High-Throughput Determination of RNA Structures

Abstract

The broad roles of RNA in performing and regulating fundamental life processes are being uncovered at a rapid pace. This has sparked profound questions about how RNA function is enacted through RNA structures that orchestrate processes in the complex cellular environment. These questions are being answered with technologies that use massively parallel sequencing to interrogate RNA structure at unprecedented throughput. In this article, we place the growing number of these technologies on a common conceptual framework, review their technical underpinnings, discuss the meaning and use of their data, and highlight their application to new areas of RNA biology.

Introduction

RNA molecules adopt intricate folds that underlie their central roles in regulating, maintaining and defending the genomes of all organisms^1,2. For example, messenger RNA is a key focal point for regulating gene expression, and non-coding RNA (ncRNA) functions range from transcription and translation regulation and promoting messenger stability regulation in prokaryotes, to gene silencing, epigenetic regulation, splicing regulation and molecular scaffolding in eukaryotes^3–7. RNAs also catalyze some of the fundamental chemical reactions of life, including peptidyl transfer in the ribosome⁸.

The functional diversity of RNA stems from its ability to form intricate structures that can change dynamically in response to cellular signals, such as ligands, proteins, temperature and the general cellular chemical environment^9,10. For example, the formation of simple secondary structures can directly impact gene expression by occluding regulatory binding sites. More sophisticated tertiary structures precisely position RNA helices, loops, bulges and junctions into architectures that define selective ligand binding pockets¹¹ and enzymatic active sites¹². Given the rapid pace at which new functional RNAs are being discovered², there is great opportunity to build off of this knowledge to uncover the structural principles that underlie the enormous functional diversity of RNA.

Biophysical characterization of RNA structure has yielded an atomic resolution understanding of the chemical and structural basis of RNA functions. However, experimental requirements often make these approaches low throughput. Biochemical approaches^13–20 (Figure 1) complement this limitation by sacrificing resolution for increased experimental flexibility and throughput. Efforts to establish biochemical probing as a quantitative measurement of RNA conformation^21–31 have yielded significant advances including the ability to interrogate RNA structures in the complex cellular environment^32,33. Additionally, the recent coupling of RNA structure probing with high-throughput sequencing (HTS)^34–36 has transformed the measurement of RNA structure into the ‘omics’ era by allowing thousands of RNAs, and even whole transcriptomes^37–40, to be studied simultaneously.

Figure 1 |. A brief history of biochemical RNA structure probing.

Biochemical probing of RNA structure began even before the diverse and important roles of RNA were fully appreciated¹⁶². Early methods used sequence and structure dependent cleavage by nuclease enzymes, which gave the first insights into how tRNA structures decode the genetic code¹³. Chemical probes were developed later^14,15,18, followed by methods that used RT primer extension to read out adduct position^16,17. Almost two decades later, biochemical probing data was incorporated into RNA folding algorithms to give more accurate models of RNA structure¹⁶³. The development of SHAPE chemistry in 2005²⁴ launched efforts to make RNA structure probing quantitative. The coupling of enzymatic^34,35,41 and chemical³⁶ probing to high throughput sequencing, and the development of probes that penetrate cell membranes^32,33 brought RNA structure probing into the ‘omics’ era. This allowed rapid advances in probing entire transcriptomes^37–40, high resolution modelling of tertiary structures⁴⁸, mapping of nascent RNA cotranscriptional folding pathways⁷⁵, and uncovering the role of RNA structures in mammalian RNA processing¹⁶⁴. Rapid technological advances such as new adduct mapping methods^44,54 promise to continue to increase the accuracy and power of these methods.

This review seeks to present a unified framework for high throughput RNA structure probing by describing recent developments and applications of these techniques in the context of their common conceptual core. We start with an overview of the experimental elements shared by all techniques and discuss how the resulting probe ‘reactivity’ data is generated. We next discuss several ways in which reactivities can be directly analyzed as well as incorporated within computational RNA structure prediction methods to generate more accurate models of RNA structures. We then discuss recent applications and practical considerations for using these techniques in several exciting new areas of RNA biology. Finally, we end with a roadmap for combining these techniques into an even more powerful suite of methods for uncovering the role of RNA structure in some of the most fundamental aspects of biology.

Probing RNA structures with sequencing

All techniques that use high throughput sequencing to characterize RNA structures share a common conceptual framework to encode RNA structural information into a pool of complementary DNA (cDNA) molecules that are sequenced and analyzed to recover that information (Figure 2) (Table 1)^34–55: in vitro or in vivo folding of RNA, structure-dependent RNA modification, recording of modifications by reverse transcription, sequencing and sequencing read alignment to map modification positions, and using these modification positions to calculate a ‘reactivity’ for each nucleotide in the RNA that reflects the underlying RNA structure. The use of HTS and bioinformatics enables simultaneous probing and mapping of modifications in complex mixtures of RNA molecules and whole transcriptomes. In the following sections, we discuss the different experimental choices for each step to enable users to adjust these techniques to ask different biological questions.

Figure 2 |. A common conceptual core for determining RNA structures with high throughput sequencing.

High throughput sequencing allows thousands of RNAs in complex mixtures to be structurally interrogated in a single experiment. In vitro or in vivo RNA mixtures (a) are first probed with chemical reagents that covalently modify the RNAs (red pin), or enzymes that modify the RNAs through cleavage (SI Figure 1), in a structure-dependent fashion (b) (Box 1). Modified RNAs are then converted into DNA (c) through reverse transcription that either stops or causes a mutation at each modification position (Fig. 3). The resulting pool of DNA molecules thus encode the original probe positions, and are formatted and sequenced to map the distribution of probe modifications using sequence alignment algorithms (d) (Fig. 4). The frequencies of mapped positions are used to estimate a reactivity value for each nucleotide of each RNA molecule (Box 2). High reactivities indicate nucleotides that are unstructured, while low reactivities indicate nucleotides that are constrained by RNA structures, tertiary interactions, ligand binding, or protein interactions. Reactivity information is then used to infer RNA structures (a) using a variety of computational methods.

Table 1 |. High Throughput RNA Structure Probing Techniques.

High throughput RNA structure determination methods combine enzymatic or chemical probes of RNA structures with high throughput sequencing to map structural information for up to thousands of RNAs in a single experiment. There has been a recent explosion in these techniques, which share a common conceptual core (Figure 2) to address a range of biological questions from determining RNA ligand interactions, uncovering cellular RNA structures and even cotranscriptional RNA folding pathways. The table below summarizes each of these techniques, the experimental choices and data analysis methods used in each, the level of RNA structure probed with each technique, and demonstrated applications. While many techniques were originally presented as a specific combination of experimental steps, due to their conceptual similarities many techniques are compatible with different probes and experimental strategies. Many techniques are also applicable to a wide array of RNA structural biology questions beyond their demonstrated applications. RNA structure modelling approaches that have used data from these techniques to predict RNA structures are listed in Supplementary Table 1.

