3D models of fungal chromosomes to enhance visual integration of omics data

Abstract

The functions of eukaryotic chromosomes and their spatial architecture in the nucleus are reciprocally dependent. Hi-C experiments are routinely used to study chromosome 3D organization by probing chromatin interactions. Standard representation of the data has relied on contact maps that show the frequency of interactions between parts of the genome. In parallel, it has become easier to build 3D models of the entire genome based on the same Hi-C data, and thus benefit from the methodology and visualization tools developed for structural biology. 3D modeling of entire genomes leverages the understanding of their spatial organization. However, this opportunity for original and insightful modeling is underexploited. In this paper, we show how seeing the spatial organization of chromosomes can bring new perspectives to omics data integration. We assembled state-of-the-art tools into a workflow that goes from Hi-C raw data to fully annotated 3D models and we re-analysed public omics datasets available for three fungal species. Besides the well-described properties of the spatial organization of their chromosomes (Rabl conformation, hypercoiling and chromosome territories), our results highlighted (i) in Saccharomyces cerevisiae, the backbones of the cohesin anchor regions, which were aligned all along the chromosomes, (ii) in Schizosaccharomyces pombe, the oscillations of the coiling of chromosome arms throughout the cell cycle and (iii) in Neurospora crassa, the massive relocalization of histone marks in mutants of heterochromatin regulators. 3D modeling of the chromosomes brings new opportunities for visual integration of omics data. This holistic perspective supports intuition and lays the foundation for building new concepts.

Introduction

What if it were possible to see all the details of chromosomes inside the nucleus of a cell? In eukaryotic cells, the nucleus is a dynamic organelle which is highly organized and characterized by extensive compartmentalization of structural components in its three-dimensional space (see (1–4) for examples of reviews). In such a crowded environment, the arrangement of chromosomes is constrained and requires the formation of multiple chromatin domains to limit gene positions to preferred locations within the nuclear space (5). The spatial organization of chromosomes is of great interest, helping molecular biologists represent the objects they work with, and understand their interactions. Even if immense progress has been made in cell imaging, biological molecules (e.g. DNA, RNA, proteins) are too small to be individualized with optical microscopes (6,7) and consequently, the interior of the cell (and the interior of a nucleus even more) remains largely invisible to the human eye. Alternative solutions are based on molecular-scale techniques like X-ray crystallography (8), Nuclear Magnetic Resonance (NMR) microscopy (9) or electron microscopy (10). By analyzing atom arrangements in molecules, these techniques produce informative views of macromolecular complexity (11,12), but they require complex technical skills and expensive equipment. Furthermore, it is important to keep in mind that in the end, these images are still artificial representations of reality. In other words, they are ‘models’.

The use of models is widespread in biology. From model organisms to model systems, their interest is to understand a phenomenon in a simplified context, in order to, later, generalize to more complex situations. In cell biology for instance, structural models are used to represent cell components (membranes, nucleus, cytoplasm, etc.), to understand their organization, and to describe their constituent molecules (13). The work of David Goodsell provides an emblematic example (14). His drawings representing cellular compartments and their molecular actors are so striking because of the unexpected density of molecules and the complexity of their organization. Goodsell's illustrations have been featured as ‘Molecule of the Month’ on the Protein Data Bank website for over twenty years and the scientific journal Nature chose one of his paintings to make the cover of a special issue on COVID-19 (August 20, 2020 issue). Models make it possible to represent and summarize, in an intuitive but still scientifically rigorous way, the massive knowledge of cell molecular structures. Creating them thus represents a stimulating challenge, at the crossroads of multiple disciplines (biology, physics, computer science, art) (15).

Modeling chromosomes is challenging because they belong to the mesoscale, i.e. a length-scale that is larger than discrete molecular complexes yet still remains intracellular (16). As a consequence, they are both too small to be observed with precision under optical microscopes, and too large to be fully modeled at the atomic scale. In eukaryotes, chromosomes are long DNA molecules, tightly packed to fit within the nucleus, a space only a fraction of their length. To this end, DNA is wrapped around histone proteins to form nucleosomes, stacked to form chromatin fibers, themselves arranged into higher-order chromatin architecture (5). In this study, we are specifically interested in seeing the spatial organization of fungal chromosomes. Our laboratory has long-standing expertise in functional genomics projects in yeasts (17–19) or filamentous fungi (20–22), and spatial genome organization in fungi has already been investigated, for model species like Saccharomyces cerevisiae (23–25), Schizosaccharomyces pombe (26–29) and Neurospora crassa (30,31). Notably, the nuclear architecture of N. crassa, a multicellular fungus that grows as a mycelium with a network of hyphae (32), has structural homology (thanks to the existence of heterochromatin and euchromatin) with the human genome (31). This makes the N. crassa genome a cost-efficient model to study chromosome conformation. S. cerevisiae and S. pombe are distantly related yeast species (evolutionary distance of at least 400 Mya) that represent very different models of unicellular eukaryotes (33,34). Their genomes, of similar size (∼12 Mb), are organized into different sets of chromosomes (16 chromosomes for S. cerevisiae and 3 chromosomes for S. pombe). These yeast genomes are more than three times shorter than the genome of N. crassa (genome size is 41 Mb, with 7 chromosomes ranging from 4 to 10 Mb). Altogether, N. crassa, S. cerevisiae and S. pombe represent an interesting diversity of genomic situations for which much knowledge of nuclear organization is available in the literature (to view representative data from several publications, see Supplementary Figure S1).

The most emblematic feature highlighted in these articles is the ‘Rabl conformation’. This is a particular positioning of chromosomes in which their centromeres and telomeres are respectively clustered at distinct peripheral locations, inside the nuclear envelope. The Rabl conformation has been observed in N. crassa, S. cerevisiae and S. pombe (Supplementary Figure S1). It has also been found in plants (35), mice (36,37) and flies (38), underpinning the idea that it represents a conserved constrained genome structure to limit topological entanglement of chromosomes (39). The clustering of centromeres and telomeres at distinct peripheral locations inside the nuclear envelope induces chromosomes to be organized in a ‘clothespin-like’ structure, meaning that they are folded in half at the centromere, their arms against each other, their ends slightly bent. In S. pombe, the clustering of centromeres and telomeres is reinforced by protein anchors to the nuclear envelope and heterochromatin epigenetic marks (Supplementary Figure S1). At the molecular level, 3D genome organization in yeasts is determined by cohesin and condensin mediated structures: cohesin seems to form 40 to 50 kb globules (40,27) and condensin 300 to 500 kb large domains (27,29,41). This highlights how 3D genome structures are organized at successive, nested scales of condensation patterns. Another important feature relates to the definition of ‘chromosome territories'', where chromosomes are folded within their own space in the nucleus (42). Those territories are another well described example of the functional importance of chromatin structure and are observed in many different organisms. In yeasts, chromosome territories exist (23) but are ‘weak’ compared to the canonical version described in plants and animals. Chromosome territories are still debated in N. crass (31). Nonetheless correlations between 3D structuration and heterochromatin/euchromatin epigenetic marks were revealed by Hi-C and ChIP-seq studies (35,39,42). Centromeres and telomeres are heterochromatic (as in yeasts) and euchromatic regions are condensed into 20–40 kb globules (consistent with yeast cohesin globules).

