Real-Time Single-Molecule Studies of RNA Polymerase–Promoter Open Complex Formation Reveal Substantial Heterogeneity Along the Promoter-Opening Pathway

Graphical abstract


Introduction
Transcription initiation is the first and most regulated step in gene expression in all organisms. The expression of most bacterial genes commences with the binding of RNA polymerase (RNAP)-r 70 holoenzyme to the promoter DNA. 1 The initial RNAP-promoter closed complex (RP C ) undergoes large conformational changes leading to a RNAP-promoter open complex (RP O ), which is capable of RNA synthesis. These conformational changes are of paramount importance, since their modulation by promoter DNA sequence, protein transcription factors, and small-molecule ligands strongly affects the number of active open complex, and thus the transcription efficiency. 2 Further, several antimicrobials, including clinically used drugs rifampicin 3,4 and fidaxomicin, 5 exert their effect by blocking RNAP from proceeding during a specific step of transcription initiation. 6 However, despite substantial progress in defining the structural basis of transcription initiation mechanism, 7-10 the identity, sequence, and kinetics of conformational changes leading to RP O formation remain elusive.
At the initial step of the RP O formation pathway, the RNAP-r 70 holoenzyme recognises the promoter by forming sequence-specific contacts with the À35 element, and sequence-independent contacts upstream from the À35 ("upstream sequence") as well as around the À10 element [reviewed in 2,11] In this RP C state, the promoter remains fully double-stranded, but is bent by $17°a t the À10 element, thus positioning the downstream promoter DNA above the DNA-binding cleft of the RNAP. 10 Studies using footprinting, [12][13][14][15] atomic force microscopy 13 and ensemble FRET 17 have indicated additional extensive bending and wrapping of the promoter upstream sequence (between the À35 element and À82); this bending, which is driven by the C-terminal domains of the two RNAP a-subunits (aCTDs) interacting with the promoter upstream, brings the upstream DNA to the RNAP surface, and strongly facilitates RP O complex formation. 14,18,19 The isomerisation of the RP C towards RP O complex begins with the flipping of non-template DNA (ntDNA) À11 conserved adenine base from the duplex DNA to a specific pocket in r 70 . 20,21 The promoter melting then somehow propagates downstream until the full transcription bubble in the RP O complex covers positions À11 to +2. 9 The bubble melting is coupled with the loading of downstream DNA duplex into the RNAP cleft, and the loading of single-stranded template DNA (tDNA) into the RNAP active site. Structural 9,10 and biochemical 2 studies have identified several putative intermediates on the path from the RP C to RP O ; however, the number and structural properties of the intermediates detected appear to heavily depend on the promoter sequence, transcription factors, and experimental conditions. The mechanism discussed above describes the formation of a uniform RP O complex on a standard linear reaction pathway. A more complete description of the transcription initiation, however, needs to consider several studies that suggested that individual RP O molecules are not identical, and they instead differ in functional properties. [22][23][24][25] One of the most notable variation among RP O complexes is their tendency to perform abortive initiation, i.e., the premature release of short RNAs synthesised by promoter-bound RNAP (reviewed in 26). In fact, it has been estimated that >50% of the RP O complexes are permanently locked into the abortive initiation mode and cannot produce full-length RNA. [22][23][24][25] The presence of at least two different RP O classes -one productive and one non-productive (abortive) -raises the possibility that the RP O pathway is also not linear, but instead branches to allow the formation of structurally and functionally different RP O molecules. It has been further suggested that the ratio of productive and non-productive RP O complexes can be modulated by transcription factors and thus offers a layer for gene regulation in the cell. 27 On the other hand, recent single-molecule studies revealed longlived pausing, backtracking and arrest of initially transcribing bacterial and mitochondrial RNAPs that could potentially explain the productive and abortive RNA synthesis by a single type of RP O complexes. [28][29][30] The RP O formation pathway branching -its occurrence and mechanism -thus warrants further study.
Here, we utilise single-molecule techniques to resolve asynchronous, multi-step and branched reaction mechanisms during r 70 -dependent RP O formation on a well characterised consensus lac promoter. Our results strongly suggest that the RP O formation pathway is indeed branched both at the step of initial promoter melting and the step of open transcription bubble stabilisation. Furthermore, aCTD interactions with the promoter upstream sequence strongly stimulate bubble initiation and tune the reaction pathway towards more stable RP O complexes. The RNAP cleft loops (and especially the b 0 rudder one), play a key role in stabilising the open transcription bubble.

Direct formation of surface-immobilised catalytically active open complexes
To be able to monitor RNAP-promoter open complex (RP O ) formation in real-time at the singlemolecule level, we used FRET to look at the changes in distances between two points, i.e., positions À15 and +15 relative to the transcription start site (position +1) on a promoter DNA fragment. A fluorophore pair incorporated in positions À15 (donor) and +15 (acceptor) produces FRET signatures that vary depending on the transcription bubble conformation; this pair has been employed before to monitor conformational changes in populations of single transcription complexes, 31,32 conformational dynamics of RP O complexes 33 and conformational changes after the formation of RP O complex 29 on a consensus lac promoter (lacCONS) ( Figure S1(A, B)).
