principles underlying chromatin organization (Takei et al., 2021b).

Another approach, in-situ Genome Sequencing (IGS) (Payne et al., 2021), was introduced for simultaneous sequencing and imaging of intact interphase genomes of human male PGP1f fibroblasts and mouse embryos at different developmental stages. IGS resolved genome-wide chromosome positioning, territories and folding. Combining in-situ sequencing with genotype information enabled to spatially resolve the maternal and paternal genomes in mouse embryos (Figure 2B). DAPI and immunofluorescence staining of CENP-A and Lamin-B1 were then used to characterize the association of imaged loci with nuclear landmarks, such as centromeres, the nuclear lamina and the nucleolus precursor bodies. Moreover, IGS was used to characterize parent-specific changes in genome organization associated with transitioning between embryonic states and predicted a Rabl-like configuration of chromosomes in early mouse embryos. The method also revealed heritable correlations in global chromosome positioning within early embryos and suggested a process of epigenetic memory transmission within clonal lineages. Another integrative imaging analysis (Su et al., 2020) focused on high-resolution tracing of large chromosomes and multimodal imaging of whole genomes together with nascent RNA transcripts and several nuclear bodies and landmarks within the same cell (Figure 2C). This integrative strategy provided key insights into how chromatin organization regulates transcription. High-resolution tracing of chromosome conformations was achieved by multiplexed FISH imaging with sequential hybridization, producing large chromosome structures at tenfold higher resolution (50kb pair resolution) than previously achieved (Wang et al., 2016). This improvement revealed a more complex and variable structures of large chromosomes. For instance, chromatin domains with the same genomic size varied considerably in volume between cells Large chromosomes revealed a lower degree of spatial segregation between compartments, likely due to cell-to-cell variability of the epigenome. This could point to an imperfect delineation of ensemble derived A/B compartment boundaries in single cells. However, it may also indicate that domain formation is not strongly coupled to compartment identity, confirming antagonistic processes for loop extrusion and compartment formation (Mirny et al., 2019; Nuebler et al., 2018). Chromatin domains also showed dramatic variations in long-range interactions between cells, while the strength of domain interactions appeared modulated by their A/B composition and sequence separation.

For tracing of whole genome structures, the authors devised a combinatorial strategy based on Multiplexed Error-Robust FISH (DNA-MERFISH), a rapid method to identify, in each hybridization round, many more loci than sequential imaging. Each oligo probe contains both a primary region binding to a specific target locus and a region containing a combination of readout sequences, chosen from 100 predefined barcodes. Each locus is then defined by a sequential combination of barcodes, which is read out by sequential hybridization of fluorescently labeled readout probes, each binding to only one of the 100 barcode sequences. The locus identity is determined by the sequential order of detected (and undetected) barcodes in each round of hybridization. This combinatorial procedure allowed to image ~1,041 genomic loci (~30kb in size and uniform coverage) across 5400 individual cells with only 10 rounds of hybridizations (10-fold fewer rounds than sequential imaging). In addition, the transcriptional activity of all imaged genes was simultaneously measured with RNA-MERFISH, which uses a similar combinatorial strategy to image nascent RNA transcripts. Locations and shapes of nuclear speckles and nucleoli were imaged by sequential staining of protein markers, and positions of lamina were inferred from the nuclear envelope.

The study revealed that transcriptional activity was highly correlated with the local enrichment of active chromatin around a gene (i.e. A/B ratio). For instance, the A/B ratio at transcription start sites was higher for the same gene when it was transcribed and genes with increased A/B ratio showed higher transcriptional firing rate. They also found a link between a gene’s activity and its associations with nuclear bodies: an increased association frequency to nuclear speckles correlated with increased gene transcription, while inhibition of transcription for the same genes reduced their speckle association. Interchromosomal interactions occurred preferentially among active chromatin regions. These observations point to a non-trivial joint molecular mechanism for interactions between active loci, both for long-range cis as well as trans chromatin interactions. These examples and others (Mateo et al., 2019; Zhuang, 2021) showcase the potential of integrated imaging to characterize jointly the nuclear transcriptome and genome structures for a better understanding of regulatory processes.