This knowledge of the spatial organization of N. crassa, S. cerevisiae and S. pombe chromosomes was mainly built on Hi-C data analyses. The initial 3C technique evolved rapidly with NGS technologies, and recent technical improvements allow single cell analysis, with better spatial resolution (43) or haploid investigations (44,45) (see also (46,47) for reviews). Various types of Hi-C data are thus available in public databases, differing in accuracy and resolution, for numerous species (48), far beyond fungal models. Their processing steps are now well described (see (47) for review), and bioinformatics tools like the Juicer suite (49) or the HiC-Pro pipeline (50) are available and widely used by the community. Generally, the analysis of Hi-C data ends with the representation of a ‘contact map’ (symmetrical heatmap). A contact map shows the frequency of contacts observed between different portions of a genome, whose sizes depend on a predefined resolution (generally around 5 to 10 kb). Contact maps present the advantage of summarizing in a single image, all pairwise distances between genomic regions, both at short and long-range. Contact maps thus reveal chromosome clusters, their organization into domains and the formation of DNA loops. They are valuable because of their precision (resolution can go down to a few kb) and the information density they convey. Still, their biological interpretation is not trivial and, in the end, contact maps remain abstract representations of the spatial organization of chromosomes.

3D modeling of Hi-C contact maps is an interesting strategy to boost interpretation of Hi-C experimental results (51,52). It consists in calculating 3D coordinates (x, y, z) for all genomic intervals shown in a contact map, so that their pairwise Euclidean distances remain consistent with their contact frequencies in the original map. 3D modeling of the spatial organization of chromosomes is not trivial, but several software packages based on different strategies exist. For instance, ‘distance-based methods’ convert the number of contacts stored in the contact maps into distances and then resolve an optimization problem to fit 3D coordinates to those distances (53,54). Such methods often require fulfilling physical constraints related to known structural features of the genomes (for instance centromere and telomere clustering in opposite locations inside the nucleus or rDNA clustering outside the overall structure to form the nucleolus). This is how the 3D models already proposed for N. crassa, S. cerevisiae and S. pombe were obtained (Supplementary Figure S1, black stars). Alternatively, ‘probability-based strategies’ model contact numbers using random variables. This offers the advantage of taking into account the nature of Hi-C measurements (an average over a cell population) (55). Unlike other methods, the Pastis-NB method (implemented in the Pastis software (56)) implements a probability-based strategy based on negative binomial random variables. This distribution properly models count data generated by sequencing technologies, as illustrated by the wide adoption of negative binomial distributions to model RNA-seq data (57). The model therefore accounts for over-dispersion, performs well in low coverage settings and doesn’t require any initial physical constraints. Once a 3D model has been generated, a large panel of tools exists to visualize the 3D coordinates of genome models: Genome3D, GMOL, GenomeFlow, HiC-3DViewer, Csynth, WashU Epigenome Browser (58–63). Initially limited to the context of viewing biomolecules (nanoscale), visualization software was thus extended to representation of the 3D organization of chromosomes at the mesoscale.

Surprisingly, recent Hi-C studies on S. cerevisiae, S. pombe and N. crassa do not highlight 3D models (31,27,64). In a context where (i) an increasing number of Hi-C studies are generated, (ii) methods for 3D modeling of Hi-C contacts exist and (iii) software for 3D visualization is routinely available in structural biology, one may wonder why contact maps are not associated with 3D models more often. In the present study, we explored the potential of using 3D models as complementary information Hi-C contact maps to better understand the spatial organization of fungal chromosomes and enhance visual integration of omics data. For this purpose, we created a workflow called 3D Genome Builder (3DGB). Open source and freely available on GitHub, 3DGB links the creation of contact maps with the buildiing of 3D models. It also adds further processing of the 3D model output as PDB files, suitable for advanced visualization with molecular viewer software. Using 3DGB, we created several models of the S. cerevisiae, S. pombe and N. crassa genomes, starting from Hi-C data available in public databases. Different strains were analyzed (wild type and mutants for heterochromatin organization or structural proteins), at different stages of the life cycle. Our models showcase known characteristics of the chromosomal organization of these genomes. But they also reveal, thanks to the visual integration of omics data, important properties of regulatory proteins with critical functions for the maintenance of spatial organization. We thus illustrate the interest of 3D modeling of Hi-C contacts for studies of genome organization.

Materials and methods

3DGB implementation and workflow management

Technical details

3DGB was orchestrated with the open-source workflow management system Snakemake (65), which automates the different steps of a data analysis in a human-readable, Python-based language. 3DGB is structured in multiple individual rules depicted in Supplementary Figure S1. Software environments were deployed and isolated in a Conda environment (for Pastis and all custom Python scripts) and a Singularity container (for HiC-Pro).

3DGB inputs

3DGB requires the following inputs: Hi-C FASTQ files (provided as SRA IDs or as local FASTQ files), a reference genome sequence provided in a FASTA file (without mitochondrial DNA), one or several enzyme restriction site motifs, and finally, values for the output resolution to draw the contact map and generate 3D models.

Main steps in the analysis

The first steps of the workflow format the necessary information for the read mapping step: (i) FASTQ files are downloaded and compressed if not already provided by the user; (ii) a list of fragments derived from the genome FASTA file and the enzyme restriction site motif is generated by HiC-Pro (50) version 3.1.0; (iii) the size of each chromosome from the genome FASTA file is extracted with a custom script and (iv) the reference genome is indexed with Bowtie2 (66) version 2.4.4. Next, HiC-Pro generates a set of ‘valid pairs’, which are relevant reads used to generate and normalize contact frequencies (or counts). For a given resolution, these values are then used to compute contact maps (heatmaps) and to serve as inputs for Pastis (55) (version as of 21 July 2021) to build a consensus 3D model. In 3DGB, the Pastis-NB method is used (56). It iteratively computes 3D models of the organization of chromosomes, through negative binomial contact count modelization. For better reproducibility and portability, HiC-Pro is used in a Singularity image provided by the authors of the software. Pastis is installed in a conda environment. In 3DGB, a final step to refine the 3D model provided by Pastis is available. It consists in predicting the coordinates for additional beads, for which initial calculations were missing in the 3D model from Pastis. Missing coordinates usually occur in centromeric and telomeric regions. Predictions are performed by interpolation, using monotonic cubic splines, as implemented by the pchip method in the Scipy Python library (67). Note that beads with missing coordinates localized at the extremities of chromosomes are discarded and beads with aberrant coordinates are filtered out based on a threshold applied to the Euclidean distance value calculated between neighboring beads.

