Guidelines for SHAPE reagent choice and detection strategy for RNA structure probing studies

Abstract

Chemical probing is an important tool for characterizing the complex folded structures of RNA molecules, many of which play key cellular roles. Electrophilic SHAPE reagents create adducts at the 2'-hydroxyl position on the RNA backbone of flexible ribonucleotides with relatively little dependence on nucleotide identity. Strategies for adduct detection such as mutational profiling (MaP) allow accurate, automated calculation of relative adduct frequencies for each nucleotide in a given RNA or group of RNAs. A number of alternative reagents and adduct detection strategies have been proposed, especially for use in living cells. Here we evaluate five SHAPE reagents: three previously well-validated reagents (1M7, 1M6, and NMIA), one more recently proposed (NAI), and one novel reagent (5NIA). We clarify the importance of carefully designed software in reading out SHAPE experiments using massively parallel sequencing approaches. We examine SHAPE modification in living cells in diverse cell lines, compare MaP and reverse transcription-truncation as SHAPE adduct detection strategies, make recommendations for SHAPE reagent choice, and outline areas for future development.

Graphical Abstract

Introduction

Many RNA molecules fold into complex structures that drive key processes, such as catalysis(1), viral packaging, protein binding(2), splicing(3), polyadenylation(4), transcription termination(5), transcript degradation(6), and translation efficiency(7). For decades, chemical probing has been an important strategy for mapping structural features of nucleic acids. The pursuit of reagents and a technology that report on the local structure of all four ribonucleotides in a concise experiment led to the development of SHAPE (for selective 2′-hydroxyl acylation analyzed by primer extension), a straightforward approach using an electrophilic reagent to modify the backbone of conformationally dynamic nucleotides, read out by truncations during reverse transcription (RT) to identify adduct locations and quantify their relative abundances(8). SHAPE has been applied to study diverse RNAs: including small riboswitches(9-11); bacteriophage, bacterial, and eukaryotic mRNAs(12-16); and the RNA genomes of complete human viruses (17-20).

Multiple efforts then focused on increasing the throughput and scope of SHAPE experiments by replacing or supplementing the adduct detection step with methods integrated with massively parallel sequencing. The sequencing reads generated from adduct-containing RNAs are then quantified with automated software to generate per-nucleotide reactivities. Most of these methods rely on counting truncations that occur when RT enzymes terminate at RNA adducts(21-24). These RT-truncation methods require follow up steps to ligate and process these cDNA products into a sequenceable library format, potentially affecting the information recorded during RT-truncation. Mutational profiling (MaP), in contrast, detects internal mutations in cDNA generated when an RT enzyme reads through RNA adducts during relaxed-fidelity DNA synthesis(25). Downstream library construction steps have little to no effect on the information recorded during the MaP read-through step. MaP has enabled rapid structural analysis of many RNAs relevant to basic biology and human health including HIV(25, 26), the Xist long non-coding RNA(27), a satellite tobacco mosaic virus genome in packaged and unpackaged states(28), ribosomal RNAs from divergent bacteria(29), hundreds of E.coli transcripts(30), human SERPINA1 mRNAs(31), and bacterial and human protein binding targets(27, 32, 33).

The SHAPE reagents 1-methyl-7-nitroisatoic anhydride (1M7), 1-methyl-6-nitroisatoic anhydride (1M6), and N-methyl-nitroisatoic anhydride (NMIA) have been extensively validated by examining their reactivity on multiple RNAs of known structure(34, 25, 35). Recent efforts have focused on improving SHAPE modification and adduct detection efficiency, especially in living cells(36, 27, 37). Additional SHAPE reagents with long half-lives have been proposed such as 2-methyl-3-furoic acid imidazolide (FAI) and 2-methylnicotinic acid imidazolide (NAI)(38). New methods for enhancing adduct detection have also been introduced including biotin enrichment approaches coupled with the SHAPE reagent N-propanone isatoic anhydride (NPIA)(23) and the clickable reagent 2-methylnicotinic acid imidazolide-azide (NAI-_N)(24). These new reagents and adduct detection methods have not been as extensively validated on large RNAs of known and complex structures as have 1M7 and MaP. Here we report results of MaP experiments using multiple SHAPE reagents across different cell types and compare MaP and RT-truncation as SHAPE adduct detection strategies. This study clarifies the importance of carefully designed software in reading out SHAPE experiments using massively parallel sequencing approaches, provides guidelines for SHAPE reagent choice, and identifies areas for future development.

