Approaches for mapping and analysis of R-loops

Abstract

R-loops are nucleic acid structures composed of a DNA:RNA hybrid with a displaced non-template single-stranded DNA. Current approaches to identify and map R-loop formation across the genome employ either an antibody targeted against R-loops (S9.6) or a catalytically inactivated form of RNase H1 (dRNH1), a nuclease that can bind and resolve DNA:RNA hybrids via RNA exonuclease activity. This overview outlines several ways to map R-loops using either methodology, explaining the differences and similarities amongst the approaches. Bioinformatic analysis of R-loops involves several layers of quality control and processing before visualizing the data. This article provides resources and tools that can be used to accurately process R-loop mapping data and explains the advantages and disadvantages of the resources as compared to one another.

INTRODUCTION:

R-loops are three-stranded nucleic acid structures consisting of a DNA:RNA hybrid and a non-template single-stranded DNA (ssDNA) (Thomas, White, & Davis, 1976). Early research promoted a perception of R-loops as inherently anomalous and recombinogenic (“pathological R-loops”). Recent studies have challenged this perception by demonstrating that R-loops are prevalent in vivo and participate in multiple regulatory roles (“physiological R-loops”), predominantly seen in immunoglobulin class switching recombination of B cells in vertebrates (Daniels & Lieber, 1995; Reaban & Griffin, 1990; Yu, Chedin, Hsieh, Wilson, & Lieber, 2003), mitochondrial DNA replication (Akman et al., 2016; Silva, Camino, & Aguilera, 2018) and in certain regulatory steps in transcription initiation and termination (Huertas & Aguilera, 2003). Today, mounting evidence supports a model in which R-loops are typically physiological under basal conditions, and pathological R-loops are an uncommon byproduct of R-loop dysregulation.

With lengths varying from 100 bp up to 2 kb and an estimated half-life of ~15 mins (Coggins, Slater, O’Prey, & Campo, 1988; Crossley, Bocek, Hamperl, Swigut, & Cimprich, 2020; Crossley et al., 2022; Sanz et al., 2016), R-loops have been observed in both head-on or co-directional collisions between the progressing replication fork and transcription machinery. Termed “transcription-replication conflicts” (TRCs), these collisions can result in replication fork stalling and genome instability. R-loops promote TRCs, but the underlying mechanism remains poorly understood. Curiously, the directionality of TRCs influences the degree of genomic instability and R-loop resolution. Co-directional (CD) R-loops are easier to resolve than those occurring due to head-on (HO) collisions, which cause fork stalling and therefore, replication stress (Hamperl, Bocek, Saldivar, Swigut, & Cimprich, 2017). Although R-loops can be resolved by several mechanisms including nucleases such as RNaseH, and helicases like senataxin and others (Drolet et al., 1995; Hausen & Stein, 1970; Hong, Cadwell, & Kogoma, 1995), defects in these pathways can result in persistent R-loops, which can become pathogenic by compromising genomic stability.

In the last decade, multiple studies have explored and characterized the role of physiological R-loops in regulating the epigenomic state of cells (Niehrs & Luke, 2020), particularly through regulation of DNA methylation (Ginno, Lott, Christensen, Korf, & Chédin, 2012), histone modification (Alecki et al., 2020; Skourti-Stathaki, Kamieniarz-Gdula, & Proudfoot, 2014; Skourti-Stathaki et al., 2019), and 3D chromatin contacts (H. Pan et al., 2020). R-loops can form across a subset of gene promoter regions and pose a barrier to their methylation either actively such as via inhibiting the activity of replication-associated DNA methyltransferase DNMT1) (Grunseich et al., 2018) or passively by promoting demethylation of CpG islands (Arab et al., 2019). In addition to their role in preventing DNA methylation, physiological R-loops also exert an epigenetic influence by promoting histone modifications.

Once used in molecular biology research as a tool for isolating nucleic acids (Born, Wittig, Wittig, & Tiedemann, 1980; Kaback, Angerer, & Davidson, 1979; Wittig & Wittig, 1979) and in experiments on transcriptional regulation and splicing (Baas, Keegstra, Teertstra, & Jansz, 1978; Coggins et al., 1988; Itoh & Tomizawa, 1980; Kadesch & Chamberlin, 1982), recent studies have transposed the utility of R-loops into the context of DNA repair. “Pathological R-loops”, firstly, by virtue of the structure’s inherent ability to tether the transcriptional machinery leading to TRCs with the replisome, and secondly, due to the presence of an exposed ssDNA loop which is vulnerable to mutagenesis, are associated with DNA damage due to double stranded breaks (DSBs) and replication stress, potentially resulting in senescence or apoptosis.

R-loops in DNA Repair and genome instability.

While accumulation of pathological R-loops can lead to genome instability (Jackson, Noerenberg, & Whitehouse, 2014; Lambo et al., 2019; Stork et al., 2016), studies showing that these structures also contribute to genome maintenance demonstrate the context-dependent consequences of R-loop formation. R-loops participate actively in DNA repair, particularly in repair of double-strand breaks (DSBs) through several mechanisms. Unlike transcription-associated R-loops formed in cis, “RNA-bridging” involves the formation in trans of an DNA:RNA hybrid via RAD52, which bridges a DSB to facilitate repair. Alternatively, “RNA-templated repair” involves RAD52-mediated RNA hybridization on break overhangs (McDevitt, Rusanov, Kent, Chandramouly, & Pomerantz, 2018). Consequently, this makes R-loops capable of promoting homologous recombination (HR) and DSB repair via recruitment of RAD52 to sites of damage (Yasuhara et al., 2018). A recent observation showed that UPF1, an RNA/DNA helicase gene required for stimulating nonsense-mediated decay (NMD), promotes R-loop formation to facilitate repair independent of NMD at sub-telomeric DSBs (Ngo, Grimstead, & Baird, 2021). Additionally, DROSHA, an miRNA processing factor, shown to be required early in the DNA repair response, has been associated with forming and retaining DNA:RNA hybrids around DSBs, facilitating DNA repair (Lu et al., 2018). These observations indicate a role for R-loops (or DNA:RNA hybrids), in participating actively in DNA repair processes. However, even within the context of DNA repair, certain studies describe an inhibitory effect of R-loops on DNA repair response (Aguilera & Gómez-González, 2017; Cohen et al., 2018). Efforts to map and analyze the formation of these structures with better spatial and temporal resolution may further elucidate their role in these biological processes. However, these potentially spurious, if very targeted events, differ from the more consistent programmatic physiological R-loops that can be detected by standard R-loop mapping methods.

