Introduction

Chromosome conformation capture methods (3 C, 4 C, Hi-C, micro-C, …) have become a prevalent approach to investigate the interplay between DNA-related metabolic processes and the 3D folding of chromosomes (e.g. gene regulation, chromosome compaction, DNA repair and rearrangements1,2,3,4,5). Computational processing of HiC data has also provided powerful and elegant solutions to several genomic limitations, in particular allowing for robust genome scaffolding6,7,8,9. Furthermore, the application of Hi-C directly to complex microbial communities allows the characterization of whole microorganism genomes, the identification of mobile genetic elements such as viruses and plasmids and their assignment to their respective hosts10,11,12,13, and characterization of prophages activity14. International consortia have emerged to orchestrate efforts to characterize chromosome conformation and nuclear organization across cell types, tissue samples and species8,15.

Genome-wide chromatin conformation capture assays, such as Hi-C, micro-C or DNAse-C16,17,18, yield lists of pairs of interacting genomic loci at a base-pair resolution (stored in pairs files, in which each record describes a single measured contact between two genomic loci), which can be further binned to a window of chosen size and stored in symmetric sparse matrix files (where consecutive columns/rows correspond to consecutive genomic bins). Hi-C data specificities thus largely differ from the typical 1D genomic file formats (e.g. ‘.bigwig‘ or ‘.bed‘ files). While a single ‘.pairs‘ file format has been formally defined by the NIH 4D Nucleome Network19, three file formats have been independently proposed to store binned matrix files, each generated by a specific processing software. Hi-C Pro generates (i) sparse matrices as three column (bini / binj / countij) text files and (ii) region files describing genomic coordinates for each bin20, Juicer produces ‘.hic‘ files21 and ‘.(m)cool‘ files are generated by the distiller pipeline22. The ‘.hic‘ and ‘.(m)cool‘ formats are binary, multi-resolution, highly compressed and indexed files and can rely on companion libraries (respectively straw and cooler) to perform random access to a subset of the data. Softwares are also developed to manipulate files in these formats (Juicer tools, Juicebox and Juicebox Assembly Toolbox for ‘.hic‘ files and HiGlass, cooltools and coolpuppy for ‘.(m)cool‘ files23,24,25,26,27, FAN-C for ‘.hic‘ and ‘.(m)cool‘ files28). These computational solutions provide a dedicated shell command line interface (CLI) or a Python API. However, they are not embedded in a larger, genomics-centric ecosystem. Other softwares, such as HiCExplorer, GENOVA, mariner or HiCUP, also provide additional toolkits for Hi-C exploratory data analysis (EDA)29,30,31.

The Bioconductor project focuses on the development of R packages to provide classes, methods and functions dedicated to genomic datasets32,33,34,35. For this reason, Bioconductor has become the reference ecosystem for in-depth genomics investigation (encompassing most genome sequencing methodologies, genome annotations, single cell omics, multi-omics, etc). However, although core methods exist to represent genomic interactions in R (defined by the GInteractions class in the InteractionSet package36), and a few packages exist to perform statistical analyses to Hi-C data (e.g. comparing chromatin interaction frequency between samples with HiCcompare37), no standard chromosome conformation capture data structure has been defined so far in R. Furthermore, data import methods to parse Hi-C processed files in R are still lacking. Overall, the lack of a unified methodology surrounding Hi-C data has limited their integration in the powerful genomics-centric Bioconductor ecosystem, particularly compared to other omics approaches. To address these limitations, we formally defined a set of classes to represent chromosome conformation capture data with Bioconductor, and developed a set of tools to process, parse, analyze and visualize this type of data in R. Compared to existing solutions, this approach allows the end user to leverage existing, powerful genomics-centric methodology already implemented in Bioconductor. Here, we cover the key aspects of the chromosome conformation capture analysis workflow and describe the packages used at each step as well as interoperability features in R. We also present an online book (https://bioconductor.org/books/OHCA/) introducing the end user to the installation of the required packages, their specific functionalities and several examples of complete chromosome conformation capture analysis workflows.