Methods

Cell culture

Mouse embryonic stem cells (SM33) were cultured on tissue culture plates coated with 0.1% gelatin and grown in high glucose DMEM with sodium pyruvate supplemented with 15% FBS, 0.1 mM nonessential amino acids (Gibco), 2 mM L-glutamine, 1000 Units/mL ESGRO leukemia inhibitory factor (Millipore Sigma), 100 Units/mL penicillin, 100 μg/mL streptomycin, and 0.1 mM 2-mercaptoethanol. C2C12 myoblast cells were cultured on tissue culture plates in DMEM media supplemented with 10% FBS, 100 units/mL penicillin, and 100 μg/mL streptomycin, and experiments were performed before cells were completely confluent. HSV40 fibroblasts were cultured on tissue culture plates in DMEM supplemented with 15% FBS, 100 units/mL penicillin, and 100 μg/mL streptomycin. LNCaP cells were cultured on tissue culture plates in RPMI-1640 media supplemented with 10% FBS, 100 units/mL penicillin, and 100 μg/mL streptomycin. Jurkat T cells were grown in suspension culture flasks using the same media as used for LNCaP cells.

Mammalian cell lysis and protein digestion

C2C12 myoblasts, H4SV fibroblasts, LNCaps, or mouse embryonic stem cells were grown in two 10-cm dishes to ~80% confluency. Both plates were washed once in PBS before scraping and lysis in 2.5 mL of ice-cold cytoplasmic lysis buffer [40 mM Tris, pH 8, 175 mM NaCl, 6 mM MgCl₂, 1 mM CaCl₂, 256 mM sucrose, 0.5% Triton-X 100, 0.5 unit/μL RNasin (Promega), 0.45 Units/μl DNase I (Roche)]. Jurkat cells (grown in suspension culture) were pelleted at 3000 ×g for 5 min, washed once with PBS, then resuspended in lysis buffer. Cells were lysed for 5 min on ice with intermittent mixing. Cell nuclei were pelleted at 3000 ×g for 5 min at 4 °C (nuclear fraction), and the resulting supernatant (cytoplasmic fraction) was transferred to a new tube. A volume of 2.5 mL of proteinase K buffer [40 mM Tris, pH 8, 200 mM NaCl, 1.5% sodium dodecyl sulfate, and 0.5 mg/mL proteinase K] was added to the nuclear pellets. The supernatant (cytoplasmic fraction) volume was increased to 5 mL with appropriate solutions to result in final concentrations of 200 mM NaCl, 1.5% sodium dodecyl sulfate, and 0.5 mg/mL proteinase K (Thermo Fisher). Proteins were digested for 45 min at 23 °C with intermittent mixing.

E.coli cell lysis and protein digestion

A 25-mL aliquot of E. coli cells at an OD₆₀₀ of 0.5 were pelleted at 8000 ×g, 4 °C for 10 min. Cells were lysed in 16.5 mL E. coli lysis buffer [15 mM Tris, pH 8, 450 mM sucrose, 8 mM EDTA, 0.4 mg/mL lysozyme] for 5 min at 23 °C, then 10 min at 0 °C. Protoplasts were collected at 5000 ×g and then resuspended in 2 mL proteinase K buffer [50 mM HEPES, pH 8, 200 mM NaCl, 5 mM MgCl₂, 1.5% SDS, 0.2 mg/mL proteinase K], vortexed for 10 s, then incubated at 23 °C for 5 min and 0 °C for 10 min.

RNA extraction and treatment with 1M7, 5NIA, or NAI

Nucleic acids were extracted twice with 1 volume of phenol:chloroform:isoamyl alcohol (25:24:1) that was pre-equilibrated with either 1.1× mammalian RNA folding buffer [110 mM HEPES, pH 8, 110 mM NaCl, 5.5 mM MgCl₂] or 1.1 × E. coli RNA folding buffer [110 mM HEPES, pH 8, 110 mM NaCl, 11 mM MgCl₂]. Excess phenol was removed through two subsequent extractions with 1 volume chloroform. The final aqueous layer was buffer exchanged into 1.1 × RNA folding buffer using PD-10 desalting columns (GE Healthcare Life Sciences). The resulting RNA solution (7 mL for mammalian cells, 3.5 mL for E. coli) was incubated at 37 °C for 20 min before being split into two equal volumes. To mammalian samples, 350 μL of 10× SHAPE reagent (250 mM 5NIA, 100 mM 1M7, or 1 M NAI) in DMSO was added to one sample, and 350 μL of neat DMSO was added to the other. For E. coli samples, the volume added was 175 μL. Samples were incubated at 37 °C for 10 min with 5NIA (CAS: 4693-02-1; AstaTech catalog 69445) and NAI and for 5 min with 1M7 before quenching the reagent with 1 M DTT (550 μL for mammalian cell samples and 225 μL for E. coli samples). RNA was precipitated with 1/10 volume of 2 M NH₄OAc and 1 volume of isopropanol. After one wash with 75% ethanol, the resulting pellet was dried and resuspended in 88 μL of water and 10 μL of 10× TURBO DNase buffer and 4 units of DNase (TURBO DNase, Thermo Fisher) were added. The mixture was incubated at 37 °C for 1 h. RNA was purified using a 1.8× ratio of RNA-binding magnetic beads (Agencourt RNAClean XP, Beckman Coulter) and eluted into 20 μL of nuclease-free water.