Integrated computational methods for genome structure modeling

The availability of data from sequencing and imaging technologies provides unique opportunities for producing quantitative 3D models of entire nuclear organizations. However, simulating accurate genome structures with all their complexities is a daunting task.

The coarse-grained chromatin fiber is typically represented by a chain of monomers, each representative of a specific chromatin region, depending on the model of granularity (i.e., base-pair resolution). Monomers can be assigned to different classes to account for differences in epigenetic or biochemical identities. The interactions among such monomers are described by a Hamiltonian (Parmar et al., 2019), which is used to either optimize genome structures to best reproduce experimental data or simulate the time-evolution of dynamic processes. In both cases, an ensemble of 3D structures is generated, which are examined to derive structure-function correlations and make quantitative predictions of structural features and their cell-to-cell variabilities. Overall, we distinguish between two main modeling concepts: mechanistic and data-driven (Figure 3).

Fig. 3.

(A) Mechanistic modeling approach relies on prior knowledge about a folding hypothesis that can explain the experimental observations. The polymer model is defined to simulate the hypothesis. If the collected structures do not agree with experimental observations, the initial hypothesis and/or the model parameters are modified and new simulations are performed until the resulting structures reproduce the data. (B) Data-driven methods use experimental data explicitly to build 3D models that fully reproduce it and are used for quantitative structure analysis. Adjustments of data interpretation may be needed to ensure better agreement with experimental data. (C) Comparison of resampling (left) and population-based (right) data-driven modeling approaches, both inferring contact restraints from Hi-C data. In resampling approaches, each structure in the population is generated from the same combination of data by independent simulations. In population-based approaches, each structure in the population contains a subset of all chromatin contacts, the summation of which recapitulates the ensemble Hi-C data.

Mechanistic modeling.

Mechanistic modeling (Figure 3A) relies on prior knowledge about a folding hypothesis, either from experimental evidence or intuition, to explain experimental observations; the polymer model is then defined to simulate such hypothesis. If the simulated structures do not agree with experimental observations, the initial hypothesis and/or the model parameters are refined, and new simulations are performed until the experimental data is matched. Mechanistic modeling has been very successful in validating loop extrusion as a primary process for chromatin loop and TAD formation (Banigan and Mirny, 2020; Fudenberg et al., 2016; Sanborn et al., 2015). Other studies performed mechanistic simulations to study the role of phase-separation and micro-phase compartmentalization of genomic regions by block copolymer (Brackey et al., 2020; Falk et al., 2019; Jost et al., 2014; MacPherson et al., 2018) or the role of cognate binding factors in establishing chromatin structure (Barbieri et al., 2012; Bianco et al., 2020). These studies also established the antagonistic nature of loop extrusion processes and those underlying phase separation in chromatin compartmentalization (Mirny et al., 2019; Nuebler et al., 2018).

Data-driven modeling.

Data-driven approaches (Figure 3B) explicitly incorporate all experimental data in the Hamiltonian by modeling the effect of each data point as pseudo-energy functions; e.g., a harmonic potential between two chromatin loci is a viable implementation of a chromatin contact determined in experiment. The chromatin structure is then a result of the collective interplay of a multitude of data-derived interaction terms. We distinguish here between two conceptually different data-driven approaches (Yildirim et al., 2021b).

Resampling methods.