Here, we modified our previous protocols to detect the nascent RNAP-promoter complex coli RNAP-r 70 holoenzyme is immobilised on the PEGylated microscope coverslip using biotinylated anti-His-tag-antibody. lacCONS promoter, which is labelled with a donor fluorophore (D, Cy3B) at non-template DNA position À15 and an acceptor fluorophore (A, ATTO647N) at template DNA position + 15, is added to the reaction buffer. The promoter binds to the RNAP and becomes visible on the coverslip surface. The initial RNAP-promoter closed complex isomerises to the open complex, which decreases the distance between the À15 and +15 dyes and increases the FRET. (B) Schematic microscope field-of-view before and after promoter addition to the reaction buffer. Data on the DD (donor excitation-donor emission) and AA (acceptor excitation-acceptor emission) channels is used to identify the RNAP-promoter complexes containing both the Cy3B and ATTO647N dyes. These molecules are highlighted with yellow circles. (RP C ) and its subsequent maturation to RP O (Figure 1). To this end, we attached molecules of the Escherichia coli RNA polymerase-r 70 holoenzyme to the surface of a coverslip and started imaging the surface using TIRF microscopy ( Figure 1(A, B)). Subsequent addition of the dual-labelled promoter DNA to the reaction buffer was expected to lead to the appearance of co-localised fluorescent spots on the donor (Cy3B label) and acceptor (ATTO647N label) detection channels of the microscope upon binding to the surface-attached holoenzyme (Figure 1(B)). The timing of the appearance of the fluorescent spots on the surface (due to DNA binding and formation of RP C complexes) is precisely defined in the single-molecule trajectories by the simultaneous appearance of Cy3B and ATTO647N fluorescence signals ("DNA binds" time point, Figure 1(C)). The À15/+15 ruler reports low FRET for the RP C complex, and intermediate FRET for the RP O complex, since the formation of the transcription bubble decreases the distance between the À15 and +15 positions in the DNA. 33 The RP C ? RP O transition in the trajectories is thus indicated by a sharp FRET increase ("DNA transcription bubble opens" time point, Figure 1(C)), which may occur in one or several steps, depending on the intermediates on the reaction pathway.
Our experimental single-molecule trajectories indeed show the expected fluorescence intensity and FRET signatures of RP C complex formation and isomerisation to the RP O state (Figure 1(D)). The moment of RNAP-promoter complex formation was precisely defined by the simultaneous appearance of Cy3B and ATTO647N fluorescence in the single-molecule trajectories (e.g., at 12.25 s and 13.5 s in the left and right panels, respectively, of Figure 1(D)). The apparent FRET efficiency (E*) of the first stable complexes was E* $ 0.2 ( Figure 1(D)), a value identical to that we obtained previously for the closed transcription bubble state. 33 After a short time, the FRET increased to E* $0.45 (at $12.6 s and $15.7 s in the traces of Figure 1(D)), a value identical to that we obtained previously for the open transcription bubble state. 33 DNA binding to the coverslip surface was strictly mediated by the RNAP, since the number of non-specific DNA binding events was negligible in the absence of RNAP on the surface (cf. Figure S2(A, B)). On the population level, the newly formed RNAP-promoter complexes displayed a bimodal FRET distribution, with mean FRET values $0.2 and $0.45 ( Figure S2 (C)) contrasting with the unimodal FRET distribution (mean $ 0.18) of the protein-free immobilised promoter DNA ( Figure S2(D)). To test whether the $0.45 FRET state is indeed a catalytically competent RP O complex, we added NTPs to the sample buffer; this addition almost eliminated the $0.45 FRET state, as expected if RP O complexes engage RNA synthesis and escape the promoter ( Figure S2 (C)).
To provide further proof for the formation of catalytically active RP O complexes in situ on the coverslip surface, we performed experiments using a promoter with a different labelling scheme, which is very effective in monitoring the progress of initial transcription (dyes at positions À15 and +20; Figure S1(C)). The scrunching of the downstream DNA towards the RNAP during initial RNA synthesis leads to a stepwise increase in FRET until RNAP escape from the promoter returns the FRET to a low level ( Figure S3(A,  B)). 24 Example trajectories in Figure S3(C) demonstrate abortive initiation and promoter-escape events occurring shortly after the formation of the RNAP-promoter complexes. However, we note that some RNAP-promoter complexes (typically 20-50% of all complexes) on the surface neither form RP O nor engage RNA synthesis; these molecules remain in stable low FRET state ($0.2) and may thus represent unproductive complexes resulting, e.g., from RNAP binding to the ends of the promoter DNA fragment ( Figure S3(D)). HMM analysis of these complexes (on both À15/+15 and À15/+20 labelled promoters) did not produce Viterbi changes, further corroborating our interpretation that the transition from low FRET state (E* $ 0.2) to a long-lived higher FRET state (E* $ 0.45) on À15/+15 promoter indeed indicates a RP O formation event instead of spurious fluorescence fluctuation. However, because the FRET sensitivity is not sufficient to confidently distinguish RP C complex from non-specific RNAP-DNA complexes, we decided to analyse further only the RNAP-promoter complexes which directly show the appearance of the FRET signature of the RP O complex (i.e., the $0.45 FRET state) on the À15/+15 labelled promoter.

Extended promoter upstream sequence stimulates RP O formation
To study the kinetics of RP O complex formation in real-time and its modulation by the aCTD-promoter upstream interactions, we performed experiments using a long (DNA extending from position À89 to position +25) and short (DNA extending from position À39 to position + 25) version of the lacCONS promoter (Figure 2(A)). We also examined the kinetics of open complex formation by using additional versions of the promoter DNAs, which were either fully double-stranded (dsLC2 promoter; Figure S1(A)) or contained a mismatch in the promoter region from À10 to À4 (a.k.a. pre-melted promoter, pmLC2; Figure S1(B)).
