Keywords

1 Introduction

The life cycle of multicellular organisms requires the coordinated control of transcriptional processes across multiple tissues and developmental stages. Regulation of gene expression generally involves two different types of cis-acting elements: the promoter , a genomic region defining the initiation of transcription, and more distal regulatory elements called enhancers. While promoters provide the essential sites of transcriptional initiation of RNAs, they are frequently not sufficient to direct appropriate developmental and signal-dependent levels of gene expression [1, 2]. This additional information is provided by enhancers, short regions of DNA that, when bound by transcription factors (TFs), enhance RNA expression from target promoters . Enhancers can reside hundreds of thousands of base pairs (bp) away from their target gene, are typically well-conserved across genomes and their function is generally considered to depend on three-dimensional enhancer–promoter interactions [3].

Active enhancers are characterized by bidirectional transcription, which results in the production of enhancer RNAs (eRNAs) believed to facilitate long-range enhancer–promoter looping [4]. Cap analysis of gene expression (CAGE) technology captures the 5′-end of transcripts and, therefore, allows the identification of transcription start sites (TSSs) of active regulatory genomic elements (Fig. 1a, [5] and Chapter 4). Indeed, CAGE has been employed as an orthogonal approach to define active enhancers in multiple datasets and cell types [1].

Fig. 1
figure 1

CAGE and RADICL-seq technologies. (a) Schematic overview of CAGE technology that identifies the 5′ end of transcripts (see ref. 5 for further details). (b) Schematic overview of RADICL-Seq technology aimed at capturing proximal RNA–chromatin interactions in a genome-wide manner (see ref. 8 for further details)

Full size image

In recent years, the development of high-throughput molecular methods has allowed the study of the three-dimensional organization of the genome of eukaryotic cells [6]. Chromatin conformation technologies determine the proximity between loci by measuring their contact frequency and comparing it with other genomic locations in an interaction matrix. These approaches have led to the identification of the regulatory activity of enhancers and their impact on the expression of target genes. However, several active enhancers do not exhibit significantly higher contact frequencies with target promoters than with surrounding chromatin [7]. Also, the functional impact of the physical proximity of two genomic regions on the regulation of gene expression is not completely understood.

Recently, we have developed RNA And DNA Interacting Complexes Ligated and sequenced (RADICL-Seq) , a novel technology to identify genome-wide RNA–chromatin interactions (Fig. 1b and [8]). RADICL-Seq employs mild fixation and a biotinylated bridge adapter to capture RNA and DNA molecules located in close proximity while preserving the nuclear structure [8]. Compared with existing methods, RADICL-Seq improves genomic coverage and unique mapping rate efficiency, thereby increasing the detection power for several transcripts including long non-coding RNAs and intronic RNAs. By employing this technology, we have mapped the genomic occupancy of multiple RNA classes in mouse embryonic stem cells (mESCs) and oligodendrocyte progenitor cells (mOPCs), identifying general and cell type-specific interaction patterns [8].

As the enhancer–promoter looping is believed to drive transcriptional activation, the spatial proximity of the nascent RNA with the enhancer region has the potential to be captured by RADICL-Seq . Detection of such RNA–chromatin interactions has the advantage to include an additional layer of functionality for the observed physical proximity that is not possible with chromatin conformation technologies.

Here we combine publicly available CAGE and RADICL-Seq data to identify the gene targets for thousands of enhancer elements. We provide a detailed computational workflow to first call the enhancers using CAGE data and subsequently assign their target genes by leveraging RADICL-Seq data.

2 Materials

The original manuscript describing the RADICL-Seq technology can be found at https://www.nature.com/articles/s41467-020-14337-6. The RADICL-Seq significant pair data and the CAGE expression data can be downloaded from GEO (https://www.ncbi.nlm.nih.gov/geo/), under the series GSE132190. All analysis was performed on an Intel quad core CPU machine with 32 GB of memory.

3 Methods

  1. 1.

    All the computations are done in R [9]. The following packages will be used: GenomicFeatures [10], InteractionSet [11], CAGEfightR [12], and tidyverse [13]. For genome annotation, mm10 and Gencode vM14 will be used, in order to be consistent with the original manuscript for the datasets used.

  2. 2.

    For the RADICL-Seq significant pair tables, we subset them on the following columns for further analysis:

    Column

    Data

    1

    Chromosome of origin for the RNA

    2

    Midpoint location for the RNA read

    6

    Sense of transcription (− refers to negative strand. + is positive strand).

    7

    Ensembl ID of the RNA

    8

    RNA class

    10

    RNA feature

    11

    Chromosome for interacting DNA read

    12

    Midpoint location for the DNA read

    14

    DNA identifier (chromosome_bin; genome has been divided in 25-kb bins)

    16

    Dataset

    17

    p-Value before correction

    18

    p-Value after correction

  3. 3.