Resampling methods (Figure 3C, left panel) express all the data by a single scoring function to describe the agreement between data and model. A large number of independent structure optimizations are performed to explore all 3D structures that equally well minimize the scoring function (i.e. maximize the agreement between data and model.) Because some of the experimental data were typically accumulated over millions of cells, each with structural variations, it is possible that ensemble data contains conflicting and mutually exclusive information. if imposed in the same structure. Some approaches solve this issue by considering only the most significant subsets of interactions that are likely to be present in a dominant structural state, for instance for local chromatin folding of chromatin regions (Baù et al., 2011; Serra et al., 2017; Umbarger et al., 2011; Yildirim and Feig, 2018). Resampling approaches have been used for integrating Hi-C with super-resolution imaging (Nir et al., 2018) or ChIP-seq data (Paulsen et al., 2017), and are also suited to generate genome models using single cell Hi-C (scHi-C) data (Nagano et al., 2013; Rosenthal et al., 2019; Stevens et al., 2017; Tan et al., 2018). Most single cell Hi-C data reveal chromosome territories and A/B compartment segregation despite the considerable cell-to-cell variability. However, due to the relatively low coverage in single cell data the detection of TADs has been challenging. Some methods have been developed (Zhang et al., 2021; Zhou et al., 2019) to infer from sparse scHi-C data features of the genome organization, such as existence of TADs (Li et al., 2021) and loops (Yu et al., 2021) their cell-to-cell variabilities, or even characterize cell cycle progression (Ye et al., 2019). More details can be found in other reviews and references therein (MacKay and Kusalik, 2020; Polles et al., 2019; Yildirim et al., 2021b; Zhou et al., 2021).

Deconvolution approaches.

Deconvolution methods (Figure 3C, right panel) divide ensemble data into individual subsets, each representative of a single structure of a population. Thus, the population of structures as a whole, rather than an individual model, is statistically consistent with the overall ensemble data. An approach by Giorgetti et al. (Giorgetti et al., 2014) modeled a chromatin domain with a chain of beads. The energy function of each interaction pair was modeled by a spherical-well potential, optimized to reproduce 5C contact frequencies. Modeling of the mouse Xic locus on the inactive X chromosome revealed high structural fluctuations in TAD conformations, which were linked to transcriptional variability.

Zhang et al. (Zhang and Wolynes, 2015) developed an iterative algorithm, based on the maximum entropy principle, that used Hi-C contact frequencies to approximate a chromosomal energy landscape. An ensemble of conformations for human chromosome 12 recapitulated Hi-C data and revealed highly variable chromosome configurations, in which TADs played a key role to locally rigidify the chromosome chain. A subset of TADs showed two-state transitions, possibly to modulate transcriptional activity. The formulation was later expanded to accommodate chromatin type interactions, loop formation terms (Di Pierro et al., 2016; Lin et al., 2021) and centromere interactions to generate models of whole diploid genomes (Qi and Zhang, 2020).

Population-based genome structure modeling.

Population-based genome structure modeling uses a different, optimization strategy. The method deconvolutes ensemble data into a population of single-cell genome structures, which recapitulate all input information (Boninsegna et al., 2021; Hua et al., 2018; Tjong et al., 2016). The method allows for different conformational states between cells, as conflicting data points can be allocated in different structures in the population (Figure 4). Thus, the method can describe an unbiased representation of genome structure variability, which is especially suitable for modeling whole genome structures (Yildirim et al., 2021a).

Fig. 4.

[Adapted from (Tjong et al., 2016)] Schematic of population-based genome structure modeling approach. A population of genome structures is simulated, in which the contacts between chromatin regions across structures are statistically consistent with the contact probability matrix derived from Hi-C data. This problem is formulated as a maximum likelihood problem and solved via iterative A/M steps. The process jointly optimizes the genome structures and the “contact indicator tensor”, a latent binary third-order tensor, which details which contacts between homologue chromosome copies are present in which structures of the population.