RP O formation was inefficient in the case of short dsLC2; in fact, we could identify only 5 real-time promoter-binding events indicating RP O complex formation (3% of all promoter-binding events, N = 167); even after prolonged incubation of the RNAP-promoter complexes ($5 min) on the surface, the RP O complex (i.e., the FRET species with E* $ 0.45) was nearly absent from the population histogram ( Figure S4(A)). In contrast, the RP O complex formed efficiently on the long promoters, as well as on the short pre-melted promoter, as seen in the E* histograms ( Figure S4 (B-D)) and individual trajectories (Figure 2(B)).
We then performed Hidden Markov modelling (HMM) of the trajectories to extract the dwell times in the RP C state (E* $ 0.2) before transcription bubble opening and RP O complex formation (Figure 2(B)). The observed distribution of dwell times in the RP C state for the long dsLC2 promoter (Figure 2(C)) was fitted to a monoexponential decay function to determine a mean lifetime for the RP C complex of 1.43 ± 0.09 s (±SE; amplitude parameter was 35.2 ± 1.5). We also tried to fit the dwell-time distribution using a biexponential equation, but rejected this more complex kinetic model because the fit parameters were poorly defined as evident from large SE (27% for lifetimes and 13-45% for amplitudes).
Using a similar analysis, we estimated the RP C complex lifetime as 0.35 ± 0.04 s on the long premelted LC2 (Figure 2(D)) and 0.39 ± 0.03 s on the short pre-melted LC2 (Figure 2(E)), respectively. These values indicate that the aCTD interactions with the upstream sequence (À89 to À40) significantly enhance the isomerisation rate of the RP C to RP O complex; however, this happens only on a fully double-stranded promoter. Because the introduction of the pre-melted region (À10/À4) to the promoter nearly equalised the rate of RP C isomerisation to the RP O complex on the short and long promoters, the aCTD-promoter interactions appear to predominantly contribute to the lowering of the activation energy of initial transcription bubble nucleation.

A subpopulation of RP O complexes form via a kinetically significant intermediate
Close inspection of the HMM fit to the RP O formation FRET trajectories revealed that, even though most bubble-opening events were described by a two-state model, (i.e., the promoter conformation in the initial complex changed to the RP O state in a single step; Figure 2 Specifically, the RP i complex was identified in 20% (exact 95% binomial confidence interval 34 : 11-30%), in 14% (8-23%) and in 14% (8-23%) of all trajectories in the case of long dsLC2, long pmLC2 and short pmLC2 promoter, respectively. The arithmetic mean lifetime of the RP i state was 0.32 s (95% CI by bootstrapping (10,000 iterations): 0.18-0.49 s, N = 15), 0.15 s (0.07-0.26 s, N = 13) and 0.17 s (0.10-0.23 s, N = 13) on the long dsLC2, long pmLC2 and short pmLC2 promoter, respectively. Similarity of the estimates suggests that the RP i lifetime does not strongly depend on the used promoter type.
The sporadic appearance of the RP i in the trajectories could indicate either heterogeneous reaction step (bubble opening with or without RP i intermediate), failure to detect most of the RP i states (if RP i was an obligatory intermediate) or false positive HMM state assignment (if RP i does not really exist). To evaluate these scenarios, we used simulated smFRET trajectories and estimated the efficiency of RP i detection by the HMM routine. To obtain a conservative estimate, the trajectories were simulated using a short mean RP C lifetime (0.16 s) and FRET signal noise that was 15-50% higher than in experimental data ( Figure S5(A-F), example trajectories; Figure S5 (G-L), trajectory noise). As expected, the detection efficiency increased with the length of the RP i dwell: 1, 2-4 and !5 frame dwells (1 frame = 20 ms) were detected with $15%, $60% and $80-90% efficiency, respectively ( Figure S5 (M)). The false positive rate of RP i on trajectories simulated without this state was only 1% (4 events in 360 trajectories) and does not thus affect conclusions. To obtain a rough estimate of how much the detection efficiency contributes to the estimated RP i lifetime, we pooled and binned all RP i dwells on different promoters and corrected the bins for the missed RP i dwells. The fit of corrected and uncorrected dwell time histograms to mono-exponential function estimated the RP i lifetime as 0.15 ± 0.02 s (±SE) and 0.22 ± 0.4 s, respectively ( Figure S5(N)). Taking into account the dwell length distribution and detection efficiency, we would expect to identify the RP i state in $60% of trajectories if the RP C ? RP O transition would always proceed via this intermediate. The fact that this number is significantly larger than the measured 14-20% occurrence of RP i raises the intriguing possibility that the RP O formation pathway on lacCONS promoter is branched, i.e., one path is a direct RP C ? RP O transition while the other path involves the RP i as an intermediate between these states.
To compare the mean FRET of the RP i intermediate to that of the RP C and RP O complexes, we extracted FRET efficiency histograms from HMM-segmented trajectories for each of the three states ( Figure 3(B)). The mean FRET values, obtained as the centres of the fit Gaussian distribution (Eq. (2)), were found as 0.19 6 ± 0.001 (±SE), 0.318 ± 0.002 and 0.448 ± 0.001 for the RP C , RP i and RP O complex, respectively. The mean FRET values suggest that the average distance between the À15 and +15 labels in the RP i state has become shorter than in the RP C complex but remains still longer than in the mature RP O complex.