In-cell SHAPE treatment and RNA extraction

Adherent cells were grown in 3-cm culture plates as described above. At ~80% confluency, cells were washed with PBS and replenished with 2.7 mL of fresh media. To the cells was added 300 μL of SHAPE reagent in DMSO; the same volume of DMSO was added to the control cells. Jurkat cells were grown to ~80% confluency and pelleted at 3000 ×g for 5 min, followed by one wash with PBS and resuspension in fresh growth media. A 900-μ aliquot of this mixture was transferred to a well (35-mm diameter) of a six-well culture plate and incubated at 37 °C. After 10 min, 100 μL of SHAPE reagent (1M7, 1M6 and NMIA at 100 mM; 5NIA at 250 mM; NAI at 1 M) in DMSO was added to the well; controls were treated with DMSO. Cells were incubated at 37 °C for 15 min. Cells were either quenched at 15 min using 125 mM DTT or left unquenched. Cells were pelleted and washed once with PBS, followed by RNA extraction (TRIzol, Invitrogen). RNA pellets were dried and resuspended in 88 μL nuclease-free water. Treatment with DNase (TURBO DNase, Thermo Fisher) and affinity purification (Agencourt RNAClean XP magnetic beads) was performed as described above for cell-free RNA extraction.

MaP of total RNA

From each sample of RNA (defined above, E. coli or murine total RNA, extracted and SHAPEmodified or untreated), 1-3 μg were subjected to MaP reverse transcription [requiring Superscript II and addition of Mn²⁺ to the RT buffer(25, 39)] using random nonamer primers. The cDNA generated was buffer exchanged (Illustra microspin G-50 columns, GE Healthcare) and the volume increased to 68 μL. For second-strand cDNA synthesis, 8 μL of 10× buffer (Second Strand Synthesis Reaction Buffer, NEB) and 4 μL enzyme (Second Strand Synthesis Enzyme mix, NEB) were added to the cDNA product and incubated for 2.5 h at 16 °C. Double-stranded cDNA was fragmented and amplified with sequencing indexes (Nextera XT library prep kit, Illumina). Nextera PCR products were affinity purified (using a 0.8× ratio of Agencourt AMPure XP beads, Beckman Coulter) and eluted in 20 μL of nuclease-free water.

U1 snRNA MaP

From each sample of eukaryotic RNA, 1-3 μg were subjected to MaP reverse transcription [requiring Superscript II and addition of Mn²⁺ to the RT buffer(25, 39)] using primers specific for either mouse or human U1 RNA. The cDNA generated was buffer exchanged (over Illustra microspin G-50 columns, GE Healthcare). Output cDNA (5 μL) was used as a template for 25 μL PCR reactions (Q5 Hot-start polymerase, NEB) with primers made to amplify U1 and add adapter sequences [1× Q5 reaction buffer, 500 nM each primer, 200 μM dNTPs, 0.02 units/μL Q5 Hot-start polymerase].Step 2 PCR proceeded as follows: 98 °C for 2 min, 20 cycles of [98 °C for 10 s, 72 °C (decreasing by 1 °C each cycle until 64 °C) for 30 s, 72 °C for 20 s], and 72 °C for 2 min. Step 1 PCR products were affinity purified (using a 1× ratio of Agencourt AMPure XP beads, Beckman Coulter) and eluted in 20 μL of nuclease-free water. Purified PCR products (2 ng) were used as a template in 50 μL of PCR to add multiplex indices and remaining sequences necessary for Illumina sequencing [1× Q5 reaction buffer, 500 nM each index primer, 200 μM dNTPs, 0.02 units/μ Q5 Hot-start polymerase]. Step 2 PCR proceeded as follows: 98 °C for 2 min, 10 cycles of [98 °C for 10 s, 66 °C for 30 s, 72 °C for 20 s], and 72 °C for 2 min. Step 2 PCR products were affinity purified (using a 0.8× ratio of Agencourt AMPure XP beads, Beckman Coulter) and eluted in 20 μL of nuclease-free water.