Results

Data representation

The HiCExperiment package implements the ContactFile class (encompassing the CoolFile, HicFile and HicproFile classes) to connect to a contact matrix stored on disk in one of these three formats (Fig. 1, Fig. S1), supporting Hi-C, micro-C and other 3C-related data binned at a fixed resolution. A ContactFile instance also lists the resolutions available in the matrix file and metadata relevant for biological analysis. The import method provides random access to a ContactFile, to only import relevant chunks of data from large Hi-C matrix files. This instantiates a HiCExperiment object containing binned genomic contacts of interest stored as GInteractions (Fig. 1). Raw counts and normalized contact frequencies (if available) are automatically imported and stored in a list of scores. Additional methods are available to move around the Hi-C map (refocus), dynamically change its resolution (zoom), subset interactions or add qualitative or quantitative metrics (using the standard ‘[‘ subsetting operator and the ‘$‘ column accessor), and set/get general information related to the contact matrix (e.g. seqinfo, anchors, bins, topologicalFeatures, metadata). The HiCExperiment package also defines a PairsFile class to efficiently import ‘.pairs‘ files in R as GInteractions.

Fig. 1: Overview of the ContactFile and HiCExperiment class.
figure 1

The HiCExperiment package defines the ContactFile class, which acts as a pointer to a disk-stored contact matrix in.(m)cool,.hic, or HiC-Pro formats, and the PairsFile class, a pointer to a disk-stored tabular pairs file. Both ContactFile and PairsFile connections can be parsed in R using the generic import function. Importing a ContactFile instantiates a HiCExperiment object. The main functions and methods provided by the HiCExperiment packages to manipulate ContactFile, PairsFile and HiCExperiment objects are written in bold, and the data structure returned by each function/method is indicated between chevron. Interoperability with other R packages is possible thanks to coercing methods provided by the HiCExperiment package.

Full size image

All the classes implemented in HiCExperiment directly extend core Bioconductor classes and generic methods, including BiocFile, GenomeInfoDb and GenomicRanges, ensuring seamless parsing, manipulation and genomic representation of locally stored Hi-C contact matrices in R. Importantly, a HiCExperiment object can be seamlessly coerced as a GInteractions, a data.frame tabular object or a (optionally sparse) matrix. This facilitates its interoperability and the integration of Hi-C processed data with other pre-existing packages.

Thanks to ever-decreasing sequencing costs and improving technology, the average size of chromosome conformation capture datasets is continuously increasing, both in sequencing depth and in resolution. HDF5-derived ‘.(m)cool‘ and binary ‘.hic‘ files both efficiently store such large-scale data, and HiCExperiment objects instantized from these file formats benefit from efficient parsing libraries based on C code, optimized for speed. Furthermore, because random access is supported for these file formats, contact matrices can be partially imported in R, allowing manipulation of large datasets – such as deeply sequenced micro-C datasets – even on personal laptops with standard hardware configuration (e.g. 4 CPUs and 8–16 Gb RAM).

Data processing

Integrated workflows such as nf-core/hic, open2c/distiller-nf or Juicer20,21,38,39 efficiently use high-performance computing (HPC) environments to process chromosome conformation capture data. The large number of indirect operations they perform (e.g. container caching, sanitary check-ups and additional quality controls) results in a significant overhead and an increase in storage and memory requirements. These large workflows are therefore less suitable for processing on local workstations, where setting up dependencies is often cumbersome. The HiCool package was developed with these limitations in mind. HiCool is an R package that automatically sets up a basilisk-managed conda environment40 linked to hicstuff, a multipurpose lightweight Hi-C processing Python library41. This environment allows HiCool to align paired-end sequencing data to a genome reference, parse them into a standard ‘.pairs‘ file, filter out invalid pairs42 and PCR duplicates, bin them into multi-resolution balanced ‘.mcool‘ and ‘.hic‘ matrix files and automatically generate an HTML report of the processing (Fig. 2a, b). The implementation of HiCool as a Bioconductor package enables its efficient integration with other local Hi-C analysis packages (Fig. S1), and unlocks access to genomic databases, e.g. to automatically retrieve and cache cloud-hosted pre-built genome reference indexes by using a genome ID string (e.g. “mm10”, “GRZc10”, …), accelerating local Hi-C data pre-processing and direct import in R.