Method	Demonstrated RNA Structure Level Probed	Probe	Modification Detection	Priming Strategy	Library Prep	Demonstrated Applications
Enzymatic Probes
PARS 34,165 and PARTE166	In vitro Secondary	RNase V1 (dsRNA) and S1 (ssRNA)	RT mapping of fragment ends	Linker defined	RNA Fragmentation, Ligation, Gel size selection	Transcriptome-wide in vitro RNA structure probing in S. cerevisiae34, Impact of SNPs on in vitro RNA structure of human transcriptomes115,165, Effect of temperature on RNA structures166
FragSeq35	In vitro secondary	P1 nuclease (ssRNA)	RT mapping of fragment ends	Linker defined	Ligation, Gel size selection	Transcriptome-wide in vitro RNA structure probing in M. musculus cell lines
dsRNA-seq and ssRNA-seq41,42	In vitro secondary	RNase ONE (ssRNase) and RNase V1 (dsRNase)	Read coverage of fragments from digested RNAs	Linker defined	RNA fragmentation, Gel size selection, Adapter ligation	Transcriptome-wide in vitro RNA structure probing in Arabidopsis thaliana, Drosophila melanogaster, and Caenorhabditis elegans
Small Molecule Probes
SHAPE-Seq36,43,50,75,85	In vitro/in vivo secondary/tertiary, RNA-ligand/RNA/protein interactions	1M7, BzCN, DMS	RT-Stop	Defined43,50, Linker defined43,75, Multiple defined85	Ligation, Bead size selection43 PCR Selection 50,75	In vitro equilibrium refolding of target RNAs36,43 in vivo probing in E. coli50,96,167, in vitro cotranscriptional RNA folding75,76,79, RNA-guided assembly of mammalian RNA nuclear export assemblies168
DMS-seq37	In vitro and in vivo secondary	DMS	RT-Stop	Linker defined	RNA fragmentation, gel size selection, adapter ligation, cDNA circularization	Transcriptome wide in vivo and in vitro RNA structure probing in S. cerevisiae and H. sapiens cell lines37, in vivo structures of G-quadruplexes in mammalian cells and E. coli128, structure of operon mRNAs in E. coli127
Mod-seq40	In vivo secondary	DMS	RT-Stop	Linker defined	RNA fragmentation, custom adapter ligation to detect RT-stops due to fragmentation, bead enrichment, circularization	Transcriptome wide in vivo RNA structure probing in S. cerevisiae, RNA-protein interaction footprinting
CIRS-seq46	In vitro secondary	DMS, CMCT	RT-Stop	Random	Ligation, Gel size selection	Transcriptome wide in vitro RNA structure probing in M. musculus cell lines
Structure-Seq38,52,54,84	In vivo secondary	DMS	RT-Stop38,52,84, RT-Stop and RT-Mutate54	Random	Ligation38 with hairpin adapter improvements52, Gel size selection or biotin-streptavidin pull down52	Transcriptome wide in vivo RNA structure probing in A. thaliana38 and Rice52, Connection to mRNA structure and protein structure in A. thaliana169. In vivo probing of RNA structure of Xist in M. musculus54,84
ChemModSeq45,53	In vivo secondary, RNA-protein interactions	DMS, 1M7, NAI	RT-Stop	Random	Ligation, Gel size selection, Random barcode to remove PCR duplicates	RNA-protein interactions governing ribosome biogenesis in S. cerevisiae
MAP-Seq47,80,170	In vitro secondary	DMS, CMCT, 1M7	RT-Stop	Defined	Ligation, Bead purification allowing 96-well plate formats, Fluorescent quantification of libraries and ligation efficiency pre-sequencing	In vitro folding of model natural47 and synthetic RNAs170, direct detection of in vitro RNA secondary structures with mutate-and-map read out (M2-Seq)145
SHAPE-MaP44	In vitro/in vivo secondary/tertiary, RNA-ligand/RNA/protein interactions	1M7, 1M6, NMIA44, DMS81	RT-Mutate	Defined, Random	No ligations if using defined priming, PCR or second strand synthesis to add all adapter sequences. Random priming requires adapter ligation.	In vitro folding of targeted RNAs, viral genomes44,86, in vivo RNA-protein interactions59, in vivo lncRNA structures and interactions (Xist)60, rRNA structure171
RING-MaP81	In vitro tertiary structure, RNA structure subpopulations	DMS	RT-Mutate	Defined	Same as SHAPE-MaP	In vitro folded model RNA systems and RNA-ligand interactions
icSHAPE39	In vivo secondary	NAI-N3	RT-Stop	Linker Defined	Biotin pull down of modified RNAs, RNA fragmentation, circularization, gel size selection	Transcriptome wide probing of RNA structure in M. musculus, effect of RNA modification on RNA structure
MOHCA-Seq48	In vitro tertiary	Targeted OH Radicals	RT-Stop	Linker Defined	Ligation, Bead purification	Tertiary reconstruction of in vitro folded model RNA systems
SHAPES172	In vitro secondary	NPIA	RT-Stop	Defined, Random	Biotin pull down, Ligation	In vitro probing of RNase P and 16S rRNA
DMS-MaP-Seq51	In vivo secondary structure	DMS	RT-Mutate	Linker defined	RNA fragmentation and gel size selection, linker ligation for RT, circularization, Random barcode to remove PCR duplicates	Global transcriptome analysis of S. cerevisiae RNA structure, Targeted RNA structures in Drosophila ovaries and mammalian cell lines51, analysis of 3’ UTR structure in mammalian genes164
Ligation-Based Methods
RNA Proximity Ligation134	In vivo RNA-RNA interactions	T4 RNA Ligase 1	Mapping of ligated junctions	Linker defined	TruSeq RNA-Seq library preparation	Transcriptome wide analysis of S. cerevisiae and H. sapiens cell lines, targeted analysis of select RNAs
LIGR-Seq136	In vivo RNA-RNA interactions	AMT crosslinking and circRNA ligase	Mapping of ligated junctions	Linker defined	RNA fragmentation, Size selection, Adapter ligation	Transcriptome-wide in vivo analysis in H. sapiens
PARIS135	In vivo RNA-RNA interactions	AMT crosslinking and T4 RNA ligase I	Mapping of ligated junctions	Linker defined	RNA digestion and 2D gel purification, Adapter ligation, Size selection	Transcriptome-wide in vivo analysis in H. sapiens and M. musculus
SPLASH137	In vivo RNA-RNA interactions	Biotinylated psoralen (biopsoralen) crosslinking and T4 RNA ligase I	Mapping of ligated junctions	Linker defined	RNA fragmentation, Bead enrichment, Size selection, Adapter ligation	Transcriptome-wide in vivo analysis in H. sapiens, E. coli, and S. cerevisiae

Abbreviations:

1M7, 1-methyl-7-nitroisatoic anhydride

1M6, 1-methyl-6-nitroisatoic anhydride

AMT, 4’-aminome-thyltrioxaie

BzCN, Benzoyl Cyanide

CIRS, Chemical Inference of RNA Structures

COHCOA, Closure-based *OH correlation analysis

CMCT, 1-cyclohexyl(2-morpholinoethyl)carbodiimide metho-p-toluene sulfonate

DMS, dimethyl sulfate

icSHAPE, in vivo click SHAPE

LIGR, LIGation of interaction RNA

MAP, Multiplexed Accessibility Probing

MaP, Mutational Profiling

MOCHA, Multiplexed *OH Cleavage Analysis

NAI, 2-methylnicotinic acid imidazole

NMIA, N-methyl isatoic anhydride

PARS, Parallel Analysis of RNA Structure

PARIS, Psoralen Analysis of RNA Interactions and Structures

PARTE, Parallel Analysis of RNA Structure with Temperature Elevation

REEFFIT, RNA Ensemble Extraction From Footprinting Insights Technique

RING, RNA Interaction Groups

SHAPE, Selective 2’-Hydroxyl Acylation Analysed by Primer Extension

SHAPES, SHAPE Selection

SPLASH, Sequencing of Psoralen crosslinked Ligated And Selected Hybrids

Different probes for different biological questions

The design of a high-throughput RNA structure probing experiment begins with the biological question being asked. This determines the experimental context, which in turn influences the choice of probe or combination of probes used.

Biochemical RNA structural probes fall into two broad categories: enzymatic and chemical probes. Enzymatic probing uses nuclease enzymes that cleave RNAs in structure- and base-specific patterns of fragments. Chemical probes react with RNA molecules to encode the structure of the RNA as a pattern of covalent adducts that can then be mapped (Figure 2). Since chemical probes are much smaller than comparatively bulky enzymatic probes, they often offer higher resolution RNA structural information^56,57. In addition, chemical probes have been tailored to allow probing of RNA structures within cells^{19,20,32,33,39,50,58–63}, which is not the case for enzymatic probes due to the difficulty of transporting bulky nuclease enzymes across cell walls in a controlled fashion. The reactivity of chemical probes can also be manipulated, giving rise to a wide array of probe choices that can be tailored to specific questions (Box 1). We therefore focus on chemical probes in this review, while enzymatic probes are covered in other reviews^56,57,64.

Box 1 |. A palette of chemical probes interrogate many layers of RNA structure.

Chemical probes offer great flexibility in designing experiments that interrogate many facets of RNA structure. Different chemical probes interrogate multiple contexts of RNA structure including base pairing, local nucleotide dynamics and solvent accessibility. In the figure, levels of structural information accessed by chemical probes are mapped onto the structure of the aptamer domain of the T. petrophila fluoride riboswitch (PDB: 4ENC) that is shown alongside chemical structures of the bases (part a). The chemical structures of an array of commonly used probes, along with the structure of the adduct formed (red), the leaving group (blue), and reagent half-life (t_1/2) where appropriate, are shown in a table (part b).