Illumina sequencing

Libraries were pooled and sequenced on an Illumina Miseq instrument, outputting 2X300 paired-end datasets. For U1 libraries, 2×150 paired-end datasets were used.

MaP analysis of highly expressed mouse transcripts

Initial transcript target sequences were retrieved from Gencode release M15 (GRCm38.p5)(40); there were 131,100 transcripts. Transcripts corresponding to a single parent gene were limited to the first listed transcript (in most cases the best-supported spliced variant), giving 52,553 transcripts. The transcript for Gapdh (ENSMUST00000118875.7, Gapdh-203) was manually chosen based on read counts mapped to each exon. Missing sequences for the 5.8S, 18S, and 28S rRNAs (accession numbers NR_003280.2, NR_003278.3, and NR_003279.1, respectively) were appended. Experimental reads were pseudomapped against these transcripts with Kallisto(41). Transcripts were selected with at least 1000 pseudomapping reads (including reads mapping to multiple loci), yielding 77 transcripts. Reads were filtered to only those pseudomapping to any of these 77 targets. The ShapeMapper pipeline (v2.1.2) was run on these filtered reads against each of the 77 target transcripts separately (this includes a mapping step using Bowtie2(42)). Transcripts with coverage greater than 1000 reads over at least half of the sequence in both plus reagent and minus reagent samples were retained, yielding 13 transcripts. Two exactly duplicated transcript sequences were removed. Pairwise sequence alignment was performed for the 11 remaining targets using Needle(43) with default parameters. Hierarchical clustering was performed in python using the scipy.cluster.hierarchy package based on [1 - (alignment score/alignment length)], with the resulting distribution of scores scaled from 0 to 1. Linkage was calculated with hierarchy.linkage with method=‘average’. Clusters were assigned with hierarchy.fclusterwith criterion=‘distance’and a threshold of 1. Each cluster of highly similar sequences was reduced to the single best-annotated transcript, first eliminating sequences whose GENCODE information indicated no human curation, then by eliminating all but the longest sequence in a given cluster. This process eliminated two apparent 7SK sequences, one apparent 18S rRNA sequence, and a shortened sequence highly similar to CT010467.1, yielding seven final transcript sequences. Finally, a single ShapeMapper analysis run was performed with the pseudomapping reads as input and these seven transcripts as targets.

Analysis of RT-truncation data for highly expressed mouse transcripts

Sequence reads for RT-truncation datasets were retrieved from the NCBI Sequence Read Archive accessions SRR1534952, SRR1534953, SRR1534954, and SRR1534955(24). Reads were pseudomapped against seven transcripts of interest using Kallisto(41), and any reads that mapped were used as inputs for a published RT-truncation analysis pipeline (icSHAPE_pipeline.pl)(44)). Calculated reactivity values below 0 and above the 95th percentile were not clipped. Analysis of one transcript did not yield any usable data, possibly due to its short length (57 nt); this transcript is therefore not included in Table S1.

Read depth

Across all MaP datasets reported here, for a given position to be included in analysis, read depth was required to be greater than 5000 in both untreated and SHAPE-modified samples.

Mutation detection and adduct inference

A significant part of the information contained in a MaP experiment lies in insertions, deletions, and sequence changes more complex than simple mutations relative to the reference sequence, and it is important to account for this signal accurately. ShapeMapper2 (35) locally realigns ambiguous sequence deletions and inserts such that a deletion or insert is placed at the 5′-most position out of the set of compatible alignment locations. Mutations separated by five or fewer unchanged reference positions are grouped and treated as arising from a single chemical adduct. For each group of mutations, adduct location is inferred to be the reference position 5′ of the unchanged reference position 3′ of the mutation group. See Fig. 1D and Fig. S2 in ref. (35) for examples, implementation details are provided in the software documentation.

Figure 1. SHAPE data read out by MaP or one-step electrophoresis-detected reverse transcription.

