Mutation signature filtering enables high-fidelity RNA structure probing at all four nucleobases with DMS

Abstract Chemical probing experiments have transformed RNA structure analysis, enabling high-throughput measurement of base-pairing in living cells. Dimethyl sulfate (DMS) is one of the most widely used structure probing reagents and has played a pivotal role in enabling next-generation single-molecule probing analyses. However, DMS has traditionally only been able to probe adenine and cytosine nucleobases. We previously showed that, using appropriate conditions, DMS can also be used to interrogate base-pairing of uracil and guanines in vitro at reduced accuracy. However, DMS remained unable to informatively probe guanines in cells. Here, we develop an improved DMS mutational profiling (MaP) strategy that leverages the unique mutational signature of N1-methylguanine DMS modifications to enable high-fidelity structure probing at all four nucleotides, including in cells. Using information theory, we show that four-base DMS reactivities convey greater structural information than current two-base DMS and SHAPE probing strategies. Four-base DMS experiments further enable improved direct base-pair detection by single-molecule PAIR analysis, and ultimately support RNA structure modeling at superior accuracy. Four-base DMS probing experiments are straightforward to perform and will broadly facilitate improved RNA structural analysis in living cells.


INTRODUCTION
RNA molecules fold into complex base-paired secondary and tertiary structures that play critical roles in RNA function ( 1 ).One of the oldest and most scalable methods for interro gating RN A structur e ar e chemical probing experiments (2)(3)(4).These experiments, which can be performed both in vitro and in cells, use small molecule chemical reagents to selecti v ely modify fle xib le nucleotides, yielding per-nucleotide reactivity measurements that report on local RNA structure.Probing data are useful as standalone measurements and can also be used to guide structure modeling algorithms to reconstruct global RNA structure ( 5 , 6 ).Recently, we and others have introduced singlemolecule chemical probing experiments as an even more powerful strategy for characterizing complex RNA systems ( 7 ).By measuring correla ted modifica tion e v ents, single molecule probing experiments enable direct measurement of secondary base-pairing ( 8 , 9 ), tertiary interactions ( 10 ), and ensembles comprising multiple structural states (11)(12)(13)(14).Both traditional and single-molecule experiments have provided critical insights into RN A biolo gy (15)(16)(17)(18)(19)(20)(21)(22), and increasingly serve as foundational technologies for RNA functional characterization and therapeutic de v elopment ( 19 , 23-26 ).
Numerous strategies utilizing di v erse chemical reagents hav e been de v eloped to probe different aspects of RNA structure ( 4 ).The oldest and still one of the most used probing reagents is dimethyl sulfate (DMS) ( 27 ).DMS methyla tes the Wa tson-Crick face of single-stranded adenine (A) and cytosine (C) nucleobases at the N1 and N3 positions, respecti v el y.DMS is readil y cell permeable and can be used to achie v e high le v els of modification, which is essential for single-molecule probing applications.Howe v er, the inability of DMS to probe uracil (U) and guanine (G) bases has r epr esented a key limitation.We r ecently showed that mildly alkaline conditions enable DMS to effectively probe U structure ( 8 ).These conditions also support probing of G structure in vitro at lower but still useful accuracy.Howe v er, DMS remained unable to probe G structure in cells.Probing G is particularly desirable because of the outsized role G nucleotides play in stabilizing RNA structure.Alternati v e chemical reagents permit probing of G and U (28)(29)(30), and SHAPE (selecti v e 2 -hydroxyl acylation analyzed by primer extension) reagents support probing of all four bases (31)(32)(33)(34), but to date these reagents have proven less amenable for single-molecule analyses.Enabling robust DMS probing at all four bases has the potential to br oadly impr ove both pernucleotide and single-molecule RNA structural analysis.
A seminal advance in chemical probing technology has been the de v elopment of mutational profiling (MaP) ( 10 , 35 ).MaP uses specialized re v erse transcription protocols to read through and encode chemical modifications as mutations in cDNA, which can then be measured via highthroughput sequencing.Compared to alternati v e strategies, MaP permits precise quantitation of e v en rare modification e v ents.Critically, MaP also permits measurement of multiple modifications per individual molecule, enabling detection of the correlated modification e v ents that underpin single-molecule probing strategies ( 7 ).MaP may also enable distinguishment between different chemical modifications tha t genera te distinct muta tional signa tures, although this application is relati v ely une xplored.
In recent years, multiple re v erse transcriptase enzymes have been adapted f or perf orming DMS-MaP experiments.G-and U-sensiti v e DMS probing and most single-molecule probing analyses have relied on a relaxed-fidelity Superscript II (SSII) re v erse-transcription protocol ( 8 , 10 , 11 ).Newer MaP protocols that use T GIRT-III (T GIRT) and Mar athonRT (Mar athon) re v erse transcriptase hav e been reported to be more processi v e and e xhibit lower backgr ound err or rates (36)(37)(38).Howe v er, these enzymes hav e not been evaluated for their ability to measure DMS modifica tions a t G and U nucleotides, nor for their ability to support direct base pair detection via single-molecule correlation analysis.
In this work, we sought to evaluate different MaP protocols for their ability to support DMS probing at all four bases.Strikingl y, our anal yses re v ealed that re v erse transcriptases decode DMS-induced N 1 -methylguanine and N 7 -methylguanine chemical modifications via distinct muta tional signa tures.Le v eraging this discov ery, we de v elop a new strategy that solves prior limitations to enable highfidelity DMS probing at all four nucleotides in cells.Four-base DMS-MaP conveys more information than other compar able structur al probing experiments, facilitates more sensiti v e single-molecule analysis, and ultimately enables improved RNA structural modeling, r epr esenting an all-inone strategy for complete RNA structural analysis.