Fig. 2: Bioconductor workflow to import and analyze Hi-C data in R.
figure 2

a Raw Hi-C.fastq files can be pre-processed in R with HiCool to generate pairs and binned matrix files. Note that other standard workflows exist to pre-process Hi-C.fastq files in a command-line interface. Binned contact matrices can be imported in R as HiCExperiment objects using the import method defined in the HiCExperiment package. From there, Hi-C data can be directly visualized, or analyzed from two different perspectives, to perform matrix-centric analyses (e.g. matrix balancing, aggregation, comparison, etc.) or interaction-centric analysis (e.g. cis/trans contact ratios, distance-dependent interaction frequency, interaction profiles, etc.). b The HiCool package enables Hi-C.fastq processing in R by automatically setting up and temporarily activating a self-managed conda environment linking to hicstuff, a multipurpose lightweight Hi-C processing Python library. It internally generates the pairs and binned matrix files, saves a visual summary of the dataset and the processing workflow, and outputs a ContactFile and a PairsFile objects to facilitate the import of the pre-processed data in R.

Full size image

Hi-C visualization

The HiCExperiment object inherits methods from the core GInteractions and GRanges classes to provide a flexible representation of Hi-C data in R. The HiContacts package leverages these inheritances to explore HiCExperiment objects, focusing on four main topics: Hi-C visualization, contact matrix-centric analysis, interactions-centric analysis and structural feature annotation (Fig. 2A, list of functions in Table S1, interoperability illustrated in Fig. S1).

Hi-C exploratory data analysis is instrumental in generating hypotheses, discovering patterns, and directly answering biological questions. A generic plotMatrix function is provided in the HiContacts package and can operate on HiCExperiment, GInteractions or standard matrix objects, with extensive Hi-C-related customization options (Fig. S2). These include single matrix visualization (Fig. S2A, B), side-by-side comparison of two matrices (Fig. S2C), visualization of ratio, observed vs. expected (O/E) and correlation matrices (Fig. S2E–G), support for horizontal Hi-C maps (Fig. S2G), annotation of structural features and alignment with genomic tracks (Fig. S2F), and visualization of aggregated matrices (Fig. S2H). All visualization functions provided by HiContacts return ggplot objects that can be easily customized, e.g. to change the scaling, range or hue of the color map or to add additional details or labels and generate publication-ready figures.

Contact matrix-centric analysis

In a Hi-C analysis workflow, a preliminary requirement is the normalization of contact matrices. A well-established approach for matrix normalization is the matrix balancing approach42,43. By default, HiCool processing performs such normalization automatically, but the end user may need to manually normalize existing contact matrices. The HiContacts package implements the balancing of a HiCExperiment object, calculating weight scores for each bin and adding a new normalized score metric to each genomic interaction (Fig. 2a).

Several basic matrix operations can be applied to HiCExperiment objects. HiContacts defines operators to subset or merge Hi-C maps, or to subtract, divide or sum two Hi-C maps. HiContacts also provides a random subsampling method of Hi-C interactions that preserves sample-specific distance-dependent interaction frequencies.

Other calculations can be performed on HiCExperiment instances. HiContacts can estimate the overall expected signal for an imported contact matrix and compute the ratio of observed vs. expected (O/E) interaction frequency (Fig. S2F). HiContacts can also compute correlation matrices, to reveal a stereotypical plaid pattern in which interactions are enriched between chromosome segments belonging to the same compartment (AA or BB)18 (Fig. S2G).

Finally, an operation frequently performed when analyzing Hi-C matrices is the aggregation of matrix snippets, e.g. matrix subsets centered at all topological domain boundaries or all chromatin loops44. The AggrHiCExperiment class stores and averages the Hi-C signal across a set of snippets of interest. Because AggrHiCExperiment is an extension of the core HiCExperiment class, it inherits all the methods available to HiCExperiment instances, including visualization functionalities (Fig. S2H).

Interactions-centric analysis

The proportion of cis (intra-chromosomal) and trans (inter-chromosomal) interactions per chromosome can be calculated with HiContacts to investigate the propensity of chromosomes to form chromosomal territories with limited intermixing45 (Fig. S2I).