Histograms show normalized SHAPE reactivities, using 1M7, as a function of detection approach. Dots below the nucleotide sequence indicate unpaired nucleotides according to accepted structure models. Inset: Receiver operator characteristic (ROC) curves for reactivity profiles compared to base pairing status. Sweeping a threshold over all ranked reactivity values, true positive rate (TPR) is calculated as the number of unpaired nucleotides with reactivities above a given threshold divided by the total number of unpaired nucleotides, and the false positive rate (FPR) is calculated as the number of base-paired nucleotides with reactivities above a given threshold divided by the total number of base-paired nucleotides. Areas under ROC curve for the TPP riboswitch are 0.86 and 0.82 for MaP and electrophoresis, respectively; for tRNA-Phe, areas are 0.82 and 0.86 for MaP and electrophoresis, respectively. Reactivities read out by MaP are shown on the accepted secondary structure models. It is expected that some nucleotides in regions shown as single stranded in the secondary structure are unreactive, as these regions are constrained by non-canonical and tertiary interactions. MaP(25) and electrophoresis data(45) were reported previously. Both RNA elements are based on E. coli sequences, and were transcribed in vitro and refolded, prior to probing.

Reference structures

Accepted secondary structure models for an E. coli TPP riboswitch and tRNA-Phe were based on crystallographic structures, as previously reported(45). Structures for E. coli ribosomal RNAs (both large and small subunits, including pseudoknots, corresponding to GenBank sequence accession J01695), and for Mus musculus 18S rRNA and 12S mitochondrial rRNA were retrieved from the Comparative RNA Website(46). A structure for Mus musculus 7SK snRNA was retrieved from Ensembl for transcript ENSMUST00000083103(47). The secondary structure for Homo sapiens U1 was based on a crystal structure(48).

Results

MaP accuracy

A critical criterion for any structure probing strategy is that chemical adducts created on the RNA are reported accurately and quantitatively. There is close agreement between previous direct, one-step electrophoresis-based primer extension readouts and the massively parallel sequencing-based MaP strategy (Figure 1). Prior work includes extensive comparisons between electrophoresis readouts and MaP, based on analyses of long and complex transcripts such as bacterial ribosomal RNAs and diverse smaller structured RNAs, and on benchmarking analyses that show that MaP reads out SHAPE experiments with the same or greater accuracy as simpler (but low throughput) methods(25, 35).

Detection accuracy by MaP is closely linked both to performing the experiment under the right conditions and to the design and quality of the algorithms used to call adduct-induced sequence changes. MaP conditions have been carefully optimized for detection of SHAPE adducts(25, 39). Some prior reports that have attempted to evaluate MaP have used ad hoc enzymes and reverse transcription conditions unlikely to produce optimal signal over background(49, 50). Both point mutations (which are relatively easy to detect) and more complex deletions, insertions, and other sequence changes contribute to MaP signal. Indeed, approximately 70% of the MaP signal is due to these more complex features (see Fig. 1C in reference (35)). Some software packages for reading out MaP experiments have ignored these components of the signal(49, 51), potentially raising sequencing coverage requirements for accurate readout or leading to systematic errors in SHAPE adduct detection and quantification. Currently, ShapeMapper2 (https://github.com/Weeks-UNC/shapemapper2) is one of the few software packages that carefully accounts for deletions and complex multi-nucleotide mutations (see Methods) and integrates this signal into SHAPE adduct detection (35).

A variety of SHAPE reagents are used in RNA structure probing experiments. All five of the examined reagents, 1M6, 1M7, NMIA, NAI, and 5NIA, report on base-pairing with comparable accuracies when read out by MaP, as shown by similar areas under the receiver operator characteristic (ROC) curve calculated with respect to unpaired versus paired nucleotide classification in accepted structures of E. coli rRNA and human U1 snRNA (Figure 2). An area under the curve of 1 indicates perfect agreement between nucleotide reactivity and pairing status, and a value of 0.5 indicates that nucleotide reactivity reports no base pairing information. SHAPE reactivities are primarily measurements of nucleotide conformational flexibility. Base pairing is a major, but not the only, contributor to nucleotide conformation and flexibility; therefore, SHAPE reactivities show close but not perfect agreement with base pairing. When coupled with best-practices thermodynamic pseudo-free energy change minimization strategies(45,52), data obtained from probing cell-free E. coli rRNA with each of the five reagents, read out with MaP and analyzed with ShapeMapper2, resulted in accurate RNA secondary structure models (Figure 3).

Figure 2. MaP readout for SHAPE reagents compared to base-pairing status.