3DGB outputs

3D models are stored in PDB files. From the raw model produced by Pastis, 3DGB annotates chromosomes by numbers, based on the reference sequence specified in the 3DGB inputs. Examples of models, fully annotated with 3DGB, are provided as PDB files in Supplementary Data (see source code and data availability). In PDB files, the chromosome annotation is present in the residue id (1, 2, 3…), in the chain id (A, B, C…) and in the residue name (C01, C02, C03…). This adaptation of the standard residue and chain fields in PDB files is particularly useful, allowing immediate visualization of chromosomes with visualization software. Eventually, quantitative values can be associated with the beads of the 3D model (for instance ChIP-Seq signal intensity), using the B factor field in PDB files. Most PDB viewers can display B factors on top of the structures. This is the case for Mol*, used to create images presented in this article.

Analyses of experimental datasets in fungal species

Access to raw data from SRA database

All the Hi-C seq, Micro-C XL seq and ChIP seq raw FASTQ files were downloaded from the SRA database, see Table 1 and Supplementary Table S1 (second tab for ChIP seq) for all SRA ids. The reference genomes of S. cerevisiae and S. pombe were obtained from the NCBI Genome database, respectively version R64 (S288C) and ASM294v2.19. The nc14 version of the N. crassa genome was downloaded from the supplementary data of Rodriguez et al. (31) on NCBI GEO database, GEO dataset GSE173593. FASTA files with chromosome sequences only were generated as 3DGB input.

Table 1.

Main characteristics of raw data used in this study to create 3D models of fungal chromosomes

Model	Fungal species (genome version)	Strain	Hi-C data sources	SRA IDs	Final number of valid read pairs	Number of beads in the final model	Figure (model resolution in bp)
1	N. crassa (nc14)	WT	Rodriguez et al. 2022 (31)	SRR14362684	107 636 304	3976	2  (10 000)
				SRR14362685
				SRR16761088
				SRR16761089
				SRR16761090
				SRR16761091
				SRR16761092
2		WT	Galazka et al. 2016 (30)	SRR2105869	5 030 121	800	5  (50 000)
				SRR2105870
				SRR2105871
				SRR2105872
3		hpo		SRR2105876	2 922 526	808
				SRR2105877
				SRR2105878
4	S. pombe (ASM294v2.19)	WT	Tanizawa et al. 2017 (27)	SRR5149251	15 511 107	1223	4  (10 000)
				SRR5149252
5				SRR5149253	14 377 600	1211
				SRR5149254
6				SRR5149255	14 394 280	1204
				SRR5149256
7				SRR5149257	15 497 423	1200
				SRR5149258
8				SRR5149259	19 035 344	1213
				SRR5149260
				SRR5942526
9				SRR5149261	13 841 928	1224
				SRR5149263
10				SRR5149264	21 936 238	1201
				SRR5149265
				SRR5942527
11				SRR5149266	21 727 635	1189
				SRR5149267
				SRR5942528
12	S. cerevisiae (R64, S288C)	WT	Costantino et al. 2020 (64)	SRR11893084	16 847 912	2332	2 and 3  (5000)
				SRR11893085
13		Mcd1 depleted		SRR11893086	14 180 445	2323
				SRR11893087

The sources of Hi-C data (original articles and SRA identifiers) are given. All datasets were obtained on wild type strains, with the exception of models #3 and #13 which correspond to the N. crassa hpo mutant (used in Figure 5) and S. cerevisiae mcd1 depleted (used in Figure 3). The number of valid pairs of reads gives an estimation of the overall quality of the data and indicates the density of the contact frequencies measured in the experiment. The number of beads in the 3D model is determined by the chosen resolution during Hi-C data analysis and the length of the reference genome.

Application of 3DGB to create 3D models

3DGB was applied with default parameters (FASTQ files, see Table 1 for all details) and configuration files can be accessed in Zenodo and Github. For S. cerevisiae, 2 models were generated at 5 kb resolution from 4 FASTQ files (total of 31 028 357 valid pairs) and ChiP-Seq information from 2 FASTQ files were integrated. For S. pombe, 8 models at 10 kb resolution were built, from 19 FASTQ files (total of 136 321 555 valid pairs). For N. crassa, 3 models were created at 50 and 10 kb resolutions, from 14 FASTQ files (total of 115 588 951 valid pairs) and ChiP-Seq information from 9 FASTQ files were integrated. The chosen resolutions of the models were defined according to the technique used (Hi-C or Micro-C XL), the number of reads and their quality.

Evaluation of 3D model stability to random noise

The hpo mutant N. crassa contact frequency matrix was used as a ‘reference’ to generate ‘noisy contact frequency matrices’. The hpo mutant was selected because it has the lowest number of valid read pairs (29 22 526) and this sample is the most likely to be sensitive to noise. Let x_ij be the number of counts in the contact frequency matrix of this sample, with i and j the indices of the beads ranging from 1 to n = 825, with n the total number of beads for this sample. To create an ‘altered’ or ‘noisy’ N. crassa contact frequency matrix, it was set y_ij = x_ij + e_ij the number of counts in the noisy contact frequency matrix, where e_ij is a value sampled from a Poisson distribution with parameter λ, independently for each pair of beads (i, j). Note that the contract frequency matrices are symmetric and e_ij = e_ji for each pair of beads. The Poisson distribution was used, because it models shot noise, i.e. the variability in read counts that is due to the read sampling process rather than to biological variations (68). The level of noise (i.e. the mean parameter λ of the Poisson distribution used to generate shot noise) varies among the following values 0.1, 5, 10, 15, 20 up to 110. For each value of, 50 replicates were generated, i.e. 50 noisy contact frequency matrices. To assess the stability of 3D structures, the RMSD were computed between the 3D structure inferred from the reference N. crassa contact frequency matrix (the raw contact matrix with counts x_ij) and the 3D structure inferred from the altered N. crassa maps (the noisy contact frequency matrix with counts y_ij). This procedure leads to 50 values of RMSD for each value of λ. The Supplementary Figure S4 represents the RMSD values for λ ranging from 0.1 to 50. Above the value λ = 50, the RMSD values reach a threshold and do not increase with the value of λ. The RMSD between each one of the seven chromosomes of N. crassa were also computed, for each value of λ and each of the 50 replicates.

