Skip to main content
NAR Genomics and Bioinformatics logoLink to NAR Genomics and Bioinformatics
. 2026 Nov 12;8(4):lqag136. doi: 10.1093/nargab/lqag136

GOntact: using chromatin contacts to infer target genes and Gene Ontology enrichments for cis-regulatory elements

Alexandre Laverré 1,2, Eric Tannier 3,4, Philippe Veber 5,✉, Anamaria Necsulea 6,✉
PMCID: PMC13754754  PMID: 43021986

Abstract

Cis-regulatory elements (CREs) can be efficiently predicted genome-wide, but identifying their target genes remains challenging. Regulatory interactions between genes and CREs can take place over long genomic distances, often bypassing genes. Inferring CRE targets based on genomic proximity, as traditionally done in genomic studies, can thus be misleading. Thanks to chromosome conformation capture techniques, chromatin contacts between CREs and gene promoters can be assayed at the genome-wide scale, thus permitting more accurate predictions of CRE target genes. Here, we present a standalone computational tool and webserver named GOntact, which infers CRE target genes using chromosome conformation capture data. GOntact can be used to derive Gene Ontology (GO) enrichments for CRE sets, thus providing a basis for functional interpretation. We apply GOntact on enhancers active in several embryonic tissues, using Promoter Capture Hi-C data to infer chromatin contacts between genes and CREs. We show that GOntact predicts functional annotations that are coherent with enhancer activity. Compared to genomic proximity, GOntact predicts consistent but more specific functional annotations. With the increasing availability of high-resolution chromatin contact data, we believe that GOntact can provide better-informed target genes and functional predictions for CREs. GOntact is available at https://gontact.univ-lyon1.fr (webserver) and https://gitlab.in2p3.fr/anamaria.necsulea/GOntact (standalone command-line tool).

Graphical Abstract

Graphical Abstract.

For image description, please refer to the figure legend and surrounding text.

Introduction

In multicellular organisms, gene expression is regulated by complex interactions involving numerous trans-acting factors and cis-regulatory elements (CREs). These elements, which are by definition situated on the same chromosome as their target genes, can activate or repress gene expression (in the case of enhancers or silencers, respectively). The genomic positions and patterns of activity of enhancer elements can now be efficiently assayed with various molecular techniques, including chromatin immunoprecipitation followed by sequencing [1, 2], DNase I hypersensitivity assays [3], or assays for transposase-accessible chromatin using sequencing [4]. These techniques are commonly used to predict CRE positions. Genome-wide maps of CREs are now readily available, in particular for model organisms [5]. However, identifying CRE target genes remains a difficult question, even when their localization and molecular function have been ascertained.

The most commonly used approach to predict CRE target genes is based on an assumption of genomic proximity [6]. This assumption is undoubtedly reasonable, at least for enhancer elements, which were originally demonstrated to act on their neighboring genes [7]. However, there is now solid evidence that enhancers can be situated at large genomic distances from the genes they regulate. A now-classical example is the ZRS enhancer of Shh, which is situated 1 megabase (Mb) away from its target gene [8]. Through fluorescence in situ experiments, it was demonstrated early on that distal enhancers are brought in physical proximity with target gene promoters in the nucleus, likely through the formation of chromatin loops [9, 10]. The physical proximity between DNA segments that are distant on the linear genome can be detected through chromosome conformation capture techniques [11, 12]. Applications of these techniques have confirmed that long-range regulatory interactions between enhancers and target genes are frequent [13–15]. Additional evidence comes from perturbation experiments, which indicated that a large fraction of regulatory interactions (up to 30%) occur between enhancers and genes that are not their nearest neighbors [16].

Recent technical developments have enabled high-resolution mapping of chromatin contacts between promoters and distal CREs. The Hi-C approach was designed to investigate chromatin contacts between all possible pairs of genomic regions, at a genome-wide scale [12]. However, this approach requires great sequencing depth to achieve high resolution and sensitivity [17]. The Micro-C technique, in which the genome is digested with a micrococcal nuclease rather than with restriction enzymes [18], enables better resolution [19]. Other Hi-C variants offer better resolution and sensitivity for a restricted class of chromatin interactions, such as those mediated by a given protein or in which the interacting regions carry a specific histone mark [20]. For promoter-CRE interactions, capture Hi-C and capture Micro-C approaches are particularly well suited, as they can be designed to target chromatin contacts that involve gene promoters, without additional requirements on the chromatin modifications of the promoter or on protein mediators [15, 19, 21–23]. The resulting data are enriched in regulatory interactions between promoters and CREs and can thus provide a better-informed prediction of CRE target genes.

To take advantage of the increasing availability of high-resolution chromatin contact data, we developed a new tool named GOntact, which uses this type of data to infer associations between CREs and genes. Importantly, GOntact can be used to derive Gene Ontology (GO) enrichments for sets of CREs from these associations, thus providing a basis for functional interpretation. Here, we apply GOntact on experimentally validated enhancer datasets for human and mouse, using Promoter Capture Hi-C (PCHi-C) data for chromatin contact inference. We show that GOntact generates functional annotations that are coherent with the activity patterns of the enhancers. We find that there is substantial overlap between functional predictions obtained based on genomic proximity [6] and based on chromatin contacts, but that the latter approach pinpoints more specific GO categories. GOntact can thus prioritize testable hypotheses for CRE functional characterization.

Materials and methods

The GOntact tool and webserver

GOntact is a standalone command-line tool developed in OCaml. It can be installed using the OCaml package manager (opam) or by directly compiling the source code. Alternatively, it can be used through Docker and Apptainer containers. The source code, the Docker container, as well as detailed installation and usage instructions, are available in a public GitLab repository:

Manual pages are accessible with the “–help” argument for the command line tool.

A GOntact webserver is also available (https://gontact.univ-lyon1.fr). The webserver enables quick GO enrichment analyses for human and mouse using chromatin contact data, genome annotations, and GO data stored on the server. While the server is limited at the moment to these two species, the command-line tool can be used for any species for which all required data types are available. Schematic descriptions of the required input files and of the output generated by GOntact is given in Supplementary Figs S1 and S2.

Chromatin contact data processing

Chromatin contact data can be provided to GOntact in iBed and/or washU formats (defined in the CHiCAGO vignette [24] and further explained in the GOntact GitLab repository). This can be done for multiple samples at the same time by specifying several input files separated by a comma (“ibed-path” and “washu-path” arguments). GOntact enables filtering chromatin contacts based on the minimum and maximum distance between interacting regions (“min-dist-contacts” and “max-dist-contacts” arguments, default values 25 kb and 1 Mb, respectively), on the minimum interaction score (“min-score” argument, default value 5.0, which corresponds to the score cutoff proposed in CHiCAGO [24]), and on the minimum number of samples in which the contact is observed (“min-samples” argument, default value 1). We note that the quality of the chromatin contacts provided as an input can affect GOntact results. CHiCAGO was recently shown to perform well for the identification of chromatin contacts [25], alongside ChiCMaxima [26]. We recommend using one of these tools for chromatin contact data preparation, prior to running GOntact.

For the analyses presented in this manuscript, we ran GOntact using PCHi-C data from a previous study, which processed and combined 26 human and 14 mouse samples [27]. The PCHi-C technique aims to identify chromatin contacts between a predefined set of restriction fragments, that are “baited” using RNA molecules and the entire genome. Baited restriction fragments were selected to target gene promoters [15, 21]. Interactions were scored using the HiCup and CHiCAGO tools on the human hg38 and mouse mm10 genome assemblies [24, 28]. Note that we did not evaluate GOntact’s ability to correctly predict gene-CRE associations on PCHi-C data processed with other pipelines.

Chromatin contact annotation

Genome annotations have to be provided to GOntact as General Transfer Format files. They are used to annotate the gene promoters that are involved in chromatin contacts. For PCHi-C data, gene promoters are targeted by the “baited” restriction fragments, whose coordinates are provided in the first three columns of the iBed and WashU files. Following this principle, GOntact considers that the first genomic interval provided for each chromatin contact is the “baited” region. It then annotates chromatin contacts by identifying the transcription start sites (TSSs) that are found close to the coordinates of the first genomic interval, within a maximum distance defined by the user (default value 1 kb). For chromatin contact data that does not have “baited” regions, GOntact can force chromatin contacts to be symmetrical. Specifically, if a contact is listed between genomic regions A and B, with the “force-symmetry” option GOntact will add a contact between genomic regions B and A if this contact is not already present in the list. Conversely, GOntact can discard interactions that occur between two “baited” genomic regions (default; the remove-bait-bait option can be set to “false” to disable this behavior).