(A) ROC curves for MaP data obtained from E. coli 16S and 23S rRNAs probed with five different SHAPE reagents under protein-free conditions. True positive rate and false positive rate calculated as in Figure 1, with respect to base pairs in accepted CRW structures(46). Inset: Area under the ROC curve with respect to basepairing status. This analysis directly compares SHAPE-MaP reactivity profiles and accepted structures and does not involve SHAPE-directed structure modeling. (B) 1M7 reactivities shown on a representative region of the accepted 23S rRNA structure. (C) ROC curves for MaP data from the U1 snRNA isolated from Jurkat cells and probed with four SHAPE reagents under protein-free conditions. (D) 1M7 reactivities shown on the accepted U1 secondary structure. Data for 1M6, 1M7, and NMIA-probed rRNAs were reported previously(34). See Methods for details on other datasets.

Figure 3. SHAPE-guided structure modeling accuracies for cell-free E. coli rRNA.

Plots of sensitivity and positive predictive value (PPV) calculated by comparison of secondary structures modeled using data from the indicated SHAPE reagent to the accepted structure(46). Structures were modeled using the Fold module of RNAstructure(52) as described(60, 45, 25). For scoring, non-canonical base pairs and reference pairs >600 nucleotides apart in primary sequence were excluded. Sensitivity: the number of pairs shared between the accepted and modeled structures (allowing offsets of ±1 nucleotide) divided by the number of pairs in the accepted structure. PPV: the number of pairs in the modeled structure also present in the accepted structure (allowing offsets of ±1 nucleotide) divided by the number of pairs in the modeled structure. Regions in which SHAPE reactivities do not support the formation of specific helices(60, 45, 25) were excluded.

Reagent biases

1M7 was identified as an especially useful general purpose SHAPE reagent because of its short, but manageable, reaction half-life and its ability to report on the flexibility of all four ribonucleotides with similar reactivity(53). We examined the per-nucleotide reactivity of 1M7 relative to background using a >4000 nt dataset. Per-nucleotide 1M7 reactivities (detected as MaP mutation rates) for cell-free E. coli ribosomal RNA confirmed the low nucleotide bias of this reagent (Figures 4, top). There is clear separation between the distributions of unpaired (red) and paired (blue) mutation rates, indicating that 1M7 reacts preferentially with unpaired nucleotides with very little bias for a particular nucleotide. We next evaluated whether the other four reagents had nucleotide biases (Figures 4 and S1). NAI showed a markedly different pattern of MaP mutation rates from that of 1M7. NAI effectively distinguished unpaired and paired adenosine residues but less effectively distinguished unpaired from paired guanosine and cytosine (Figures 4, middle, and S2). We suggest that the bias observed with NAI primarily reflects differences in nucleotide reactivity, but adduct detection rates and adduct reversal (54) may contribute. 5NIA showed excellent discrimination for all four ribonucleotides but reacted with adenosine at high relative rates compared to the other three nucleotides (Figures 4, bottom row, and S2). Because adenosine residues are over-reactive with 5NIA, we recommend rescaling 5NIA SHAPE reactivity profiles on a per-nucleotide basis using factors estimated from the median background-subtracted mutation rates of unpaired nucleotides (Figure S2). Rescaling facilitates visual interpretation of reactivity profiles and yields structure models of equivalent accuracy for the rRNAs examined here. Consistent with prior work(55, 34), choice of reagent can have large effect on reactivity profile.

Figure 4. Nucleotide-specific base pair discrimination for various SHAPE reagents.

Plots of mutation rates from SHAPE-treated versus background samples as calculated by ShapeMapper2(35). For a given RNA position, mutation rate is calculated as mutation count divided by local read depth. Data shown correspond to protein-free E. coli ribosomal RNA (both large and small subunits, same data as Figure 2A). Contour lines show smoothed density, similar to a histogram. Data for the 1M6 and NMIA reagents are reported in Figure S1; distributions of per-nucleotide reactivity for all reagents are shown in Figure S2.

MaP versus click-selective RT-truncation

We performed SHAPE on total RNA extracted from mouse B-lymphocytes, probing with the NAI reagent and reading out adducts with MaP. We compared this profile with a previously published NAIN3 click-selective RT-truncation dataset of poly-A enriched extracted mouse RNA(24) (see Methods). SHAPE data read out by MaP and by click-selective RT-truncation generally showed poor to modest correlations (Spearman r values spanned 0 to 0.5; Table S1). The observed low correlation is consistent with a study comparing truncated and internally mutated cDNA products generated by RT of DMS-modified RNA under conditions not optimal for MaP(50). Despite only modest correlation, for transcripts with well-determined secondary structures, both readout methods showed comparable agreement with accepted base-pairing models, based on AUC analyses for the few RNAs for which overlapping datasets are available. In general, very few RNAs with known structures have been examined with good read depth in their entirety using RT-truncation methods, limiting this analysis.