When choosing which chemical probe to use to address a particular biological question, care must be taken to properly account for the detailed nuances of each probe’s chemistry to ensure the resulting reactivity data can be properly analyzed, interpreted and compared between different conditions and experiments. Several details are particularly important. First, while probes generally fall into broad categories of chemical reactivity patterns, each probe chemistry is specific and it is generally being established that the most informative experiments use multiple probes to obtain more complete RNA structural information^{25,48,72,80,151}. Second, probes can often react with base specificities other than that commonly reported. For example, DMS can also react with the N7 position of G’s⁶⁵, and some reports detect DMS modification of the N3 position of U and N1 and N3 positions of G^81,82. In addition, SHAPE reagents such as NMIA and 1M7 show minor, but significant, dependencies on base identity¹⁵⁹, though these measurements have not been performed for the newer SHAPE probes. Third, the measured reactivity can depend on the timescale of the probe reactions (Box 2). For probes that need to be quenched with specific quenching reactions (see part b) this is particularly important, as the measured reactivity will increase with the longer duration of the probing step before quenching. For probes that naturally quench through the reaction with water, the t_{1/2,hydrolysis} (t_1/2 in the figure) is an important parameter that governs the timescale of RNA folding being probed. Specifically, t_{1/2,hydrolysis} can directly govern the measured reactivity of a specific position if the timescale of ribose puckering transitions are much slower than SHAPE adduct formation²⁵ (Box 2). This fact can be exploited to reveal particularly slow conformational changes in RNA^72,74 and lead to more accurate models of RNA dynamics¹⁵¹. In addition, t_{1/2,hydrolysis} can be influenced by ionic strength and pH for some SHAPE reagents such as NMIA⁷¹, so care must be taken to account for this when designing folding conditions. It is also generally unknown how side reactions with other cellular components may influence the general timescales, and therefore observed reactivities of in vivo probing with SHAPE reagents, which should be a topic of future research. Fourth, the formation of covalent adducts with the RNA could change the underlying RNA structure being probed^82,93, especially under conditions where multiple modifications are being sought⁸¹, and control experiments or careful data interpretation are needed to take this into account. Finally, some probing reagents are toxic or have toxic side products so care must be taken to perform probing experiments safely.

Chemical Structures

	Probe	Chemical Structure (Adduct Red, Leaving Group Blue). Letters refer to Box_l_ProbeStructures.	Primary Modification Sites	Half-Life	Reference
	N-methyl-isatoic anhydride (NMIA)	a	2’ OH, all nts	430 s	Merino et al., 2005
	1-methyl-7-nitroisatoic anhydride (1M7)	b	2’ OH, all nts	14 s	Mortimer and Weeks, 2007
	1-methyl-6-nitroisatoic anhydride (1M6)	c	2’ OH, all nts	31 s	Steen et al., 2012
SHAPE	Benzoyl cyanide (BzCN)	d	2’ OH, all nts	0.25 s	Mortimer and Weeks, 2008
	2-methylnicotinic acid imidazolide (NAI)	e	2’ OH, all nts	33 min	Spitale et al., 2013
	2-methyl-3-furoic acid imidazolide (FAI)	f	2’ OH, all nts	73 min	Spitale et al., 2013
	2-methylnicotinic acid imidazolide azide (NAI-N3)	g	2’ OH, all nts	33 min	Spitale et al., 2015
	Dimethyl sulfate (DMS)	h	G N7, A N1, C N3	N/A	Peattie and Gilbert, 1980; Tijerina et al., 2007
Base Specific	N-cyclohexyl-N’-(2-morpholinoethyl)carbodiimide metho-p-toluenesulfonate	i	G N1, U N3	N/A	Gilham, 1962; Ho and Gilham 1967
	Kethoxal, other 1,2-dicarbonyl compounds	j	G N1 and N4	N/A	Litt and Hancock, 1967; Mitchell et al., 2018
Solvent Accessibility	Hydroxyl Radical	k	backbone	N/A	Latham and Cech, 1989; Wang and Padgett, 1989
	Nicotinoyl Azide (Naz)	l	G C8, A C8	N/A	Feng et al., 2018

Chemical probes fall into two main classes (Box 1): base-specific probes and generalist probes. Base-specific probes react with base moieties and are therefore directly sensitive to base-pairing interactions or solvent accessibility. For example, dimethyl sulfate (DMS) reacts preferentially with the Watson-Crick face of adenine (N1 position) and cytosine (N3 position) as well as the N7 position of guanine⁶⁵, 1-cyclohexyl(2-morpholinoethyl)carbodiimide metho-p-toluene sulfonate (CMCT) reacts with the Watson-Crick face of guanine (N1 position) and uracil (N3 position)⁵⁶, kethoxal reacts with the N1 and N2 positions of guanine to form a new ring structure⁵⁶, and glyoxal derivatives react with the amidine moieties on the Watson crick faces of G, A and C⁶². On the other hand, N,N-(dimethylamino)dimethyl chlorosilane (DMAS-Cl)⁶⁶ reacts with the N2 position of guanine and nicotinoyl azide (NAz) ⁶³ react with the C8 position of guanine and adenine to measure solvent accessibility (Box 1). While there is a large palette of base-specific probes⁶⁷ that enable direct interrogation of the structural state of the base, it is necessary to use multiple reagents if structural information for every nucleotide is desired.

Generalist probes react with the RNA backbone to interrogate RNA structure and can therefore be used to monitor every nucleotide simultaneously (Box 1). For example, hydroxyl radicals (*OH) probe solvent accessibility through backbone cleavage^18,68. SHAPE probes (Selective 2’-hydroxyl acylation analyzed by primer extension)²⁴ are electrophilic reagents that interrogate local nucleotide dynamics by reacting with the ribose 2’-hydroxyl of each backbone position^25,27,28,69. SHAPE probes self-quench by reacting with water and are typically characterized by hydrolysis half-life (t_{1/2,hydrolysis}) which governs the timescale of RNA dynamics probed, the type of environment they can be used in (in vitro vs. in vivo), and the ultimate meaning of the reactivity measured from these probes (Box 1,2). The mechanism of SHAPE probing can make reactivity interpretation challenging. For example, unpaired bases in an RNA hairpin loop can still be constrained by base stacking interactions that make them weakly reactive to SHAPE probes⁷⁰. However, if interpreted appropriately this can reveal higher order structures and tertiary contacts which can manifest themselves as regions of intermediate to weak reactivities that change when different reagents or divalent cation concentrations are used in the folding conditions ^71,72.

The availability of diverse chemical probes enables tailoring of probe choice to specific biological questions. Probes that react with RNA rapidly are well-suited to studying RNA folding dynamics. For example, the renaturation folding pathway of RNase P RNA has been studied with the fast SHAPE reagent benzoyl cyanide (BzCN) which revealed that the forming of a single stacking interaction is the rate limiting step of folding this RNA^73,74. BzCN was also used to study how nascent RNAs fold as they exit RNA polymerase cotranscriptionally^75,76. More recently, a fast probing technique called light-activated structural examination of RNA (LASER) was developed ⁶³. Here, UV light is used to activate nicotinoyl azide (NAz) into an electrophilic aroyl nitrenium ion, with both excitation and quenching occurring on a ps timescale, which promises to open new doors to accessing even faster timescales of RNA loop and interhelical dynamics ¹⁰with chemical probes.

There is also a wide pallet of probes that can be used to study RNA folding in vivo. Probes that react with RNA slowly naturally allow this, since their long reaction time allows them to penetrate cellular and nuclear membranes. Specifically, probes such as DMS, 1-methyl-nitroisatoic anhydride (1M7), 2-methyl-3-furoic acid imidazolide (FAI), 2-methylnicotinic acid imidazolide (NAI), NAI-N3 and glyoxal have been used in a number of studies to ask fascinating questions about the role of the cellular environment on RNA folding and function (see below)^{19,20,32,33,39,50,58–62}. In addition, since NAz can pass through cellular membranes and be subsequently activated with light, LASER joins hydroxyl radical footprinting ⁷⁷ as a fast in vivo probe of solvent accessibility, tertiary structure and RNA-protein interactions though without the need to use synchrotron radiation. We discuss additional applications of the diverse chemical probes available to biological questions below.