Probing of HEK293 cells
DMS probing.Human HEK293 cells (ATCC; CRL-1573) were maintained in DMEM supplemented with 10% FBS and 100 U / ml Pen / Strep at 37 • C and 5% CO 2 .∼1 × 10 6 cells were seeded in a 6 cm culture dish and grown to ∼75% confluency.Prior to DMS treatment, media was exchanged with 2.8 ml fresh media, followed by addition of 800 l of 1 M bicine (pH 8.3 a t room tempera ture), 1 M sodium cacod yla te (pH 7.2), or nuclease-free water, and equilibrated for 3 min at 22 • C. Cells were modified by adding 400 l of 1.7 M DMS solution in ethanol (or 100% ethanol for control reactions) and incubating at 37 • C for 6 min.Reactions were quenched by addition of 4 ml ice-cold 20% 2mercaptoethanol (v ol / v ol in PBS).Cells were scraped and pelleted by centrifugation at 1000 g for 5 min at 4 • C, followed by RNA extraction with 1 ml TRIzol reagent (Ther-moFisher).Genomic DNA was removed by addition of 6 U of TURBO DNase (Ambion) and incubation at 37 • C for 45 min.RNA then was purified (RNA Clean & Concentrator, Zymo), quality assessed by Ta peStation anal ysis (Agilent), and concentration quantified by UV absorbance (NanoDrop, ThermoFisher).2A3 probing.Probing was done as described by Marinus et al. ( 34 ).HEK293 cells were maintained as described above.Prior to 2A3 treatment, media was removed, cells washed with 1 × PBS, followed by addition of 1 ml trypsin and 2 min incubation at 37 • C to detach cells from the dish.Trypsin was neutralized by 2 ml of media and cells pelleted by centrifuga tion a t 1000 g for 5 min a t 22 • C .The cell pellet was resuspended in 90 l of PBS. 10 l of 1 M 2A3 (Tocris Bioscience) was added to the cells followed by 15 min incuba tion a t 37 • C with occasional tapping to mix.2A3 reactions were quenched by addition of 100 ml of ice-cold 20% 2-merca ptoethanol.RN A then was extracted as described above.

DMS probing of E. coli RNA
Cell-fr ee experiments .Cell-free DMS probing of E. coli K-12 MG1655 total RNA was performed as previously described ( 8 ).148 ml LB was inoculated with 2 ml of an overnight culture and grown at 37 • C until OD 600 ≈0.5.16.65 ml of 187.5 g / ml rifampicin was added followed by incubation for 20 min at 37 • C to chase assembly of RNAprotein complexes ( 39 ).Cells were pelleted and resuspended in 32 ml of lysis buffer [15 mM Tris-HCl (pH 8), 450 mM sucrose, 8 mM EDTA (pH 8)], followed by addition of 1.28 ml of 1 mg / ml lysozyme, and 10 min incubation on ice.Total RNA was extracted by  C, 1 volume of 1.7 M DMS solution in ethanol (or 1 volume of 100% ethanol for control reactions) was added to 9 volumes of total RNA and reacted for 6 min a t 37 • C .Ten volumes of ice-cold 20% 2-mercaptoethanol was added to quench reactions, followed by purification (RNeasy Midi, Qiagen), DNase treatment (TURBO DNase, ThermoFisher), and a final purification (RNeasy Midi, Qiagen).RNA was then quantified for quality (TapeStation, Agilent) and concentration (Nan-oDrop, ThermoFisher).
In-cell experiments.In-cell DMS probing of E. coli total RNA was performed as described ( 8 ).Cells were grown and treated with rifampicin as described for cell-free experiments, then pelleted by centrifugation at 4000 g for 5 min at 22 all pr epar ed at room temperature).4.5 ml of cells was added to 0.5 ml of 7 M DMS in ethanol (or 0.5 ml 100% ethanol for control reactions) and reacted for 6 min at 37 • C. DMS reactions were quenched by addition of 20 ml of ice-cold 20% 2-mercaptoethanol and the cells placed on ice.Cells were pelleted, resuspended in 1 ml of 1mg / ml lysozyme, and incubated on ice for 5 min.Total RNA was extracted using TRIzol reagent, DNase treated (TURBO DNase, Invitrogen), and then purified (RNeasy Midi, Qiagen).

Reverse transcription
Four different mutational profiling (MaP) re v erse transcription (RT) protocols were evaluated, which we refer to by the RT enzyme used: Superscript II (SSII) ( 8 ), MarathonRT (Marathon) ( 37 ), T GIRT-III (T GIRT) ( 36 ) and e volv ed HIV RT (eHIV) ( 40 ).Gene-specific priming was used for human RNase P and RMRP, and E. coli tmRNA, and random priming was used for ribosomal RNAs (Supplementary Table S3).2A3-probed RMRP and RNase P samples wer e r e v erse transcribed using the SSII protocol.All RT reactions were purified using magnetic beads (Mag-Bind Total Pure NGS, Omega Bio-Tek).
SSII. 1 l of 10 mM dNTPs and 1 l of either 2 M specific primer or 200 ng / l random 9-mer was added to 1-2 g RNA in a total volume of 10 l and incubated at 65 • C for 10 min followed by 4 • C for 2 min.Subsequently, 9 l of SSII MaP buffer [final concentration 50 mM Tris-HCl (pH 8), 75 mM KCl, 6 mM MnCl 2 , 1 M betaine, 10 mM DTT] was added to the solution, followed by 1 l of SSII (Invitrogen).The reaction was then incubated according to the temper ature progr am: 25  C for 10 min.Some r eactions wer e also performed using 2 l of 10 mM dNTPs, with a total step 1 volume of 11 l, which had no appreciable impact on results.
Marathon. 1 or 2 l of 10 mM dNTPs and 1 l of either 2 M specific primer or 200 ng / l random 9-mer was added to 1-2 g extracted RNA in a total volume of 5.8 l and incuba ted a t 65 • C f or 10 min f ollowed by 4 • C f or 2 min.Afterward, 12.2 l of Marathon MaP buffer [final concentration 50 mM Tris-HCl (pH 8.3), 200 mM KCl, 5 mM DTT, 1 mM MnCl 2 , 20% glycerol] and 2 l of Marathon (Kerafast) were added to the solution.The solution then was incubated at 42 • C for 3 h followed by 95 • C for 1 min.
TGIRT. 1 or 2 l of 10 mM dNTPs and 1 l of either 2 M specific primer or 200 ng / l random 9-mer was added to 1-2 g RNA in a total volume of 11 l and incubated at 65 • C for 10 min followed by 4 • C for 2 min.8 l of TGIRT-III RT buf fer [final concentra tion 50 mM Tris-HCl (pH 8.3), 75 mM KCl, and 3 mM MgCl 2 ] and 1 l TGIRT-III (InGex) were added to the solution, followed by incubation at 60 • C for 2 h.
eHIV. 1 l of 10 mM dNTPs and 1 l of 5 M specific primer was added to 7 l containing 2 g of RNA and incuba ted a t 70 • C for 2 min followed by 4 • C for 2 min.9 l of 5 × eHIV RT buf fer [final concentra tion 200 mM Tris-HCl (pH 8.3), 400 mM KCl, 20 mM MgCl 2 ] and 2 l of eHIV enzyme were added to the solution.The reaction was then incuba ted a t 42 • C f or 3 h f ollowed by 95 • C f or 1 min.The eHIV enzyme was expressed and purified from pET30-RT1306 (gift from Bryan Dickinson; Addgene plasmid # 131521) following published protocols ( 40 ).