Overview of R-loop mapping techniques

Approaches to identify and map R-loops are predominantly based on sequestering the hybrid structures with moieties that can specifically bind to them. Current techniques to map R-loops are thus broadly classified to either those that use a monoclonal antibody S9.6, or methods that incorporate a catalytically “dead” version of the RNase H1 nuclease (referred to as “dRNH1” hereon), which renders the enzyme inactive but allows it to bind R-loops. Continuous advancements in R-loop mapping methods have led to a broad array of techniques as listed in Table 1. Nevertheless, R-loop mapping remains an empirical approach for assessing the locations of R-loops genome-wide via high-throughput sequencing.

Table 1.

Read quality control tools for R-loop mapping data.

Tool	Description	Source	Used in
FastQC	Assesses sequence data quality for high-throughput pipelines preemptively.	https://www.bioinformatics.babraham.ac.uk/projects/fastqc/	(Bayona-Feliu et al., 2017; L. Chen et al., 2017; Cohen et al., 2018; Ellis et al., 2021; Rassoulzadegan, Sharifi-Zarchi, & Kianmehr, 2021)
fastq-mcf	Detects adapters, determines clipping parameters, performs clipping, skewing, filtering.	https://github.com/ExpressionAnalysis/ea-utils	(Ginno et al., 2012; Sanz et al., 2016; Villarreal, Mersaoui, Yu, Masson, & Richard, 2020)
skewer	Efficient adapter trimming for Illumina sequences with dynamic programming algorithm.	https://sourceforge.net/projects/skewer	(Abakir et al., 2019; Crossley et al., 2020)
fastx-toolkit	Toolkit for processing FASTA and FASTQ files.	https://github.com/agordon/fastx_toolkit	(Park et al., 2021)
cutadapt	Removes adapters, primers, poly-A tails from sequencing reads.	https://cutadapt.readthedocs.io/en/stable/index.html	(Crossley et al., 2020; H. Li et al., 2022; Park et al., 2021)
Trimmomatic	Performs versatile trimming tasks for Illumina paired-end/single-end data.	http://www.usadellab.org/cms/index.php?page=trimmomatic	(Alecki et al., 2020; Liu et al., 2021; Rassoulzadegan et al., 2021)
TrimGalore	Automated trimming, adapter removal, quality control for sequencing data.	https://www.bioinformatics.babraham.ac.uk/projects/trim_galore/	(Briggs et al., 2018; Fang et al., 2019)

S9.6-based methods

The S9.6 antibody was first isolated and characterized in 1985 in an attempt to raise antibodies against DNA:RNA hybrids (Boguslawski et al., 1986). Whereas DNA duplexes were difficult to raise antibodies against, RNA and DNA:RNA hybrids were immunogenic. Due to its high affinity to hybrids, S9.6 soon became a standard reagent to immunoprecipitate DNA:RNA hybrids (Hu, Zhang, Storz, Gottesman, & Leppla, 2006). DRIP-seq (“DNA:RNA immunoprecipitation sequencing”),the first and most commonly used R-loop mapping technique was described in 2012 and used the S9.6 antibody to immunoprecipitate restriction enzyme-digested genome fragments containing DNA:RNA hybrids (Ginno et al., 2012). This work was critical in demonstraing that R-loops are prevalent genome-wide, particularly in promoter regions along with evidence that R-loops suppress promoter methylation.

A drawback identified during the development of DRIP -seq was the lack of strand-specificity for the DNA template from which an R-loop originates. Several techniques were thus developed in the subsequent years to allow strand-specific, RNA-based R-loop mapping. “DNA:RNA immunoprecipitation sequencing” or RDIP-seq permitted researchers to demonstrate that a sizeable proportion of R-loops occur in intergenic regions (Nadel et al., 2015). Unlike DRIP-seq, RDIP-seq uses sonication to digest genomic DNA, which has the potential to result in unbiased fragmentation, though it could also disrupt some native R-loops. Shortly thereafter, a similar protocol called DRIP-RNA-seq was described, which was utilized to reveal the co-localization of R-loops with the Tip60-p400 complex (P. B. Chen, Chen, Acharya, Rando, & Fazzio, 2015). Later, “DNA:RNA immunoprecipitation followed by cDNA conversion” sequencing or DRIPc-seq, was employed to reveal the colocalization of R-loops with regions of open chromatin, permissive histone marks, and regions of active transcription (Sanz et al., 2016). Development of these high-throughput sequencing protocols introduced new, higher-resolution R-loop mapping techniques and further reinforced the notion that R-loops are prevalent and play physiological roles.