Approaches to detecting probe modifications

Following probe selection, a method for RNA modification detection must be chosen. Current strategies for chemical adduct detection use reverse transcription (RT) to record modifications in the resultant cDNA sequences as either truncations (RT-stop) or mutations (RT-mutate) (Figure 3c).

Figure 3 |. Reverse transcription strategies for detecting RNA modifications.

Chemical probing experiments detect the presence of RNA modifications using reverse transcription (RT). Three important variables to consider are the RNA processing steps used before RT, the RT priming strategy, and the method used to detect modifications (Table 1). Optional processing steps before RT include fragmenting the RNA, linker ligation for downstream RT priming, or modifying adducts to allow purifications (a). Several RT priming strategies can be used (b). RNA defined priming uses an RT priming site within the RNA, while linker defined primes off of the ligated linker. Multiple RT primers are used for long RNAs, and random primers can be used to tile across long RNAs, or to probe complex mixtures of RNAs and transcriptomes. Two methods are used to detect modification positions (c). In RT-stop methods, RT stops one nucleotide before the RNA modification, leading to a truncated cDNA product. In RT-mutate methods, RTs read through the modification and cause a mutation in the cDNA. Both methods result in premature truncations and background mutations in both the probed (+) and control (−) reactions that should be accounted for in downstream data analysis (Box 2).

RT-stop methods are based on the tendency for chemical adducts to halt reverse transcriptases one nucleotide prior to the modification^24,67. Historically, RT-stop has been the detection strategy of choice¹⁷ because truncated cDNAs are easily separated by gel electrophoresis. The first sequencing-based strategies also relied on RT-stops to encode RNA modifications³⁶ but used HTS to map cDNA ends thereby enabling multiplexed probing. RT-stop methods are amenable to the least expensive HTS kits because they require very short DNA sequencing reads to map each fragment. However only one modification can be detected per cDNA because RT has a propensity to stop at the first modification encountered. Therefore, when using RT-stop methods, probing conditions should be optimized to balance the desire for single-hit kinetics while still modifying the RNAs sufficiently to allow good data quality⁷⁸. RT-stop methods also require signal decay correction during data analysis since longer cDNA fragments are naturally less abundant due to the higher chance of encountering a previous stop ^79,80 (Box 2).

RT-mutate methods are based on the tendency for reverse transcriptases to misincorporate at chemical adducts under specific reaction conditions⁴⁴. Modification detection by RT-mutate is specifically enabled by HTS due to the requirement of sequencing to read mutations. In these Mutational Profiling (MaP) methods, the encoding of modifications as mutations rather than truncations during RT is promoted by the use of Mn²⁺ as the divalent cation in place of Mg²⁺. Because misincorporation under these conditions does not tend to halt RT, multiple RNA modifications can be detected per molecule^44,51. This allows for more advanced data analysis approaches that separate sequencing reads into different groups according to mutation pattern before downstream analysis to uncover the signatures of tertiary interactions, or multiple subpopulations of RNA structures in the probed ensemble⁸¹, or even to detect paired bases⁸² (see “Towards Tertiary Interactions” below). RT-mutate methods have also be used to detect structural changes in an RNA due to single-nucleotide differences by splitting reads according to these differences before structural analysis ⁸³. When using RT-mutate methods, it is important to use DNA sequencing read lengths that cover the entire RNA region of interest to obtain complete information on its structure.

While the properties of the two modification detection methods may appear to favor RT-mutate methods, two recent reports suggest that both may suffer from biases in their ability to detect specific RNA modifications. Sexton et al.⁵⁴ and Novoa et al.⁵⁵ performed in cell DMS modification and detected both RT truncations and mutations. They found several important features of the resulting data: (i) RT stops are detected even in conditions that favor mutations; (ii) distinct biases are observed for each encoding strategy, with RT-mutations favoring DMS-modification detection at C and RT-stops favoring detection at A (observable in many previous DMS probing datasets^37,38,45,81); and (iii) biases in modification recording are both reverse transcriptase and sequence context dependent. Overall this suggests that RT-stop and RT-mutate methods should be used simultaneously to give the most accurate mapping of DMS modifications^54,55 (Box 2). More work is needed to confirm that these conclusions apply broadly to the palette of chemical probes, and to study these effects in the context of other well-defined RNA folding benchmarks.

A key aspect of experimental design for both RT-stop and RT-mutate approaches is the choice of RT priming strategy (Figure 3b). Defined priming uses DNA primers that bind to specific parts of the RNAs under study, and is well suited to studying isolated RNAs, or specific RNAs that are present in a complex mixture^{36,40,43,44,47,49,50,54,58–60,80}. Here, the priming sites can be either internal to the RNA^{48,50,51,54,59,84} or within a defined region that can be included in the RNA 3’ end and that is present during folding and probing^44,47. Ligation of a linker sequence to the RNA after probing enables the addition of a defined priming site after probing^{40,43,48,49,51,75}. For RT-stop methods, each priming site can be used to generate useful information for a ~300–400 nt window beyond which the signal becomes too weak due to signal decay⁸⁵. For RT-mutate methods, longer regions are possible but are limited by current sequencing read lengths. For RNAs that exceed these length limitations, multiple defined priming sites can be used and reactivity data stitched together within the different windows^84,85. Multiple priming can also be performed using random RT primers that can bind at many positions throughout a mixed population of RNAs, which is particularly useful for transcriptome-wide studies^38,45,46,52, or studying long RNAs^44,60,85,86. The use of stronger binding LNA primers can be advantageous for random priming⁴⁴. We note that some methods perform RNA fragmentation before or after RNA modification and before reverse transcription to address issues with RT drop off in long fragments^37,39,40. If this is done, special care must be taken to account for the presence of RT stops due to the fragmentation process instead of adduct formation, which can confound the calculation of reactivities⁴⁰. Fragmentation can also be used with RT-mutate approaches to study long RNAs⁵¹.

Finally, for both RT-stop and RT-mutate approaches, another key aspect of experimental design is the inclusion of a no-reagent control sample (often referred to as a (−) channel sample). This RNA sample is folded in the same conditions, but treated with a control solvent without modifying reagent prior to RT in the exact same steps as the modified ((+) channel) RNA sample. The inclusion of the (−) channel allows data analysis to remove false positive reactivity signal that could be due to natural RT fall off or mutations in the specific sequence context of the RNAs being studied (see Box 2 below).

Multiplexing with high throughput sequencing

The conversion of modified RNA into cDNA enables sequence analysis by HTS. The use of HTS to directly sequence cDNAs immediately expanded the scope of previous structure probing experiments (Figure 1), as the DNA sequencing information could be used to bioinformatically distinguish reads that arise from thousands of different RNA molecules all probed simultaneously.

While powerful, care must be taken when designing experiments using HTS as many additional experimental steps are required to ‘format’ cDNAs into ‘libraries’ that are compatible with specific sequencing platforms. Specifically, most HTS platforms require that cDNAs contain platform-specific adapter sequences on either end of the molecule that are used by the instrument for a variety of purposes such as localizing DNA molecules to imaging surfaces, facilitating amplification, or serving as sequencing priming sites⁸⁷. Adapter sequences can be added during reverse transcription as an RT primer tail, through ligations, by PCR, or by a combination of these approaches (Figure 4).

Figure 4 |. Strategies for next generation sequencing library preparation.

Several biochemical steps are needed in order to convert cDNA products from reverse transcription (Fig. 3) into a format compatible with sequencing instruments. After reverse transcription, a series of ligation, circularization and PCR steps are performed in different combinations to append adapter sequences needed for the sequencing process (a), and an optional set of barcodes and indices that can be used for multiplexing and downstream analysis (b). Adapter sequences are often required on both ends of the cDNA. One adapter can be conveniently included as a tail of the RT primer. The second adapter can then be ligated on through a single stranded DNA ligation, or through a circularization if the RT primer tail contained elements of both adapter sequences. Note that ligations/circularizations must be used for RT-stop techniques since each RNA modification position leads to a different cDNA end which must be mapped to recover the position. In RT-mutate methods, the presence of full length products due to RT read through allows PCR to be used to add the necessary adapter sequences. Many techniques contain additional enzymatic processing (c) and purification (d) steps throughout library preparation that get rid of excess oligonucleotides or unwanted side products that can impact downstream steps or the number of informative reads from the library. Each additional step has potential biases that must be considered when estimating reactivities (Box 2).