For GO enrichment analyses, only genes and transcripts annotated as “protein-coding” are retained, as GO annotations are still generally lacking for non-coding genes. For CRE target gene predictions, this behavior can be controlled with the “filter-biotypes” argument (possible values “protein_coding” or “all”, default value is “protein_coding”).

In this manuscript, we used genome annotations prouvided by the Ensembl database release for human hg38 (Ensembl release 115) and mouse mm10 (Ensembl release 102) genome assemblies.

Associating cis-regulatory elements and target genes

The first function of GOntact is to associate CREs with potential target genes. This function is accessible through the “gontact annotate” sub-command.

In the default run mode (“contacts”), GOntact associates CREs to genes based on chromatin contacts. CRE coordinates have to be provided to GOntact in BED format. GOntact identifies CREs that are close to genomic intervals involved in chromatin contacts within a certain maximum genomic distance defined by the user. Only the second genomic interval provided for each chromatin contact is used, following the principle described earlier for PCHi-C data. For chromatin contact data that do not have “baited” regions, GOntact can enforce chromatin contacts to be symmetrical. In addition to chromatin contacts, GOntact will associate CRE to genes if the CREs are found in a basal regulatory domain around the gene TSS, consisting of an upstream interval and a downstream interval whose size is defined by the user (“upstream” and “downstream” arguments, default values 25 kb for both).

GOntact also provides a “proximity” run mode, in which the association between CREs and genes is performed based on genomic proximity. In this mode, first a basal regulatory domain is defined, as described earlier and as proposed by McLean et al. 2010. The basal regulatory domain is then extended on both sides of the TSS for a maximum distance defined by the user (“extend” argument, default value 1 Mb) and at most up to the boundary of the basal regulatory domain of the neighboring genes. This corresponds to the “basal + extension” run mode of GREAT [6]. To associate CRE to genes with a simple genomic proximity approach, in which CREs are attributed to all genes whose TSS is found within a given maximum genomic distance, the “upstream” and “downstream” argument should be set to the desired maximum distance, and the “extend” argument should be set to 0.

To test the influence of the immediate neighborhood of the gene TSS on CRE functional annotation, we enabled the possibility of masking a genomic interval centered on the gene TSS in “proximity” mode for GO enrichment analyses. The size of the masked genomic region can be defined with the “masked-size” argument (default value is 0, corresponding to no masking).

For regulatory domain definition, only one TSS is used for each protein-coding gene. This TSS corresponds to the isoform with the longest exonic length (including coding and untranslated regions).

Gene Ontology enrichment analyses

The second function of GOntact is to perform GO enrichment analyses for CREs. This function is accessible through the “gontact enrich” sub-command.

To use this function, GOntact users must provide a file containing the structure of the GO database (“ontology” argument), in OBO format. They must also provide a file containing the GO annotation of the studied species (“functional-annot” argument) in GAF format.

In this manuscript, we downloaded the GO database and GO annotations for human and mouse from geneontology.org, for the 2025-10-10 database release. For human, we used annotations generated by the EBI GO Annotation database, and for mouse, we used annotations generated by Mouse Genome Informatics, also from the 2025-10-10 release of the GeneOntology database. We used the “basic” version of the ontology, which is acyclical.

GO annotations are typically provided only for the most precise term of a given ontology branch, but genes are assumed to inherit the annotations of all the parent terms. GOntact processes the structure of the ontology and the provided gene annotations to propagate annotations through the term hierarchy. The GO domain to be analyzed can be specified with the “domain” argument (possible values “biological_process,” “molecular_function,” “cellular_component”; default value “biological_process”). In the analyses presented in the manuscript, we focused on the “biological process” GO domain.

GO annotations are transferred to CREs from their predicted target genes. Gene-CRE associations can be obtained based on chromatin contacts or on genomic proximity, as described earlier. A given CRE can be associated only once with a given GO category, even if it is associated with several genes that have this GO annotation.

GO enrichment analyses are performed by comparing a foreground set of CREs (“foreground” argument, compulsory) with a background set of CREs (“background” argument, optional). If the background set is provided, the expected frequencies of GO categories are the frequencies observed in the background set.

If the background set is not provided, the expected frequencies of GO categories are computed on the entire set of chromatin contacts or on the entire set of regulatory domains, depending on the selected mode. GOntact will compute for each GO category the total number of nucleotides that are associated with this GO category. In “contacts” mode, we consider that individual nucleotides are associated with a GO category if they belong to genomic intervals that are in contact with genes associated with this category, or to basal regulatory domains associated with these genes (if a basal regulatory domain was defined with the “upstream” and “downstream” arguments). The background GO frequency is then computed as the ratio between the number of nucleotides associated with the focal GO category and the number of nucleotides that are associated with at least one GO category. Each nucleotide is counted only once. In “proximity” mode, the background GO frequency computation is similar, but based only on basal and extended regulatory domains.

For each GO category, GOntact will compare the frequency observed in the foreground CRE set with the expected frequency computed based on the background set, using a binomial test. Adjusted P-values that control the false discovery rate are computed with the Benjamini–Hochberg procedure. Output files contain several tab-separated columns, with rows corresponding to GO categories. For each GO category, output files provide its identifier and its description, the number of CREs associated with it in the foreground set, the observed GO frequency in the foreground set, the number of CREs (or nucleotides) associated to it in the background set, the expected GO frequency computed based on the background, the enrichment value (i.e. the ratio between the observed and the expected frequencies), the P-value and the adjusted P-value obtained with the Benjamini–Hochberg procedure.

Comparison with other methods for predicting gene-enhancer associations

To evaluate the ability of GOntact used with chromatin contact data to correctly predict gene-enhancer regulatory associations, we performed comparisons with other prediction methods, in addition to the genomic proximity (GREAT) approach. First, we downloaded gene-enhancer associations derived from a CRISPRi enhancer perturbation assay coupled with single-cell RNA sequencing in the K562 human cell line [16]. These datasets consisted of 5779 candidate enhancers, which were perturbed with the CRISPRi method [29]. We converted enhancer coordinates from the hg19 to the hg38 human genome assembly using the liftOver tool provided by the UCSC Genome Browser [30]. The screen identified 664 gene-enhancer regulatory relationships. We filtered these relationships to retain only those that included protein-coding genes targeted in our PCHi-C dataset with a maximum enhancer-TSS distance of 1 Mb. This reduced the dataset to 465 gene-enhancer relationships.

Second, we used gene-enhancer associations predicted with the activity-by-contact (ABC) method [31]. Data were available for five cell types or tissues (GM12878, K562, LNCAP, NCCIT, and liver). We converted enhancer coordinates from the hg19 to the hg38 human genome assembly using liftOver [30].

We used GOntact to predict gene-enhancer associations for the enhancers included in these datasets, in “contacts” mode with our PCHi-C dataset (only chromatin contacts were used in the prediction, with no basal regulatory domain), and in “proximity” mode (GREAT regulatory domains, with a basal regulatory domain of 5 kb upstream and 1 kb downstream of the TSS and an extension of maximum 1 Mb, as proposed in the original article [6]). Results are provided in Supplementary Dataset 2.

Running GOntact on lower quantities of PCHi-C data

To analyze the robustness of our results with respect to the PCHi-C data used as input, we performed two different subsampling analyses. First, we randomly split the human PCHi-C dataset into two subsets of samples derived from different cell types. We repeated the random subsampling procedure 10 times. This procedure enables us to test how the set of PCHi-C samples used by GOntact affects the prediction of gene-CRE relationships and the GO functional annotation. Second, we randomly drew fixed numbers of chromatin contacts (ranging from 100 000 to 900 000) from the complete human PCHi-C dataset, combined across all biological samples. We performed three random replicates for each number of chromatin contacts. This procedure enables us to test how the number of PCHi-C contacts used as an input by GOntact affects the prediction of gene-CRE relationships and the GO enrichment analysis.