Visual integration of omics data

The same pipeline was used for integration of ChIP-Seq data from N. crassa and S. cerevisiae. To perform alignment, Bowtie2 was used with default parameters (Supplementary Table S1). The program samtools (69) was used to sort and index the SAM files into bam files, and the program bamCoverage was used to generate 5 kb or 50 kb .bedgraph files with RPKM normalization (for S. cerevisiae and N. crassa respectively). Each bin in the bedgraph corresponds to one bead of the 3D model. Note that the resolution of the bedgraph and of the 3D model are the same, allowing the integration of the data into the PDB file. For S. cerevisiae, the threshold was set to 80 counts after normalization, converting the continuous ChIP-seq signal into a binary signal highlighting the high residency CARs (64). For N. crassa, the ChIP-seq data was kept continuous to highlight the epigenetic mark distribution. The values below 20 counts are discarded for visualization.

Results

Part 1. Simplifying 3D model creation for Hi-C data analysts with a robust and reproducible workflow

General overview of 3D Genome Builder

To simplify the creation of 3D models for researchers interested in Hi-C data analysis, we created 3D Genome Builder (3DGB), a bioinformatics workflow that streamlines the generation of 3D models to visualize and explore the spatial organization of chromosomes, based on Hi-C experimental results. A general overview of the workflow is presented in Figure 1. Starting from Hi-C raw FASTQ files, 3DGB automatically performs the critical bioinformatics steps required to (i) compute Hi-C contact frequencies, (ii) infer associated 3D models of the chromatin organization and (iii) annotate and control the quality of the 3D models. These 3D models are stored in standard PDB files, so they can be further investigated with complementary visualization tools (see below). 3D models can optionally be enriched with supplementary quantitative omics data, such as ChIP-seq or RNA-seq signals (see next section). 3DGB has been designed to remain as simple as possible. Therefore, it requires only four basic inputs to be specified by the user (Figure 1, green boxes): FASTQ file identifiers, a FASTA file with the sequence of the reference genome, the list of restriction sites for the enzymes used during the Hi-C experiments, and the targeted resolutions for the Hi-C data analysis. This final parameter has an impact on the final 3D models, i.e. the smaller the value (specified in bp), the more detailed the 3D model.

Figure 1.

General overview of the 3D Genome Builder (3DGB) workflow. Inputs and outputs are respectively represented with green and orange boxes (left and right panels). The workflow is detailed as a sequence of blue boxes (middle panel). Each blue box represents a task of particular interest. The 3DGB workflow is orchestrated with Snakemake (see Supplementary Figure S2 and Materials and methods). All tasks are automatically chained one after the other until target outputs are produced. Final results comprise quality control reports generated by HiC-Pro, contact heatmaps and PDB files with the 3D models. PDB files are enriched with additional annotations (see main text for more details) and can be visualized with any PDB viewer software (mol* in our case)

The eight main steps required for Hi-C data processing are represented in blue boxes in Figure 1 and rely on two state-of-the-art software packages. The first step utilized HiC-Pro (50). It processes raw FASTQ files, performs quality control and generates normalized contact counts and associated figures (presenting important statistics to evaluate the quality of read mapping and justify potential read filtering). Then, Pastis (56) iteratively computes 3D models of the organization of chromosomes, through an original negative binomial contact count modelization (referred to as Pastis-NB). Interestingly, it outputs a consensus model, which, by its uniqueness, greatly simplifies downstream analyses and interpretations. Around those two main components (HiC-Pro and Pastis), 3DGB centralizes the configuration, performs contact calculations, generates contact heatmaps, builds 3D models and adds further processing of the 3D model output as PDB files. 3DGB is open source, available in GitHub (https://github.com/data-fun/3d-genome-builder) and archived in Software Heritage (swh:1:dir:bb48f33b57ce3cf7b00696a43eaac04f0c4047c8).

Enriched output PDB files for advanced visualization with molecular viewer software

The main 3DGB outputs are 3D models of genome organization. These 3D models are composed of beads for which 3D coordinates (x, y, z) were inferred from contact information (see Materials and methods). One bead represents several thousand base pairs corresponding to the chosen resolution during the Hi-C data analysis (see previous section). To facilitate the study of these models, we enriched the structures produced by Pastis as PDB files with a four-step procedure. First, 3DGB formats and enriches the 3D model output by Pastis for visualization and data integration by annotating each bead with the chromosome number. This is useful to distinguish chromosomes when viewing complete structures (see Figure 2 for illustrations). Second, it automatically reconstructs beads for which no coordinates could be calculated by Pastis, by interpolating missing coordinates from the existing ones (see Materials and methods). Note that we do not extrapolate missing coordinates, meaning that beads with missing coordinates located at the extremities of the model are discarded. Third, outlier beads, i.e. beads placed outside the overall model, are filtered out and deleted based on a threshold value which can be specified by the user. Four, quantitative values, such as ChIP-seq data, can be used to color the model, allowing visual integration of omics data on the 3D structure of the genome (see next section for detailed examples of such integration of omics data). All these additional functionalities were implemented in Python (70) scripts integrated within the Snakemake workflow (see Materials and methods).

Figure 2.

3D modeling of fungal chromosomes in wild type situations, in the model yeasts S. cerevisiae and S. pombe and the filamentous fungus N. crassa. (A) 2D representations of chromosomes in each species. They are displayed with the same scale, showing the diversity of genomes used in this study to create 3D models. (B) 3D models obtained with the 3DGB workflow. The source of the Hi-C data used to create them are presented in Table 1. They are visualized with HiC3D-Viewer (left) and Mol* (right). Note that HiC3D-Viewer links beads artificially, whereas with Mol*, beads remain individualized (they each represent a 5 or 10 kb chromatin region, depending on the species, see Table 1). (C) Key features of genome organization are highlighted. The Rabl configuration, the hypercoiling of chromosomes and their distribution into separate territories are well observed on the 3D models obtained with 3DGB and presented in (B) (see circles and rectangles).

Altogether, these enhancements provide convenient 3D models to be viewed and explored by Hi-C data analysts. For chromatin 3D model file formats, we provided the 3D model structure in the PDB format, a traditional file format widely used to store coordinates of molecular structures. Most software used in structural biology to view and manipulate structures can handle the PDB format. In this paper, we preferentially used Mol* (71), a ubiquitous viewer for large scale molecular structures, with a user-friendly web interface allowing visualization and customization of 3D models with only a few mouse clicks. We also used HiC3D-Viewer (61), a viewer especially developed for chromatin structures. Note that HiC3D-viewer requires conversion of the PDB file format into the G3D format. 3DGB also provides 3D models in the G3D file format, allowing users to choose the viewing software they prefer. As illustrations, images produced by Mol* and HiC3D-viewer are presented in Figure 2, for three different 3DGB models inferred in yeasts and a filamentous fungus (see next section).