The chromosome-wide distance-dependent interaction frequency (a.k.a. P(s)) and its slope are valuable metrics that can be used to infer physical properties of individual chromosomes18,46. For example, in yeast entering mitosis, there is a significant decrease in interactions in the 20–30 kb range and a downward shift in the P(s) slope beyond this range (Fig. S2J). These features have been successfully used to accurately model the reorganization of the nuclear genomic content into mitotic condensed chromosomes47,48,49. Distance-dependent interaction frequency can also be summarized in scalograms (Fig. S2K): the median interaction genomic distance (±25%, or other quantiles specified by the end user) can be plotted along a linear axis representing a chromosome segment. This is often useful for deciphering the behavior of chromatin interactions along chromosomes50.

Finally, on a smaller scale, one may also be interested in studying interactions between an discrete viewpoint (a.k.a. bait) locus (e.g. a single or cluster of regulatory elements) and neighboring genomic features (e.g. other regulatory elements, gene bodies, repeats, etc). Profiles of contacts between such a viewpoint and the rest of the genome, sometimes referred to as virtual 4 C plots, can be computed with HiContacts (Fig. S2L). This feature is an efficient way to summarize and compare interactions between a number of different loci at once.

Structural feature annotations

A key step in using chromosome conformation capture data to explore the functional organization of chromosomes is the annotation of structural features. HiContacts implements methods to identify A/B chromosome compartments using eigenvector decomposition18, topologically associated domains (TADs51), using a diamond insulation score52 et chromatin loops using computer vision53 (Table S1). It is nonetheless advised to investigate structural features using a range of different methods. For instance, A/B compartments can be identified in R with HiTC and HiCDOC packages, while finer sub-compartments can be annotated using CALDER54,55,56. To allow end-users to use best-suited existing R packages, HiCExperiment objects can be coerced into the specific data structures such as matrices, data frames or GInteractions.

Data integration

Hi-C has gained traction in several fields related to genome biology, and several consortia have developed large-scale programs based on this technique. The NIH 4D Nucleome Program hosts a data portal that lists > 500 chromosome conformation capture experiment sets performed in humans and a variety of model organisms15. For each experiment set, Hi-C contact matrices, pairs files, coverage tracks and downstream analysis files are publicly available. The DNA Zoo consortium is using Hi-C scaffolding to generate genome references for hundreds of animals, plants, fungi and microorganisms. The polished genome sequences and corresponding Hi-C contact maps are directly accessible on a dedicated website8. Two R packages, fourDNData and DNAZooData, provide gateways to these databases. A list of available experiment sets and their metadata is provided within each package, and the actual data files (‘.mcool‘, ‘.hic‘, ‘.bw‘, ‘.bed‘, ‘.pairs‘ and ‘.fasta‘ files) associated with an experiment of interest can be seamlessly downloaded by providing a sample ID, and are locally cached for reuse across independent R sessions. For example, the mouse and chicken Hi-C data presented in Fig. S2 and Fig. 4 were retrieved directly in R using the fourDNData package. Providing programmatic access to existing databases will accelerate investigation in genome biology and open new avenues of research.