For each of the two subsampling analyses described earlier, we evaluated the number of gene-CRE relationships that were inferred by GOntact for ENCODE and Vista forebrain enhancers. We performed GO enrichment analyses for the Vista forebrain enhancers, comparing them with the full set of ENCODE-predicted enhancers. We evaluated the concordance between the enriched GO categories identified with the subsampled PCHi-C dataset and the enriched GO categories identified with the entire PCHi-C dataset. For the first subsampling analysis, we evaluated the concordance between the GO enrichment results obtained with a subset of PCHi-C samples and the GO enrichment results obtained with the remaining set of PCHi-C samples. Results are provided in Supplementary Dataset 3.

Running GOntact with Hi-C and Micro-C data

Most of the analyses presented in the manuscript were performed by running GOntact on a collection of PCHi-C data, which offers high resolution and high detection power for promoter-CRE contacts. To test how GOntact performs with lower-resolution, genome-wide chromatin data, we used Hi-C data [32] and Micro-C data [33]. We downloaded processed data from the 4D nucleome database [34], specifically loop calls in “looplist” format and in “bedpe” format. Both file formats can be read by GOntact with the “loops-path” parameter. Hi-C data were available for eight human cell types (GM12878, HeLA, HMEC, HUVEC, IMR90, K562, KBM7, NHEK), with two replicates for GM12878. Micro-C data were available for H1 embryonic stem cells and HFF fibroblast cells.

We used this data to identify gene-enhancer associations for the human ENCODE and Vista enhancer datasets and to extract enriched GO categories. Results are provided in Supplementary Dataset 4.

Gene expression analyses

To further test the validity of the gene-enhancer relationships inferred with GOntact, we examined the expression patterns of the genes associated with Vista enhancers. To do this, we downloaded gene expression levels from [35] for human heart and brain across multiple developmental stages. We then compared the expression levels of the genes associated with heart and forebrain Vista enhancers by GOntact and by GREAT with the expression levels of protein-coding genes that were not associated with such enhancers by neither of the two methods. Gene expression levels were provided as RPKM values.

Statistical analyses and evaluation metrics

All statistical analyses (except the GO enrichment tests implemented in GOntact in OCaml) were performed in R. For comparisons between GOntact, GREAT, and additional approaches for gene-CRE association prediction, we computed the percentage of the reference set of gene-CRE associations that are recovered by GOntact or GREAT. This corresponds to the recall metric, defined as follows: recall = number of recovered relevant items / number of relevant items. Here, the relevant items are the reference gene-CRE associations.

For additional pairwise comparisons between gene-CRE sets or between sets of enriched GO categories, we used the Jaccard index, defined as follows: Jaccard index = size of the intersection between the two sets / size of the union of the two sets.

Results

Predicting CRE targets and functions based on chromatin contacts with GOntact

We developed GOntact, a tool that aims to provide an easy means of predicting CRE target genes based on chromatin contacts. Given a set of CREs, a genome annotation, and a list of chromatin contacts, GOntact identifies chromatin contacts that link gene promoters to CREs (“Materials and methods” section, Supplementary Figs S1 and S2). The GOntact tool has two sub-commands: the first one (“gontact annotate”) outputs a list of CRE-gene associations, including the evidence on which the association was inferred; the second one (“gontact enrich”) performs a GO enrichment analysis for a given set of CREs with respect to a background set of CREs or to the entire genome (“Materials and methods” section). GOntact enables filtering of the input chromatin contact dataset and allows users to fine-tune the parameters used for gene-CRE associations (“Materials and methods” section).

To enable comparisons between results obtained with chromatin contacts or based on genomic proximity, we re-implemented within GOntact the main principle of the widely used GREAT (Genomic Regions Enrichment of Annotations Tool) method [6, 36]. GOntact also enables using a simple genomic proximity approach, in which CREs are associated to genes if the distance to gene promoters is below a user-defined threshold (hereafter, this method is referred to as the “fixed window” approach). Users can switch between the approaches based on chromatin contacts or based on genomic proximity by setting the “mode” argument in the command line.

Inferring CRE target genes based on chromatin contacts and genomic proximity

We illustrate the different approaches for predicting CRE target genes with GOntact focusing on the genomic region around the human POU6F1 gene (Fig. 1A) for ENCODE-predicted enhancers. The simplest inference method based on genomic proximity attributes to a given gene all CREs found in its vicinity, that is, within a fixed window centered on the gene TSS. In the widely used GREAT method, a more complex principle is proposed (Fig. 1A). For each gene, a basal regulatory domain is constructed, which spans a certain distance upstream and downstream of the gene TSS [default values 5 kilobases (kb) and 1 kb, respectively]. These domains are then extended on each side, at most reaching the basal regulatory domains of the neighboring genes (on either DNA strand), and no further than a given genomic distance (1 Mb by default). As illustrated in Fig. 1A, this results in regulatory domains of variable sizes depending on the local gene density. The associations between genes and CREs strongly depend on the quality of the genome annotations: a gene missing from or wrongly included in the genomic annotation may affect the attribution of CREs to its neighbors. Of note, when using GREAT to infer enriched GO categories for sets of CREs, genes are filtered to discard those that lack GO annotations before regulatory domain construction.

Figure 1.

Graphics representing associations between genes and cis-regulatory elements obtained with various computational methods, with subfigures labeled A to C.

Gene-enhancer associations. (A) Illustration of gene-enhancer associations around the POU6F1 locus in the human genome. Top panel: position of predicted enhancers (ENCODE) and depiction of PCHi-C chromatin contacts between the POU6F1 gene promoter and predicted enhancers. Orange rectangles represent enhancers associated with the POU6F1 gene through PCHi-C contacts; red rectangles represent enhancers associated with the POU6F1 gene through the “regulatory domains” (GREAT) approach (McLean et al., 2010). Second panel: depiction of the basal regulatory domains proposed by McLean et al., 2010, comprising 5 kb upstream and 1 kb downstream of the TSS. Third panel: depiction of extended regulatory domains, which are limited by the basal regulatory domains of upstream and downstream neighboring genes, within a maximum distance of 1 Mb. Only protein-coding genes are shown. (B) Boxplots representing the distribution of the number of enhancers associated with each gene through PCHi-C contacts (orange), regulatory domains (red), and simple genomic proximity (fixed 100 kb window, black). Only genes associated with enhancers with all three approaches were considered. The Y-axis limit is set at 200 for better legibility, though higher values exist. (C) Distribution of the distances between genes and associated enhancers, as predicted with different approaches (PCHi-C contacts, regulatory domains, and fixed 2 Mb window). For this panel, the size of the fixed window was set to 2 Mb for consistency with GOntact and GREAT, which (by default) allow a maximum gene-CRE distance of 1 Mb.

In contrast, CRE target gene predictions based on chromatin contacts do not depend as much on the quality of genome annotations, but instead strongly depend on the quality of the chromosome conformation capture data used as input. In Fig. 1A, we illustrate the enhancer elements that are attributed to POU6F1 by GOntact based on PCHi-C data derived from multiple cell types and conditions [27]. They are situated >100 kb away from the POU6F1 TSS and bypass at least two other gene promoters. With the genomic proximity approach (GREAT regulatory domains), four enhancers are associated with POU6F1, at distances ranging from 700 bp to 34 kb (Fig. 1A). There is no overlap between the sets of CREs attributed to POU6F1 with the two approaches. This may be in part due to the fact that the PCHi-C approach cannot efficiently detect short-range contacts.

The discrepancy between the regulatory domain approach and GOntact for gene-enhancer association inference is particularly strong for genomic regions where genes are tightly clustered together. At the human HOXD locus, which contains multiple genes separated by short intergenic distances, regulatory domains are short, and thus few enhancers are attributed to HOXD genes (Supplementary Fig. S3). In contrast, GOntact infers multiple distant enhancers for the same genes, many of which are situated >500 kb away from gene promoters. This is in agreement with previous knowledge regarding HOXD gene regulation [37]. The ENCODE-predicted enhancers attributed to HOXD genes by the two approaches are listed in Supplementary Table S1.