Part 2. Seeing the spatial organization of fungal chromosomes with 3D Genome Builder (3DGB)

Creation of wild type models and visual consistency with the literature

To test the performances and the relevance of 3DGB, we chose three emblematic fungal species: Saccharomyces cerevisiae, Schizosaccharomyces pombe and Neurospora crassa. We wanted to create 3D models of their genome organization and confront the models with current knowledge found in the literature. The S. cerevisiae Hi-C data are from a recent study from Constantino et al. (64), the S. pombe Hi-C data are from Tanizawa et al. (27) and the N. crassa Hi-C data are from Galazka et al. (30). Detailed information regarding the source of the data is given in Table 1. As a result, we obtained new models of S. cerevisiae, S. pombe and N. crassa genomes. Three of them, which correspond to wild type situations, are shown in Figure 2. As expected, the key features of the spatial organization of the three species’ genomes were observed, i.e. the Rabl conformation, the hypercoiling of chromatin fibers and the chromosome territories. Finding these well-described characteristics was an important step in validating the relevance of the 3DGB workflow, especially considering that 3DGB uses a probabilistic-based strategy to infer 3D models (Pastis-NB), which is free of initial constraints, such as the relative position of chromosomes.

Added value of 3DGB models for visualization: examples ranging from a detailed view to a more general view of chromatin organization

Example #1: Spatial alignment of cohesin binding sites along chromosomes in S. cerevisiae

Ten years after the first 3D model of the S. cerevisiae genome (see (23) and Supplementary Figure S1), Costantino et al. (64) produced new Hi-C and ChIP-seq data, in order to better understand the relationship between the identified patterns of chromatin domains (information derived from Hi-C data analysis) and the cohesin residency regions (information derived from ChIP-seq data analysis). They used cells arrested in mitosis, and studied both the wild type strain and mutants altered for cohesin or its regulators. Notably, to finely describe chromatin loop formation, they used a recent improvement of the Hi-C technique, called Micro-C XL (43). This method has the advantage of greatly improving the detection of short-range interactions (at the scale of nucleosomes), while still allowing the detection of whole genome chromatin interactions. This is of particular interest for a species like S. cerevisiae, in which chromosomes are very small compared to the other species (S. cerevisiae chromosomes range from 0.23 to 1.53 Mb only, see Figure 2A). As a result, they showed that in S. cerevisiae mitotic cells, high residency cohesin anchor regions (named ‘CARs’ and detected with ChIP-seq) correspond to the genome-wide boundaries of chromatin loops (named ‘CAR domains’ and detected with Micro-C XL). This was an important observation which (i) supports the idea that the yeast genome is organized into recurrent defined chromatin structures delimited by cohesin and (ii) that this spatial organization of chromosomes can impact genome functions, as in mammals. Still, in their original article, the authors essentially show Hi-C contact heatmaps to support their interpretations.

To go further in this and enrich visualization, we collected their Micro-C XL data and reanalyzed them with 3DGB, to create an updated 3D model of the spatial organization of the S. cerevisiae chromosomes (see Materials and methods). Considering the very high precision of the data, we expected to observe both the general configuration (chromosome territories and Rabl conformation) and the fine chromatin domains (coiling), by zooming in and out with visualization software. Our results are presented in Figure 3. Note that the overall model (Figure 3A) is the same as the one shown in Figure 2B, but with different (color-blind friendly) colors used to identify the 16 chromosomes of the S. cerevisiae genome. ChIP-seq data were also reanalyzed (see Materials and methods) to locate the CARs on the new 3D model. They are shown in yellow in Figure 3. As expected from Hi-C contact maps presented in the original paper of Costantino et al. (64), we observed a continuous distribution of CARs along the S. cerevisiae chromosomes (Figure 3A). To complete this observation, we zoomed in on the chromosomes individually. An example of chromosome 13 is shown Figure 3B. Notably, we could observe that CARs, which are still represented by yellow beads on the 3D models (Figure 3D), were aligned in space, and this, independently of the length of the chromatin loops whose boundaries they represent (Figure 3C). This observation was true for all chromosomes (another example can be found in Supplementary Figure S3). 3D modeling of Hi-C contacts is therefore of significant interest here, revealing the existence of a cohesin skeleton, ensuring structural stability of the spatial organization of the S. cerevisiae chromosomes.

Figure 3.

Revealing the 3D skeleton formed by cohesin in S. cerevisiae chromatin. (A) 3D model of S. cerevisiae chromatin, obtained with 3DGB, using the Micro-C XL data from Costantino et al. (64). Each bead represents a 5 kb chromosomal region. Yellow color indicates genomic regions with a high value of ChIP-seq signal intensity of the Mcd1p cohesin subunit, in mitotically arrested cells. Alternative chromosomes coloring is presented in Supplementary Figure S6. (B) Isolation of chromosome 13 from the overall structure, showing the spatial alignment of the DNA binding sites of cohesin. Cohesin structures a chromosomal skeleton on which chromatin loops can form. (C) Juxtaposition of the 3D model and the associated contact map, for the beads which are located between 430 and 580 kb. On the 3D structure, the cohesin backbone is highlighted in yellow. Viewing this spatial alignment of cohesin is complementary to the information contained in the contact map, i.e. the frequency of contacts is very high inside genomic regions delimited by yellow beads. (D) Schematic representation of the cohesin complex in S. cerevisiae. The protein Mcd1p, depleted in cells used to create the model shown in (E) and (F), is indicated with a black star. Genomic regions, which are attached by cohesin are referred to as CARs, i.e. high residency cohesin anchor regions, and are visualized in yellow on the 3D models. (E) Isolation of chromosome 13 from the overall structure, obtained with the mcd1 mutant (see the main text). The yellow beads correspond to the cohesive backbone as it was originally identified in the wild type (B). (F) Juxtaposition of the 3D model and the associated contact map, for the beads which are located between 430 and 580 kb. This is the same genomic region exposed in (C).

To verify the relevance of this observation, we also reanalyzed the Hi-C data obtained with an S. cerevisiae strain in which the gene encoding the subunit Mcd1 of the cohesin complex was depleted (Figure 3D, black star). Results are presented in Figure 3E. This time, the 3D models showed more irregular coiling of chromosomes and globule-like structures. Again, this observation is relevant to the original article of Costantino et al. in which the authors found that in the absence of cohesin, the mitotic chromatin is not entirely disorganized, but rather structured into globular domains, dependent on histone modifications (64). To visualize what happens, in this context, to the cohesin skeleton previously observed in the wild type strain, we therefore mapped the genomic positions of CARs, as they were defined based on the ChIP-seq data in the wild type strain, on the 3D ‘mutant’ model. Interestingly, we observed a complete destabilization of the cohesin backbone initially detected in wild type (Figure 3E, yellow beads).

In summary, we illustrate here the interest of placing known 2D patterns of chromatin loops in a 3D context. Our observations reinforce the idea that a cohesin backbone stabilizes the spatial organization of chromosomes in S. cerevisiae. 3DGB brings a new perspective to the visualization compared to Hi-C contact maps: the flat loops are arranged in a coil supported by a backbone of aligned CAR regions.