To probe the structural origin of the RP i state, we included in the reaction buffer the RNAP inhibitor myxopyronin B (Myx). Biochemical and structural studies using Myx 35 and structural studies using corallopyronin A (Cor), 9 a Myx analogue, have suggested that this class of RNAP inhibitors block the formation of RP O complex by preventing the loading of template DNA strand into the active site cleft. We observed that the FRET distribution of the RNAPpromoter complexes preformed in the presence of Myx was described by two Gaussians with mean FRET values 0.231 ± 0.001 and 0.307 ± 0.010 (  Figure S6(A)). Interestingly, a sub-fraction (N = 19/58) of these molecules stochastically sam-pled a very short-lived, i.e., typically 20-40 ms (1-2 frames), higher E* state ( Figure S6(B)). The second class of molecules involved potential nonproductive complexes as indicated by a stable E* $ 0.2 value (N = 37, 38%, Figure S6  binding trajectories in the presence of Myx inhibitor demonstrated the formation of initial RP C complex (E* $ 0.2) and its subsequent isomerisation to E* $ 0.3 state ( Figure S6(F)) while the remaining 68% (N = 45) of the nascent RNAP-promoter complexes maintained the E* $ 0.2 state for the entire duration of the trajectory ( Figure S6(G)). The increasing trend in observed FRET values as the RNAP-promoter complexes react towards RP O is consistent with structural modelling data; the distance between the À15 and +15 labels decreases from 98 A in the RP C complex, to 87 A in the Cor-stabilised RNAP-promoter intermediate, and further to 66 A in the RP O complex ( Figure S6 (H)).

Transcription bubble opening leads to static and dynamic RP O complexes
We next analysed the transcription bubble behaviour immediately after initial RP O complex formation; our observation span for these measurements was 1.  To evaluate the bubble conformations accessed by the dynamic RP O complexes, we analysed the E* distribution of these complexes, which was fit well by two Gaussian functions with mean E* values of 0.265 ± 0.004 and 0.467 ± 0.001 (Figure 4(D)). In contrast, the E* histogram of the static RP O showed only a single distribution centred at E* of 0.448 ± 0.001 (Figure 3(B)). These values suggest that the dynamic RP O complexes do not sample either the RP C (E* $ 0.20) or the on-pathway intermediate RP i (E* $ 0.32) states. Instead, it is more likely that the dynamic RP O complexes sample an offpathway state, which is characterised by E* $ 0.27 and which we coin as RP ISO .

The role of RNAP-promoter upstream interactions in the RP O pathway selection
We next evaluated how RNAP aCTD-promoter upstream sequence interactions contribute to the relative formation of stable and dynamic RP O complexes and the kinetic parameters of the transcription bubble dynamics ( Figure S7(A)). To this end, we prepared RP O complexes at 37°C using either a short or a long dsLC2, challenged them with heparin, i.e., a DNA competitor, and immobilised them on the coverslip surface for smFRET analysis. In this protocol, RP O complexes also form on the short dsLC2 promoter fragments, allowing direct comparison to the long dsLC2 promoter. HMM-based classification of the trajectories indicated that the dynamic RP O complexes were 1.7-fold more prevalent (25 ± 3.6% vs 15 ± 2.7% of all complexes; mean and SD of three independent experiments) on the short promoter compared to the long promoter ( Figure S7(B)). A two-sample T-test (p = 0.035) confirmed that the observed difference in the relative number of dynamic complexes on the two promoters is statistically significant. Further, the observation span for the measurements was 2.0-25 s (median 8.0 s, N = 348) on the long promoter and 1.4-25 s (median 6.9 s, N = 435) on the short promoter, suggesting that the higher prevalence of dynamic RP O complexes on the short promoter is not explained simply by longer trajectories that are expected to accumulate more state transitions.
As an additional control, we estimated the identification accuracy of dynamic/static RP O complexes in simulated trajectories, which had state lifetimes and FRET noise levels similar to the experimental data. HMM analysis results indicate that the detection efficiency of dynamic RP O complexes is $95% when the trajectory length is !4 s (a length observed for 84% of experimental trajectories) and remains at 84% when the trajectory length is !3 s (a length observed for 92% of experimental trajectories) ( Figure S7(C)). These numbers indicate that the trajectory length variation in the experiment does not significantly bias the static/dynamic RP O estimations. Noteworthy, the analysis of simulated trajectories did not produce any false positive dynamic RP O 's.
Kinetic analysis indicated that the lifetimes the RP ISO (85-90 ms, Figure S7(D, F)) and RP O (560-660 ms, Figure S7(E, G)) states were similar on both promoters ( Figure S7(H)). Collectively, our results suggest that the aCTD-promoter interactions steer the RP O pathway selection towards the static RP O complex; however, the effect is moderate, and significant number of dynamic RP O complexes form also on the long promoter. The similarity in the timescales of transcription bubble dynamics on the short and long promoters indicates that the bubble isomerisation rates are independent of the status of the aCTD-promoter upstream sequence interactions.