SHAPE modification in cells

Mutation rates obtained by SHAPE-MaP probing in cells are generally lower than from probing cell-free RNAs, in agreement with previous reports(38, 37, 56). This difference varies by reagent and cell line (Figure 5). The modification efficiency is likely affected by reagent half-life, reagent diffusion rate, and cell permeability. A few general trends are clear. All reagents evaluated here modify RNA in cells, based on a single reagent treatment, in the context of conscientiously performed experiments. Reactivity above background was clearly detected using a MaP readout for all reagents. 5NIA and NAI are more effective than 1M7 at modifying RNA in cells for all examined cell lines. Adherent cell lines (C2C12, H4SV, ESC, and LNCaP in Figure 5) typically had lower modification efficiencies with all reagents. Extracellular matrix or a reduced available cell surface area may reduce in-cell SHAPE modification. Conversely, cells that grow in suspension (Jurkat) were readily modified by all reagents (Figure 5).

Figure 5. SHAPE reagent performance in live cells.

SHAPE-MaP reactivities for U1 snRNA probed in total RNA isolated under non-denaturing conditions (cell free) and in cells. U1 snRNA reactivities were determined by MaP using gene-specific primers(39) (see Methods). Box plots span the central 50% of the data, the interquartile range [IQR, from quartile 1 (Q1) to quartile 3 (Q3)]; center line indicates median; whiskers indicate the most extreme points within (Q1 – 1.5 × IQR) to (Q2 + 1.5 × IQR).

Discussion

Experimental considerations

The design of an RNA structure probing experiment requires weighing the following criteria: The reagent should react readily under structurally informative or physiologically relevant conditions. Results from a probing experiment should be repeatable and tolerant of slight changes in conditions, and the procedure should be concise and contain no unnecessary steps. The experiment should accurately report on base pairing or other structural information and should ideally report on all four nucleotides with similar reactivity and high signal above background regardless of local sequence context. Importantly, data for evaluating these criteria should be obtained from carefully executed experiments with equivalent sequencing coverage across experiments and attentive adherence to published methods.

For RNA probing and structure analysis and modeling investigations focusing on in vitro transcribed and cell-free extracted RNAs, SHAPE experiments using 1M7 modification and MaP readout broadly satisfy these criteria. 1M7 remains notable for its relatively fast, but manageable(53), and unbiased reactivity toward all four nucleotides (Figure 4). SHAPE probing in living cells requires a nuanced balancing of the strengths and weaknesses of each reagent(56). Two broad trends are apparent from this and prior work. First, bacterial cells(57, 58, 30) and eukaryotic cells grown in suspension culture (for example, Jurkat cells; Figure 5) show higher levels of reactivity for all reagents. Second, longer-lived reagents like NAI and 5NIA exhibit better in-cell reactivity than faster-reacting reagents; however, these reagents have nucleotide biases, and for NAI a quench step is necessary. Individual investigators and laboratory teams will balance these factors in different ways, but we offer the following recommendations (Table 1):

Table 1. SHAPE reagent properties.

Half-life refers to reagent half-life in aqueous solution at pH 8.

Short name	Name and structure	Solubility	Half-life at 37 °C	Treatment concentration (mM)	Notes
1M7	1-methyl-7-nitroisatoic anhydride	modest	17 s	10	General purpose reagent for cell-free studies. Very even per-nucletide reactivity.
1M6	1-methyl-6-nitroisatoic anhydride	modest	31 s	10
NMIA	N-methylisatoic anhydride	modest	260 s	10	Readily commercially available. Not recommended for in-cellprobing.
NAI, NAI-N3	2-methylnicotinic acid imidazolide[-azide]	high	~30 min	100-200	High modification rate in cells.Bias against guanosine and cytidine. Requires specific quench step.
5NIA	5-nitroisatoic anhydride	modest	~100 s	25	Fast-acting. Self-quenching.Highly reactive with adenosine.Useful for in-cell probing.

• 1M7 or 1M6 for general use, especially for experiments performed under cell-free conditions.

• NMIA for general use if 1M7 or 1M6 are not readily available; however, reaction times will be extended, increasing the risk of RNA degradation.

• 5NIA for in-cell probing, with rescaling by per-nucleotide relative reactivity factors (Figure S2).