Libr ary pr epar ation
Small RNAs.Libraries were prepared using a two-step PCR strategy (Supplementary Table S3) ( 6 , 8 ). 1 l of cDNA was input into PCR1 using the following temperature cycles: 98 • C for 30 s, 18 cycles of [98 • C for 10 s, 60 • C for 30 s, 72 • C for 20 s], and 72 • C for 2 min.PCR1 products were purified (Mag-Bind Total Pure NGS, Omega Bio-Tek) using a 0.7x bead ratio. 1 ng of PCR1 product was used as input for PCR2 using the following temperature cycles: 98 • C for 30 s, then 12 cycles of [98 • C for 10 s, 66 • C for 30 s, 72 • C for 20 s], and 72 • C for 2 min.PCR2 products were purified using a 0.7x bead r atio.Libr aries were sequenced on an Illumina MiSeq instrument using either 2 × 250 (v2 chemistry) or 2 × 300 (v3 chemistry) paired-end sequencing (Supplementary Table S4).
rRNAs.Libraries from randomly primed total RNA were pr epar ed using the xGen NGS RNA (Integrated DNA Technologies) kit for HEK293 and the Nextera XT (Illumina) kit for E. coli cells.cDNAs were converted to double stranded DN A (dsDN A) by NEBNext second-strand synthesis module (New England Biolabs) using a 2 h incubation a t 16 • C .dsDNA was purified and size selected using ma gnetic beads (Ma g-Bind Total Pure NGS, Omega Bio-Tek) using a 0.65 × bead r atio.Libr aries were then generated following the Nextera XT manufacturer protocol, followed by purification and size-selection by magnetic beads (Mag-Bind Total Pure NGS, Omega Bio-Tek) using a 0.56 × bead r atio.Libr aries were sequenced on an Illumina MiSeq instrument using 2 × 300 paired-end sequencing (v3 chemistry) (Supplementary Table S4).

Mutation signature analysis
RMRP and RNase P pr obing data were initially pr ocessed using ShapeMapper 2.1.5with theoutput-countedmutations flag to tabulate mutation types observed at each sequence position.Area under the recei v er operating characteristic curves (AUROC) were calculated with Scikit-Learn (0.24.1) in Python using background-subtracted muta tion ra tes with pairing sta tus of each position deri v ed from known r efer ence structur es (41)(42)(43)(44)(45).To identify the G muta tion signa ture filter , A UROC was calculated for all possible combinations of m utation types, w hich revealed that including only G-to-C and G-to-T substitutions yielded the highest AUROC for all enzymes.

ShapeMapper 2.2
Building on our mutation signature analysis, we incorporated se v eral ne w features into ShapeMapper to automate four-base DMS-MaP processing.This ne w v ersion of ShapeMapper (v2.2) is available for download at https://github.com/Weeks-UNC/shapema pper2 .DMSspecific processing is invoked using the -dms flag.
Mutation signature filtering.G-to-A single-nucleotide misma tches, G multi-nucleotide misma tches and insertions and deletions at all nucleotides are ignored (set internally to 'no data').

DMS reactivity normalization.
Because four-base DMS modifica tion ra tes vary significantly based on nucleotide identity, r eactivities ar e normalized on a nucleotide-specific basis.The DMS modification rate is calculated as the difference between the modified and untreated mutation rates, or simply as the modified mutation rate if no untreated sample is provided.Normalization factors for each nucleotide type n are computed as where r n [ P 90 , P 95 ] denotes the mean of 90th-95th percentile modifica tion ra tes, and P 75 ( r n > 0 .001 ) denotes the 75th percentile of modification rates > 0.001.This scheme is more robust for RNAs such as the ribosome where most nucleotides ar e unr eacti v e. Final normalized reacti vities are then obtained by dividing the modification rate by the nucleotide-specific normalization factor.The normalized r eactivities ar e output dir ectly as text files with the suffix .dms .

Final data processing
Four-base DMS probing data were processed using ShapeMapper 2.2 with thedms andoutput-parsedmutations options.Amplicon libraries from small RNAs were processed using theamplicon option, and total RNA (rRNA) libraries were processed using therandom-primerlen 9 flag.Unfiltered (standard DMS) and 2A3 data were processed using ShapeMapper 2.2 without thedms flag.

Expected structural information
Inspired by metrics for quantifying sequence information content ( 46 ), we de v eloped the Expected Structural Infor-mation (ESI) metric to intuiti v ely quantify the total information provided by a probing experiment.Each nucleotide can adopt two possible structural states: base-paired ( b ) or unpaired ( u ).In the absence of any probing data, we assume each nucleotide has an equal probability of being paired or unpaired ( p ( b) = p ( u ) = 0 .5 ).A reactivity measurement ( r i ) reduces the structural uncertainty, which can be quantified as where SI ( r i ) denotes the structural inf ormation con veyed by reactivity r i , and H ( s i ) denotes the Shannon entropy of position i : ) and p( u | r i ) are determined from the empirical reactivity distributions for paired and unpaired positions based on the known structure: The distributions ˆ p ( r i | b ) and ˆ p ( r i | u ) are estimated by fitting the paired and unpaired reactivity data for each nucleotide type to double-gamma mixture models.The expected structural information (ESI) is then obtained as the av erage ov er all nucleotides n in the molecule (excluding primer binding sites and other positions with low-quality data): For computing ESI of DMS at only A and C nucleotides, SI ( r i ) of G and U nucleotides was set to 0.

PAIR-MaP analysis
We incorporated se v eral minor updates to PairMapper analysis to maximize performance on four-base DMS datasets.PairMapper previously required nucleotide windows to have a minimum of 50 co-modification e v ents to be considered for PAIR correlation analysis ( 8 ).We reevalua ted this co-modifica tion threshold across the range of 5-50, finding optimal performance with co-modification count cutoff of 10.We also updated the reactivity thresholds for primary and secondary PAIRs to 0.2 and 0.4, respecti v ely.
Positi v e predicti v e value (ppv) and PAIR-MaP sensiti vity (sens) were computed relati v e to accepted reference structures as previously described ( 8 ).RNA regions lacking DMS data were excluded from ppv and sens calculations.