Overall, individual genes are assigned substantially more enhancers based on PCHi-C contacts than based on genomic proximity (Fig. 1B). The median number of ENCODE-predicted enhancers assigned to a gene is 22 with GREAT and 50 with GOntact for human (Wilcoxon rank sum, P-value < 1e-10, Fig. 1B). The median number of enhancers attributed to genes is similar between GREAT and a simple fixed 100 kb window approach (21 enhancers), but the two distributions are significantly different (Wilcoxon rank sum, P-value < 1e-10, Fig. 1B). Conversely, the distributions of the numbers of genes associated with each enhancer are also significantly different between the three approaches (Wilcoxon rank sum, P-value < 1e-10 for all three comparisons, Supplementary Fig. S4A). With GOntact, 55% of enhancers are associated with at most two genes, while with GREAT this is true for 99% of enhancers by construction (extended regulatory domains cannot overlap for more than two genes). Importantly, the sets of enhancers attributed to each gene differ considerably among methods. The median Jaccard index (which measures the ratio between the size of the intersection and the size of the union of two sets) is equal to 0.01 for the comparison between GREAT and GOntact (Supplementary Fig. S4B). There is more similarity between GREAT and the fixed 100 kb window approach (median Jaccard index 0.35, Supplementary Fig. S4B), as expected given that both methods are based on genomic proximity. The numbers of gene-enhancer associations inferred by GOntact can be modulated by changing the contact filtering parameters, in particular the minimum contact score (which corresponds in this case to the contact confidence score attributed to PCHi-C contacts by CHiCAGO, [24]) and the minimum number of PCHi-C samples in which contacts were detected (Supplementary Fig. S4C and D).

The genomic distances between promoters and enhancers predicted to be in regulatory interactions vary considerably among methods. With GREAT, 53% of all gene-enhancer pairs are separated by <100 kb, and only 8% are separated by >500 kb (Fig. 1C). With GOntact, distances between pairs of enhancers and promoters are generally larger: only 18% are separated by <100 kb, and 20% are separated by >500 kb (Fig. 1C). In contrast, regulatory associations inferred from PCHi-C data are less frequent in close proximity to genes, which may be due to the technical difficulty of distinguishing genuine from spurious chromatin contacts at small distances on the linear genome [24].

Comparison with other approaches for gene-enhancer association inference

To further test the validity of the results obtained with GOntact, we compared its predictions with those of two other methods for gene-enhancer association inference. First, we analyzed experimentally predicted gene-enhancer regulatory relationships obtained with the CRISPRi enhancer perturbation approach [16]. Second, we analyzed the gene-enhancer associations inferred with the ABC method [31]. This method relies on enhancer activity patterns and on the physical proximity between gene promoters and enhancers, as inferred with genome-wide Hi-C data (or extrapolated from the genomic distance between the two elements) to infer gene-enhancer associations [31]. We filtered the results of both methods to retain only genes whose promoters were captured in the PCHi-C experiments and only gene-enhancer pairs found within a maximum distance of 1 Mb (“Materials and methods” section).

The majority of CRISPRi-detected gene-enhancer relationships take place within short genomic distances (52% at distances below 50 kb, Supplementary Fig. S5). These relationships are recovered by GOntact with PCHi-C data in only 14% of the cases, but well recovered by GREAT, in 93% of the cases (Supplementary Fig. S5). However, long-range associations are better recovered by GOntact than by GREAT: out of 91 gene-enhancer pairs separated by >250 kb, GOntact recovered 18 (19.7%), while GREAT recovered 5 (5.5%). Due to the skew toward short-range regulatory interactions in this dataset, overall GREAT recovers a higher proportion (69%) of CRISPRi gene-enhancer pairs than GOntact (29%).

In contrast, GOntact recovers a higher fraction of ABC-inferred regulatory relationships than GREAT (21% versus 10.7% in the GM12878 cell line). This is largely due to its better ability to infer long-range interactions (Supplementary Fig. S6). We note that ABC infers considerably more gene-enhancer associations than GOntact and GREAT: in the same cell line, for the same enhancers, there are 1 073 004 gene-enhancer associations inferred with ABC, 318 508 with GOntact and 282 987 with GREAT. Because neither GREAT nor GOntact require evidence of co-activity between genes and enhancers, this discrepancy arises solely from the distance distribution imposed by the different methods. The frequency of gene-enhancer associations decreases more rapidly as a function of the genomic distance for GOntact (Fig. 1) than for ABC (Supplementary Fig. S6). Overall, GOntact appears to be more stringent than ABC for the prediction of gene-enhancer associations, with this PCHi-C dataset.

Gene Ontology enrichment analyses for cis-regulatory elements

GO enrichment analyses contrast a focal gene set (e.g. genes that are up-regulated in a given condition) with a background gene set (e.g. all analyzed genes). They are traditionally used as a basis for functional predictions. Such analyses can be used to formulate new hypotheses that can be validated through further experiments. Similar analyses can be performed for sets of CREs by transferring GO annotations to CREs from the genes they are predicted to regulate. This principle has been implemented within GREAT [6]. We applied the same principle in GOntact by transferring GO categories to CREs from the genes with which they form chromatin contacts (“Materials and methods” section).

We used three different methods to infer gene-CRE associations and thereby GO-CRE associations: GOntact applied on PCHi-C chromatin contacts, GREAT regulatory domains, and a simple genomic proximity method (CREs are attributed to genes if found within a fixed 100 kb window around the TSS). We performed GO enrichment analyses for three sets of experimentally validated enhancers, which are active in the embryonic forebrain, limb, or heart. Enhancer coordinates and activity patterns in embryonic tissues were downloaded from the VISTA Enhancer browser [38]. We performed GO enrichment analyses contrasting each of these enhancer sets with the full set of ENCODE-predicted enhancers (“Materials and methods” section). This comprehensive enhancer dataset was obtained using a large number of cell types, tissues, and biological conditions. Alternatively, we performed GO enrichment analyses using the whole genome as a background, with similar results (Supplementary Dataset 1; “Materials and methods” section).

Overall, we found good agreement among the three methods for the top GO categories (Fig. 2 and Supplementary Figs S7 and S8). The top GO categories were related to neuron differentiation and brain development for forebrain enhancers (Fig. 2), to digit and limb morphogenesis for limb enhancers (Supplementary Fig. S7), and to cardiac muscle development for heart enhancers (Supplementary Fig. S8). All three methods thus propose functional categories that are consistent with enhancer activity patterns. However, we note that the top GO categories put forward by each method are not identical. For example, for forebrain enhancers, the second GO category found significant by GOntact is “forebrain radial glial cell differentiation,” which does not appear in the top 15 GO categories with the two other approaches (Fig. 2). Conversely, the two approaches that rely on genomic proximity show significant enrichment of GO categories related to transcriptional regulation, which do not appear at the top of the list for GOntact (Fig. 2). Likewise, for embryonic heart enhancers, GOntact tends to bring forward categories related to the early development of the organ (e.g. lateral mesoderm formation, embryonic heart tube elongation), whereas genomic proximity methods seems to prioritize categories related to the physiology of the embryonic and adult organ, such as the regulation of striated muscle contraction or the regulation of blood circulation (Supplementary Fig. S8).

Figure 2.

Graphics showing gene ontology enrichment results obtained with different methods, with subfigures labeled A to E.

GO enrichments for Vista forebrain enhancers. (A) Barplot representing log10-transformed adjusted P-values for the top 15 significantly enriched GO categories, for a GO enrichment analysis performed based strictly on PCHi-C chromatin contacts (we did not include a basal regulatory domain). Dot sizes and colors represent enrichment values. (B) Same aspanel (A), for the “regulatory domains” approach (McLean et al., 2010). (C) Same as panel (A), for the simple genomic proximity approach (fixed 100 kb window). (D) Venn diagram representing the overlap between the GO categories found to be significantly enriched (FDR < 0.01) with the three approaches. (E) Boxplot representing the number of genes associated with enriched GO categories, for the three methods. For all panels, enrichment analyses are performed using as a background the set of human ENCODE-predicted enhancers. Analyses performed using the whole genome as a background are provided in Supplementary Dataset 1.

The numbers of significantly enriched GO categories (FDR < 0.01) differ considerably among the three methods, with far larger numbers observed with GREAT for all three enhancer sets (Fig. 2 and Supplementary Figs S7 and S8). Interestingly, the enriched GO categories are generally associated with larger numbers of genes for the two methods based on genomic proximity than for GOntact (Fig. 2 and Supplementary Figs S7 and S8). This suggests that GOntact can prioritize more specific biological processes. We note that, for both methods, the expression patterns of the genes associated with Vista enhancers are coherent with enhancer activity patterns: genes predicted to be regulated by heart enhancers tend to be highly expressed in the heart, and likewise for the brain (Supplementary Fig. S9).