Example #2: Oscillations in coiling of chromosome arms during the cell cycle in S. pombe

In the previous section, we were able to observe details of S. cerevisiae chromosome organization by zooming into the 3D model we built with 3DGB. This model was associated with cells arrested in mitosis and therefore represents a static view of the chromosomal organization at a very particular point in the yeast cell cycle. The goal of our second example was to evaluate the possibility of seeing, with 3D models, changes in the organization of chromosomes during the cell cycle. Indeed, chromosomes are subject to major structural constraints and undergo rearrangements at the different stages of the cell cycle (G2, M, G1 and S). We chose to explore this question using the data of Tanizawa et al. (27). The authors applied an in situ Hi-C protocol to follow the organization of the fission yeast S. pombe genome throughout the cell cycle. The authors observed that during mitosis, chromosomes are structured into large (300 kb to 1 Mb) and small (30–40 kb) domains, which are respectively structured by condensin and cohesin protein complexes (27). Based exclusively on their interpretation of Hi-C contact maps, they showed that if the mitotic organization into large domains gradually dissolves across the cell cycle, small domains remain relatively stable. They also hypothesized that the Rabl conformation was stable in interphase but disrupted during mitosis. With this dataset, we assessed our ability to observe these structural features in the 3D models obtained with 3DGB, at different stages of the S. pombe cell cycle. We collected Hi-C data from the original article (Table 1) and produced 3D models (see Methods), showing the organization of chromosomes at different time points of the S. pombe life cycle, after an initial synchronization of cells in phase G2. Our results are presented in Figure 4.

Figure 4.

Dynamics of the spatial organization of S. pombe chromosomes during the cell cycle. 3D models obtained with 3DGB using Hi-C data from Tanizawa et al. (27). Each bead represents a 10 kb chromosomal region and is colored according to the chromosome to which they belong (from 1 to 3). Eight models are shown here. They correspond to different organization of S. pombe chromosomes during the fission yeast cell cycle. After the synchronization of cells in phase G2, the following time points were analyzed: 20, 30, 40, 50, 60, 70, 80 and 120 min. Their correspondence with the cell cycle phases is shown in the middle. Two time points are thus in the G2 phase (20 and 120 minutes), three time points are in the M phase (30, 40 and 50 min), two time points are in the G1 phase (60 and 70 min), and finally one time point is in the S phase. Centromere and telomere regions are annotated. Centromeres of all chromosomes are constantly clustered, whatever the cell cycle time point. This is consistent with previous observations (see the main text). Coiling of chromosome arms, as described in the literature (40,27), is visible on the 3D models. Zooms are provided in (A) and (B) boxes, which correspond to time points (respectively 120 and 50 min) for which the intensity of coiling is minimal and maximal (see the main text for further explanation).

As expected, the Rabl conformation is visible, particularly during interphase (G1, S and G2 models, between 60 and 120 min). However, it becomes somewhat disorganized at the transition between G2 and Mitosis (Figure 4, 20 min model), even though the cluster of centromeres remains stable during the entire cell cycle. Telomeres of chromosomes 1 (yellow color) and 2 (orange color) also remain associated, whereas chromosome 3 (red color), in which ribosomal DNA repeats are found, occupied an external position consistent with compartmentalization of rDNA in the nucleolus. We also observed that the level of chromatin compaction varies over the cell cycle, with an observed minimum in G2 (Figure 4A) and an observed maximum in late mitosis (Figure 4B). Indeed, the 120 min model shows uncoiled chromosome arms, with loose globule-like folding (Figure 4A) whereas the 50 minute model chromosome arms are structured into regular coils of ∼250 kb (Figure 4B). This more compact (see the zoom-in 1 Mb region Figure 4A, B) and structured (regular coiling) organization can be seen gradually appearing on the 20, 30 and 40 minute models (Figure 4, right arrow). Finally, it is important to keep in mind that the Hi-C data of each model were obtained independently, reinforcing the biological relevance of the observed similarities in relative position of chromosomes and level of compaction. Therefore, the highly organized mitotic structure gradually fades though the G1 and S phases to return to the Rabl conformation (Figure 4, left arrow).

In summary, the 3D models obtained with 3DGB consistently illustrate the oscillations in coiling of chromosome arms during the cell cycle in S. pombe. This strengthens the original study by Tanizawa et al. (27) and opens new perspectives for further analysis, for example omics data integration.

Example #3: Massive relocalisation of histone marks in N. crassa heterochromatin regulator mutants

In the N. crassa genome, heterochromatic regions are a major component of the chromosome conformation (30). Constitutive and facultative heterochromatin are respectively genomic regions that contain few genes with little transcription (constitutive heterochromatin) and genomic regions that contain genes with regulated gene repression (facultative heterochromatin). At the molecular level, constitutive and facultative heterochromatin can be distinguished by the presence of H3K9me3 or H3K27me2/3 histone marks (30). Interactions between constitutive and facultative heterochromatin are complex and an important subject of discussion in the literature. In that respect, an emblematic observation is that the reduction of H3K9me3 in constitutive heterochromatin, causes the redistribution of H3K27me2/3. In particular, in a genetic context in which the heterochromatin protein 1 (Hp1, which recognize H3K9me3) is lost, H3K27me3 is depleted from facultative heterochromatin and H3K27me2 is gained at constitutive heterochromatin (72). Our objective was to assess the ability to see this phenomenon at the scale of the complete genome. To do so, we started by creating 3D models with 3DGB, from the wild type (WT) and HP1-deficient (hpo) strains. These models were then used for visual integration of ChIP-seq data, referring to the genomic location of histones with H3K9me3 and H3K27me2/3 post translational modifications (see Materials and methods). Results are shown Figure 5A.

Figure 5.

Visual integration of ChiP-seq data using 3D models of the N. crassa genome. (A) 3D models obtained with 3DGB using Hi-C data from Galazka et al. (30), respectively in the wild type (WT) and HP1-deficient (hpo) strains (see Table 1 and Materials and methods section for technical details). Each bead represents a 50 kb chromosomal region and is colored according to the chromosome to which it belongs (from 1 to 7). The colocalization of centromeres and telomeres is observed in both models (centromeres on the bottom and telomeres on the right). (B) Same models as shown in (A), using the ChiP-seq signal intensity as a color code for beads (see Methods). ChIP-seq data are from Basenko et al. (75) and refer to the genomic location of histones with H3K9me3 and H3K27me2/3 post translational modifications, respectively. In the WT model, the intensity of the ChIP-seq signal for H3K9me3 is particularly high (yellow color) around centromeres (compared to the rest of the genome, dark red color), whereas the intensity of the ChIP-seq signal for H3K27me3 is particularly high at telomeres. These observations are relevant to known functions of constitutive and facultative heterochromatin (see the main text). In the hpo strain, important changes in the intensities of ChIP-seq signals are observed with, notably, a massive relocalization of H3K27me3 high intensity signal to the centromeres (blue arrow). (C) Isolation of chromosome 4, better highlighting the changes in H3K27me2/3 ChIP-seq signal intensity between WT and hpo strains.