Structure modeling
Four-base DMS pseudo-energy parameterization.We followed a previously described strategy ( 8 , 47 ) to derive four-base-DMS-optimized pseudo-energy potentials for structure modeling in RNAstructure (v6.3) ( 48 ).Nucleotidespecific reactivity likelihood functions for paired and unpair ed bases wer e fit using a double gamma mixture to normalized four-base DMS data collected on the cell-free E. coli 23S rRNA.These 23S rRNA deri v ed parameters serve as universal folding parameters for all RNAs.Four-base DMS potentials only vary modestly from our previous nucleotide-specific DMS potentials ( 8 ), but more strongly penalize pairing of reacti v e G and U nucleotides.Fitted model parameters are provided in Supplemental Table S5.For structure modeling purposes, we replaced the DMSdist nt.txt file in RNAstructure / data tables / dists with our new four-base DMS-specific file.We plan to make these parameters automatically available in future releases of RNAstructure .
RNAstructure modeling.Four-base DMS and SHAPEdirected modeling of RMRP, RNase P, and tmRNA was performed using iterati v e ShapeKnots folding to enable modeling of multiple pseudoknots ( 8 , 49 ).foldPK.py, the automa ted script tha t facilita tes this itera ti v e folding strategy, is available for download at https://github.com/MustoeLab/StructureAnal ysisTools .Folding of rRN As was performed using Fold with the -mfe and -md 600 options.Default SHAPE and four-base DMS parameters were used for all Fold and ShapeKnots modeling.For four-base DMS modeling, PAIR r estraints wer e additionally passed using the -x option.Structure modeling was not possible for human 28S rRNA due to memory overflow errors in RNAstructure .
Quantification of model accuracy.The positi v e predicti v e value (ppv) and sensitivity (sens) of modeled structures were computed relati v e to accepted r efer ence structur es as pr eviously described ( 8 ), using all Watson Crick and GU pairs allowing for one-position register shifts and ignoring singleton pairs.Accepted r efer ence structur es wer e obtained from refs (41)(42)(43)(44)(45). Modifications to tmRNA and RMRP structur es wer e included as pr eviously described ( 8 ).

Existing DMS-MaP strategies are unable to probe G structure in cells
To evaluate the ability of different MaP strategies to measure DMS modifica tions a t all nucleotides, we genera ted MaP datasets from identical DMS-probed RNA inputs using published SSII ( 8 ), TGIRT ( 36 ), and Marathon ( 37 ) MaP protocols (Figure 1 A).We additionally evaluated an HIV-1 re v erse transcriptase that was e volv ed to MaP N 1methyladenosine modifications (eHIV) ( 40 ), which has not been previously tested on DMS modified samples.DMS probing experiments were performed in duplicate on living HEK293 cells, under mildly alkaline buffer conditions that support multiple-hit DMS modification at all four nucleobases (see Materials and Methods) ( 8 ).An amplicon strategy was then used to obtain targeted DMS-MaP datasets for the Ribonuclease P (RNase P) and RnaseP RMP (RMRP) non-coding RN As, w hich adopt well-defined, known structures.
Consistent with prior studies ( 36 , 37 ), analysis of untreated control samples re v ealed significant differences in the background error rates of different MaP protocols.While all protocols exhibit low median background rates (0.001, 0.001 and 5 × 10 −4 for SSII, Marathon, and TGIRT respecti v ely), SSII samples feature significantly more positions with high background rates (95th percentiles of 0.02, 0.008, 0.005, respecti v ely) (Supplementary Figure S1A).eHIV featured a 3-fold higher background rate than SSII (median 0.003; Supplementary Figure S1A), leading us to focus on the established MaP protocols for subsequent analyses.
Despite differences in background error profiles, background-subtr acted DMS modification r ates measured at adenosine (A), cytidine (C) and uridine (U) bases are broadly consistent across all enzymes ( R > 0.9; Figure 1B; Supplementary Figure S1B).As expected, Us are a pproximatel y 5-fold less reacti v e than A and Cs (Fig- ur e 1 C).U r eactivities also vary mor e in SSII samples compared to Marathon and TGIRT, reflecti v e of greater background noise in SSII samples.SSII and Marathon consistently measure higher modification rates compared to TGIRT (Figure 1 B, C), indicating differences in random nucleotide incorporation and enzyme drop-off.SSII also generates an increased fraction of indels compared to Marathon and TGIRT (26%, 2.8%, 7.4%, respecti v ely; Supplementary Figure S1C).Nonetheless, each enzyme is similar ly accur a te a t distinguishing single-stranded versus base-pair ed nucleotides (ar ea under the r ecei v er operating characteristic curve [AUROC] ≈0.79 for A; ≈0.93 for C; ≈0.83 for U; Figure 1D; Supplementary Table S1).These da ta corrobora te our prior observa tion ( 8 ) tha t DMS is a highly specific probe of U pairing status and validate that all established MaP protocols r eliably measur e U modifications.
In contrast to A, C and U nucleotides, all three MaP protocols exhibited minimal-to-no ability to measure G nucleotide structure (AUROC ≤ 0.6; Figure 1D; Supplementary Table S1).G mutation rates significantly increase upon DMS trea tment, indica ting tha t DMS is modifying G bases (Figure 1 C).Howe v er, these modifica tions occur a t similar rates in both single-stranded and base-paired nucleotides.Surprisingly, G r eactivity measur ements vary 10-fold across MaP protocols, with SSII, Marathon, and TGIRT reporting median modification rates of 0.002, 0.007 and 8 × 10 −4 , respecti v ely.Despite yielding the lowest overall modification rates, TGIRT does identify se v eral highly reacti v e single-stranded Gs, but not with sufficient sensitivity to be useful.Thus, consistent with our prior studies ( 8 ), existing DMS pr obing pr otocols are unable to reliably probe G pairing status in cells.