A major consideration when selecting a library formatting strategy is the biochemical biases that can be introduced at every stage of library preparation, including during the linker ligation used for RT priming⁸⁸, the RT priming strategy itself⁷⁸ (Figure 3), the ligation/circularization step used to add additional adapter sequences⁸⁹ (Figure 4) and PCR amplification⁴³. Bias in DNA adapter ligation is of particular importance in RT-stop methods, since the cDNA 3’ end that the adapter is ligated to is precisely the sequence that needs to be accurately mapped in order to recover the location of RNA modification (Fig. 3c). Common concerns for each of these steps is that the reactions do not proceed to completeness, and have biases towards specific sequences that would allow those contexts to be unequally sampled, which would lead to biased reactivity data. To address this, several studies have undertaken careful optimizations of some of the key steps of library preparation including using quality control analysis to directly assess ligation efficiency⁴⁷, and the particularly promising approach of using optimized adapter formats and conditions to reduce adapter ligation bias^52,89. Lastly, while PCR amplification of sequencing libraries enables the use of low amounts of input material, there is a concern that amplification could lead to bias. The limited studies to directly assess this have concluded this is not the case⁴³, though extensive comparisons for all variations of library preparation were not made. A particularly promising approach for eliminating PCR bias has recently been used by several studies that incorporate a random barcode within the library adapter sequences^51,55 (Fig. 4B). This barcode is designed to be theoretically long enough to uniquely tag each cDNA product for bioinformatic filtering so that each molecule is only counted once.

Another important consideration is the removal of unwanted side products. Many library preparation steps use excess primer, linker or adapter oligonucleotides to help drive reactions to completion. This excess can interfere with downstream library preparation steps or result in adapter dimers that use up valuable sequencing reads. Gel extraction protocols have been used to remove these excess oligonucleotides and side products. However, this approach is laborious and inefficient. To remedy this several bead selection approaches have been developed to select against unwanted products by targeting specific sequences of interest through complementary oligonucleotides to the target RNA ⁴⁷, streptavidin binding to biotinylated positions in the cDNA added during RT⁵² or streptavidin binding to biotin positions introduced through click chemistry approaches on specially designed SHAPE reagents such as NAI-N3³⁹. Finally, PCR selection approaches can be used whereby PCR primers are designed to include a mismatch in unwanted side products that prevents their amplification⁵⁰.

Experimental challenges

Since all high-throughput RNA structure probing techniques share this conceptual umbrella, they share several common challenges that should be considered to understand the complexities and limitations of the data generated. Perhaps the most significant is that all techniques inherently project the sophisticated three-dimensional nature of RNA structures onto a one-dimensional vector of reactivity information. This means that in order to recover or infer models of RNA structures from this data, additional information needs to be added – either through computational methods or multiple experiments, which are both discussed below. In addition, these techniques are inherently ensemble measurements – they probe a population of RNA molecules and structural states, meaning the resulting data reflects an average over this population. Computational methods^90–92 or recently developed experimental strategies⁸¹ are needed to extract information about different RNA structures that may exist in the population. Furthermore, the fact that these techniques probe structure by directly chemically modifying RNA has two important implications: covalent adducts can alter the underlying RNA structure being probed ^82,93 meaning care must be taken in interpreting data resulting from multiple modifications per molecule, and these techniques are inherently destructive meaning that there is no way to continually probe the time trajectory of RNA folding of the same sample. Finally, while some work has been done to identify what sequencing depth of coverage is required for accuracy^44,78,94,95, more work is needed to fully establish best practices, particularly in the case of transcriptome-wide experiments. Nevertheless, the benefits of determining RNA structures with high throughput sequencing outweigh these challenges and the rapid pace of innovation is already starting to improve the technique.

Data analysis: defining and calculating reactivities

The ultimate goal of RNA structure probing experiments is the quantitative measurement of RNA structural information. This measurement is most naturally reported as ‘reactivity’ that describes how reactive each RNA nucleotide is to a given probe. Since probes tend to react with unpaired positions, high reactivities generally correlate with unstructured regions of the RNA, while low reactivities correlate with structured regions. However, the definition of reactivity more deeply links the nuances of the probe chemistry to the pattern of observed DNA sequencing reads (Box 2), and informs the interpretation of reactivities for structural insights and modeling.

‘Reactivity’ links chemistry with RNA structure

Chemical and enzymatic probing are governed by the chemical reaction of the probe with an RNA. Consequently, the quantitative definition of reactivity lies within the chemical kinetics that govern a probing reaction. The inherent fact exploited by RNA probing methods is that the rate of modification is structure-dependent. Nucleases depend on the propensity for specific enzymes to cleave either single or double stranded RNA^34,35,56, whereas chemical probes tend to react faster with unstructured nucleotides. For some chemical probes, models of the chemical kinetics have been developed that directly link reactivity to the rate constants that describe the structural context of a nucleotide and the chemical reaction itself^23–25 (Box 2). The implication is that if the rate constants of the chemical reaction are known, then reactivities can be interpreted as quantitative kinetic parameters about nucleotide-level RNA structural properties. For example, when defined in this way, reactivity values that are in between the extreme high and low ends of the scale reveal information about nuanced structural contexts such as base stacking⁷⁰, slow structural fluctuations in specific regions of the molecules^25,74, regions of the molecule which undergo melting transitions^69,96,97, tertiary interactions^69,71,72,98, and others, which highlight the richness of chemical probing reactivity data.

Calculating reactivities from sequencing reads

Chemical kinetics naturally establishes the reactivity of a nucleotide as the fraction of molecules with a modification at that position at the end of the probing reaction (Box 2). Thus, reactivity information is inherently encoded in the distribution of modifications across an RNA molecule – if many modifications are observed at a given nucleotide, the reactivity will be higher at that position and vice versa. In the absence of complicating factors, the reactivity of a nucleotide is the number of modifications observed at that position divided by the number of RNA molecules probed, or the fraction of adduct formed at that nucleotide. However, because current methods for modification detection are indirect, this calculation must be adjusted to account for experimental factors in order to estimate the true reactivity value (Box 2).

Experimental factors can be accounted for by calculating reactivities as a corrected formula for the fraction of adduct formed at each position (Box 2), and represent a statistical estimate of the true reactivity values determined by chemical kinetics. Consequently, the kinetic parameters that describe RNA structural context should be quantitatively extractable from these reactivity estimates. Furthermore, these estimates define a natural and absolute scale for reactivity that could be used to make comparisons across different experiments, laboratories, and even conditions⁸⁰ if the kinetic parameters and biases of different experimental strategies are taken into account. Many methods propose normalization schemes whereby highly reactive positions are normalized to a value of one. While normalized data can still be used for relative comparisons within a molecule, the normalization breaks the link to the chemical kinetics definition of reactivity. Furthermore, alternative methods for calculating reactivity that do not estimate fraction of adduct formed also break the links to the underlying chemistry of the probing reaction.

Data analysis: interpreting and using reactivities

Structural probing reactivities can be analyzed in different ways. Most directly, reactivity values can be interpreted based on their magnitude to roughly identify structured and unstructured regions of the RNA. Reactivities can also be compared across different RNAs or across different experimental conditions to identify regions that change structure in these contexts. More complete models of RNA structure can be generated using computational algorithms that incorporate reactivity data in their calculations. Below we discuss each of these areas.