Consistent with the observations on the top 15 GO categories, we observed that overall, the ranks of the enriched GO categories (ordered based on the adjusted P-value associated with the GO enrichment test) are only moderately correlated among the three approaches (Supplementary Fig. S10). However, the enrichment values (i.e. the ratio between the frequencies of the GO categories in the foreground and background CRE sets) are well correlated among methods, with Spearman’s rho coefficients ranging from 0.9 to 0.97 (Supplementary Fig. S11).

Influence of GOntact parameters on GO enrichment results

To test how GOntact parameters affect GO enrichment results, we extracted two relevant GO categories for each tissue (i.e. those that were at the top of the enrichment lists with default parameters: “neuron differentiation” and “regulation of transcription by RNA polymerase II” for forebrain enhancers, “limb morphogenesis” and “embryonic digit morphogenesis” for limb enhancers, and “heart development” and “muscle tissue morphogenesis” for heart enhancers). We then analyzed their enrichment values and significance levels for GOntact runs with different parameters (Fig. 3 and Supplementary Fig. S12).

Figure 3.

Graphics representing gene ontology enrichment results obtained with different parameters for GOntact, with subfigures labeled A to C.

GO enrichment results obtained with different approaches and parameter sets. (A) Dot plot representing the log10-transformed adjusted P-value for the “neuron differentiation” GO category for GO enrichment analyses contrasting human forebrain enhancers (Vista) with ENCODE-predicted enhancers. Different columns represent method variants, in order: PCHi-C contacts, requiring that chromatin contacts be observed in at least 1, 2, and 3 samples, respectively; PCHi-C contacts + basal 10 kb domain, with contacts observed in at least 1, 2, and 3 samples; PCHi-C contacts + basal 50 kb domain, with contacts observed in at least 1, 2, and 3 samples; regulatory domains (basal + extended domains, maximum 1 Mb); modified regulatory domains, in which a 10 kb window centered on the TSS is “masked” (meaning that enhancers found within <5 kb of the gene TSS are excluded); modified regulatory domains, in which a 20, 50, and 100 kb window centered on the TSS is “masked”; fixed 50 kb window centered on the TSS; fixed 100 kb window centered on the TSS. (B) Same aspanel (A) for the “limb morphogenesis” GO category for human limb enhancers. (C) Same aspanel (A) for the “heart development” GO category for human heart enhancers. Dot sizes and colors represent enrichment values, as in Fig. 2.

We first wanted to evaluate how the quantity and the quality of the input chromosome conformation capture data may affect the results. The GOntact results presented earlier were performed using chromatin contacts derived from 26 human PCHi-C samples [27]. These included numerous sample-specific contacts. We reasoned that contacts shared across multiple biological samples may be more relevant to define target genes for sets of CREs that are derived from unrelated tissues, as these shared contacts may include pre-formed interactions observed in the absence of CRE and gene activity [39]. Overall, similar GO enrichment results were observed when requiring contacts to be observed in 2 or 3 samples (Supplementary Dataset 1). However, we observed a decrease in detection sensitivity: the significance level associated with the most relevant GO categories decreased in this setting (Fig. 3 and Supplementary Fig. S12). The decrease in sensitivity is likely due to the fact that fewer chromatin contacts are used for target gene prediction. Interestingly, this pattern was not observed for all GO categories: for the “heart development” GO category, a more significant enrichment was observed with shared PCHi-C contacts (Fig. 3).

Because significant interactions at short genomic distances are depleted in chromatin contact data, the default parameters of GOntact exclude chromatin contacts occurring at distances below 25 kb. As there is a genuine enrichment for CREs in close proximity to gene promoters [16], we wanted to test whether adding a proximal regulatory domain to PCHi-C contacts could improve our results (“Materials and methods” section). Overall, the significance levels of the GO enrichment tests for the relevant GO categories were stronger upon adding a 10 kb or a 50 kb basal domain centered around the TSS, for the three analyzed tissues (Fig. 3 and Supplementary Fig. S12).

We next wanted to test to what extent the GO enrichment results obtained with GREAT can be explained by the abundance of CREs in the immediate vicinity of gene TSS. To do this, we masked a genomic region centered on the TSS in the GREAT analysis. That is, rather than incorporating a basal regulatory domain, we exclude gene-CRE associations at distances smaller than a certain threshold (ranging from 5 to 50 kb, Fig. 3, Supplementary Fig. S12). We observed that the significance of the GO enrichment test was lower following the masking procedure for all analyzed GO categories and for all three enhancer sets. Notably, the drop was larger for some GO categories, such as “heart development” for embryonic heart enhancers. Conversely, the fixed window approach generally provided stronger significance values for larger window sizes, although we note that the increase in statistical significance was only moderate for some GO categories (e.g. “limb morphogenesis” for embryonic limb enhancers, Fig. 3).

Influence of input chromatin contact data on GOntact performance

We next wanted to evaluate how the quantity and quality of chromatin contact data used as an input for GOntact affects its performance. To do this, we performed several analyses using human data. First, we randomly divided the 26 PCHi-C samples used here into two complementary subsets, ensuring that the two sample subsets do not share cell types (“Materials and methods” section). We performed the subsampling procedure 10 times. We inferred gene-enhancer associations for ENCODE and Vista enhancers and performed GO enrichment analyses contrasting Vista forebrain enhancers with ENCODE-predicted enhancers. Despite the distinct samples and cell types included in the two complementary subsets, we observed that gene-enhancer associations and enriched GO categories were often shared between the two (Jaccard indexes between 0.4 and 0.6, Supplementary Fig. S13). The enriched GO categories were related to neuron differentiation and forebrain glial cell differentiation (Supplementary Fig. S14), as also observed for the complete PCHi-C dataset (Fig. 2).

Second, we subsampled fixed numbers of PCHi-C contacts from the human PCHi-C data, combined across all samples. We drew numbers of contacts ranging from 100 000 to 900 000, in three replicates. We then compared the results obtained by GOntact with the subsampled datasets with those obtained by GOntact with the complete PCHi-C dataset, which contains 910 069 unique chromatin contacts. As expected, we observed that the numbers of gene-enhancer associations predicted by GOntact increase with the number of PCHi-C contacts used as an input (Supplementary Fig. S15). The increase is not linear, but we do not reach a plateau within this range of chromatin contact numbers. For draws including at least 300 000 PCHi-C contacts, we observe good agreement with the complete PCHi-C dataset in terms of significant GO categories (Jaccard indexes above 0.6) and of rank correlations between significant GO categories (Spearman’s rho coefficient above 0.6, Supplementary Fig. S15). Likewise, GO enrichment analyses display the expected significant GO categories related to neuron and glial cell differentiation if at least 300 000 contacts are used (Supplementary Fig. S16).

Third, we ran GOntact with lower-resolution chromatin contact data instead of PCHi-C data. We used chromatin loops detected from human Hi-C data [32] and from Micro-C data [33]. While these datasets offer high resolution and detection power compared to classical Hi-C experiments [32, 33], they remain underpowered to detect individual chromatin loops or contacts compared to PCHi-C data. We were able to analyze 38 428 unique chromatin loops derived from 9 Hi-C samples and 62 830 unique chromatin loops derived from 3 Micro-C samples, at 5 kb resolution. As expected given the low numbers of input chromatin contacts, only a small fraction of the gene-enhancer associations detected with our PCHi-C data were recovered with Hi-C data (6% for both ENCODE and Vista enhancers, Supplementary Fig. S17) or with Micro-C data (7% for ENCODE, 8% for Vista enhancers, Supplementary Fig. S18). Nevertheless, GO enrichment analyses performed with Hi-C data recovered meaningful functional predictions for Vista forebrain and heart enhancers with Hi-C data (GO categories related to cell differentiation in the brain, Supplementary Fig. S17) and for Vista limb enhancers with Micro-C data (GO categories related to limb development, Supplementary Fig. S18).