As expected from the literature, we observed a slight difference between the two model organizations, essentially the intensity of chromatin compaction and the relative position of chromosomes 6 and 7 (Figure 5A). To extend this observation, we calculated for each 50 kb region represented by a bead in the 3D models, the intensity of ChIP-seq signals arising from experiments targeting H3K9me3 and H3K27me2/3 post-translational histone modifications (see Methods). Results are shown Figure 5B (whole genome) and Figure 5C (chromosome 4). In WT, we could observe as expected an accumulation of H3K9me3 histones in centromeres, relevant to constitutive heterochromatin, and an accumulation of H3K27me2/3 histones in sub-telomeric regions, relevant to facultative heterochromatin. In hpo, ChIP-seq signals appeared to be quite disorganized, especially for H3K27me2/3 marks, and massively relocalized to centromeres (blue arrow, Figure 5B). This observation was expected according to the literature, but for the first time it can be seen clearly, in a simple and integrated way.

This example illustrates the interest of 3D models in the context of multi-omics data integration. If Hi-C data allows us to better understand the organization of chromosomes, other omics data (such as ChIP-seq here) bring insights on the mechanisms of genome function. It is important to keep in mind that the images presented abstract a very large quantity of data, i.e. several million measurements (frequencies of contacts, probability of interactions, quality values, etc.). These 3D models can confirm or invalidate hypotheses on the overall functioning of genomes, and possibly lead to the formulation of new hypotheses.

Additional benefits of using 3DGB

Detection and correction of assembly errors

In the context of our analyses of Hi-C data in N. crassa, we faced an unexpected situation with respect to the reference genome sequence for this organism. Indeed, the N. crassa genome assembly, originally published in 2003 (32), is composed of 7 chromosomes and 13 supercontigs, all available in GenBank under the accession ‘assembly nc12’. In 2016, Galazka et al. (30) identified an inversion of the contig named ‘12.304’, corresponding to a large region in chromosome 6. In 2022, Rodriguez et al. (31) further improved the original assembly ‘nc12’ based on Hi-C data analysis. The updated assembly that integrates the improvements of Galazka et al. and Rodriguez et al. is available in the GEO database (73) under the accession ‘assembly nc14’. When we started using 3DGB to create a 3D model of the wild-type N. crassa chromosome (Figure 2), we used as reference genome the ‘nc12’ assembly, available from the GenBank database (74). Surprisingly, we observed in our inferred 3D model, an inconsistency in the order of the genomic regions, with respect to the succession of our beads (Figure 6). Our 3D model thus directly highlighted the inverted contig 12.304 on the chromosome 6, originally found by Galazka et al. (30). At 10 kb resolution, the 95 beads used to represent this region were placed in the 3D space according to the information of Hi-C contacts only, independently of the reference genome assembly. This explains why we were able to highlight a mismatch between the chaining of beads in the 3D models and the numbering of these beads according to the genome sequence (Figure 6A). Note that when using the ‘nc14’ assembly of the genome as reference, the obtained 3D model of chromosome 6 we obtained was coherent both spatially (order of the beads) and sequentially (order of the genomics regions). From this experience, we developed an automated procedure (integrated to 3GDB as an option) to detect and correct inverted contigs by comparing the chaining of beads in the 3D model (based on the distance between adjacent beads, Figure 6B) and the numbering from the genomic sequence written in the FASTA file. This method also produces a corrected version of the 3D model and of the genome assembly (Figure 6C).

Figure 6.

Inconsistency between the spatial organization of chromosome 6 in N. crassa and the reference genome assembly ‘nc12’. (A) 3D model of chromosome 6 obtained with 3DGB using Hi-C data from Galazka et al. (30) and the ‘nc12’ assembly of the N. crassa genomic sequence as reference. Each bead represents a 10 kb chromosomal region. The numbers label the beads, according to the genomic region (DNA sequence) with which they are associated. Therefore, the bead number 1 corresponds to the interval]0, 10] kb of the linear genomic sequence, the bead number 2 corresponds to the interval ]10, 20] kb, the bead number 3 corresponds to the interval ]20, 30] kb, etc. Unexpected transitions between the pairs of beads 185–279 and 186–280 are shown with different colors (blue and red). (B) Euclidean distances between 3D beads with successive numbers as defined with genomic intervals on the DNA sequence. Unexpected high values are observed between bead pairs 185 - 186 and 279–280. They are in line with the structure shown in (A). (C) Based on the inconsistency revealed in the Euclidean distances, it is possible in 3DGB to automatically correct the genomic sequence used as a reference.

Quantification of the stability of 3D models to random noise with multiple uses of 3DGB

The advantage of using a workflow like 3DGB is the possibility of automating tasks. In this part, we assess the stability of the 3D organization of the chromosomes with respect to potential inaccuracies in the contact frequencies measured with Hi-C (see Materials and methods). Our strategy was to alter the original values of contact frequencies (issued from the Hi-C data) by adding a ‘shot noise’ on the contact frequencies, and then run 3DGB to create an associated 3D model. If the intensity of the shot noise was low, the output model was expected to be close to the original model (obtained from the original data). The resulting RMSD score, calculated by comparing the 3D structure inferred from the reference N. crassa contact map and the 3D structure inferred from the altered N. crassa map, was also expected to be low. Altogether, we assessed 23 levels of intensity of the shot noise (see Materials and methods) and generated, automatically, a total of 1150 models. Our results are summarized in Supplementary Figure S4. As expected, we observed an increase in the RMSD scores, when the intensity of the shot noise increases. This was more striking for the global structure (Supplementary Figure S4B, green boxplots) than for individual chromosomes (Supplementary Figure S4B, blue boxplots). This observation underlines the greater sensitivity of the organization of the chromosomal territories compared to the internal organization of the individual chromosomes.

Discussion

The objective of this work was to explore the interest of creating 3D visualizations of chromosome organization to complement the classic contact maps producing during Hi-C data analysis. We found that, although methods to create 3D models exist, they are not very often exploited. Emblematic 3D representations of the complete genome of S. cerevisiae had been obtained >10 years ago (23,25), but only partial structures were presented in publications for S. pombe and N. crassa (see Supplementary Figure S1). Our first intention was therefore to create, from existing bioinformatics tools and Hi-C datasets, complete models for these three species. In addition to what has already been done for S. cerevisiae by Duan et al. for instance, we wanted to build models without specifying initial physical constraints regarding chromosome positioning. The Pastis-NB method was particularly suitable for this task. To automate the link between calculation of contact maps and generation of 3D models, we built 3DGB, a high-throughput Hi-C data processing workflow (Figure 1). With this workflow, we provided enriched PDB files for advanced visualization with molecular viewer software (Figure 2) and we created several hundred models in reasonable computing times (a few hours per model) (Supplementary Figure S4 and S5). The 3DGB workflow is available to other scientists who would like to add 3D models to their analyses of Hi-C data.