Interpreting reactivity information

The direct interpretation of reactivity values can be a powerful starting point when analyzing high throughput structure probing data. Highly structured or unstructured regions can frequently be observed by visual analysis of the reactivity patterns across the molecule. For simple RNA structures, characteristic reactivity patterns often emerge. For example, strong hairpins often result in a ‘low-high-low’ pattern where low reactivity regions correspond to the stem and the high reactivity pattern indicates the loop (Figure 5). However, complex RNA structures often yield reactivity patterns that are difficult to directly interpret and may necessitate computational modeling for a deeper understanding (see below). Nevertheless, these complex patterns can often be understood through the lens of additional structural information. For example, comparing reactivity patterns to known crystal structures can reveal how base stacking or non-canonical base pairing are manifest as reactivity signatures^{27,28,36,44,48,70,72}. More recently, comparison of reactivity data to detailed molecular dynamics simulations revealed that reactivities correlated with features of RNA structural dynamics such as how frequent the ribose backbone samples the C2’-endo configuration, and suggested mechanistic roles for SHAPE reagent stacking with the backbone and specific nucleotide bases to facilitate reagent binding prior to reaction ⁹³. It is important to note that the interpretation of intermediate reactivities will depend on the reactivity scale and any normalization schemes used (Box 2). As community efforts to share RNA structural mapping datasets grow^99–101 this creates an exciting opportunity for a deeper analysis of reactivity patterns to aid in their direct interpretation.

Figure 5 |. Applications of high-throughput structure probing.

The development of high-throughput RNA structure probing methods has enabled new and diverse applications to previously inaccessible biological questions. The development of new chemical probes and analysis frameworks has enabled the characterization of in vivo RNA structure as it pertains to RNA protein interactions⁵⁹, RNA functional states, and post-transcriptional interactions^39,50 (a). The extension of in vivo chemical probing experiments to transcriptome wide studies has identified functional roles for RNA structure in mRNA cleavage and polyadenylation¹⁶⁴ and the structural organization of operon mRNAs¹²⁷, and has found that G-Quadruplexes are globally unfolded in eukaryotic cells¹²⁸ (b). The combination of high-throughput RNA structure probing and multiplexed in vitro transcription allows measurement of nascent RNA folding pathways⁷⁵ and immunoprecipitation of Pol II has been used to isolate cellular nascent RNAs for structure probing¹⁴⁶(c). MOHCA-Seq ⁴⁸ and RING-MaP⁸¹ leverage the information richness of high-throughput RNA structure probing to infer higher order RNA structure (d). Structural subpopulations can be identified through spectral clustering of RING-MaP data⁸¹ and by multiple structure prediction using SLEQ computational analysis methods ¹²⁰ (e).

Comparative reactivity analysis

A strength of chemical probing experiments is their flexibility, which permits the controlled manipulation of probing conditions in vitro such as temperature, salt concentration, and the presence of interacting ligands and proteins^{32,49,69,72,96,102}, and even allows probing in cellular conditions^{32,33,37–40,50,51,58–61,96}. A direct comparison between reactivities in different conditions is therefore a powerful way to study how these conditions affect RNA structure. Comparative reactivity analysis typically involves subtracting values between specific conditions to find RNA regions that are higher or lower reactivity (less structured or more structured) in particular conditions. For example, several studies have applied in vitro probing to study how ligand binding changes riboswitch structure^48,49,72,102, and to find differences between kinetically folded and equilibrium refolded RNAs^75,76. Comparative methods have been recently developed that make use of sliding window averages and statistical tests to give more confidence in the discovery of protein-RNA interactions⁵⁹. Comparative methods have also been coupled to high-throughput targeted mutagenesis to dissect RNA structural ensembles and the mechanism of ligand-mediated structural switching in a regulatory RNA ¹⁰³. A particular exciting area is merging comparative reactivity analysis with multiple sequence alignment, which has been used to identify structural RNA regulatory elements within viral genomes⁸⁶. While powerful, care should be taken in these analyses to separate the effects of changes in reactivity due to structural changes of the RNA, or to ligand binding and other cellular interactions.

Using reactivity information to model RNA structures

The accuracy of computational RNA secondary and tertiary structure prediction can be improved by incorporating reactivity data¹⁰⁴. Reactivities can be used within RNA secondary structure prediction methods as restraints to return structures that are more consistent with the data, by either using reactivities to modify the underlying model parameters in the algorithm^104–109, using data to select structures from a set of generated possibilities ^92,110, or a combination of both^90,91,111. For example, SHAPE data incorporation into thermodynamic RNA folding models have be done through addition of defined pseudo-free energy terms ¹⁰⁴, or iteratively perturbing the energy model to generate structures that better match the SHAPE data ¹⁰⁶. Specifically, pseudo-free energy terms that use reactivities (r) when calculating free energies of nucleotides involved in base stacking have typically taken the form:

where is the pseudo-free energy term at nucleotide i, m and b are constant parameters, and r(i) is the reactivity at nucleotide i ¹⁰⁴. With m positive and b negative, this term effectively penalizes nucleotides within structures that have high reactivity, and thus favors nucleotides with low reactivities to be in structured positions. As an alternative to specifying a functional form for a free energy derived from reactivities, iterative approaches add small perturbations to each free energy term in order to minimize the difference between estimated experimental base pairing probabilities derived from reactivities and predicted base pairing probabilities using the pseudo-free energy terms ¹⁰⁶.

Once reactivity information is incorporated into the energy model, there are a variety of computational methods to predict the experimentally-restrained structural state. Minimum free energy (MFE) approaches return the most thermodynamically favored conformation, which is theoretically the structure most occupied by an RNA in equilibrium¹¹². On the other hand, maximum expected accuracy (MEA) methods use a probabilistic model to select the most probable structure across the ensemble ^105,113,114. Multiple structures consistent with the experimental data can be predicted with these two types of methods by reporting the next best MFE or MEA structure(s)¹⁰⁹. Computational approaches that sample multiple RNA structures from the equilibrium thermodynamic probability distribution can also be used to model an ensemble of structures as a way to extract population level structural information from reactivity data, and can be used to look at how structural populations shift when conditions are changed or SNPs are introduced^{83,90,91,96,115}. There are also methods that can use reactivity information to restrain tertiary structure models^48,116–120. For more details, there are excellent reviews of RNA structure prediction algorithms that incorporate experimental structure probing data ^109,121.

Many of the methods that incorporate experimental probing data have been shown to improve predictive accuracy ^49,122. To date, most accuracy assessments are done by comparing the base pairing patterns of structural predictions to solved RNA crystal structures since the latter reveal both secondary and tertiary structure base pairing partners ^43,123. In addition, experimentally restrained structural predictions have been made to structures derived from comparative sequence analysis to determine the accuracy of these methods on highly conserved RNAs ¹²². In each case, accuracy is quantified through measures of sensitivity (true positives divided by the sum of true positives and false negatives) and positive predictive value (PPV, true positives divided by the sum of true positives and false positives), with current techniques improving each measure broadly from around 60% for predictions without experimental data to above 85% when experimental data is included ¹⁰⁵. For all methods, the accuracy improvement of each algorithm varies based on the RNA, and to date there is no algorithm that can use experimental data to predict the structure of every RNA with perfect accuracy. This is due to several factors. First there are inherent limitations on the computational structural models that may limit accuracy. These include the fact that free energy parameters are optimized to predict structures in specific in vitro folding conditions (temperature, salt concentration, pH) that may not match well to the experimental probing conditions, and the fact that certain non-canonical base pairing interactions are not included in most models. Second, predictive accuracy is likely influenced by experimental data quality, though the relationship between data quality and prediction accuracy has been mostly assessed using simulated data. Finally, accuracy assessments have only been done on a limited set of benchmark RNAs that are not representative of all biologically relevant RNAs across all organisms ^{100,101,124,125} which could reveal important classes of structures where method improvement is needed. The continual improvement of these data-informed RNA structure modeling methods, and an extension into modeling RNA folding dynamics ⁹³, is an exciting area of research and will likely evolve with the improvements in experimental techniques.

Enabling new RNA structural biology

The advent of techniques to probe RNA structures with HTS has given rise to fundamentally new ways to interrogate the role of RNA structures in a wide range of biological processes from transcription to macromolecular assembly. Below we highlight several areas of RNA structural biology that are being revolutionized and in some cases pioneered by these techniques.