Here we briefly summarize the basic principles of the method. The structure population is defined as a set of diploid genome structures X = {X₁, … , X_S}. The input dataset $\mathcal{D}=\left\{D_{\mathit{\text{exp}}1},D_{\mathit{\text{exp}}2},\dots ,D_{\mathit{\text{exp}}N}\right\}$ can include a variety of multi-modal data from different experimental sources, for instance, Hi-C or laminB1 DamID (Boninsegna et al., 2021; Li et al., 2017) (Figure 5A). Ensemble data typically provide information averaged over a large population of cells and do not indicate which data points (for instance Hi-C chromatin contacts) co-exist in which individual structure of the population. Moreover, often these data are unphased and do not reveal the identity of homologous chromosome copies. To represent such missing information for each data type at single cell level, latent variables ${\mathcal{D}}^{\prime }$ are introduced, see also (Figure 4). For example, the latent variable of ensemble Hi-C data is a contact indicator tensor detailing whether a contact exists or not in each structure for all diploid chromatin pairs. A population consistent with all data modalities $\left(\mathcal{D},{\mathcal{D}}^{\prime }\right)$ is identified by jointly optimizing the latent variables ${\mathcal{D}}^{\prime }$ and structure coordinates X, such that the probability $\mathit{\text{log}}P(\mathcal{D},{\mathcal{D}}^{\prime }\lceil X)$ that the experimental data are expressed is maximal (Boninsegna et al., 2021; Hua et al., 2018; Tjong et al., 2016). Since there is no closed form solution for this high-dimensional optimization problem, a variant of the Expectation-Maximization (EM) method is used to iteratively optimize local approximations of the log likelihood function (Li et al., 2017; Tjong et al., 2016). Each iteration consists of two steps. The assignment step optimizes the latent variables by maximizing the log-likelihood over all possible values of ${\mathcal{D}}^{\prime }$, given the current population of structures, and produces the optimal allocation of data, i.e. which data instances are most likely present in the same structure. The modeling step then performs a structure optimization by maximizing the log-likelihood over all structures, given the current estimated latent variables ${\mathcal{D}}^{\prime }$. These steps are iteratively repeated until the coordinates of all structures and the latent variables are fully optimized.

Fig. 5.

(A) The Integrated Genome Modeling (IGM) platform uses Hi-C, Lamina DamID, SPRITE, and 3D FISH HIPMAp data as input to generate a population of genome structures (Boninsegna et al., 2021). (B) Generated structures are analyzed to identify the nuclear microenvironment of genes (Yildirim et al., 2021a). Genome structures provide information about the genes’ nuclear positions, distance to nuclear bodies, local chromatin fiber decompaction, and their cell-to-cell variability. Models also predict experimental data including speckle, lamina, and nucleoli TSA-seq (Chen et al., 2018), lamina DamID (Guelen et al., 2008), and GPseq (Girelli et al., 2020).

Population-based modeling has been successfully applied to generate predictive structures for a variety of cell types and organisms (Dai et al., 2016; Hua et al., 2018; Kalhor et al., 2012; Li et al., 2017; Tjong et al., 2016; Yildirim et al., 2021a; Zhu et al., 2017).

In a recent study, population-based modeling accurately predicted the functional microenvironments of genes in human GM12878 cells on a genome-wide scale (Yildirim et al., 2021a). The genome structures, generated from Hi-C data alone, predicted folding and subnuclear compartmentalization of the genome, along with the locations of nuclear speckles, nucleoli and lamina (Figure 5B). These models predicted both the mean spatial distances and the association frequencies of all genes to nuclear bodies, and showed high correlations with data from SON and laminB1 TSA-seq experiments, as well as trans A/B ratios, speckle and lamina association frequencies from DNA-MERFISH imaging. Thus, these models provided a detailed description of the nuclear microenvironment of genes, which was directly linked to its functional potential in gene transcription, DNA replication timing and subnuclear compartmentalization. Some chromatin regions were characterized by their strong preference to a single microenvironment (either transcriptionally active or silenced), due to strong associations to specific nuclear bodies. Other chromatin regions showed highly variable microenvironments between cells and lacked specific preferences, an indication that these regions are functionally ambiguous across cells. Overall, these models, generated from Hi-C data alone, showed high predictive values for several different experimental data, and can provide a first approximation for these data that otherwise would only be available through extensive experimental effort.