GO enrichment for human-specific conserved element deletions (hCONDELs)

Finally, we applied GOntact to a set of ~10 000 conserved elements that experienced human-specific deletions [40]. These elements were named hCONDELs [40]. Sequences that are conserved across large evolutionary distances but accelerate specifically in the human lineage have attracted considerable attention, as they may include elements that are responsible for the emergence of human-specific phenotypes [41, 42]. Many of these elements are involved in gene expression regulation [43]. An analysis performed with GREAT showed that genes near hCONDELs are enriched in transcriptional regulatory functions, but also in neural development [40]. We performed GO enrichment analyses on this set of elements with the three approaches described earlier, using the whole genome as a background. As observed for VISTA enhancer analyses, overall the three methods were in good agreement, confirming the enrichment in GO categories associated with transcriptional regulation and with developmental processes (Fig. 4). However, here again, the three approaches highlight different GO categories at the top of the enrichment results. Notably, among the top 15 GO categories obtained with GOntact, two (ranked first and ninth) are associated with cranial nerve morphogenesis. The top GO category is related to the morphogenesis of the glossopharyngeal nerve, which has sensory and motor functions associated with the pharynx and the larynx. This specific GO category may be relevant for the emergence of vocalization-related processes. In contrast, the GO categories that are at the top of the significance list with approaches based on genomic proximity are more general and relate to the regulation of transcription or to system development. Thus, in this situation, GOntact can provide a different list of candidates for further functional exploration than proximity-based approaches.

Figure 4.

Graphics representing gene ontology enrichment results obtained for human-specific conserved element deletions with different computational methods, with subfigures labeled A to E.

GO enrichments for human-specific conserved element deletions (hCONDELs). (A) Barplot representing log10-transformed adjusted P-values for the top 15 significantly enriched GO categories, for a GO enrichment analysis performed based strictly on chromatin contacts (we did not include a basal regulatory domain). Dot sizes and colors represent enrichment values. (B) Same as panel (A), for the “regulatory domains” approach (McLean et al., 2010). (C) Same as panel (A) for the simple genomic proximity approach (fixed 100 kb window). (D) Venn diagram representing the overlap between the GO categories found to be significantly enriched (FDR 0.01) by the three approaches. (E) Boxplot representing the number of genes associated with enriched GO categories, for the three methods. For all panels, enrichment analyses are performed using the whole genome as a background. (A–C) Dot sizes and colors represent enrichment values.

Discussion

Here, we propose a new tool for CRE target gene predictions and functional enrichment analysis, which uses chromatin contacts as a basis. Evaluating the performance of GOntact in a purely computational setting is a difficult task, as no ground truth is available. Nevertheless, we rely on an extensive body of literature, which demonstrates that promoter-CRE chromatin contacts are enriched in genuine regulatory interactions [15, 21, 44, 45]. We performed all analyses with GOntact and GREAT, a widely used software that implements CRE target gene predictions based on genomic proximity [6, 36]. We compared the performances of GOntact and GREAT with those of additional approaches for the prediction of gene-enhancer regulatory relationships: an experimental method based on enhancer perturbation with CRISPRi [16] and a computational method that incorporates enhancer activity and physical proximity, as inferred with Hi-C or extrapolated based on genomic distance [31]. Our results show that long-range interactions are well recovered by GOntact with PCHi-C data, while GREAT is able to better predict short-range interactions.

In addition to gene-CRE association prediction, both GOntact and GREAT aim to predict putative biological functions for a set of CREs of gene expression. We applied these methods on sets of experimentally validated enhancer elements whose pattern of activity was ascertained in transgenic mouse embryos [38]. We thus expect an enrichment for biological processes associated with the development and the physiology of the organs in which enhancer activity was detected. Reassuringly, all the methods that we tested here show results that are coherent with these expectations. GOntact and GREAT provide consistent GO enrichment results, although these two methods rely on very different approaches for CRE target gene inference. We believe that this is due to the fact that both methods capture functionally relevant gene-CRE associations: GREAT is good at capturing short-distance regulatory relationships, which represent the majority of gene-CRE associations [16], while GOntact captures long-distance regulatory interactions. We thus recommend that both types of methods be applied and compared, as they each may prove to be relevant in different contexts. Moreover, given that GOntact used with PCHi-C data performs poorly for short-range gene-enhancer associations, we recommend that a basal regulatory domain be included. Our Vista enhancer analyses show that including a basal regulatory domain comprising 25 kb upstream and 25 kb downstream of the gene TSS improves GOntact’s ability to recover meaningful GO categories.

Beyond the overall agreement between GOntact and proximity-based approaches, some differences are notable. In particular, we find that proximity-based approaches output larger numbers of significantly enriched GO categories, but that these categories are more general (that is, they are associated with larger numbers of genes) than those put forward by GOntact. This is particularly evident in the hCONDELs analysis, where the GO categories at the top of the enriched list pinpoint specific functions related to cranial nerve morphogenesis, while proximity-based approaches highlight more general functions related to transcriptional regulation. The fact that GOntact outputs narrower lists of enriched GO categories may be an advantage in exploratory analyses, as they can help in the decision process for further experimental validation.

A further advantage of the regulatory relationships inferred by GOntact is that they are to a large extent independent of the quality of gene annotations, in contrast with the methods implemented in GREAT. In well-annotated genomes such as human and mouse, GREAT undoubtedly performs very well. However, genome annotation quality can affect the target gene inference as performed by GREAT. A missing gene or a wrongly assigned TSS can have an inordinate impact on the definition of the regulatory domains of neighboring genes and thus on CRE target gene prediction. We note that GOntact, used with PCHi-C data, is also to some extent sensitive to genome annotation quality, as knowledge of gene promoters is required in the design of the capture experiment. Beyond genome annotation quality issues, some genomic organization features, such as gene clustering, also have a strong impact on GREAT predictions. This is well illustrated by the example of human HOXD genes (Supplementary Fig. S3), which are tightly clustered together and thus are assigned very short regulatory domains. In these cases, GREAT can only predict proximal gene-CRE relationships, while GOntact predicts large numbers of long-range regulatory interactions. We note that gene-CRE relationships predicted by GOntact are also likely influenced by functional genome organization features, in particular by the organization of the genome in topologically associating domains (TADs), which are megabase-scale genomic regions where chromatin contacts preferentially occur [46]. The great majority of gene-CRE chromatin contacts detected with high-resolution PCHi-C data were found to occur within TADs in human cells [21]. These domains are thought to play an important role in the evolution of gene regulation [47]. The presence of functionally related genes (e.g. genes derived from duplication events [48]) within the same TAD could thus contribute to the functional enrichment signal extracted by GOntact for sets of CREs. Thus, part of the signal obtained by GOntact may reflect broader patterns in the functional organization of the genome, in addition to individual promoter-CRE relationships.

GOntact is dependent on the quantity, quality, and specificity of the input data, in particular the set of PCHi-C chromatin contacts. Here, we used a previously published set of chromatin interactions [27], which were obtained with a computational pipeline based on HiCUP [28] and CHiCAGO [24]. A recent comparative study showed that these data analysis tools perform well for the detection of individual chromatin contacts [25]. Alternatives exist for all steps of the computational processing pipeline, e.g. using tools such as HiC-Pro [49] or Juicer [50] for raw data processing, or ChiCMaxima [26] for interaction calling. The quality of the resulting chromatin contacts dataset is likely to depend on the tools and on their sensitivity and specificity settings at all processing steps. We note that, in this manuscript, we did not evaluate the effects of using alternative computational pipelines for PCHi-C data processing and chromatin interaction calling on GOntact predictions.

The PCHi-C dataset used in this manuscript was derived from multiple cell types and conditions [27]. A non-negligible fraction of these interactions were only observed in a single cell type [27]. Ideally, chromatin interactions should be provided for the same biological conditions (tissues, cell types, developmental stages, etc.) in which CRE activity is assessed. In the absence of perfectly matched chromatin contact and enhancer activity data, one solution is to use as an input a set of chromatin interactions that are shared across multiple biological conditions, which are thus more likely to be relevant in the biological context that is being examined. However, we note that this filter greatly reduces the number of usable interactions and thus results in a decrease in sensitivity. With our data, 300 000 PCHi-C contacts are sufficient to obtain meaningful GO enrichment results for human Vista enhancers, although this may vary in other settings depending on the quality of the input data and of the characteristics of the analyzed CRE sets.