Examining how the cellular environment affects structure

The ability to perform RNA structure probing within different experimental conditions has allowed some of the first insights into how the complex environment in the cell impacts RNA folding (Figure 5A). These studies make use of the fact that some older chemical probes^32,37,38 and some newly developed probes^33,39,62 can traverse cellular membranes. By comparing in vivo and in vitro measurements, several studies have found general impacts of the cellular environment on RNA structure, with some concluding that RNAs in vivo are more³², or less³⁷ structured than in vitro. Other studies have found that RNA structure in cells depends on the context³⁹, for example in regulatory regions where low reactivities were observed upstream of splice sites for unspliced RNAs³⁸. Other studies have started to uncover clues as to how specific factors such as protein and ligand binding, or RNA modifications alter RNA structures in the cell. These include the observation of protein binding to sites in regulatory and catalytic RNAs such as snRNPs, SRP RNA, rRNAs and RNase MRP^{45,53,58,59,126}, hints that m⁶A modification affects RNA structures³⁹, views on the role of structure on differential translation rates within mRNA operons¹²⁷, and intriguing suggestions that G quadruplexes are globally unfolded in eukaryotic cells¹²⁸. When performing cellular probing, care should be taken when deciding which probe to use⁶¹, the priming strategy to target specific RNAs or interrogate structures transcriptome-wide, and the depth of coverage needed to accurately estimate reactivities^78,94,129.

RNA ‘structuromics’: probing RNA structure transcriptome-wide

All RNAs have an inherent structure and thus studying RNAs transcriptome-wide could provide a systems level view of RNA structure (Figure 5B). The first transcriptome-wide structure probing experiments used in vitro enzymatic methods^34,35,41,42 and the subsequent application of chemical probing^37–39,130 enabled these measurements to be made within the cell. These methods have been applied across diverse systems and provide insight into the global regulation of RNAs through structure. Comparison of in vivo and in vitro measurements has started to address questions such as how RNA binding proteins affect RNA structure and the structural accessibility of coding sequences at transcriptome scale^37–39,131. Developments in in silico computational screens for conserved RNA structures have also been used to enhance the discovery of functional structured RNAs expressed across the genome and have been shown to agree with in vitro structure probing data ^132,133. Alternative to direct structural probing, sequencing-based proximity ligation methods have also been developed to more directly determine RNA-RNA interactions^134–137 and those mediated by RNA binding proteins^138–141. Depth and coverage are a large consideration for transcriptome-wide studies since not only do reads need to span the transcriptome, there should also be a high coverage of reads to accurately estimate reactivities for each RNA. New computational methods that incorporate the estimation of transcript isoform abundance along with reactivities are an important advance for the analysis of experimental transcriptome-wide studies¹⁴². In addition, the ability to use SHAPE-MaP to detect structural differences between different RNA alleles ⁸³ may become important for understanding how genome sequence variation impacts cellular functions through changes in RNA folding.

Towards Tertiary Interactions

RNA chemical probing has emerged as a tool that can provide insights into complex RNA folds using simple biochemical measurements that are compatible with diverse experimental contexts. Interestingly, even though chemical probing is lower resolution than biophysical methods, recent work has leveraged the information richness of HTS to infer higher order structure. In particular, the insight that RT-mutate methods can extract the locations of chemical adducts at multiple positions within a single RNA has enabled the implementation of a probing experiment to identify RNA interaction groups by mutational profiling (RING-MaP)⁸¹. RING-MaP heavily modifies a target RNA using DMS, which results in a pattern of mutations in each cDNA. Nucleotide positions that are closer to each other in space can cause modification patterns to correlate with each other, for example in a breathing motion where multiple positions may become reactive. The identification of such RNA interaction groups thus facilitates tertiary structure modeling (Figure 5D). Alternatively, if cDNAs are first clustered into groups before reactivity calculation, structural information about RNA subpopulations may be extracted from the data⁸¹. Another approach to uncovering tertiary interactions jointly applies Multiplexed OH (radical) Cleavage Analysis with paired-end sequencing (MOHCA-seq) to infer tertiary contacts by measuring solvent accessibility of positions⁴⁸. MOHCA-Seq uses 2’-NH₂ modified RNAs coupled to isothiocyanobenzyl-Fe(III)•EDTA treatment such that hydroxyl radicals generated by fenton chemistry cleave positions that are near the modified base. By using HTS to map both the modification position and the cleavage position, through-space connections can be mapped. The multidimensional chemical mapping (MCM) pipeline combines MOHCA-Seq with the mutate-and-map approach^98,143–145 to first infer a secondary structure model through systematic mutagenesis and chemical probing, which then creates a rich dataset that can be used with three-dimensional modeling up to 1 nm resolution (Figure 5D). For RING-MaP, it is important for the RNA to be sufficiently modified and sequenced at sufficient depth to extract the most higher order structural information. Similarly, the size of the RNA target is an important consideration when applying the MCM pipeline as its multidimensional nature makes the experimental cost scale quadratically with RNA length⁴⁸.

Uncovering the fleeting structures of nascent RNAs

The folding of RNA molecules into functional structures begins during transcription, however technical barriers have limited the measurement of nascent RNA intermediates at nucleotide resolution. Because the coupling of chemical probing to HTS enables the interrogation of RNA mixtures in solution, it is ideal for mapping intermediate nascent RNA folds (Figure 5C). We recently developed methods to simultaneously probe every intermediate transcript of a target RNA in an in vitro transcription reaction^75,76. The basis of cotranscriptional RNA structure probing is the stable distribution of transcription elongation complexes across every position of a DNA template prior to rapid chemical modification. The resulting data can be used to infer the folds that a nascent RNA could pass through during transcription and its relation to its function, and exciting new computational techniques are able to use this data to estimate the proportion of RNAs that undergo a particular folding trajectory within the population ¹²⁰ (Figure 5E). New methods that couple these techniques to RNA polymerase pull downs are also giving promising early views on nascent RNA structure in vivo¹⁴⁶ (Figure 5C). While studies of nascent RNA structure using high-throughput chemical probing are in their infancy, early results emphasize the importance of metastable intermediate structures, particularly in the context of riboswitch-mediated transcription regulation. The compatibility of chemical probing with the cellular environment makes in vivo nascent RNA structure probing an intriguing prospect. However, stably halting cellular transcription for the duration of chemical modification while minimally perturbing nascent RNA structure poses a significant challenge. Given the dynamic nature of cotranscriptional folding it is also important to consider the timescale of the probing reaction when designing experiments and interpreting data.

Future perspective and conclusions

Massively parallel DNA sequencing technologies have transformed efforts to understand the molecular nature of life by enabling information rich biochemical measurements at unprecedented scale¹⁴⁷. This is especially true for RNA biology, where high throughput RNA structure determination technologies are quickly becoming an engine for discovering the RNA structures and interactions that mediate fundamental cellular functions. When applied in well-designed and innovative experiments, these tools have generated a renewed appreciation for the importance of gaining a complete understanding of the physicochemical basis of RNA function, which must include knowledge of cellular RNA folding.

This next chapter of RNA biology though is not without challenges. As the application of high-throughput structure probing increases, a major challenge is to improve experimental accuracy. To do this, high quality benchmarks to systematically evaluate these techniques are needed. Current benchmarks^{43,104,122,148,149} use a panel of natural RNAs with sophisticated structures, and evaluate techniques by whether reactivities can be incorporated into computational structural models to accurately predict those structures. While important, this is an indirect assessment of experimental accuracy that is confounded by the computational modeling approaches. In contrast, a more direct method that evaluates accuracy at the level of the reactivity data itself is needed. For example, developing a panel of simpler RNAs to directly test the ability of techniques to uncover the chemical kinetic parameters of the probing reaction (Box 2), and verify these using complementary approaches, would be a simpler and more direct measure of accuracy. Once established, this benchmark would provide a route to solving the challenge of reducing bias and maximizing the accuracy of measured reactivities. While there has been progress to this end, no study to date has directly compared all of the protocol choices on the same pool of modified and unmodified RNA. There is also the challenge of the depth of sequencing needed to accurately recover reactivity information^78,94,129, which varies by approach. Finally, in vivo studies must address how the timescale of chemical probing can impact measurements made in the dynamic cellular environment and in the presence of cellular factors.