In the following years, multiple new DRIP-based techniques were introduced to improve the accuracy of R-loop mapping. In 2016, a DRIP-seq protocol was introduced which includes a preliminary S1 nuclease digestion to remove single-stranded DNAs (“S1-DRIP-seq”) (Wahba, Costantino, Tan, Zimmer, & Koshland, 2016).” Bisulfite-converted DRIP-seq” or bisDRIP-seq, was developed as a modification of DRIP-seq incorporating bisulfite conversion of ssDNA to enhance the resolution of R-loop maps to a near-nucleotide resolution (Dumelie & Jaffrey, 2017). A method for strand-specific, DNA-based DRIP-seq, termed “ssDRIP” was presented in the same year (Xu et al., 2017). The authors observed that only the DNA:RNA hybrid strand from the triple-stranded R-loop was recovered with ssDRIP. More recently, a strand-specific, sonicated, and spike-in-normalized enhancement of DRIP-seq, termed “quantitative DRIP-seq” (qDRIP-seq) was described (Crossley et al., 2020). sDRIP-seq is a sonication-based, strand-specific version of the original DRIP-seq protocol (Smolka, Sanz, Hartono, & Chédin, 2021). Most recently, subsequent addition of spike-in nucleic acids to allow quantitative analysis of R-loop data was described (Crossley & Cimprich, 2022). Each new modality addressed limitations of preceding iterations and provided novel insight into the biology of physiological R-loops.

RNase H1-based methods

The use of a catalytically inactive RNase H1 nuclease (dRNH1) to isolate R-loops has become a gold standard The RNase H1 enzyme recognizes and cleaves the RNA strand of an DNA:RNA hybrid. This specificity ensures that only RNA-containing structures, such as R-loops, are targeted, leaving the DNA intact, which is crucial for accurately mapping R-loops without introducing additional DNA breaks or artifacts. After the development of ex vivo R-loop mapping techniques, there remained uncertainty regarding how faithfully these mapping approaches can represent “native” R-loop dynamics which occur in vivo. This uncertainty was addressed through in situ R-loop mapping approaches which can detect R-loops in native chromatin. These techniques utilize overexpression of D210N mutant RNase H1 in cells to isolate R-loops with subsequent immunoprecipitation. DRIVE-seq (“DNA:RNA In Vitro Enrichment sequencing”), a technique that utilizes dRNH1 for immunoprecipitation (Ginno et al., 2012) was developed in parallel with DRIP-seq. Expressing dRNH1 followed by strand-specific amplification of immunoprecipitated (IPed) DNA (termed R-ChIP), was demonstrated as an efficient method to capture R-loops associated with gene promoters and dynamically correlated with transcriptional pausing (L. Chen et al., 2017). Due to the selective loss of the ssDNA in these isolated R-loops, the profile obtained appears relatively strand specific. “RR-ChIP,” a modified version of the R-ChIP protocol that involves sequencing the RNA moiety in a strand-specific manner to create high-resolution R-loop maps was introduced subsequently (Tan-Wong, Dhir, & Proudfoot, 2019). This method also incorporates a mutant “WKKD” RNase H1 (W43A, K59A, K60A and D210N), which neither binds to nor cleaves R-loops, as a negative control. MapR, a dRNH1-based in situ mapping modality which involves treating permeabilized cells with a micrococcal nuclease-dRNH1 fusion protein to cleave and release R-loops for sequencing, eliminates the additional step of creating the stably transfected cell lines required for the R-ChIP and RR-ChIP protocols (Q. Yan & Sarma, 2020; Q. Yan, Shields, Bonasio, & Sarma, 2019). A new MapR-based modality was subsequently described, which incorporates bisulfite conversion to achieve higher resolution, termed “BisMapR” (Wulfridge & Sarma, 2021). Using BisMapR, a subset of bidirectionally-transcribed enhancers was identified, which paradoxically show unidirectional R-loop formation.

Ex vivo vs in situ R-loop mapping

At present, nearly all S9.6-based R-loop mapping modalities are ex vivo, i.e., the hybrid structures are bound by the antibody post nucleic acid extraction and deproteinization, while nearly all dRNH1-based R-loop mapping modalities are in situ, where the R-loops are bound by dRNH1 within cells, prior to nucleic acid extraction (Figure 1). In comparing DRIP-seq and R-ChIP, several discrepancies were observed: R-ChIP-mapped R-loops tend to be nearly an order of magnitude smaller than DRIP-seq-mapped R-loops (L. Chen et al., 2017). They also occur more frequently in promoter and intergenic regions, unlike DRIP-seq-mapped R-loops which occur predominantly in the gene body. MapR performs similarly to R-ChIP and shows similar discrepancies with ex vivo mapping approaches (Q. Yan et al., 2019). In situ approaches generally do not map R-loops in gene bodies, which raises the possibility that in situ approaches only detect certain subtypes of R-loops that may indicate differences in accessibility or that ex vivo techniques detect other types of nucleic acid structures that are not R-loops. This led some to question whether these techniques detect different R-loops because dRNH1 and S9.6 detect R-loops differently or, rather, because of differences between ex vivo and in situ detection. To address this question, an in situ S9.6-based R-loop mapping method and an ex vivo dRNH1-based R-loop mapping technique were developed R-loop CUT&Tag is an approach that combines a DNA–RNA hybrid sensor consisting of the RNase H1 hybrid binding domain (HBD) (GST-His6–2 × HBD) with a pA-Tn5-based technique that requires far lesser starting material as compared to current R-loop mapping methods (H. Wang, Li, & Liang, 2022; K. Wang et al., 2021). The authors of that study also applied this approach using S9.6 as the R-loop sensor to enable in situ detection of R-loops via S9.6. Finally, they used the HBD to develop a dRNH1-based version of DRIPc-seq which enabled dRNH1-based ex vivo R-loop mapping. In developing the optimal protocol for both R-loop CUT& Tag and DRIPc-seq, the authors were able to demonstrate that S9.6 and dRNH1 detect R-loops similarly when comparing the same ex vivo to ex vivo or in situ to in situ conditions - overall, R loop signals generated from all the different approaches were very similar, with their densities highly and positively correlated with each other. Thus, it remains likely that discrepancies between R-loop mapping studies are due to the processing involved with ex vivo versus in situ approaches and not due to the choice of dRNH1 or S9.6 for binding and immunoprecipitating DNA:RNA hybrids. Considering that the ex vivo methods generally involve a protease treatment, a likely explanation is that DNA:RNA hybrids, particularly intragenic ones, are less accessible in situ. On the other hand, hybrids made available following the protease treatment in ex vivo approaches may result in the R-loop probe (S9.6 or dRNH1) to become the limiting factor.