    We convert the subsetted data tables into GInteractions objects, after extending each DNA and RNA midpoint positions by 1 kb on either side. In effect, we will be comparing 2 kb regions for the rest of the analysis when using RADICL-Seq significant pairs. After the conversion, the first anchor of the resulting GInteractions object would be the RNA hit regions, and the second anchor would be the DNA regions.

    For each data table, the conversion can be performed by:

    radicl_set <- map (radicl, function(tab) {   rna <- GRanges(seqnames=tab$chrom.R,                  ranges=IRanges(start=tab$pos.R - flank,                   end=tab$pos.R + flank),                   strand=tab$strand.R,                   seqinfo=genomeInfo) %>% trim()   dna <- GRanges(seqnames=tab$chrom.D,                   ranges=IRanges(start=tab$pos.D - flank,                  end=tab$pos.D+flank),                   seqinfo=genomeInfo) %>% trim()   interaction <- GInteractions(rna, dna)   mcols(interaction) <- tab   interaction })

    where “radicl” is the list object containing the significant pair tables, “flank” is the flanking region to be added (1 kb), and “genomeInfo” is the Seqinfo object [14] containing the mm10 chromosome information.

  4. 4.

    The CAGE data can also be downloaded from GEO, under the series GSE132191. The data is provided as BED files containing CAGE transcription start sites (CTSS). CAGEfightR package is used for de novo promoter and enhancer identification and quantification. For details on how to process CAGE data using CAGEfightR, refer to [12].

  5. 5.

    The CTSS BED files are first converted to BigWig files using rtracklayer package’s import and export functions.

    bed <- rtracklayer::import(ctss_file) bed <- GenomicRanges::GRanges(bed, seqinfo=genomeInfo) bed_plus <- bed[ bed@strand=="+", ] bed_minus <- bed[ bed@strand=="-", ] rtracklayer::export(object=bed_plus, filename, format="BigWig" )

  6. 6.

    The converted BigWig files are used to load and quantify CTSS reads into RangeSummarizedExperiment objects and calculate tags per million (TPM ) values.

    CTSSs <- quantifyCTSSs(plusStrand=bw_plus,                        minusStrand=bw_minus,                        design=sample_info,                        genome=genomeInfo) %>%    trim() %>%   subsetBySupport(inputAssay='counts', outputColumn='support',                    unexpressed=0, minSamples=1) %>%   calcTPM(inputAssay='counts', outputAssay='TPM ') %>%    calcPooled(inputAssay='TPM ')

    where “bw_plus” and “bw_minus” are plus and minus strand BigWig files, respectively.

  7. 7.

    The quantified CTSSs are now used to detect unidirectional and bidirectional CTSS clusters, which will be taken as promoters and enhancers, respectively. Unidirectional clusters within 20 bp of one another are merged together. The clusters are then assigned transcript IDs accordingly. Any bidirectional clusters that overlap existing promoter annotations within 500 bp up- and downstream, as well as those that are annotated as promoter /5’ UTR/exons are filtered out. Once the de novo promoters and enhancers are identified, we quantify them by summing the number of CTSS reads that overlap those regions.

    uni_TCs <- clusterUnidirectionally(supported_CTSSs,                                     pooledCutoff=3, mergeDist=20) %>%   trim() %>%   assignTxID(txModels=txdb, outputColumn='txID') %>%   assignTxType(txModels=txdb, outputColumn='txType') %>%   assignTxType(txModels=txdb, outputColumn='peakTxType',                 swap='thick') bi_TCs <- clusterBidirectionally(CTSSs, balanceThreshold=0.9) bi_TCs <- calcBidirectionality(bi_TCs, samples=CTSSs) bi_TCs <- assignTxType(bi_TCs, txModels=txdb,                         tssUpstream=500, tssDownstream=500,                         outputColumn='txType') bi_TCs <- subset(bi_TCs, !txType %in% c('promoter ','fiveUTR','exon')) # quantify the de novo promoters TSSs <- quantifyClusters(CTSSs, clusters=uni_TCs,                           inputAssay='counts') %>%   calcTPM() %>%   calcPooled() # quantify the de novo enhancers enhancers <- quantifyClusters(CTSSs, clusters=bi_TCs) %>%   calcTPM() %>%   calcPooled()

    where txdb is the TxDb object containing the Gencode vM14 annotations.

  8. 8.