Future innovation in high-throughput RNA structure probing will combine advances in probe chemistry, adduct detection, and multi-level computational models to address new frontiers in RNA biology. Techniques for direct RNA sequencing¹⁵⁰ that could provide a direct measurement of adduct position could greatly simplify the complexity, and cost needed to uncover reactivity information. In addition, the new emphasis on using multiple probes to extract maximal information^44,48,72,151 with advanced modeling⁴⁸ creates an exciting prospect for in vivo probing to uncover higher resolution cellular RNA structure models. It is also foreseeable that the creation of troves of probing data^99–101 will create a rich resource for machine learning approaches to infer these higher resolution RNA structure models.

There is also great potential for combining high throughput RNA structure probing with complementary approaches to enhance and expand our view of the RNA structure/function relationship. For example, merging these methods with single molecule biophysical approaches could link the structural signatures uncovered by chemical probing to RNA folding dynamics^152–154. In vivo probing techniques could also be merged with optical methods for nucleic acid detection^155,156 that could create enormously rich views of not only what, but where RNA structures are in the cell.

The adaptation of high throughput RNA structure probing methods to new biological and application contexts also has great potential to address longstanding biological questions and facilitate medical advances. For example, we anticipate the application to studying RNA folding dynamics will create a new frontier in RNA biology to uncover the importance of nascent RNA structure. There is also vast potential for these techniques to uncover therapeutic RNA targets and help engineer treatments as evidenced by a new wave of emerging biotechnology companies that are treating RNAs as druggable targets, or using engineered antisense oligonucleotides as therapeutic drugs themselves^157,158. An efficient future pipeline for developing RNA centric therapies will likely involve a combination of high-throughput characterization and screening approaches to identify targets and uncover antisense oligonucleotide design principles.

The questions about how RNA structure mediates biological function are some of the oldest of molecular biology. While our ability to ask and answer these questions has been transformed by high throughput RNA structure determination technologies, it is likely that many more waves of innovation are yet to come. Given the history of discovery in RNA biology^1,2,6, we are likely on the cusp of another expansion in our understanding of the roles that RNA molecules play in life.

Box 2 |. Estimating chemical kinetic ‘reactivities’ from statistical analysis of sequencing reads.

The concept of ‘reactivity’ quantitatively links the chemical probing reaction to the underlying RNA structures. Reactivity is therefore defined from the coupled kinetics of RNA structural fluctuations¹⁰ and probe chemistry (Box 1), but must in practice be statistically estimated from the observed sequencing reads. This creates two views of reactivities that should be linked to allow the maximum amount of RNA structural information to be obtained.

Chemical kinetics view

Under a coarse-grained model²⁵, different bases fluctuate between reactive (open), and unreactive (closed) conformations with rates k_open and k_close, respectively (part a). The open state can react with the probe (red) with a rate constant k_add, and the reaction proceeds for a time t_rxn (set by either t_{1/2,hydrolysis} or by the quenching time (Box 1)). The reactivity of site j (denoted r_j) is then defined as the fraction of RNA molecules in the population that is modified at site j, and can be written as a function of all the rates, t_rxn and the concentration of reagent used, S₀. Under this model, the reactivity is linked to RNA structure through the dependence on k_open and k_close²⁵. For example, when S₀k_add is slower than k_open and k_close, r_j can be calculated to be²⁵

$$r_j=1-e^{-k_{\mathit{open}}/(k_{\mathit{open}}+k_{\mathit{close}})(k_{\mathit{add}}/k_{\mathit{hydrolysis}})S_0}$$

though this form can change and be exploited to gain deeper structural information^25,72,74. More work is needed to incorporate features of the nuanced mechanisms of probe reactions into this model^27,28,93. This definition of reactivity has been established through modelling SHAPE probes^24,25,28. However, since any RNA nucleotide can dynamically fluctuate¹⁰, similar considerations could apply to all probes.

Read statistics view

Reactivities must ultimately be estimated from counts of the sequencing reads observed in a chemical probing experiment ⁷⁹. This is most rigorously done by using statistical methods to estimate the percentage (or fraction, f) of adducts observed at any given position to give an accurate estimate of reactivity at that given position j (r_j^) that should correspond to the chemical kinetic reactivity (r_j) (part b) ¹⁶⁰. While the most simple estimate of reactivity would be to divide the number of RT-stops or RT-mutations at a specific position by the total number of observed reads, more accurate statistical estimates must account and correct for the many steps of sequencing library construction used (Figure 4) that can give rise to multiple sources of bias and noise that must be accounted for:

Bias in adduct detection: An adduct can either cause RT to stop (stop_j) or introduce a mutation (mut_j) at a specific position¹⁶⁰, and the propensity towards one or the other changes based on RT enzyme and conditions^54,55. It was recently suggested that a more accurate estimate of total reactivity should use both signatures to count adducts ^54,55, which can be rigorously incorporated as a sum of the reactivities due to stops (r^_j(stop)) plus that due to mutations (r^_j(mut)) ¹⁶⁰.

RT drop off bias: RTs prematurely terminate due to encountering an adduct, a sequence/structure context, or other processes. This causes fewer long cDNAs, which requires a correction for single ^79,161 or multiple⁸⁵ RT priming positions, and should also be used for random priming. Specifically, to estimate the fraction of adduct observed at each position, f_j, adduct counts stop_j and mut_j are divided by the observed read depth at that position, depth_j, rather than the total reads observed ¹⁶⁰.

Experimental noise due to background RT stops and mutations: This noise can be corrected for by performing a control reaction (−) where no adduct is used, and subtracting percent adduct estimates from the reaction with adduct (+).

Context biases in library construction: These could be reduced through experimental optimization or techniques to measure and correct for specific biases⁴⁷.

Reactivities calculated from sequencing data are estimates to the true chemical kinetics reactivity, and the specific interpretation of what constitutes the values of ‘high’, ‘low’, and intermediate reactivities is nuanced. By estimating reactivities that correspond to the fraction of adduct observed ¹⁶⁰, reactivity values should most closely align with the kinetics of the chemical probing reaction, which should allow a deeper understanding of data from high throughput RNA structure probing experiments.

Glossary

Chemical Probe

A small molecule that chemically reacts with RNA molecules in a structure-dependent fashion. Reactions produce adducts which can be detected to give a measure of an RNA’s structure

cDNA

The complementary single stranded DNA molecule that is produced when RNA is converted to a DNA molecule by a reverse transcriptase enzyme. This conversion allows RNAs to be indirectly sequenced by HTS

High Throughput Sequencing (HTS)

A suite of technologies that can be used to sequence millions to billions of DNA molecules simultaneously. Many experiments can be performed at once since bioinformatics can be used to distinguish signals between experiments

Positive Predictive Value (PPV)

A measure of accuracy of RNA structure prediction equal to the number of true positive pairs predicted divided by the sum of true positive and false positive pairs predicted. PPV is often used in combination with sensitivity to assess predictive accuracy of RNA structure models. PPV is equivalent to one minus the false discovery rate

Reactivity

A measure of a chemical probing reaction that contains RNA structural information. Typically high reactivities indicate unstructured regions while low reactivities indicate structured regions

Reverse Transcription (RT)

The process by which RNA is enzymatically converted into complementary DNA. RT proceeds in the RNA in the 3’->5’ direction

RT-Stop

An event where reverse transcriptase stops when encountering a chemical probe adduct on an RNA. This produces a truncated cDNA that can be used to map the adduct position

RT-Mutate

An event where reverse transcriptase produces a mutation when encountering a chemical probe adduct on an RNA. This mutation can be used to map the adduct position

Secondary Structure

The pattern of base pairing interactions in an RNA that create helices, loops, bulges, junctions and single stranded regions. In addition to Watson-Crick base pairs, RNAs can pair in many non-canonical patterns

Sensitivity

A measure of accuracy of RNA strucutre prediction equal to the number of true positive pairs predicted divided by the sum of true positive and false negative pairs predicted. Sensitivity is often used in combination with positive predictive value (PPV) to assess predictive accuracy of RNA structure models

Shape

Selective 2’-hydroxyl acylation analyzed by primer extension. A class of chemical probes that modifies the RNA backbone. SHAPE probes can be used to interrogate RNA structure at single nucleotide resolution

Tertiary Structure

The three dimensional orientation of secondary structure elements and nucleotides that give rise to sophisticated three dimensional structures. Tertiary structures can be stabilized by non-covalent interactions and divalent cations