Figure 1. Simplified workflow of R-loop mapping approaches comparing in situ versus ex vivo mapping.

The upper diagram depicts the typical steps in an R-loop mapping workflow. Current approaches to recognize R-loops largely employ either a catalytically dead RNase H1 (in blue) or the S9.6 antibody (in red). Approaches that use either modality (or other probes, in dark blue), can be classified into either in situ (green) or ex vivo (magenta), based on whether the probe is introduced into the workflow prior to nucleic acid extraction and deproteination or after, respectively.

Analyzing R-loop mapping data

In parallel to the development of empirical R-loop mapping methods, there were several attempts to predict their likely locations by in silico approaches. Early biochemical experiments were instrumental in characterizing R-loopsin their functional aspects, such as their kinetic stability as DNA:RNA hybrids, which would lead to their use in isolation of nucleic acids. However, there remained a need for development of tools to predict R-loop formation. Subsequently, several strategies for prediction of R-loops were formulated such as “R-loop forming sequences” (RLFS), defined as the combination of a R-loop initiation zone, a linker region (any sequence up to 50 nucleotides in length), and a R-loop elongation zone region (Roy & Lieber, 2023; Wongsurawat, Jenjaroenpun, Kwoh, & Kuznetsov, 2012), and “SkewR,” where the algorithm analyzed G/C skew to predict regions which are favorable for R-loop formation. Predictions from these algorithms, including R-loop formation within cis-regulatory elements, and at the 3′ ends of genes (Ginno, Lim, Lott, Korf, & Chédin, 2013; Promonet et al., 2020), set a precedent for recent R-loop mapping studies, which have since then empirically corroborated these in silico results.

Since SkewR and RLFS are not based on empirical R-loop mapping, they may fail to account for other factors which influence R-loop formation in vivo. To address this potential limitation, a deep learning models such as deepRloopPre (K. Li et al., 2023) and extremely randomized trees (ET) (X. Pan & Frank Huang, 2022) have been introduced, which can predict genomic sequences that favor R-loop formation based on ssDRIP-seq and epigenomic data, respectively. Overall, R-loop prediction algorithms have provided preliminary insight into the dynamics of R-loops and their relationships with various genomic and epigenomic factors. However, these models lack the capability to reveal differences in R-loop formation across biological conditions, a task for which empirical R-loop mapping is required. Although in silico methods do not provide an exact prediction of where R-loops occur, they have proven useful as tools to evaluate likelihood of R-loops occurrence and further, as means to evaluate quality of R-loop mapping data.

Processing of R-loop mapping data

R-loop mapping data is like other types of genomic sequencing data (e.g., ChIP-sequencing), but tends to be noisier and, in many cases, no genomic input control is supplied (Ginno et al., 2012; Sanz et al., 2016). Raw data are typically provided in the form of FASTQ files, text files containing sequence reads, a unique identifier for each read, and quality information for each sequenced base. The raw data are then processed, often with standard genomics software. After processing, alignment files, coverage tracks, and peaks are obtained. Finally, downstream analysis is performed to explore, model, and visualize the data. There are no formal guidelines at present for R-loop mapping data analysis, potentially contributing to the quality control issues we review in the next section. Upstream processing of R-loop mapping data typically includes a first step to quality control the raw sequencing reads (Table 1). This involves trimming sequencing adapters, filtering low quality reads, and measuring sequence asymmetry (e.g., A-skew). Following read quality control, reads are aligned to the genome using short-read aligners like bowtie (or bowtie2) and bwa or (bwa-mem). Then, duplicate alignments are often removed using samtools (L. Chen et al., 2017; Ellis et al., 2021), Picard tools (Abakir et al., 2019; Alecki et al., 2020; Zeng, Onoguchi, & Hamada, 2021), or sambamba (Bayona-Feliu, Casas-Lamesa, Reina, Bernués, & Azorín, 2017), and the alignment files are compressed in BAM format.

After alignment, genomic coverage is typically calculated to produce coverage track files which can be viewed on a genome browser. This involves quantifying the number of reads which aligned at any given genomic location (usually aggregated into 10–200bp bins). The result is a genomic track which shows a genome-wide histogram of R-loop mapping coverage (usually in bigWig or bedGraph formats). Previous R-loop mapping studies have typically generated coverage tracks using bedtools (Crossley et al., 2020; Kotsantis et al., 2020), samtools (Ellis et al., 2021; Sanz & Chédin, 2019), or deepTools (Alecki et al., 2020; Itoh & Tomizawa, 1980; K. Wang et al., 2021; Wulfridge & Sarma, 2021). Finally, BAM files are used to calculate “peaks,” discrete genomic ranges which show robust coverage enrichment compared to background or input control. These peaks are useful for confidently identifying regions in which R-loops were detected. Most studies called peaks using MACS or MACS2 or custom peak calling models (Ginno et al., 2012; Sanz & Chédin, 2019; Sanz et al., 2016).

Additionally, there are other, uncommon R-loop mapping approaches which leverage bisulfite conversion or other non-standard mapping protocols (see descriptions in preceding section). To analyze these datasets, tools such as bismark are often employed (Dumelie & Jaffrey, 2017). In the case of ultra-long read bisulfite-converted R-loop mapping (SMRF-seq), a custom software program, footLoop, was developed for processing those data (Malig, Hartono, Giafaglione, Sanz, & Chedin, 2020).