The role of RNAP cleft loops in the RP O stabilisation
To interrogate protein structural elements contributing to the RP O stability, we deleted the b gate loop (DGL), b' rudder loop (DRL) or b' lid loop (DLL) from the RNAP and determined the effects of these deletions on the transcription bubble dynamics using preformed RP O complexes. Structurally, GL mediates the RNAP b-pincer interaction with the RNAP b'-clamp, thus forming a barrier for the DNA entry and exit from the RNAP cleft ( Figure 5(A)) 9,10 ; RL locates between the tDNA and ntDNA strands in the RP O ; and LL locates adjacent to the RL and is able to interact with the tDNA around base À6 in the RP O .
We used 2-state HMM of FRET trajectories to classify RP O complexes into static (i.e., no bubble dynamics) and dynamic ( Figure 5(B)) and found that all deletions shifted the balance of RP O formation towards the direction of dynamic RP O complexes. The effects of DGL and DLL were moderate, as they increased the fraction of dynamic RP O complexes from the wild-type (WT) RNAP level by 1.3-1.4-fold, i.e., from 26% (exact 95% binomial CI: 22-31%) in WT to 34% (CI: 29-40%) in DLL and 36% (CI: 30-43%) in DGL ( Figure 5(C)). However, the DRL effect was much stronger, 2.9-fold, making most RP O complexes, i.e., 77% (CI: 72-83%), dynamic. The addition of NTPs to DRL-RNAP-promoter complexes depopulated the RP O state as indicated by the substantial decrease in E* $ 0.45 state probability and corollary increase in $0.2 E* state ( Figure 5 (D)). Because the NTP-response was similar in the case of WT-RNAP ( Figure S2(C)), which in contrast to DRL-RNAP forms mostly static RP O 's, both static and dynamic RP O 's appear capable to initiate RNA synthesis. Kinetic analysis of the dynamic RP O 's indicated that DRL and DLL shortened the lifetime of the open bubble state by 2.8-fold, whereas DGL had no effect ( Figure 5(E,  F)). In contrast, all three mutations increased the lifetime of the RP iso state. Specifically, the RP iso lifetime increased by 2.5-, 1.8-and 1.5-fold in the case of DLL, DRL and DGL mutant RNAPs, respectively ( Figure 5(G, H)). Notably, the median observation span for these measurements was 4.26 s, 4.84 s, 5.20 s and 6.00 s in the case of WT, DLL, DGL and DRL, respectively. The variation in the observation span, however, does not significantly contribute to the classification of the molecules to the static and dynamic RP O classes or kinetic parameters because, even in the combination of shortest trajectories (median 4.26 s) and the most stable state, i.e., wild-type RP O complex (lifetime 0.5 s), the probability that RP O ? RP ISO transition does not take place within the trajectory is extremely small (0.0002). Collectively, our data indicate that the GL, RL and LL domains in the RNAP contribute to the RP O pathway branching (by changing the balance between static and dynamic RP O complexes) and affect the rates of the transcription bubble conformation changes.

Discussion
In this work, we establish the ability to look at the earliest stages of transcription initiation in real-time and at the single-molecule level. This unique capability bypasses the need to synchronise complexes and offers unprecedented access to co-existing reaction pathways and transient intermediates; as a result, we gained new mechanistic insight about the paths and intermediates used by RNA polymerase to form RP O complexes on a lac promoter derivative. Our work also provides further insight on the dynamics and heterogeneity of open complexes and their determinants.
Open complex formation may proceed via more than one path. Our data indicate that the RP C ? RP O isomerisation step involves mechanistic branching. In most cases ($60% of molecules), the isomerisation occurs in one step without observable intermediate(s) within our 20ms temporal resolution, suggesting that bubble  Possible structure of RP i . We first considered that possibility that the intermediate may have structural similarities to an RNAP-promoter complex stabilised by an antibiotic targeting RNAP. Recent cryo-EM data of Mycobacterium tuberculosis RNAP showed that corallopyronin A (a Myx analogue) stabilises a partially melted transcription bubble (region À11/À4). 9 The same cryo-EM work showed that a similar conformation was present in the absence of Myx, raising the possibility that the observed conformation may represent an intermediate on the RP O pathway. 9 However, our real-time trajectories using E. coli RNAP show an intermediate (RP i ) with a structural signature with significantly different FRET efficiency (E* $ 0.3) than that expected by the partially melted intermediate (E i2 $ 0.2). Further, the presence of Myx does not prevent full opening of the promoter DNA, as sensed by fluorescence enhancement of a Cy3 probe introduced at position +2, both at the ensemble 35 and at the single-molecule level 36 ; we note that such enhancement is expected only when the bubble opening has been complete.
Instead, the possibility we favour for the structure of the RP i is an intermediate further down the promoter opening pathway (hence, more consistent with the E* value of $0.3 we observe for RP i ), where all the melting has been completed, but the template DNA has not been fully loaded to the active site cleft; such a structure is supported by results showing that Myx does not prevent full opening of DNA on two prototypical promoters (kP R and lacCONS). 36 Regardless, at least for the lac promoter derivative used in this work, the intermediate is kinetically significant only in a subset of RP O formation events.