• NAI for in-cell experiments where reactivities with alternative reagents are very low. The long half-life of NAI facilitates diffusion into cells, but the requisite long reaction time means that RNAs likely turn over many times during the experiment. This reagent has strong per-nucleotide reactivity variation (Figure 4) and must be quenched. TRIzol treatment does not inactivate NAI (Figure S3).

Adduct detection strategy

The experimental approaches and technical requirements for MaP versus RT-truncation strategies (either with or without click chemistry enrichment) differ substantially in the number of experimental steps, amount of experimental and hands-on time, and diversity of techniques and supplies required. Both methods share some procedures, including in-cell chemical probing, RNA purification, library amplification, and sequencing. An entire MaP procedure, starting with live cells and including RT and all steps shared with click-selective RT-truncation, can be performed in 1-2 days. Click-selective RT-truncation requires additional biochemical steps including a copper-free click reaction, RNA fragmentation and end repair, adapter ligation, gel-based size selection of RNAs, streptavidin pulldown of RNA-cDNA hybrids, and further gel-based size selection of eluted cDNAs. These steps require an additional 2-3 days and may introduce readout biases. Likely due in part to these extensive differences in experimental steps, there is a low to modest quantitative correlation between experiments read out by MaP and those read out by RT-truncation (Table S1) and ref. (36)). We did find that, despite the low per-nucleotide correlation, AUC analysis revealed similar discrimination for unpaired nucleotides in the small subunit ribosomal RNA (Table S1).

Some studies have proposed that MaP and RT-truncation readouts provide complementary information(49,50). However, prior comparative analyses employed RT conditions and enzymes not optimized for MaP, very low read depth cutoffs, and (in specific cases) software that ignored deletions (49) and presentation formats that compressed data points (50). These factors introduce noise and sampling bias and impact interpretability. Moreover, the quantitative accuracy of MaP would benefit only if RT-truncations were to show substantially different patterns than mutations after accounting for convolution by sequencing library preparation steps, and if their inclusion improved correspondence with a validated external metric.

We recommend MaP over RT-truncation as a readout method because of its extensive validation based on long RNAs with complex structures(25, 35, 39), its much simpler and more concise experimental implementation, and the ability to directly examine rare cellular transcripts by target-specific RT-PCR without requiring additional steps(27, 56). The MaP readout works well for all of the reagents evaluated in this work. MaP also provides an intuitive within-sample divisor (read depth) for computing relative adduct frequencies, enabling direct comparison between reactivities of different transcripts probed in the same experiment(30), without requiring use of proxy metrics such as the Gini index(59). For example, with a MaP readout, coding RNAs can be directly compared and shown to have higher SHAPE chemical modification rates than non-coding RNAs in E. coli, suggesting that they are less structured overall(30).

Areas for future development

SHAPE-MaP is now a mature technology, but there is room for improvement in several areas. If SHAPE reagent solubility in water could be enhanced without substantially lengthening the reaction half-life, such reagents could enable higher signal above background. SHAPE reagents such as NAI enable signal enrichment in cell types in which reactivities of 1M7 and 1M6 are low, but show nucleotide biases that might be reduced through further chemical exploration. 1M7 and 5NIA show excellent overall discrimination of most unpaired and paired nucleotides; however, discrimination at cytidines is not at the theoretical limit (Figures 4 and S2) and could potentially be improved by increasing adduct formation or detection rates at these nucleotides. MaP reverse transcription enzyme and reaction conditions have been optimized for adduct readthrough and signal above background(17), but further reverse transcriptase design or screening could identify enzymes or conditions with improved efficiencies or signal-to-noise detection rates.

In summary, MaP is now a mature and well-validated strategy for RNA secondary structure probing with diverse SHAPE reagents both in vitro and in cells. The accuracy of MaP-based approaches for reading out the results of the chemical modification step matches that of direct one-step electrophoresis approaches. Compared to RT-truncation strategies, the MaP strategy is much simpler, less time-consuming, and requires less initial material. The reagents examined here in cell-free experiments, 1M6, 1M7, NMIA, NAI, and 5NIA, report on RNA flexibility and enable accurate modeling of RNA secondary structure. For live cell probing, suspension cultures are readily amenable to treatment with many SHAPE reagents, whereas some adherent cell cultures may require longer halflife reagents to enhance signal above background. However, tradeoffs displayed by longer-lived reagents, such as nucleotide biases and the potential to miss shorter time-scale information, should be weighed when designing an experiment. We think the potential of comprehensive, nucleotide resolution probing of RNA structure has just begun to be realized. We expect that there will be many creative applications that will result in deeper understandings of the wide-ranging roles of RNA structure in governing information transfer in diverse biological systems.