Mutational signature filtering enables robust DMS probing of G nucleotides
DMS is known to methylate G bases at two positions ( 2 , 8 ): DMS predominantly modifies G at the N 7 position with minimal dependence on Watson-Crick pairing sta tus; a t much lower ra tes, DMS can methyla te singlestranded, deprotonated G nucleobases at the N 1 position.While re v erse-tr anscriptases are gener ally considered to be insensiti v e to N 7 -G modifications, we hypothesized that MaP may detect these modifications at low rates, convoluting an y inf ormati v e N 1 -G signal in DMS-MaP data.Further, we hypothesized that the differences in G modification rates measured by alternati v e MaP protocols reflect differences in N 7 -G detection efficiency.
To explore these hypotheses, we more closely analyzed the G muta tional signa tures yielded by each enzyme.DMS-MaP anal ysis traditionall y considers all mutation types as conv eying equi valent information.Strikingly, howe v er, we observed major differences in the structural specificity of dif ferent muta tion types.The eleva ted G modifica tion ra te measured by Marathon is dri v en by G → A substitutions, which occur at equivalent frequencies (median ≈ 0.005) in single-stranded and paired nucleotides (Figure 2A; Supplementary Figure S2).SSII and TGIRT datasets similarly exhibit a surplus of structurally non-specific G → A substitutions, although at lower frequencies (median < 0.001).By contrast, DMS-dependent G → C and G → T substitutions occur almost e xclusi v ely in single-stranded nucleotides for  S1 for the complete list of probed RNAs.all three enzymes (Figure 2 A).Thus, our DMS data are consistent with MaP measuring N 7 -G modifications specifically as G → A substitutions, whereas structurally informati v e N 1 -G modifications are decoded as G → C and G → T substitutions.
To validate this mutational signature, we used MaP to measure natural N 1 -G and N 7 -G modifications in untreated E. coli and human ribosomal RNAs ( 50 , 51 ).Consistent with our DMS data, N 7 -G modifications are detected by Marathon with low efficiency (1-3%), but overw helmingl y as G → A substitutions ( > 90% of mutations) (Figure 2 B).By comparison, the natural ribosomal N 1 -G modification is read out with high efficiency (65%) as a mixture of G → C and G → T substitutions (89% of mutations) (Figure 2 B).Similar results were also observed for SSII (Supplementary Figure S3).Together, these data confirm that MaP decodes N 1 -G and N 7 -G modifica tions via distinct muta tional signatures.We can also extrapolate from the detection efficiencies of natural modifications to estimate the DMS modifica tion ra tes of N 7 -G and N 1 -G in our experiments; we estimate a mean modification rate of 0.2 − 0.8 for N 7 -G and ∼0.001 for N 1 -G, compared to ∼0.02 for A and C and 0.006 for U calculated across both paired and single-stranded nucleotides.
Le v eraging this mutational signature, we implemented a refined bioinformatics pipeline within ShapeMapper ( 52 ) that filters out G → A substitutions and other uninformati v e mutation types in DMS probing data (see Materials and Methods).This refined pipeline resulted in dramatic improvements in the structural specificity of G DMS reactivity, with AUROC increasing from < 0.6 to > 0.7 for all thr ee r e v erse-transcriptase enzymes (Figure 2 C).Benchmarking on an expanded panel of RNAs from human and E. coli cells confirmed that our pipeline enabled accurate DMS probing of G base-pairing status across di v erse systems (Supplementary Table S1).Marathon consistently yielded superior AUROC at G nucleotides (Figure 2 D, Supplementary Table S1), leading us to select it as the optimal re v erse transcriptase for DMS-MaP experiments.This improved strategy also reduced the importance of background muta tion ra te subtraction (Supplementary Figure S4), although the use of an untreated control still offers minor increases in probing accur acy.Over all, we conclude mutationsignature filtering enables r obust DMS pr ofiling of RNA structure at all four nucleotides, which we term four-base DMS-MaP.

Appropriate buffering is essential for measuring U and G reactivities
The ability of DMS-MaP to measure structure-specific modifica tions a t U and G r equir es transient deprotonation of N 1 -G and N 3 -U ( 8 ).We previously reported that bicine buffer (pH 8.0) is critical for robust DMS modification of G and U nucleotides in vitro ( 8 ).Howe v er, the e xtent to which extracellular buffering impacts DMS modification in cells is unclear.Cells work to maintain pH homeostasis, but changes in extracellular pH can affect intracellular pH, particularly on short time scales ( 53 , 54 ).DMS treatment may also perturb the plasma membrane or induce stress responses that compromise internal pH control.We ther efor e investigated the impact of extracellular buffering by DMS probing E. coli and HEK293 cells at a variety of extracellular pHs: no supplemental buffer, which results in rapid acidification of the media (pH < 6); neutral pH 7.2 (sodium cacod yla te buf fer); and across the bicine buf fering range (pH 7.7, pH 8.0, and pH 8.7).Both U and N 1 -G modification rates strongly depend on extracellular buffering, increasing ∼10-fold at pH 8 compared to unbuffered conditions (Figure 3 A, B).Interestingly, the reactivity rate of U and N 1 -G plateaus above pH 8 in E. coli , suggesting that bacterial cells b uffer a gainst major deviations from pH neutrality.The increase in G and U modification rate coincides with a significant increase in AUROC (from mean 0.76 to 0.91 for U, and 0.43 to 0.72 for G, respecti v ely; Figure 3 C, D).By contrast, minimal changes in mutation rate or AUROC are observed at A or C nucleotides.The rate of N 7 -G modifications , measured by G → A substitutions , also minimally changes with pH (Figure 3 A, B).Thus, these data establish that pH --whether it is acidic due to DMSinduced acidification in the absence of buffer ( 10 ), neutral, or basic --modulates DMS reactivity in cells and emphasize that proper buffering is essential for four-base DMS probing.These data also provide further support for the deprotonation mechanism of DMS modification at N 1 -G and N 3 -U.