Possible sources for heterogeneity in open complex formation pathways. Since the RP i is detectable only in a subset of trajectories, it is conceivable that, for those trajectories, a structural module of the RNAP delays downstream tDNA loading to the active site cleft. One candidate for such a structural module is the RNAP b 0 switch-2 segment. Based on mutational analysis, structural studies, and observed Myx effects, it has been established that the switch-2 region can adopt two different conformations. 37 The conformation dominant in the absence of Myx is compatible with tem-plate loading to the active site; in contrast, the alternative conformation (which is stabilised by Myx and specific mutations in the switch-2) blocks template loading to the active site. If the switch-2 was at the moment of DNA bubble initiation in the blocking conformation in a fraction of RNAPs, the loading of template DNA to the active site cleft would be delayed by the necessary switch-2 refolding.
The RP i heterogeneity may also result from alternative promoter discriminator (region À6/ +1) conformations in different RNAP molecules. This hypothesis is supported by the previous finding that G À6 G À5 G À4 and C À6 C -5 C/T À4 motifs in the ntDNA stabilised in crystallo two distinct discriminator conformations and imposed in solution one base-pair difference in the predominant transcription start site. 38 GTG motif, which is found in our promoter, had transcription start site statistics halfway between the GGG and CCC/T motifs, consistent with the assumption that a promoter with this motif can readily adopt either of the two discriminator conformations.
RP O complexes appear to have different stability immediately after their formation. A longstanding question in the transcription field is whether all RP O on a given promoter are the same or differ in their structural and functional properties. [22][23][24][25] Our results show that indeed there is another layer of heterogeneity as indicated by the differing stability of the transcription bubble, even immediately after RP O formation (as judged by the appearance of the E* $ 0. 45  The analysis of mutant RNAPs suggest that the main difference between the dynamic and stable RP O complexes arises from the RNAP interaction with the single-stranded template DNA in the active site cleft. The deletion of rudder loop, which presses against the template DNA positions À7 to À9, tripled the amount of dynamic RP O (from 26% to 77%; Figure 5(C)) and decreased 3-fold the open bubble lifetime in the dynamic RP O complexes ( Figure 5(F)). The deletion of lid loop, which interacts with the template DNA base À6, had a similar effect on the open bubble lifetime. We previously found that deletion of the r 70 3.2 region (the r "finger", which interacts with the template DNA strand from bases À3 to À6), destabilised the RP O . 33 These interactions with the template DNA form late in the RP O mechanism, i.e., when the bubble forms fully and the template DNA strand loads to the active site cleft. Our data suggest that these final interactions form by two alternative ways generating "tight" and "loose" template DNA binding modes: the tight binding mode gives rise to the stable RP O complexes, and the loose binding mode gives rise to the dynamic RP O complexes.
It will be interesting to investigate in future singlemolecule studies whether such significant differences in RP O stability have functional consequences, and whether they are related to reports of non-uniform RP O function. Specifically, a subset of RP O complexes (on many promoters) appear to be locked in an abortive initiation mode, in which they repetitively synthesise short RNA products (<12-mer, with the exact sequence depending on the specific promoter), whereas another RP O subset escape the promoter efficiently and synthesise full-length RNA products. [22][23][24][25] The failure of promoter escape may also result from long-lived backtracking and arrest of initially transcribing RNAPs. 28,29 The RP O 's apparently locked in the abortive mode are also referred as "moribund" complexes, and they apparently have a role in transcription regulation in the cell. 27 Mechanistically, the dynamic RP O could be candidates to form such moribund complexes, since unstable template DNA binding to the active site is likely to enhance the dissociation probability of short RNAs, leading to abortive initiation. Consistent with this reasoning, the D3.2 r 70 mutant (which increase RP O dynamics substantially) released 4-7-mer RNAs more efficiently compared to the WT. 24 However, it has also recently been suggested that the intermediate RP i3 and not the stable RP O is the productive initiation complex on the kP R promoter. 39 Role of the promoter upstream interactions on the RP O formation and stability. The RNAP aCTDs interact with the promoter upstream sequences either by specifically recognising the promoter UP element 14 or via sequenceindependent interactions 14,19 ; both interactions are important for RP O formation. We found that the upstream part of the lacCONS promoter (from À40 to À89; Figure S2(A, B)), which does not contain a full UP element but is partially similar to the distal UP subsite, 41 facilitates transcription bubble melting in the context of fully double-stranded pro-moters. In fact, the short double-stranded promoter (which lacks aCTDs interactions) failed to form RP O under our experimental conditions, which involve measurements at room temperature. On the other hand, if the requirement for the DNA melting nucleation was bypassed (e.g., by using a pre-melted promoter), the aCTD interactions with upstream sequence no longer increased the rate of RP O formation (Figure 2(D, E)). This finding is consistent with previous biochemical studies showing that the aCTD interactions with the UP element enhance both the initial promoter binding and subsequent isomerisation to competitor-resistant conformation. 14, 19 We also found that the presence of upstream sequence interactions did not substantially change the ratio of initial bubble opening events that occur in one step or in two steps (i.e., via the RP i ), and did not significantly change the rates of transcription bubble dynamics in the pre-formed RP O . However, the dynamic RP O complexes formed more often on the short promoter in comparison to the long promoter (25% vs. 16%), suggesting that the aCTD-promoter interactions, instead of being fully decoupled from mechanistic steps occurring after the bubble nucleation, have a role in the late steps of RP O pathway branching; the exact mechanism of such modulation is unclear, but it may involve the bending of the upstream sequence on the RNAP surface, as observed by Record and collegues, 17 and subsequent interactions that affect RNAP conformation dynamics in a way that it influences bubble dynamics. Our promoter sequence has only a partial similarity to the distal UP element subsite, predicting a non-ideal (if any) sequence-specific interaction with aCTDs. It has been shown that promoters with a full UP element drive stronger transcription activity than promoters with partial UP element or non-UP element sequence. 41 To generalise our findings, further studies on many different promoters are needed to establish the UP element and general sequence-dependence of the promoter upstream control over the formation of static vs dynamic RP O 's.