Interoperability between Hi-C packages

The HiCExperiment class provides interoperability between Hi-C packages. To illustrate this point, we present here a typical workflow to analyze seven Hi-C yeast datasets obtained from WT cells or mutants of the cohesin complex, synchronized in either G1 or G2/M, a stage in which chromosomes are compacted into arrays of loops57. For each library, Hi-C reads are aligned to the yeast reference genome and binned into contact maps using HiCool. Visual inspection of Hi-C maps suggests an increase in contact over longer distances in G2/M vs G1, enhanced in the absence of wpl1 and wpl1/eco1 (Fig. 3a), and this is confirmed by the P(s) curves (Fig. 3b). That the replicates are coherent is demonstrated by both the strong overlap of the P(s) curves (Fig. 3b) and by the stratum-adjusted correlation coefficients (SCC) calculated using HiCRep58. Here, we use the HiCExperiment coercion methods to convert the imported Hi-C maps into dense matrices, the input class required by HiCRep. SCC scores show that replicates for WT G2/M and wpl1 are overall correlated, while the two replicates for WT G1 are slightly more divergent (Fig. 3c). The stratum-dependent correlation between G1 and G2/M replicates decreases dramatically at short distance (10–30 kb), corresponding to the range of cohesin-mediated chromatin loops along G2/M chromosomes in yeast (Fig. 3d). In contrast, stratum-dependent correlation with wpl1 single mutant and wpl1/eco1 double mutant decreases at mid-range (50–100 kb) and mid-to-long-range (50–200 kb) respectively, consistent with the independent roles of Eco1 and Wpl1 factors in chromatin loop formation. We took advantage of the replicates to perform differential interaction (DI) analysis using the multiHiCcompare package59. Using HiCExperiment, we imported wpl1 and WT replicates chromosome XI Hi-C data and seamlessly coerced them into the multiHiCcompare-specific tabular format. The contact frequency fold-change and adjusted p-values computed by multiHiCcompare are injected back into the original HiCExperiment objects to visually represent these metrics in Hi-C maps (Fig. 3e) or as volcano plots, separating inter-arm and intra-arm interactions over chrXI (Fig. 3f). This analysis highlights that the increase in contact frequencies over longer distances occurs specifically for intra-arm contacts in wpl1 compared to WT, while contacts spanning the acrocentric chrXI centromere decrease. Interoperability with other packages in R is further illustrated in the following page of the companion online book: https://bioconductor.org/books/devel/OHCA/pages/interoperability.html.

Fig. 3: Comparison of distance-dependent interaction frequency across samples.
figure 3

a Hi-C contact matrix over chromosome X for four different yeast samples: WT G1 merged duplicates, WT G2/M merged duplicates, wpl1 G2/M merged duplicates and wpl1/eco1 single replicates (1 kb bins). b Distance-dependent contact frequency (P(s)) for each of the seven individual replicates. c Overall stratum-corrected correlation (SCC) scores between each of the seven individual replicates, computed with HiCRep. The line represents the average SCC across yeast chromosomes. The shaded ribbon represents the 90% confidence interval. d Distance-dependent correlation scores between WT G2/M replicate 1 and the other datasets, computed with HiCRep. e Contact matrix over chromosome X (2 kb bins). The color code represents the interaction frequency fold-change between wpl1 G2/M and WT G2/M (in log2 scale), computed with multiHiCcompare. f Volcano plot showing the statistical significance (-log10 p-value) versus the magnitude of the interaction frequency fold-change between wpl1 G2/M and WT G2/M within the chromosome X. P-values and fold-change estimates are computed with multiHiCcompare using a negative binomial exact test without correction for multiple testing. Intra-arm and inter-arm (centrosome-spanning) interactions are shown in two separate facets. Significant differential interaction (p-value ≤ 0.05, absolute fold-change ≥ 2) are highlighted.

Full size image

Delivering new biological insights using Hi-C

To illustrate how HiCExperiment can be leveraged to raise new biological hypotheses, we investigated a time-course Hi-C dataset of chicken cells released from a G2 block into mitosis60. We used the fourDNData gateway package to retrieve the data processed by the 4D consortium, and HiContacts to annotate compartments. Hi-C maps of chr3 (Fig. 4a) illustrate the progressive loss of compartment organization following G2 release, resulting in a rod-like polymer organization as early as 10 min after release when cells are in prophase and followed by the emergence of a second broader diagonal corresponding to helical coiling of chromosomes60. We generated saddle plots for each time point with HiContacts and noted that although AA and BB interactions are comparably lost at 10 min of release onwards, at 5 min BB interactions seem to be specifically depleted compared to AA ones (Fig. 4b). Correlation matrices over a magnified section of chr4 at G2, and 5 min and 30 min after release further confirmed that at this locus AA interactions are retained at 5 min while BB interactions disappear (Fig. 4c). We quantified the average contact frequency between pairs of genomic loci at these three time points, revealing a similar trend genome-wide (Fig. 4D). These results suggest that within minutes after G2 release, the B compartment - corresponding to heterochromatin - is affected faster than A compartment. This is consistent with the model whereby H3 S10 phosphorylation, occurring in late G2 first at chromocenters, initially induces HP1 eviction from H3K9me361 and heterochromatin dissolution62, and then spreads across entire chromosomes to allow for mitotic condensation63.