Finally, we note that generating chromatin contact data is becoming increasingly accessible. Hi-C experiments are now typically performed for genome assembly projects and for analyses of broad-scale chromosome conformation features, including TADs [46]. While these data lack the resolution and sensitivity of PCHi-C and of other capture-based approaches, with sufficient sequencing depth they provide chromatin contacts at the kilobase scale [19, 32, 33]. Several computational methods were proposed to detect chromatin loops from Hi-C data [51–53], which can be used as an input for GOntact. Our analyses revealed that running GOntact with chromatin loops detected with Hi-C and Micro-C data can provide meaningful functional predictions for CREs, although these approaches remain underpowered compared to PCHi-C for the detection of individual promoter-CRE contacts. Numerous Hi-C and Micro-C experiments can now be accessed in publicly available resources such as the 4D nucleome database [34]. Going forwards, GOntact can facilitate leveraging this wealth of chromatin contact data to provide better informed predictions of gene-CRE relationships. We thus believe that GOntact can provide a practical alternative to approaches based purely on genomic proximity for CRE target enrichment and functional enrichment analyses.

Supplementary Material

lqag136_Supplemental_Files

Acknowledgements

We thank C. Berthelot, Y. Ghavi-Helm, F. Picard, C. François, and D. Mouchiroud for discussions and advice on the project. We thank Stéphane Delmotte for help with webserver deployment. Computational analyses were performed using the computing facilities of the CC LBBE/PRABI and the Core Cluster of the Institut Français de Bioinformatique (IFB) (ANR-11-INBS-0013).

Author contributions: Alexandre Laverré (Conceptualization [equal], Formal analysis [equal], Methodology [equal], Software [equal], Validation [equal], Visualization [equal], Writing—review & editing [equal]), Eric Tannier (Conceptualization [equal], Project administration [equal], Supervision [equal], Writing—review & editing [equal]), Philippe Veber (Conceptualization [equal], Investigation [equal], Methodology [lead], Software [lead], Writing—review & editing [equal]), and Anamaria Necsulea (Conceptualization [equal], Data curation [equal], Formal analysis [equal], Funding acquisition [lead], Methodology [equal], Project administration [lead], Software [equal], Supervision [equal], Validation [equal], Visualization [lead], Writing—original draft [lead], Writing—review & editing [equal]).

Contributor Information

Alexandre Laverré, Universite Claude Bernard Lyon 1, LBBE, UMR 5558, CNRS, VAS, Villeurbanne 69622, France; Department of Ecology and Evolution, University of Lausanne, 1015 Lausanne, Switzerland.

Eric Tannier, Universite Claude Bernard Lyon 1, LBBE, UMR 5558, CNRS, VAS, Villeurbanne 69622, France; Centre de recherche Inria de Lyon, 69603 Villeurbanne, France.

Philippe Veber, Universite Claude Bernard Lyon 1, LBBE, UMR 5558, CNRS, VAS, Villeurbanne 69622, France.

Anamaria Necsulea, Universite Claude Bernard Lyon 1, LBBE, UMR 5558, CNRS, VAS, Villeurbanne 69622, France.

Supplementary data

Supplementary data is available at NAR Genomics and Bioinformatics online.

Conflict of interest

None declared.

Funding

This work was funded by the French National Research Agency (ANR-17-CE12-0019-01 “LncEvoSys”).

Data availability

The source OCaml code and the R scripts used to generate graphics and statistical analyses are freely available in a GitLab repository: https://gitlab.in2p3.fr/anamaria.necsulea/GOntact.

Detailed installation and usage instructions are provided in GitLab. For reproducibility, we provide GOntact binaries in a Docker container. The data used in the manuscript and the results of the GOntact runs analyzed in the manuscript are available as Supplementary Datasets in Zenodo with the following DOI: https://doi.org/10.5281/zenodo.21685753.