Visualization of processed data.

Following upstream data processing, a downstream analysis is performed using custom scripts in R (L. Chen et al., 2017; Sanz et al., 2016; Wahba et al., 2016), python (Crossley et al., 2020; Dumelie & Jaffrey, 2017; Ellis et al., 2021), MATLAB (Xu et al., 2020), or perl (L. Chen et al., 2017; Sanz et al., 2016). The simplest approach for exploring R-loop mapping data is to visualize it using a genome browser - the Integrative Genomics Viewer (IGV) (Robinson et al., 2011) is a genome browser widely used in previous R-loop mapping studies (Crossley et al., 2020; Dumelie & Jaffrey, 2017; Xu et al., 2020). Another commonly used browser is the UCSC Genome Browser (Jenjaroenpun, Wongsurawat, Sutheeworapong, & Kuznetsov, 2017; Lin et al., 2022; Miller et al., 2022; Wittig & Wittig, 1979). These browsers allow users to query the co-localization of R-loops with epigenetic marks, TF binding sites, and other features of interest by providing public datasets from sources like ENCODE and therefore, offer an effective visualization technique. However, this approach does not provide a robust test of genome-wide colocalization. To address this limitation, bioinformaticians use statistical approaches that measure the genome-wide enrichment of R-loops within various genomic features, such as enhancers and introns. Within previous R-loop mapping studies, this type of analysis has been implemented using deepTools (Alecki et al., 2020; Crossley et al., 2020; Wahba et al., 2016; Wongsurawat et al., 2012; Xu et al., 2020, 2017), bedtools (Alecki et al., 2020; L. Chen et al., 2017; Crossley et al., 2020; K. Wang et al., 2021)), HOMER (P. B. Chen et al., 2015; Kotsantis et al., 2020; Xu et al., 2017), or using custom scripts.

Feature enrichment analysis.

To further analyze genomic feature enrichment in R-loop mapping data, bioinformaticians can use qualitative or quantitative approaches. For qualitative analysis, the most common approach is to produce “metaplots” that show the enrichment of R-loops around features of interest, such as within transcription termination regions downstream of poly-adenylation sites (Stork et al., 2016). Another common style of qualitative analysis involves ranking the intensity of R-loop mapping signal within genomic features of interest and comparing them to other pertinent signals. Usually generated with deepTools (Ramírez, Dündar, Diehl, Grüning, & Manke, 2014), the resulting “tornado plots” depict both R-loop enrichment within a feature of interest and co-localization of R-loop signal with other signals of interest. This type of analysis has been recently used to show that R-loops form at bivalent enhancers in pluripotent stem cells in regions which show low nascent transcription (Miller et al., 2022).

In addition to these more qualitative approaches, quantitative analyses are typically preferable, but rarely conducted. The simplest approach involves calculating the proportion of R-loop peaks that overlap with the relevant genomic features of interest and creating a pie or bar chart to visualize those proportions. This method was applied to demonstrate that most R-loops mapped in situ by R-ChIP are detected within promoter regions (L. Chen et al., 2017). A more robust approach involves using statistical methods like Fisher’s exact test to determine the significance and effect size of R-loop enrichment within genomic features, such as repetitive elements such as ribosomal RNA (rRNA) (P. Yan et al., 2020). Additionally, permutation testing can be used to measure the enrichment of R-loop peaks within and around genomic features. R packages like regioneR (Gel et al., 2016) can perform this type of analysis.

Differential analysis.

R-loop abundance can change under various biological conditions. Like feature enrichment analysis, there are more qualitative and more quantitative approaches for assessing differential R-loop localization and abundance. The simplest qualitative approach is to determine the overlap between R-loops mapped in different conditions and use a Euler diagram to show the proportion of overlapping and non-overlapping R-loops. This approach was successfully applied to reveal that knock-down of the splicing factor SRSF1 leads to an increase in R-loops (Promonet et al., 2020). As an integration of differential and feature enrichment analyses, overlaid metaplots that display results from different biological conditions are also used for qualitative comparisons of R-loop coverage around features of interest, recently employed to reveal that pharmacological inhibition of SF3B1 (a core splicing factor) decreases R-loop coverage within gene bodies (Han et al., 2022). Like differential gene expression analysis, “Differential R-loop Abundance” (DRLA) analysis involves calculating peak count matrices, fitting a negative binomial distribution, and calculating significance using a Wald test. R-loop mapping studies have employed this approach using R packages like DiffBind (Abakir et al., 2019; Bayona-Feliu et al., 2017) and DESeq2 (Crossley et al., 2020; Xu et al., 2020). In addition to the above approaches, there are several less-common analysis approaches used throughout R-loop mapping studies. One prominent example is motif analysis, which has been applied in a few studies to date (Briggs, Hamilton, Crouch, Lapsley, & McCulloch, 2018; Nadel et al., 2015; Wulfridge & Sarma, 2021; Xu et al., 2017).

Bioinformatics resources for R-loop research.

R-loop software packages.

While there is a growing interest in R-loop mapping, there is a dearth of software packages for the analysis of R-loop mapping data. “DRIP-optimized peak annotator” or DROPA, introduced in 2019, provides users with the capability to automatically annotate DRIP-seq peaks with various genomic features (Russo et al., 2019). When supplied with RNA-Seq data, it also provides the capability to predict DRIP-seq strand assignment. In addition to DROPA, footLoop, a perl-based utility was developed for processing SMRF-seq data (Malig et al., 2020). While DROPA and footLoop have high utility within certain contexts, they do not address several typical approaches for R-loop mapping data analysis.