F our-base DMS conv e ys greater information than comparable probing experiments
To facilitate structural analysis, we implemented a strategy to correct for differences in DMS modification rates across bases and normalize reactivities to a common 0 to ∼1 scale, denoting unreacti v e and reacti v e nucleotides, respecti v ely (Methods).Consistent with the AUROC analysis above, these normalized four-base DMS reactivities provide a precise map of RNA structure in cells with significantly incr eased r esolution compar ed to traditional DMS data (Figure 4 A).
We sought to understand how four-base DMS reactivity data compare to SHAPE data, which r epr esent the gold-standard for measuring structure at all four nucleotides ( 35 ).Prior in vitro studies have suggested that DMS and SHAPE data can convey similar amounts of structural information ( 8 , 47 ), but direct comparisons of in-cell DMS and SHAPE data are lacking.Recently, 2A3 was introduced as an improved SHAPE reagent for in-cell structure probing, and high-quality 2A3 datasets are available for human and E. coli ribosomal RN As (rRN As) and E. coli tmRNA ( 34 ).We also collected our own 2A3 SHAPE-MaP data for RMRP and RNase P in HEK293 cells (Supplementary Figure S5).We note that we relied on published SHAPE protocols without pursuing further optimizations.Visual analysis indicates that SHAPE and four-base DMS reactivity profiles are qualitati v ely similar, although DMS provides a more binary measure of structure compared to a more continuous measure provided by SHAPE (Figure 4 A).Fourbase DMS data also typically yield modestly higher AU-ROC than SHAPE da ta (Figure 4 C , Supplementary Figure S6).
Prompted by the differences observed between SHAPE and DMS data, we sought to de v elop a better analytic frame wor k for e valua ting the informa tion conveyed by probing experiments.AUROC quantifies how well reactivity data perform as a binary classifier of a nucleobase being paired versus single-stranded, but this does not reflect how reactivity data are normally interpreted.For most applica tions, da ta ar e interpr eted probabilistically: gi v en a measur ed r eactivity, wha t is the probability tha t a base is paired versus unpaired?These probabilities are represented by the pair ed / unpair ed likelihood ratio function, with more extreme likelihoods indicating greater information content ( 55 ).The DMS likelihood ratio function is mor e extr eme for A, C and U nucleotides, indica ting tha t DMS conveys more structural information at these nucleotides, whereas SHAPE conveys more information at G (Figure 4 B).To quantify these differences, we de v eloped a new metric termed expected structural information (ESI) that measures the total per-base information conveyed by a probing experiment (see Materials and Methods).ESI has units of bits and ranges between 0 (no information) to 1 (perfect specification of pair ed / unpair ed status for all nucleotides) (Supplementary Figure S7).SHAPE experiments provide an average of 0.25 bits of ESI, whereas fourbase DMS data convey an average of 0.35 bits (Figure 4 C; Supplementary Figure S6), supporting that four-base DMS provides a more deterministic measure of base-pairing.By comparison, two-base DMS data only convey an average of 0.23 bits of ESI, less than SHAPE (Figure 4 C).Together, these analyses indicate that four-base DMS experiments typically provide more structural information than other widely used probing strategies.

Four-base DMS-MaP enables improved direct base-pair detection from single-molecule PAIR analysis
In addition to providing a per-nucleotide readout of basepairing status, DMS probing data can be analyzed at the single-molecule le v el to identify nucleotides that undergo correla ted modifica tion.PAIR analysis ( 8) is a powerful strategy for detecting RNA duplexes from single-molecule probing data, and significantly improves the confidence and accur acy of structur al analysis.We hypothesized that the improved specificity of four-base DMS-MaP would also benefit P AIR analysis.Indeed, P AIR analysis applied to four-base DMS-MaP data significantly outperformed analysis of traditional DMS-MaP data (SSII without mutationsignature filtering): the positi v e predicti v e value (ppv) of PAIR corr elations incr eased from 0.73 to 0.84, and sensitivity (sens) increased from 0.18 to 0.31, respecti v ely) (Figure 5 A,B; Supplementary Figures S8, S9; Supplementary Table S2).Four-base DMS-MaP also enabled detection of PAIRs at lower read coverages (coverage of ∼300 000 versus ∼400 000 r equir ed for DMS-MaP ( 8)) (Figur e 5 C).
Surprisingl y, PAIR anal ysis on traditional DMS-MaP data (SSII without mutation-signature filtering) performed worse than in prior studies ( 8 ).This reduced performance can be attributed to lower DMS modification rates (Sup-plementary Figure S10) and, for cell-free ribosomal RNA samples, significantly lower read depth coverage in our current experiments.We also note that many of the 'false positi v e' PAIRs observ ed in cell-free rRNA samples likely correspond to 'real' non-nati v e interactions formed under these conditions ( 8 , 56 , 57 ).Interestingly, we also observed a 4-fold greater G → A substitution rate in our current DMS-MaP datasets compared to our prior e xperiments, suggesti v e of cryptic re v erse-transcription dif ferences tha t impact detection of N 7 -G modifications and reduce PAIR performance (Supplementary Figure S10).Muta tion-signa ture filtering of SSII DMS-MaP data improved PAIR performance, but still underperformed Mara thon DMS-MaP da ta (Supplementary Figures S8 and S11).
We also performed PAIR analysis on four-base DMS-MaP datasets collected using TGIRT re v erse-transcriptase.Compared to Marathon and SSII, TGIRT data gave significantly worse PAIR results (mean ppv = 0.57, sens = 0.20; Supplementary Figure S8, Supplementary Figure S11), presumably due to TGIRT measuring fewer modifications (Figure 1 , Supplementary Figure S11).
We additionally explored whether the 2A3 SHAPE reagent supports PAIR anal ysis.Historicall y, SHAPE reagents have been unable to achieve high enough modifica tion ra tes for PAIR analysis.Howe v er, 2A3 mostly addresses this limitation, modifying RNA at comparable rates to DMS, with lower modification of A and C compensated by higher modifica tion a t U and G (Supplementary Figure S5).Deeply sequenced human RMRP and RNase P in-cell 2A3 datasets both feature multiple PAIR correlation signals, but with lower sensitivity and specificity than four-base DMS-MaP (Figure 5 A, B).Thus, SHAPE experiments can enable PAIR detection, but further optimization is needed to make SHAPE-based PAIR analysis broadly useful.