A working model for the RP O formation mechanism on a bacterial consensus promoter. We summarise our key findings in the context of the RP O formation mechanism on the lacCONS promoter in Figure 6. 2 The process begins with the RP C complex formation as the RNAP holoenzyme binds to the promoter and establishes interactions with the À35 element, À10 element and upstream sequences. Interaction of aCTDs with upstream sequences stimulate RP O formation by bending the upstream DNA around the RNAP 12-17 and coupling it energetically with bubble formation.
The initial nucleated bubble expands via two different mechanistic paths: in the first path (most common for our lac promoter derivative), the RNAP melts the entire bubble and loads the template DNA strand to the active site cleft in one apparent step without detectable intermediates; the second path, however, involves a short-lived intermediate, RP i , which features incomplete template loading to the active site cleft. We hypothesise that the intermediate appears when a mobile element of the RNAP, e.g., the switch-2 module, is initially in a conformation incompatible with template loading to the active site cleft. Template DNA loading to the active site cleft leads to the tight-binding and loose-binding states, which do not readily interconvert. Because stable and dynamics RP O complexes formed with similar probability both directly from RP C and via RP i , we assume that these pathways merge before the branching to the stable and dynamics RP O 's takes place at the template DNA loading step ( Figure 6). The tight template DNA binding mode keeps transcription bubble open whereas the loosebinding features dynamic movement of the template DNA. Template DNA interactions with the RNAP rudder loop and r finger are part of the key determinants of tight binding mode. Ongoing studies in our group aim to decipher the promotersequence dependence of the RP O formation mechanism and functional significance of the RP O heterogeneity.

Protein preparation
Escherichia coli core RNAPs were expressed in E. coli and purified as previously described. 42 The wild-type RNAP was expressed from plasmid pVS10 (T7p-a-b-b 0 _His 6 -T7p-x). 43   The reaction pathway from the promoter binding to the RP O complex has heterogeneity in two separate steps. The first step is hypothesised to depend on a mobile RNAP element, which can be either in an active or inactive conformation (green and red flaps, respectively). The inactive conformation blocks the loading of the tDNA strand into the active site cleft, resulting the formation of intermediate RP i . The isomerisation of the mobile element to the active conformation clears the block and allows progress from RP i to RP O . The second branching is related to the stability of the RNAPtemplate DNA interaction in the active site cleft. In $15% of the RP O complexes, the interaction is weak, allowing continuous dynamic movement of the template DNA and thus the downstream DNA. Because stable and dynamics RP O complexes formed both from RP C1 and RP i , we assume that these pathways merge before the next branching step leading to the stable and dynamics RP O 's. Green vs red pins depict tight and loose interactions between the tDNA and the RNAP, respectively. The numeration (1, 2 and 3) indicates the key steps in the mechanism that may define the rate and efficiency of RP O complex formation. expressed from pMT041, pHM001 and pTG011, respectively. 44 Wild-type E. coli r 70 was purified as previously described. 45 Holoenzymes were assembled by incubating 0.5 mM RNAP with 1.5 mM r 70 for 15 min at 30°C in RNAP storage buffer [20 mM Tris-HCl (pH 8.0), 150 mM NaCl, 50% (vol/vol) glycerol, 0.1 mM EDTA, 0.1 mM dithiothreitol (DTT)].

Microscope coverslip preparation
Borosilicate glass coverslips (1.5 MenzelGlä zer, Germany) were heated to 500°C in oven for 1 h to reduce background fluorescence. The coverslips were then rinsed with HPLC-grade acetone and immerged into 1% (v/v) Vectabond (product code #SP-1800, Vector Labs, CA, USA) in acetone for 10 min to functionalise the glass surface with amino groups. Coverslips were then rinsed with acetone followed by deionized water before drying them under a stream of nitrogen gas. A silicone gasket (103280, Grace Bio-Labs, OR, USA) with four reaction wells was placed in the middle of the coverslip. The coverslip surface was then simultaneously passivated by pegylation against unspecific protein/DNA binding and biotinylated to provide attachment points for specific protein immobilisation. Each well on the coverslip was thus filled with 20 ml of 180 mg/ml methoxy-PEG (5 kDa)-SVA (Laysan Bio, AL, USA) and 4.4 mg/ml biotin-PEG (5 kDa)-SC (Laysan Bio, AL, USA) in 50 mM MOPS-KOH buffer (pH 7.5), incubated for $ 3 h at room temperature and finally the wells were thoroughly rinsed with phosphate-buffered saline (PBS; Sigma Aldrich, UK). The coverslips remained functional for at least two weeks when stored at 4°C in plastic pipette tip box containing a layer of deionised water at the bottom. During the storage the coverslip wells were filled with PBS.