References

  • 1. Wei  C-L, Wu  Q, Vega  VB  et al.  A global map of p53 transcription-factor binding sites in the human genome. Cell. 2006;124:207–19. 10.1016/j.cell.2005.10.043 [DOI] [PubMed] [Google Scholar]
  • 2. Visel  A, Blow  MJ, Li  Z  et al.  ChIP-seq accurately predicts tissue-specific activity of enhancers. Nature. 2009;457:854–8. 10.1038/nature07730 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 3. Thurman  RE, Rynes  E, Humbert  R  et al.  The accessible chromatin landscape of the human genome. Nature. 2012;489:75–82. 10.1038/nature11232 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Buenrostro  JD, Giresi  PG, Zaba  LC  et al.  Transposition of native chromatin for fast and sensitive epigenomic profiling of open chromatin, DNA-binding proteins and nucleosome position. Nat Methods. 2013;10:1213–8. 10.1038/nmeth.2688 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. ENCODE Project Consortium, Moore  JE, Purcaro  MJ  et al.  Expanded encyclopaedias of DNA elements in the human and mouse genomes. Nature. 2020;583:699–710. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. McLean  CY, Bristor  D, Hiller  M  et al.  GREAT improves functional interpretation of cis-regulatory regions. Nat Biotechnol. 2010;28:495–501. 10.1038/nbt.1630 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Banerji  J, Olson  L, Schaffner  W. A lymphocyte-specific cellular enhancer is located downstream of the joining region in immunoglobulin heavy chain genes. Cell. 1983;33:729–40. 10.1016/0092-8674(83)90015-6 [DOI] [PubMed] [Google Scholar]
  • 8. Lettice  LA, Horikoshi  T, Heaney  SJH  et al.  Disruption of a long-range cis-acting regulator for Shh causes preaxial polydactyly. Proc Natl Acad Sci USA. 2002;99:7548–53. 10.1073/pnas.112212199 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Carter  D, Chakalova  L, Osborne  CS  et al.  Long-range chromatin regulatory interactions in vivo. Nat Genet. 2002;32:623–6. 10.1038/ng1051 [DOI] [PubMed] [Google Scholar]
  • 10. Amano  T, Sagai  T, Tanabe  H  et al.  Chromosomal dynamics at the Shh locus: limb bud-specific differential regulation of competence and active transcription. Dev Cell. 2009;16:47–57. 10.1016/j.devcel.2008.11.011 [DOI] [PubMed] [Google Scholar]
  • 11. Dekker  J, Rippe  K, Dekker  M  et al.  Capturing chromosome conformation. Science. 2002;295:1306–11. 10.1126/science.1067799 [DOI] [PubMed] [Google Scholar]
  • 12. Lieberman-Aiden  E, van Berkum  NL, Williams  L  et al.  Comprehensive mapping of long-range interactions reveals folding principles of the human genome. Science. 2009;326:289–93. 10.1126/science.1181369 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 13. Montavon  T, Soshnikova  N, Mascrez  B  et al.  A regulatory archipelago controls Hox genes transcription in digits. Cell. 2011;147:1132–45. 10.1016/j.cell.2011.10.023 [DOI] [PubMed] [Google Scholar]
  • 14. Smemo  S, Tena  JJ, Kim  K-H  et al.  Obesity-associated variants within FTO form long-range functional connections with IRX3. Nature. 2014;507:371–5. 10.1038/nature13138 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Schoenfelder  S, Furlan-Magaril  M, Mifsud  B  et al.  The pluripotent regulatory circuitry connecting promoters to their long-range interacting elements. Genome Res. 2015;25:582–97. 10.1101/gr.185272.114 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Gasperini  M, Hill  AJ, McFaline-Figueroa  JL  et al.  A genome-wide framework for mapping gene regulation via cellular genetic screens. Cell. 2019;176:1516. 10.1016/j.cell.2019.02.027 [DOI] [PubMed] [Google Scholar]
  • 17. Parker  SM, Davis  ES, Phanstiel  DH. Guiding the design of well-powered Hi-C experiments to detect differential loops. Bioinforma Adv. 2023;3:vbad152. 10.1093/bioadv/vbad152 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Hsieh  T-HS, Weiner  A, Lajoie  B  et al.  Mapping nucleosome resolution chromosome folding in yeast by Micro-C. Cell. 2015;162:108–19. 10.1016/j.cell.2015.05.048 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19. Lee  BH, Wu  Z, Rhie  SK. Characterizing chromatin interactions of regulatory elements and nucleosome positions, using Hi-C, Micro-C, and promoter capture Micro-C. Epigenetics Chromatin. 2022;15:41. 10.1186/s13072-022-00473-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20. Mumbach  MR, Rubin  AJ, Flynn  RA  et al.  HiChIP: efficient and sensitive analysis of protein-directed genome architecture. Nat Methods. 2016;13:919–22. 10.1038/nmeth.3999 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Mifsud  B, Tavares-Cadete  F, Young  AN  et al.  Mapping long-range promoter contacts in human cells with high-resolution capture Hi-C. Nat Genet. 2015;47:598–606. 10.1038/ng.3286 [DOI] [PubMed] [Google Scholar]
  • 22. Golov  AK, Abashkin  DA, Kondratyev  NV  et al.  A modified protocol of Capture-C allows affordable and flexible high-resolution promoter interactome analysis. Sci Rep. 2020;10:15491. 10.1038/s41598-020-72496-4 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Downes  DJ, Beagrie  RA, Gosden  ME  et al.  High-resolution targeted 3C interrogation of cis-regulatory element organization at genome-wide scale. Nat Commun. 2021;12:531. 10.1038/s41467-020-20809-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 24. Cairns  J, Freire-Pritchett  P, Wingett  SW  et al.  CHiCAGO: robust detection of DNA looping interactions in Capture Hi-C data. Genome Biol. 2016;17:127. 10.1186/s13059-016-0992-2 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25. Aljogol  D, Thompson  IR, Osborne  CS  et al.  Comparison of Capture Hi-C analytical pipelines. Front Genet. 2022;13:786501. 10.3389/fgene.2022.786501 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Ben Zouari  Y, Molitor  AM, Sikorska  N  et al.  ChiCMaxima: a robust and simple pipeline for detection and visualization of chromatin looping in Capture Hi-C. Genome Biol. 2019;20:102. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27. Laverré  A, Tannier  E, Necsulea  A. Long-range promoter-enhancer contacts are conserved during evolution and contribute to gene expression robustness. Genome Res. 2022;32:280–96. 10.1101/gr.275901.121 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Wingett  S, Ewels  P, Furlan-Magaril  M  et al.  HiCUP: pipeline for mapping and processing Hi-C data. F1000Res. 2015;4:1310. 10.12688/f1000research.7334.1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Canver  MC, Smith  EC, Sher  F  et al.  BCL11A enhancer dissection by Cas9-mediated in situ saturating mutagenesis. Nature. 2015;527:192–7. 10.1038/nature15521 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Casper  J, Speir  ML, Raney  BJ  et al.  The UCSC Genome Browser database: 2026 update. Nucleic Acids Res. 2026;54:D1331–5. 10.1093/nar/gkaf1250 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Fulco  CP, Nasser  J, Jones  TR  et al.  Activity-by-contact model of enhancer–promoter regulation from thousands of CRISPR perturbations. Nat Genet. 2019;51:1664–9. 10.1038/s41588-019-0538-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Rao  SSP, Huntley  MH, Durand  NC  et al.  A 3D map of the human genome at kilobase resolution reveals principles of chromatin looping. Cell. 2014;159:1665–80. 10.1016/j.cell.2014.11.021 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Krietenstein  N, Abraham  S, Venev  SV  et al.  Ultrastructural details of mammalian chromosome architecture. Mol Cell. 2020;78:554–65.e7. 10.1016/j.molcel.2020.03.003 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34. Dekker  J, Oksuz  BA, Zhang  Y  et al.  An integrated view of the structure and function of the human 4D nucleome. Nature. 2026;649:759–76. 10.1038/s41586-025-09890-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35. Cardoso-Moreira  M, Halbert  J, Valloton  D  et al.  Gene expression across mammalian organ development. Nature. 2019;571:505–9. 10.1038/s41586-019-1338-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36. Tanigawa  Y, Dyer  ES, Bejerano  G. WhichTF is functionally important in your open chromatin data?. PLoS Comput Biol. 2022;18:e1010378. 10.1371/journal.pcbi.1010378 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Montavon  T, Duboule  D. Landscapes and archipelagos: spatial organization of gene regulation in vertebrates. Trends Cell Biol. 2012;22:347–54. 10.1016/j.tcb.2012.04.003 [DOI] [PubMed] [Google Scholar]
  • 38. Kosicki  M, Baltoumas  FA, Kelman  G  et al.  VISTA Enhancer browser: an updated database of tissue-specific developmental enhancers. Nucleic Acids Res. 2025;53:D324–30. 10.1093/nar/gkae940 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Paliou  C, Guckelberger  P, Schöpflin  R  et al.  Preformed chromatin topology assists transcriptional robustness of Shh during limb development. Proc Natl Acad Sci USA. 2019;116:12390–9. 10.1073/pnas.1900672116 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Xue  JR, Mackay-Smith  A, Mouri  K  et al.  The functional and evolutionary impacts of human-specific deletions in conserved elements. Science. 2023;380:eabn2253. 10.1126/science.abn2253 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41. Pollard  KS, Salama  SR, Lambert  N  et al.  An RNA gene expressed during cortical development evolved rapidly in humans. Nature. 2006;443:167–72. 10.1038/nature05113 [DOI] [PubMed] [Google Scholar]
  • 42. McLean  CY, Reno  PL, Pollen  A  et al.  Human-specific loss of regulatory DNA and the evolution of human-specific traits. Nature. 2011;471:216–9. 10.1038/nature09774 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Capra  JA, Erwin  GD, McKinsey  G  et al.  Many human accelerated regions are developmental enhancers. Phil Trans R Soc B. 2013;368:20130025. 10.1098/rstb.2013.0025 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. Javierre  BM, Burren  OS, Wilder  SP  et al.  Lineage-specific genome architecture links enhancers and non-coding disease variants to target gene promoters. Cell. 2016;167:1369–84. 10.1016/j.cell.2016.09.037 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45. Jung  I, Schmitt  A, Diao  Y  et al.  A compendium of promoter-centered long-range chromatin interactions in the human genome. Nat Genet. 2019;51:1442–9. 10.1038/s41588-019-0494-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Nora  EP, Lajoie  BR, Schulz  EG  et al.  Spatial partitioning of the regulatory landscape of the X-inactivation centre. Nature. 2012;485:381–5. 10.1038/nature11049 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47. Acemel  RD, Maeso  I, Gómez-Skarmeta  JL. Topologically associated domains: a successful scaffold for the evolution of gene regulation in animals. Wiley Interdiscip Rev Dev Biol. 2017;6:1–19. 10.1002/wdev.265 [DOI] [PubMed] [Google Scholar]
  • 48. Ibn-Salem  J, Muro  EM, Andrade-Navarro  MA. Co-regulation of paralog genes in the three-dimensional chromatin architecture. Nucleic Acids Res. 2017;45:81–91. 10.1093/nar/gkw813 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 49. Servant  N, Varoquaux  N, Lajoie  BR  et al.  HiC-Pro: an optimized and flexible pipeline for Hi-C data processing. Genome Biol. 2015;16:259. 10.1186/s13059-015-0831-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50. Durand  NC, Shamim  MS, Machol  I  et al.  Juicer provides a one-click system for analyzing loop-resolution Hi-C experiments. Cell Syst. 2016;3:95–8. 10.1016/j.cels.2016.07.002 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51. Kaul  A, Bhattacharyya  S, Ay  F. Identifying statistically significant chromatin contacts from Hi-C data with FitHiC2. Nat Protoc. 2020;15:991–1012. 10.1038/s41596-019-0273-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 52. Lu  J, Wang  X, Sun  K  et al.  Chrom-Lasso: a lasso regression-based model to detect functional interactions using Hi-C data. Brief Bioinform. 2021;22:bbab181. 10.1093/bib/bbab181 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 53. Lagler  TM, Abnousi  A, Hu  M  et al.  HiC-ACT: improved detection of chromatin interactions from Hi-C data via aggregated Cauchy test. Am Hum Genet. 2021;108:257–68. 10.1016/j.ajhg.2021.01.009 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

lqag136_Supplemental_Files

Data Availability Statement

The source OCaml code and the R scripts used to generate graphics and statistical analyses are freely available in a GitLab repository: https://gitlab.in2p3.fr/anamaria.necsulea/GOntact.

Detailed installation and usage instructions are provided in GitLab. For reproducibility, we provide GOntact binaries in a Docker container. The data used in the manuscript and the results of the GOntact runs analyzed in the manuscript are available as Supplementary Datasets in Zenodo with the following DOI: https://doi.org/10.5281/zenodo.21685753.


Articles from NAR Genomics and Bioinformatics are provided here courtesy of Oxford University Press

RESOURCES