To address current limitations, we recently described RLSuite, a collection of software packages that provide an integrative R-loop bioinformatics framework (Miller, Montemayor, Levy, et al., 2023). The core components of RLSuite are (1) RLPipes, a CLI utility for automated upstream processing of R-loop mapping datasets, (2) RLHub an R/Bioconductor package for accessing processed R-loop mapping data, and (3) RLSeq a feature-rich R/Bioconductor package for the downstream analysis of R-loop mapping data. These packages provide a standardized upstream preprocessing workflow which should help to harmonize R-loop mapping data analysis approaches in future studies. Similarly, they also provide a feature-rich downstream analysis workflow which incorporates several of the typical analysis methods described above and enables users to compare their results with those obtained from public R-loop mapping samples.

R-loop web servers and databases.

At present, there are a limited number of web servers and databases for interacting with R-loop data. The first such database, R-loop DB enabled the exploration of RLFS along with 12 public R-loop mapping samples (Jenjaroenpun et al., 2017). R-loop DB also provided the capability to predict RLFS from a user-supplied sequence, and it includes a genome browser interface for viewing R-loop mapping datasets. Limitations of R-loop DB include the small number of public R-loop mapping samples, the lack of visualization options for R-loop mapping data, and the lack of in-browser analysis capabilities. Another database, R-loopAtlas, serves as a repository for data and analysis tools geared towards the study of R-loops in plant genomes (Xu et al., 2017). While R-loopAtlas is highly useful for a subset of R-loop biologists, it has limited utility for researchers outside of the field of plant biology. Like R-loop DB, it also provides a genome browser interface for exploring R-loop mapping data. More recently, R-loopBase, a database of human R-loop mapping datasets and consensus R-loop zones was described which, unlike previous databases, provides the capability to explore R-loop zones and R-loop regulators (Lin et al., 2022). However, like R-loop DB, R-loopBase only facilitates exploration of R-loop mapping data through a genome browser interface. Moreover, as a web-based platform, R-loopBase does not provide the capability to analyze user-supplied data.

To address the limitations of previous web databases, we recently developed RLBase, a database and web server for accessing, exploring, and analyzing R-loop mapping data (Miller, Montemayor, Li, et al., 2023). RLBase facilitates access and exploration of 810 R-loop mapping datasets. Moreover, it provides users with the capability to analyze their own data in the browser. RLBase also provides an interface for exploring R-loop consensus regions, viewing the association between each and gene expression. Finally, it provides a UCSC genome browser session with all human datasets included, along with consensus R-loop regions (RL Regions).

Limitations of current R-loop mapping approaches.

R-loop mapping over the last decade has advanced significantly, as evident with the development of multiple methods and growing interest in R-loop biology. However, with the progress in development of mapping methods, variation in data derived from these different methods has arisen as a confounding factor. The use of different cell lines, in varying contexts based on the question being asked by a particular study and with different experimental controls for the method(s) chosen in a study has led to some confusion in interpretation of results. The apparent bias in R-loop detection derived from in situ and ex vivo methods, which largely aligns with dRNH1 and S9.6 based methods respectively, has further fueled this confusion. However, a meta-analysis of the data in RLBase (Miller, Montemayor, Li, et al., 2023) largely allays these concerns. There are clear differences in the R-loops detected by dRNH1 and S9.6 methods, but there is significant consistency too. Moreover, R-loops can be further divided into constitutive and variable R-loops (Miller et al., 2022). The constitutive R-loops correlate with constitutively expressed housekeeping genes while the variable ones mostly denote cell type differences. The variable R-loops should not be surprising given that R-loops are, in general, co-transcriptional events, and various cell lines derived from different tissues display dissimilar transcriptional programs.

In a fairly recent development, R-loops have been classified into Class I and Class II R-loops, where Class I R-loops are associated with promoter-proximal pausing of RNA polymerase II and have been particularly highlighted by in situ dRNH1 based methods (K. Wang et al., 2021). Class II R-loops are posited to rather associate with transcription elongation with hybrids assembling along the gene body and through the transcription termination site (Castillo-Guzman & Chédin, 2021). Though it has been proposed that Class I R-loops arise more frequently than Class II R-loops, this difference may be exaggerated by the differences in R-loop detection by the in situ and ex vivo contexts; additional more comparable analyses of R-loops in different regions will need to be conducted to verify quantitative differences. Beyond a programmatic role, there is some debate about the contribution of either Class I or Class II R-loops to genomic instability. Currently the role of R-loops and genome rearrangements have been mainly correlative, identifying enrichment of breakpoints in regard to R-loop forming sites (Lambo et al., 2019). However, the potential for R-loops to cause a block to DNA replication and instigate genome instability is significant and has been suggested to be more associated with Class I R-loops both because of their perceived frequency and their proximity to promoters (Castillo-Guzman & Chédin, 2021). However, that same study suggests that Class I R-loops, shorter in length, are likely more associated with CD conflicts between an active RNAPII and a progressing replication fork, whereas Class II R-loops, the longer counterparts, are more likely to be associated with HO collisions. This leads to an apparent contradiction where work examining R-loops and DNA replication for resolution of the hybrid structures observed that HO conflicts gave rise to more stable, deleterious R-loops while the R-loops in CD conflicts were resolved more efficiently (Hamperl, Bocek, Saldivar, Swigut, & Cimprich, 2017). This has since become an area of strong interest, and there is significant effort to understand and identify the R-loop subset involved in causing genomic stability, their characteristics and processes involved in formation and resolution of deleterious R-loops. Current R-loop mapping modalities do not consider differences in transcription-DNA replication choreography, cell cycle or origin firing, though these contexts will clearly be important to better understand the contribution of different R-loops to altered genomic integrity. This would be especially significant in pathogenic phenotypes such as cancer where oncogene-driven firing of otherwise dormant replication origins can interrupt the normal, carefully orchestrated progression of replication and transcription machineries, leading to more conflicts and R-loop formation.