Another study in lymphoblastoid cells predicted chromosome-specific centromere clusters located at the nuclear interior, which played a pivotal role in chromosome positioning and stabilizing interchromosomal interactions and were confirmed by cryo-soft X-ray tomography (Tjong et al., 2016). The models also detected hundreds of frequently occurring multivalent chromatin clusters, which were enriched for the same regulatory factors (Dai et al., 2016). A different study characterized structural genome reorganizations during mouse neutrophil differentiation, leading to the discovery of chromosome supercontraction, driven by long-range heterochromatic interactions, combined with repositioning of centromeres and nucleoli (Zhu et al., 2017). Population-based modeling was also used to study major structural differences in genome organizations of cardiac myocytes and liver tissue (Chapski et al., 2019). By integrating Hi-C and lamina DamID data, population-based modeling of the Drosophila Melanogaster genome predicted location preferences for heterochromatin regions of each chromosome in the heterochromatic phase along preferred locations of the nucleolus, which were confirmed by FISH experiments (Li et al., 2017). Even though unphased Hi-C data cannot reveal interactions between chromosome copies, the models correctly showed an anti-correlation between predicted pairing frequencies for homologue chromatin regions and the enrichment of Mrg15 binding sites, a protein known to cause homologue unpairing.

The Integrative Genome Modeling (IGM) platform has recently expanded population-based modeling for large-scale multi-modal data integration to generate models of the human HFFc6 genome by simultaneously integrating data from ensemble Hi-C, lamina-B1 DamID, large-scale 3D HIPMap FISH and SPRITE experiments (Boninsegna et al., 2021). These structures predict accurately a host of orthogonal experimental observations from TSA-seq and DNA MERFISH experiments. The study provided the first systematic assessment of the benefits of multimodal data integration by comparing model accuracy from different combinations of data types. Overall, multi-modal data integration truly maximizes model accuracy, indicating that single data sources cannot fully capture all information about a genome’s structural organization. It was shown that systematic biases in some data can be compensated by adding other data sources. Indeed, artificially biased Hi-C data itself could not produce accurate models, but combining it with lamina DamID and 3D FISH data could. Also, different combinations of data sources produced models with similarly high predictivity. The study also uncovered the crucial role of low probability inter-chromosomal interactions for accurate predictions of genome organizations, including predictions of radial gene positions and chromatin compartmentalization. The IGM approach is highly versatile, as it can accommodate independent data sources expressed as spatial restraints on genomic loci. As of now, most imaging experiments have been used as validation for modeled structures; the next step is to use imaging directly as input information in the modeling process. For example, integrating volumes and shapes of nuclear bodies and nuclei from live cell imaging together with genomics data will provide more realistic descriptions of the nuclear genome. Also, super-resolution imaging can provide tentative positions of selected genomic loci, which can be incorporated with other data types. Preliminary results are very promising, and indicate that bridging genomics and imaging can be successfully tackled, as we strive for genuinely integrative genome models.

Outlook

Multi-modal data analysis provides unique opportunities for a deeper understanding of the dynamic nuclear genome structure, the underlying mechanisms that govern it and its functional importance. Integrated imaging methods probe the genome structure and transcriptome in the same cells, while sequencing-based multi-omics studies quantitatively describe detailed structural properties across multiple scales. Together, these fields provide unprecedented opportunities to reveal structure-function correlations. Computational modeling is positioned as the missing link to bridge these two fields by combining multi-modal data into one cohesive interpretation of the dynamic structure of genomes. Special care must be devoted to a faithful interpretation of experimental data, adequate model assessment strategies, and a description of model uncertainties. Then, integration of varied data modalities can overcome limitations and weaknesses of individual methods to increase accuracy and structural coverage of genome models. It is thus crucial to develop methods that faithfully incorporate imaging with genomics data. Data-driven methods are in an ideal position to address this challenge, as they can incorporate multiple data sources simultaneously. Resampling methods are well suited for studying local chromatin folding of chromatin regions and can efficiently accommodate also single-cell data for structure modeling. Population-based approaches can also use ensemble data to study the cell-to-cell variability of structures, and thus are well suited to study large-scale structures of entire genomes. Several efforts, including the NIH 4DN Nucleome initiative (Dekker et al., 2017), are key in bringing together research groups with multidisciplinary expertise in imaging, genomics and computational modeling, to study common cell types using complementary approaches. These and other joint efforts can provide synergies to further advance integrated approaches for generating quantitative models of the dynamic nuclear organization to better illuminate its functional implications.