    The identified de novo enhancers are overlapped with the DNA regions in the RADICL-Seq significant pair sets using the findOverlaps function in the GenomicRanges package. For each enhancer overlapping a given RADICL-Seq DNA region, the corresponding RNA hit regions that pair with these regions are taken as its interacting regions. We restrict ourselves to those regions that have CAGE expression support. We also merge the overlapping DNA regions to avoid double counting of DNA-enhancer overlaps.

    rad_rna <- trim(anchors(radicl_set[[n]])$first) rad_dna <- trim(anchors(radicl_set[[n]])$second) # only include those where both RNA and DNA regions have CTSS support matching_regs <- mcols(radicl_set[[n]]) %>% as_tibble() # now filter for enhancer overlap overlap <- findOverlaps(rad_dna, enhancers) ind <- unique(queryHits(overlap)) rad_dna <- rad_dna[ind] rad_rna <- rad_rna[ind] merged_rad <- GenomicRanges::reduce(rad_dna) overlap <- findOverlaps(rad_dna, merged_rad) matching_regs$merged_ind <- subjectHits(overlap)

  9. 9.

    We can calculate how many enhancers a given RADICL-Seq DNA region (2 kb) overlaps on average. The results are shown in Fig. 2.

    enh_per_dna <- Reduce(bind_rows,  map (matching_regs, function(regs) {   tab <- mcols(regs)   n <- as_tibble(tab) %>%         select(enhancer, merged_ind) %>%         group_by(merged_ind)         %>% summarise(length(unique(enhancer)))   res <- table(n[,2])   vals <- c(unique(tab$cell_type), as.numeric(res))   names(vals) <- c('cell', names(res))   vals }))

    where “matching_regs” is the list of enhancer-overlapping RADICL-Seq significant pairs for each sample as calculated from the previous step.

  10. 10.

    To tally the RADICL-Seq RNA hit regions that have CAGE expression support, we set a filter of at least 1 out of 3 replicates for each cell type having at least 1 TPM .

    exp_enh <- assays(enhancers)$TPM exp_enh <- rownames(exp_enh)[rowSums(exp_enh > 1) > 1] exp_rna <- assays(TSSs)$TPM exp_rna <- rownames(TSSs)[rowSums(exp_rna > 1) > 1]

  11. 11.

    For each overlapping enhancer, we calculate the average number of interacting genes by counting the number of unique gene IDs associated with the paired RNA hit regions. The results are illustrated in Fig. 3.

    gene_counts <- mcols(regs) %>% as_tibble() %>%     select(c(enhancer, merged_ind, gene_id.R)) %>%      group_by(enhancer) %>%     summarise_at(vars(merged_ind, gene_id.R),                   function(x) {length(unique(x))})

    where “regs” is the list of enhancer-overlapping RADICL-Seq significant pairs for one sample, i.e., a given element of the “matching_regs” list.

  12. 12.

    With the list of RADICL-Seq DNA regions that overlap enhancers and their interacting RNA hit regions, we calculate the distances between the pairs and produce a distance distribution. We repeat the process for the DNA regions that do not overlap any enhancers, and test whether the two distance distributions are significantly different using Wilcoxon test.

Fig. 2
figure 2

Distribution of number of enhancers overlapping one RADICL-Seq DNA region (2 kb window centered on the midpoint for the DNA region of a given RNA–DNA significant pair). While the number of overlapping enhancers for a given DNA region can vary, the majority overlap less than five enhancers (mESC mouse embryonic stem cells, mOPC mouse oligodendrocyte progenitor cells)

Full size image
Fig. 3
figure 3

Distribution of number of expressed genes linked to a single enhancer. A given gene is considered to be expressed if a CAGE cluster associated with the gene that has expression value of at least 1 tag per million (TPM) in at least one replicate sample. We establish the link between a given enhancer and a given gene by determining whether they overlap any of the RADICL-Seq DNA–RNA significant pairs (mESC mouse embryonic stem cells, mOPC mouse oligodendrocyte progenitor cells)

Full size image

enhancer_dist_distrib <- bind_rows(   map (matching_regs, function(regs) {     tab1 <- mcols(regs) %>% as_tibble() %>%       select(cell_type, merged_ind, RNA_pos, merged_pos) %>%        distinct() %>%       mutate(gene_radicl_dist=abs(RNA_pos - merged_pos))   })) non_enhancer_dist_distrib <- bind_rows(   map (nonenh, function(mcoltab) {     mcoltab %>%       dplyr::select(cell_type, merged_ind, RNA_pos, merged_pos) %>%        distinct() %>%       mutate(gene_radicl_dist=abs(RNA_pos - merged_pos))       })) a <- dplyr::filter(enhancer_dist_distrib$gene_radicl_dist,                     cell_type == 'mESC') b <- dplyr::filter(non_enhancer_dist_distrib, cell_type == 'mESC') wilcox.test(a$gene_radicl_dist, b$gene_radicl_dist)