Another significant controversy associated with R-loops involves the role of G quadruplex structures. Formation of R-loops is favored by the presence of G-rich sequences as the hybrid duplex forms mainly in the presence of Gs in the RNA strand (Huppert, 2008; Roberts & Crothers, 1992). As the G-rich non-template DNA strand is displaced during R-loop formation, this facilitates the assembly of a G4 structure on this single stranded DNA. Current models suggest that once formed, G4 structures can stabilize R-loop formation and enhance transcription (Lee et al., 2020; Tan, Wang, Phoon, Yang, & Lan, 2020). Conversely, the same sequences that can form a DNA G quadruplex would be present on the nascent RNA and can form an RNA Q quadruplex that would prevent DNA-RNA hybrid formation. Given that many of the same proteins that have been shown to bind DNA G4 structures are the same as those that bind RNA G4 structures, some of which have been shown to be associated with R-loops, it is easy to conjecture that the dynamics between these structures could provide the basis of a programmatic switch in R-loop prevention, formation or release. G4 complexes can also form on the template DNA strand, however, the functional implications of this on R-loops is not known, though other than promoting RNA polymerase stalling to facilitate R-loops behind the polymerase, it would be expected that these structures should prevent R-loop formation.

CONCLUDING REMARKS:

This overview details the various approaches that currently exist to recognize and map R-loops and methods employed by bioinformaticians to analyze the R-loop mapping data obtained. Detection methods for R-loops, such as high-throughput sequencing and immunofluorescence assays have several limitations in sensitivity, specificity, and resolution. Sequencing methods, while powerful for genome-wide mapping, suffer from background noise and non-specific signal, affecting the accuracy of R-loop identification. Similarly, immunofluorescence, while informative about R-loop formation qualitatively, are subject to specificity of the probe, cell cycle stage and cellular localization of R-loops (Smolka et al., 2021). Improved methods with enhanced sensitivity, specificity, and single-cell resolution are needed to accurately capture the dynamics and distribution of R-loops in diverse cellular contexts.

Understanding the dynamic regulation of R-loops in response to cellular cues is crucial for deciphering their context-dependent functions. While studies have implicated various factors in the regulation of R-loop dynamics, such as transcriptional activity, chromatin state, and DNA damage signaling, the underlying molecular mechanisms remain incompletely understood. Time-resolved studies using live-cell imaging or single-molecule techniques (Malig et al., 2020; Martin, De Almeida, Gameiro, & De Almeida, 2023) can provide insights into the temporal dynamics of R-loop formation, resolution, and turnover in different cellular contexts, shedding light on their regulatory pathways and functional outcomes. A recent study using single molecule tracking has successfully established the ability of RNA polymerase to form R-loops at double stranded breaks, independent of other factors (G. Lim et al., 2023). Investigating the interplay between R-loops and chromatin structure will be critical to understand their impact on genome function and stability. While studies have shown associations between R-loop formation and changes in chromatin accessibility, histone modifications, and nuclear organization (P. B. Chen et al., 2015; Penzo et al., 2023; Sanz et al., 2016), the causal relationships and mechanistic links remain mostly elusive. Integrative approaches combining chromatin conformation capture techniques with R-loop mapping methods can elucidate how R-loops influence chromatin architecture and gene expression regulation, providing insights into their functional consequences in health and disease. The localization of R-loops into various cellular compartments (nucleus, mitochondria and nucleolus) is another important aspect that has begun garnering attention, with recent findings tracing release of hybrids even into the cytoplasm (Crossley et al., 2022).

While significant progress has been made in understanding R-loop formation, regulation, and function, several challenges remain to be addressed to advance our knowledge in this area. Current studies largely focus on furthering our understanding of R-loop dynamics at a molecular level, while there is much to explore regarding the implications of R-loops for human diseases. It is generally understood that R-loop dysregulation is implicated in various diseases - usually observed by an increase in global levels of R-loops, often associated with a loss or mutation of an R-loop resolving factor - including cancer (Wells, White, & Stirling, 2019), neurodegenerative disorders (Cuartas & Gangwani, 2022; Kannan, Cuartas, Gangwani, Branzei, & Gangwani, 2022) and autoimmune conditions (Y. W. Lim, Sanz, Xu, Hartono, & Chédin, 2015). While correlation studies have linked R-loop accumulation to disease pathogenesis, establishing causality and mechanistic links requires further investigation. Experimental models, such as genetically engineered mice or patient-derived cell lines, can help dissect the contribution of R-loops to disease progression and identify potential therapeutic targets. Moreover, clinical studies correlating R-loop levels with disease severity or treatment response can provide valuable insights into their diagnostic and prognostic utility. On a similar vein, studying tissue-specific R-loops and differences in R-loop-mediated gene expression changes can provide mechanistic insights into the role of R-loops in disease pathogenesis (Jauregui-Lozano, Cottingham, & Hall, 2022; Yeo et al., 2014). However, as the field of R-loop biology expands, there are some significant knowledge gaps yet to be addressed - although the consensus is that R-loops appear to be programmatic and context-dependent, nonetheless, the basis of these programs and therefore the specific, contextual role of R-loops, or the consequences of their deregulation, are unknown or at the very least, not well understood.

Computational analysis of R-loop data faces challenges in data integration, interpretation, and predictive modeling. While bioinformatics tools exist for R-loop mapping and analysis, they often lack standardization and robustness, leading to variability in results (Miller et al., 2022). Moreover, predicting R-loop distribution, dynamics, and functional outcomes from genomic sequences and epigenetic features remains a complex task. Developing advanced computational algorithms and machine learning approaches tailored to R-loop biology will enhance our ability to analyze multi-omics datasets and unravel the regulatory networks governing R-loop function in health and disease.