The relevance of 3D models of the spatial organization of chromosomes is an important question of our study. Beyond the production of a ‘beautiful image’, what is the added value of these models for Hi-C data analysts? It is indeed important to keep in mind that these models are not ‘pictures’ of the interior of cell nuclei. They are based on experimental data, whose quality may vary and significantly impact the inferred model. Ultimately, they represent measurements of contact frequencies observed at the scale of cell populations and therefore, they are simplified representations of reality. This work on fungal genomes taught us that the visualization of 3D models is profitable for several reasons. First, it gives an overall representation of the organization of the chromosomes (Figure 2). While contact maps are very useful for identifying local chromatin structures (at the scale of kilo base pairs), 3D models allow a complete zoom out, at the scale of the entire chromosome, giving a global representation of all contact frequencies. Second, 3D models can both serve as the final visualization of the results of a Hi-C study, but also as a starting point for further data integration. The results we presented for the S. cerevisiae and S. pombe genomes are examples of final visualization results. On the one hand, our models revealed important features of the organization of chromosomes, respectively a backbone of cohesin (Figure 3) and hypercoiling dynamics (Figure 4), which are difficult to fully understand from the contact heatmaps alone. On the other hand, the results we presented for N. crassa are examples of original data integration (Figure 5). This time, 3D models are starting points for additional exploration. The coloring of the 3D representation of the models, based on quantitative measurements from ‘omics’ experiments, albeit a simple idea, is particularly effective for studying global phenomena, such as the massive relocalization of histone marks in a mutant, as shown here for N. crassa.

Still, it is worth considering that contact-based modeling of the genome 3D organization has some limits. They remain a 3D representation of a Hi-C based contact network, with chromatin regions modeled as independent beads, each representing several kilo base pairs. The sequencing depth of the Hi-C data and the chosen resolution for their analysis have a direct impact on the accuracy of the final model: for a given number of reads, the lower the resolution, the fewer the inter chromosomal contacts detected. We observed that when the sequencing depth is low, the chromosomes in the 3D models are more distant from each other, for example in the mitotic S. cerevisiae and S. pombe models (Figure 2). In such a situation, it is difficult to discriminate between the biological compaction of the chromatin and the limitations due to the Hi-C experimental technique, when only looking at the final volume occupied by a model. However, structural features like level of coiling are still visible (Figures 3 and 4) and can provide insight at the level of the biological compaction of chromatin, compensating for technical limitations (resulting in chromosomes which are more spaced). As stated previously, another inherent limitation of 3D models is that they are ‘static’, ‘population averaged’ pictures of a biological system (chromatin organization) known to be highly dynamic and variable from cell to cell. These artifacts will exist as long as the Hi-C data are snapshots obtained from cell populations. Still, it is important to note that Rabl conformations and weak chromosome territories were observed in all species (Figure 2), even with the N. crassa model, which is an average representation derived from Hi-C experiments performed on an asynchronous cell population. Therefore, with synchronous cells, the situation can be expected to be even better. This is what we observed with S. pombe models. Because the cells were initially synchronized, the Hi-C datasets produced at different stages of the cell cycle revealed several other interesting genome organizations in the inferred 3D models (Figure 4) and we thus managed to render the dynamics of chromatin. As for the ‘population average’ limit, single cell Hi-C strategies are developing (36), opening the promising perspective of creating 3D genome models of single cells. A final limit of current 3D models is the missing regions of the fungal genomes. As illustrations, the models presented here are built using genomic sequences in which the rDNA repeats were deleted, thus preventing the correct representation of structural features in the nucleolus (explaining the hole we can observe on chromosome 12 in S. cerevisiae models, Figures 2 and 3). An additional important simplification arises from the joint representation of sister chromatids and the complexity of working with diploid genomes (in classical Hi-C data analyses, the contact measurements cannot be distinguished between identical sequences). Strategies are emerging to solve haplotypes (44,45) and the software used in 3DGB has tuned options for this: at the mapping step, HiC-Pro can build allele-specific contact maps if SNP information is provided. At the 3D modeling stage, the Poisson-based method of Pastis has been extended to haplotype resolution. Even if this allows us to imagine creation of more complete models in the future, it is important to keep in mind that the fungal genomes examined here, (i) remain haploid during the cell cycle and (ii) have sister chromatids tightly maintained together by cohesion. We therefore believe that representation of fungal chromosomes as one string of beads is informative.

Overall, the limits underline the dependence of 3D modeling on the quality of experimental Hi-C raw data and generation methods. We are confident however, that the technical improvements that are being introduced at a rapid pace will progressively minimize these limitations. Importantly, even though the technical ceiling is not yet reached, we have already managed to highlight important fungal genome characteristics in a novel way by reanalyzing public datasets. For N. crassa, the 6 3D models produced by 3DGB condensed the information of 5 research articles and about 20 raw FASTQ data files into one rich large-scale illustration (Figure 5). They bring new opportunities for visual integration of omics data: while Hi-C analysis often implies a ‘zoom-in’ logic, focusing on precise regions of a contact map, 3D modeling completes and enhances this logic with a ‘zoomed-out’ vision.

In conclusion, we have presented a holistic approach that is favorable to intuition and new hypotheses. We explored large-scale integration of epigenetic ChIP-seq data into the 3D context of chromatin, but any suitable combination of omics datasets can be made using the genome model as a visual support. For example, we are currently exploring the interpolation of genes on the 3D genome structure. 3DGB could be adapted to provide a model at the gene (one bead each) level, to allow the mapping of gene-dependent omics data, such as RNA-seq, on the chromosome structure. Gene positions need not be accurate, but rather provide a convenient way to plot omics data on top of the 3D organization of the chromatin. Going even further, it could be of interest to integrate information from protein occupancy or transcription factor interaction networks. Seeing chromatin domains as loops of a coil instead of distinct entities could open new insights on inter-domain gene regulation.

Data availability

The 3DGB workflow is available on GitHub (https://github.com/data-fun/3d-genome-builder) and Software Heritage (https://archive.softwareheritage.org/, swh:1:dir:bb48f33b57ce3cf7b00696a43eaac04f0c4047c8). The .PDB files of the 13 models presented in this study are available on Zenodo (https://doi.org/10.5281/zenodo.7740302), as well as animated GIF for 4 of them. The .YML (config files) for generating those 13 models are available on GitHub and on Zenodo.

Supplementary data

Supplementary Data are available at NARGAB Online.

Funding

Agence Nationale pour la Recherche (MINOMICS project) [ANR-19-CE45-0017].

Conflict of interest statement. None declared.