Fig. 4: Comparison of A/B interactions during mitotic entry.
figure 4

a Hi-C contact matrix over chromosome 3 for four different chicken samples: cells blocked in G2, or 5, 10, 15 and 30 min after G2 release (100 kb bins). b Saddle plots for the respective Hi-C datasets shown in G. 250 kb-wide non-overlapping genomic windows are grouped in 38 quantiles according to their A/B compartment eigenvector value (E1), from lowest eigenvector values (i.e. strongest B compartments) to highest eigenvector values (i.e. strongest A compartments). The average observed vs. expected (O/E) interaction scores for pairwise eigenvector quantiles are computed and plotted in a 2D heatmap. c Hi-C correlation matrices over a 16 Mb-wide segment of chromosome 4, in three different samples: cells blocked in G2, or cells 5 and 30 min after G2 release. The color scale indicates correlation of interaction profiles between any pair of pixels (100 kb bins). The A/B compartments identified in cells blocked in G2 are displayed below each heatmap. Note the loss of positive correlation preferentially between pairs of B compartments 5 min after G2 release. D Boxplot of O/E interaction scores between pairs of homologous (A-A or B-B) or heterologous (A-B and B-A) compartments (250 kb bins), in three different samples: cells blocked in G2, or cells 5 and 30 min after G2 release. Only interactions between loci > 5 Mb apart are used. The number of interactions (in thousands) is indicated in italics. The lower and upper hinges correspond to the first and third quartiles (the 25th and 75th percentiles), The lower and upper whiskers extend from the hinge to the smallest/largest value no further than 1.5 * IQR from the hinge (IQR: inter-quartile range, or distance between the first and third quartiles). Outliers are not displayed here.

Full size image

Discussion

Over the past decade, dozens of Bioconductor-hosted packages have led to widely adopted functional data classes for the generation, parsing and analysis of emerging genomic technologies. These developments allow advanced multi-omics analyses in R to an extent unmatched in other programming languages. Yet, manipulating chromosome conformation capture standard file formats in R remains particularly cumbersome. Here, we present the implementation of a flexible HiCExperiment class built on the robust Bioconductor core infrastructure. The HiCExperiment class facilitates chromosome conformation capture data integration (from Hi-C, micro-C, …) into existing genomic analysis workflows in R, reducing redundancy and improving interoperability. The companion HiContacts package provides the essential toolkit to compare, aggregate and further investigate HiCExperiment objects. A detailed introduction and extensive examples of Hi-C data analysis workflows are provided in the companion website https://bioconductor.org/books/OHCA/.

This rich ecosystem has several advantages compared to existing chromosome conformation capture libraries: (1) it is embedded in the genomics-focus Bioconductor ecosystem, ensuring a rational genomic representation of C data and evolvability (an extension of HiCExperiment to support single-cell Hi-C data is currently in development); (2) it extends pre-existing generic methods used by a large community, facilitating the intuitive integration of the C data in existing genomics workflows; (3) it supports quantitative and qualitative analysis of C data, represented as a numerical matrix or as a set of genomic interactions; (4) the daily building/testing infrastructure maintained by Bioconductor assures reproducibility of chromosome conformation capture analyses. For developers, a Docker image with preinstalled development versions of HiCExperiment-related packages is available here: https://github.com/users/js2264/packages/container/package/ohca.

A tight integration of HiCExperiment within the Bioconductor ecosystem unlocks future development opportunities for Hi-C data analysis. First, HiCExperiment could adopt the “tidy” grammar recently adapted to omics data investigation64, a project spearheaded by the Bioconductor community. This would make Hi-C data wrangling more intuitive and accessible to new investigators. Secondly, Bioconductor supports a DelayedMatrix framework and a block processing mechanism, which could be used to improve summarization of Hi-C data over multiple loci. Finally, Bioconductor is currently making efforts to deploy HiCExperiment functionalities within the AnVIL cloud computing platform, a project powered by Terra to facilitate collaborative data investigation between biomedical researchers65. We hope that this will accelerate the use of Hi-C in biomedical research, e.g. to shed light on genomic rearrangement events often identified through Hi-C66,67.

Methods

All analyses were performed using R 4.3.0 with Bioconductor 3.18. Further details of how each analysis was performed can be found in the Code Availability section.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.