The availability of several mapping methods has resulted in several discrepancies cropping up in the data obtained, particularly in the number of R-loop sites sequenced, which have been recently quantified to provide a better look at the differences in R-loop mapping approaches. The numbers of peaks called differ between R-loop mapping modalities spanning multiple orders of magnitude,even when confined to, for example, dRNH1-based approaches. In particular, one DRIVE-seq sample produced only 337 peaks while R-ChIP-seq can produce ∼6000 average peaks, and in some ssDRIP-seq datasets, upwards of ∼300000 peaks were reported. Variability is seen even within modalities, suggesting that mapping R-loops via only one method may not provide the most accurate data (Miller et al., 2022). Apart from the variances arising in performing ex vivo or in situ R-loop mapping as mentioned earlier, using S9.6 or dRNH1 as a probe to identify proteins that bind to R-loops elicits similar divergences as the R-loop forming genes mapped by either modality. Comparison of four datasets of R-loop binding proteins that were either immunoprecipitated using S9.6 (Cristini, Groh, Kristiansen, & Gromak, 2018; Wu, Nance, Chu, & Fazzio, 2021) or dRNH1 (Mosler et al., 2021; Q. Yan et al., 2022), shows that only 18% of proteins, mostly associated with ribosomal RNA processes, can be identified by both modalities. However, similar to the exclusivity of R-loop regions mapped by either probe, there are far more proteins found to solely bind R-loops recognized by either S9.6 (31%) or dRNH1 (51%) exclusive of the other (Figure 2). We observe that proteins immunoprecipitated with dRNH1-bound R-loops were primarily spliceosome proteins This conforms to and validates the observation that dRNH1 chiefly binds to promoter-proximal R-loops, which is where spliceosome assembly occurs. Interestingly, proteins immunoprecipitated with S9.6-bound R-loops were seen to be predominantly bound by proteins involved in co-translational activity and protein transport, indicating post-transcriptional protein interaction with R-loops which warrants further exploration.

Figure 2.

A) Venn diagram comparing immunoprecipitation (IP) datasets using inactive RNase H or S9.6 antibody to identify R-loop binding proteins. Four datasets were used for comparative analysis. Blue signifies unique R-loop binding proteins from two combined datasets using an inactive RNase H IP method. Red signifies unique proteins from two combined datasets using an S9.6 antibody IP method. Center white in the Venn diagram signifies proteins that were found in both inactive RNase H and S9.6 IP methods. B) Biological processes pathway analysis of all proteins from A, highlighted for pathways unique to S9.6 antibody or inactive RNase H IP methods. Proteins from four datasets in panel A were combined and analyzed for the top twenty biological processes pathways, organized by P value. Blue text signifies pathways that were unique to proteins identified from two datasets using inactive RNase H IP methods. Red text signifies pathways that were unique to proteins identified from two datasets using S9.6 antibody IP methods. Black text signifies pathways shared between both inactive RNase H and S9.6 IP methods.

While studies examining R-loops in disease models such as cancers assess them as pathological structures, with this being the area of most intense R-loop research, the physiological roles of R-loops cannot be ignored. Although it is still unclear what distinguishes physiological R-loops from pathological ones, it is likely that the “pathological” consequences of R-loop are attributed to R-loops accumulating due to defective R-loop resolution, rather than the formation of de novo R-loops (Rinaldi, Pizzul, Longhese, & Bonetti, 2021). As these unscheduled R-loops often accumulate at high levels in pathological contexts and are associated with genome instability (Lambo et al., 2019; Wells et al., 2019), R-loop mapping studies provide invaluable insight into further understanding the biology of these structures. Apart from using high throughput R-loop mapping techniques, other approaches to qualitatively measure R-loop formation can be used as a blunt first step, such as immunofluorescence or dot blotting using S9.6 or the inactive RNase H1 (Crossley et al., 2021; Ramirez, Crouch, Cheung, & Grunseich, 2021; Skourti-Stathaki, 2022). However, care should be taken to employ proper controls when using these methods as there can be non-specific binding of these probes, especially the S9.6 antibody, to other RNA structures or proteins. As a correlative measure, methods that assess active or paused transcription, such as RNA-seq, RNAPII-ChIP-seq, PRO-Seq or GRO-Seq can be applied to evaluate transcription-associated R-loop formation. Long-range cleavage sequencing (LORAX-seq) is a recently developed technique to sequence backtracked RNA at sites where RNA polymerase II pauses. This technique may provide insight into a new class of R-loops who consequences are yet to be determined, as persistent backtracked RNA can contribute to R-loop formation (Yang et al., 2024).

Future Perspectives

Addressing the challenges mentioned above requires collaborative efforts among researchers from diverse disciplines, including molecular biology, genetics, biochemistry, computational biology, and clinical medicine. By leveraging technological advancements and interdisciplinary approaches, we can deepen our understanding of R-loop biology, identify R-loop associated diseases and potentially uncover new therapeutic opportunities for addressing them. Best practices for R-loop mapping techniques include use of appropriate control treatments with RNase H1 (and RNase T1, RNase III when using S9.6 antibody as a probe) to eliminate non-specific recognition of other RNA structures, critical to a clean dataset. Using both an S9.6 antibody and a dRNH-based probe within the same RNA-mapping modality will not only contribute to stronger rigor and reproducibility, but also validate data obtained. Likewise, comparing an ex vivo and an in situ technique in parallel will provide the researcher with a more accurate representation of changes in mapped R-loops than using a single modality in their experimental approach. Using adequate starting material can prove to be crucial in obtaining clean datasets as most modalities involve multiple steps during the processing of samples which can lead to loss of sample amounts. As methods to map R-loops further evolve, it is imperative for researchers to be cognizant of the aforementioned limitations and precautions. While the methods themselves can be laborious and time-consuming, and therefore require multiple quality checkpoints, the sheer variety of approaches currently employed dictates an acute need for quality-controlled studies that contribute to accurate datasets.