F our-base DMS-MaP enables impr ov ed RNA structure modeling
The end goal of many chemical probing studies is to transla te probing da ta into an accura te model of RNA struc-ture.Building on the improvements of four-base DMS-MaP, we de v eloped ne w pseudo energy functions for incorporating four-base DMS reactivities as restraints during structure modeling with RNAstructure (Methods).Integrated structure modeling guided by four-base DMS reactivities and PAIR correlations facilitated accurate modeling of di v erse, challenging RNA targets (Figure 6 ).Notably, four-base DMS meets or exceeds the accuracy of SHAPE-dir ected structur e modeling (Figur e 6 ).Four-base DMS data particularly benefits modeling accuracy for pseudoknot-containing tmRNA, RMRP, and Rnase P RNAs (Figure 6 ), enabled by PAIR correlations unavailable to SHAPE.As an exception, SHAPE (2A3) data yields more accurate models for the in-cell 16S and 23S rRNAs, consistent with the superior ESI of 2A3 versus fourbase DMS for these RNAs (Figure 4 C).Both four-base DMS and SHAPE data yielded similar ly inaccur ate models for in-cell human 18S rRNA; this inaccuracy arises from the e xtensi v e pr otein pr otections that reduce ESI of both reagents and from sequence features of the 18S rRNA ( 58 ).When the unr epr esentati v e human and E. coli rRNAs are e xcluded, av erage ppv and sens of DMS-directed models increase to 0.93 and 0.92, respecti v ely.In sum, four-base DMS-MaP supports best-in-class structure modeling accu- racy and enables reconstruction of even challenging multipseudoknotted RNA structures.

DISCUSSION
DMS has long been a favored structure probing reagent and has played an essential role in enabling next-generation single-molecule probing analyses ( 4 , 7 ).Howe v er, the inability of DMS to probe U and G nucleotides has been a critical limitation.In this work we introduced four-base DMS-MaP as a strategy for high-fidelity structure probing at all four nucleotides in living cells.Through rigorous benchmar king, we estab lished that four-base DMS-MaP e xperiments typically convey more structural information than other probing strategies and enhance single-molecule analysis, enabling accurate structure modeling of complex RNAs that challenge other methods.Four-base DMS-MaP experiments are straightforward to perform, requiring only minor changes to standard DMS probing protocols and bioinf ormatics pipelines.Thus, f our-base DMS-MaP r epr esents an 'almost for free' upgrade offering improved resolution in both conventional per-nucleotide and single-molecule probing analysis.
The success of four-base DMS-MaP experiments depends on se v eral subtle but collecti v ely critical e xperimental parameters.Most importantly, our results emphasize the need for proper buffering, with G and U reactivity strictly dependent on pH (Figure 3 ).Despite cells buffering against significant changes in intracellular pH, extracellular pH clearly modulates in-cell DMS reactivity.We also showed that different MaP protocols can impact data quality.SSII, Marathon, and TGIRT protocols performed similarly for per-nucleotide DMS reactivity analysis, although Marathon consistently performed the best at measuring N 1 -G modifications.Howe v er, the choice of MaP enzyme significantly impacted the success of single-molecule PAIR analyses, with Marathon consistently detecting more duple xes with fe wer false positi v es than other MaP enzymes.TGIRT performed the worst at single-molecule PAIR analysis, likely because of a reduced ability to MaP through highl y modified RN As.We note that our anal yses were limited to published MaP protocols and speculate that opti-mization of MaP in the context of four-base DMS-MaP may yield e v en further improvements in data quality.
Four-base DMS-MaP is also built on the insight that MaP enzymes can sim ultaneousl y encode distinct types of chemical information via different mutational signatures.DMS modifies G nucleotides at two positions ( 59 ): The bulk of modifications occur at the N 7 position, which do not report on Watson-Crick pairing, whereas only a minority of modifications occur at the informati v e N 1 position.Consistent with other studies ( 60 ), our analysis indicates tha t Mara thon (and to a lesser degr ee other r e v erse transcriptases) selecti v ely decode N 7 -G modifications as G → A mismatches, allowing us to discriminate N 1 -G modifications and measure G pairing status with high fidelity.While not the focus of our current study, DMS N 7 -G modifications can provide information about RNA tertiary structur e and G-quadruplex es ( 27 , 60 , 61 ), and further investigating the value of N 7 -G reactivity is a compelling area of futur e r esear ch.Mor e generall y, using m uta tion signa tures to decode multiple coexisting modification signals r epr esents a powerful paradigm for improving chemical probing analysis.
Our finding that four-base DMS-MaP typically conveys greater structural information than SHAPE-MaP is surprising.SHAPE chemistry holistically and unbiasedly measures nucleotide flexibility at the 2 OH ( 62 ), but this holistic measure may come with the tradeoff of reduced specificity for Watson-Crick pairing compared to direct nucleobase probing by DMS.Re v erse transcriptases may also decode 2 OH SHAPE modifications with lower fidelity.Ne v ertheless, we emphasize that the performance gap between fourbase DMS and SHAPE experiments is subtle, with both strategies performing well for most RNAs.SHAPE reagents also offer important benefits, including that they are generally less cytotoxic than DMS, are better suited for probing RNAs with modified bases, and, as noted above, holistically measure nucleotide flexibility.SHAPE reagents can further be used to probe all four nucleobases under singlehit reaction conditions, whereas four-base DMS probing of U, and especially G, nucleobases r equir es high overall modifica tion ra tes.We also note tha t unlike for DMS, we made no attempt to optimize SHAPE-MaP.For example, SHAPE-specific data processing algorithms may enable improved PAIR analysis on SHAPE datasets.More generally, we belie v e tha t systema tic ef forts to improv e MaP re v ersetranscription and bioinformatics protocols such as we pursued here for DMS will dri v e further increases in SHAPE probing resolution and accuracy.
Ultimatel y, our anal ysis demonstrates how improved chemical probing data can support further advances in RNA structure determination accuracy.The increased structural information provided by four-base DMS reactivities combined with PAIR correla tion da ta is sufficient to guide in-cell structure modeling to > 90% average accuracy for e v en v ery difficult targets such as tmRNA.Modeling accuracy is lower for rRNAs, but ribosomes are clearly exceptional cases with atypically high protein protections.Modeling accuracy is also reduced for human RNase P; followup analysis re v ealed that prediction accuracy was low e v en w hen using sim ula ted 'perfect' da ta (Supplementary Figure S12), indicating that pseudoknot modeling algorithms remain imperfect.Combining four-base DMS-MaP with new statistical-learning strategies ( 63 ) r epr esents one of several potential avenues for further improving modeling accuracy.Moving forward, we expect that focus will increasingly turn to the more difficult problem of modeling RNAs with heterogenous structures.Importantly, four-base DMS-MaP is fully compatible with emerging ensemble deconvolution analysis (11)(12)(13)(14).We anticipate that the greater information provided by four-base DMS-MaP will help propel further advances in modeling and understanding complex RNA systems.

DA T A A V AILABILITY
Raw and processed probing data have been deposited at the GEO under accession number GSE225383.Analysis codes are available at https://github.com/MustoeLaband have been archived at 10.5281 / zenodo.7808746.

SUPPLEMENT ARY DA T A
Supplementary Data are available at NAR Online.

Figure 1 .
Figure 1.Comparison of different MaP protocols for measuring DMS modifica tions a t all four RNA bases.( A ) Experimental scheme for in-cell DMS probing, re v erse transcription, and reactivity analysis using identical RNA inputs.( B ) Background-subtracted DMS-MaP reactivity profiles measured for RMRP.Gray curves shown at bottom indicate known base pairing interactions.( C ) DMS modification rates measured at each nucleotide combined across RMRP and RNase P. Background-subtracted mutation rates are shown for base-paired (filled) and single-stranded (open) bases.The y axis has a linear scale < 10 −2 (indicated by thick axis) and logarithmic scale for values > 10 −2 (thin axis).( D ) Recei v er oper ator char acteristic (ROC) curves quantifying ability of DMS reactivity to discriminate single-stranded versus base-paired nucleotides in RMRP and RNase P.

Figure 2 .
Figure 2. Mutation signature filtering enables high-fidelity DMS probing of G base-pairing status.( A ) G-specific mutation spectrums for in-cell probed RMRP and RNase P generated by SSII (green), Marathon (blue), and TGIRT (orange).Rates are shown separately for base-paired (filled) and singlestranded (open) G nucleotides.The y axis has a linear scale < 10 −2 (indicated by thick line) and logarithmic scale for values > 10 −2 (thin line).( B ) Mutation rates (top) and percentage of detected mutations (bottom) measured by Marathon MaP for naturally-occurring N 1 -G and N 7 -G modifications in untreated E. coli and human rRNA.( C ) ROC curves for muta tion-signa tur e-filter ed G r eacti vities for in-cell probed RMRP and RNase P. Curv es generated without mutation filtering for each enzyme are shown in gr ay. ( D ) Aver age AUROC across all probed RNAs quantifying the ability of muta tion-signa tur e-filter ed DMS reactivities to discriminate pairing status at each nucleotide.The best performing enzyme, Marathon, is boxed.See Supplementary TableS1for the complete list of probed RNAs.

Figure 3 .
Figure 3. Four-base DMS-MaP depends strongly on extracellular buffering.( A, B ) Mean DMS modification rates measured by Marathon MaP for single-stranded A, C, G, and U nucleobases for E. coli tmRNA and human RMRP probed in cells using different extracellular buffers.G → A substitutions, which are filtered out by muta tion-signa ture-filtering, are shown in gray.( C, D ) Corresponding AUROC values for E. coli tmRNA and human RMRP quantifying the ability of DMS reactivities to discriminate pairing sta tus.All da ta points r epr esent the mean of two independent biological replicates, with vertical bars indicating standard error.Lines are drawn between points to guide the eye.

Figure 4 .
Figure 4. Four-base DMS-MaP conveys more structural information than other probing strategies.( A ) Normalized reactivity profiles for in-cell probed RMRP gi v en by SHAPE-MaP (top), traditional DMS-MaP (SSII; no muta tion-signa ture filtering) considering only A and C nucleotides (upper middle) or all nucleotides (lower middle), or four-base DMS-MaP (bottom).Base pairing interactions in the accepted structur e ar e shown using gray shading and ar cs. ( B ) Unpair ed / pair ed likelihood ratios for SHAPE and four-base DMS-MaP reactivities measured on cell-free probed E. coli 16S and 23S rRNA.A likelihood ratio of 1 indicates that a base has equal probability of being paired or unpaired gi v en the measured reactivity.Likelihood ratios are shown for each nucleotide for four-base DMS-MaP and are aggregated for all nucleotides for SHAPE.( C ) Comparison of AUROC (top) and expected structural information (ESI; bottom) for four-base DMS-MaP and SHAPE-MaP across a di v erse panel of RNAs.2A3 SHAPE data for human rRNAs, and E. coli rRNAs and tmRNA were taken from ( 34 ).AUROC and ESI values were also calculated for only A and C DMS r eactivities, r epr esenting what is obtained by traditional two-base DMS probing experiments.

Figure 5 .
Figure 5. Four-base DMS probing improves single-molecule direct base-pair detection.( A ) PAIR analysis performed on in-cell probed RMRP using traditional DMS-MaP (left), four-base DMS-MaP (center) or SHAPE-MaP (right) datasets.PAIRs are prioritized based on strength as principal and minor signals, shown in dark and light blue respectively.The accepted structure is shown at top.Long-range helices that are masked by primer binding sites, and thus lack PAIR-MaP data, are shown in light gray.PAIR positi v e predicti v e value (ppv) and sensitivity (sens) relative to the known structure are also indicated.( B ) PAIR ppv and sens for traditional DMS-MaP (tan), four-base DMS-MaP (red) and SHAPE-MaP (gray).Data from two biological r eplicates ar e shown for in-cell RMRP, RNase P and tmRN A. 16S and 23S rRN A data are shown in open circles and demonstrate reduced ppv and sens due to low sequencing coverage and likely misfolding of these RNAs under cell-free conditions.( C ) Mean PAIR ppv and sens values across RMRP, RNase P, in-cell tmRNA, and cell-free tmRNA are shown for traditional DMS-MaP (tan) and four-base DMS-MaP (red) as a function of sequencing depth.The indicated number of sequencing reads were undersampled without replacement from the larger datasets.

Figure 6 .
Figure 6.Four-base DMS-MaP enables improved RNA structure modeling.( A ) E. coli tmRNA structure models obtained via RNAstructure modeling with four-base DMS and PAIR restraints (left) or SHAPE restraints (right) from in-cell e xperiments.True positi v e (gray), false positi v e (purple), and falsenegati v e (green) predicted pairs are shown at top, and overall model ppv and sens are shown at bottom.For four-base DMS, measured PAIR correlations are also shown at bottom.Correctly modeled pseudoknots are labeled.( B ) Structure modeling accuracy of four-base DMS-MaP and SHAPE-MaP.For in-cell RMRP, RNase P, and tmRNA, four-base DMS data from two biological r eplicates wer e pooled together.Asterisks (*) next to in-cell E. coli and HEK293 rRNAs indicate that four-base DMS structure modeling was done without PAIR r estraints, which wer e not available due to insufficient sequencing depth.2A3 SHAPE data for E. coli tmRNA and rRNAs, and human rRNAs were taken from( 34 ).