To analyse the RP O complex formation in realtime at 22°C the anti-His-tag-antibody coated coverslip was incubated $10 min with 1 nM labelfree holoenzyme in the reaction buffer, rinsed thoroughly with the reaction buffer and mounted on the microscope. 25 ml of imaging buffer [i.e., reaction buffer supplemented with 2 mM UVtreated Trolox, 1% (w/v) glucose, 0.4 mg/ml catalase (10106810001, Roche Diagnostics, Germany), 1 mg/ml glucose oxidase (G2133, Sigma Aldrich, UK)] was replaced to the imaged well. Data recorder was started to take an 80 s movie. 1 ml of 4 nM promoter in the reaction buffer was gently pipetted to the well at $8 s time-point. The formation of RNAP-promoter complexes was evident by the appearance of bright co-localised spots on the Cy3B and ATTO647N fluorescence channels. In some experiments these surfaceformed RNAP-promoter complexes were further imaged after exchanging fresh imaging buffer to the well and finding non-bleached field-of-view. The age of RNAP-promoter complexes at the moment of recording these 20 s post-binding movies was $3-7 min. In some control experiments, we monitored the initial RNA synthesis activity of RNAP by including 1 mM NTPs (ATP, GTP, CTP and UTP) in the imaging buffer.
To analyse transcription bubble dynamics in preassembled RP O complexes, 2 nM holoenzyme was incubated with 5 nM promoter in reaction buffer for 15 min at 37°C. 100 mg/ml sodium heparin (H4784, Sigma, UK) was added to disrupt nonspecific RNAP-promoter complexes and $1.3 ml of the mixture was transferred to anti-His-tagantibody coated coverslip well containing 25 ml reaction buffer. The RP O complex immobilisation at 22°C was let to continue until $50 molecules were detected on the field-of-view. The well was then rinsed with reaction buffer and finally filled with 25 ml imaging buffer. Data was recorded as 20 s movies from about ten field-of-view per well at 22°C.
Single RNAP-promoter complexes were imaged using objective-based single-molecule TIRF microscope previously described. 46 The donor (Cy3B) and acceptor (ATTO647N) fluorophores in the promoter were excited using 532 nm and the 642 nm lasers in alternating laser excitation (ALEX) mode, respectively. 47 The emission of donor and acceptor fluorophores was separated from each other and from the excitation light, using dichroic mirrors and optical filters, and recorded side-byside on an electron multiplying charge-coupled device camera (iXon 897, Andor Technologies, Northern Ireland). The frame time of the recordings was 20 ms with 10 ms ALEX excitation by each laser. The measured laser power before the dichroic mirror was set to $4 mW and $1 mW for the 532 nm and 642 nm laser, respectively.

Single-molecule data analysis
To extract the intensities of co-localised donor and acceptor fluorophores, the recorded movies were processed with custom-made TwoTone TIRF-FRET analysis software (46; see also https://groups.physics.ox.ac.uk/genemachines/ group/Main.Software.html). If the processed movies had fluorescent complexes on the surface already at the beginning of the movie, i.e., post-binding and pre-formed RP O complex movies (see above), the following Twotone-ALEX parameters were applied to select only complexes containing a single ATTO647N acceptor dye and a single Cy3B donor dye: channel filter as DexDem&&AexAem (colocalisation of the donor dye signal upon donor laser excitation, the acceptor dye signal upon acceptor laser excitation), a width limit between the donor and the acceptor between 1 and 2 pixel, a nearest-neighbour limit of 6 pixels, and signal averaging from the frames 2-40. In the case of real-time RP O complex formation movies, the nearestneighbour limit was turned off and the timewindow for the search of the surface-bound fluorescent molecules was set with the signal averaging setting (i.e. typically frames $1000-3000) to the part of the movie in which most promoter binding events took place. The trajectories selected by the TwoTone-ALEX analysis were manually sorted by eliminating all traces that displayed extensive fluorophore blinking, multi-step photobleaching indicating more than one donor or acceptor dye in the same diffraction limited intensity spot, or other aberrant behaviour.
The apparent FRET efficiency (E*) was calculated using Eq. (1) where I DD and I DA are the emission intensities of the donor (Cy3B) and acceptor (ATTO647N) dyes upon donor excitation (532 nm), respectively. 48 The trajectories were analysed using a modified version of the hidden Markov model ebFRET software. 49,29 The trajectories from the pre-formed RP O or post-binding movies were fit to 2-state HMM model followed by noise filtering by requiring an accepted dwell time to satisfy the criteria that the step (i.e., change in E*) is separated from the subsequent step by more than 3-fold the Allan deviation. 50,29 The trajectories were then classified into dynamic or static populations depending whether they displayed >2 or 2 accepted E* transitions, respectively. The dwell times were extracted from the dynamic trajectories to compile a dwell time distribution. The dwells with undefined length, i.e. the first and last dwell, were discarded at this point.
The trajectories from the real-time RP O formation movies were analysed separately for the first transcription bubble opening event, i.e., the RP C ? RP O transition, and transcription bubble dynamics after the RP O formation. The latter analysis was identical to the case of pre-formed RP O complexes with the exception that the RP C ? RP O transition at the beginning of the trajectory was trimmed away before HMM. In contrast, the analysis of the RP C ? RP O transition in the trajectories was performed after trimming away possible bubble dynamics subsequent to the first transcription bubble opening event. We fit the first bubble opening trajectories using 2-state HMM, i.e., RP C ? RP O mechanism, and 3-state HMM, i.e., RP C ? RP i ? RP O mechanism. The initial fits were filtered by requiring true state transitions to be at least 2-fold the Allan deviation. 50,29 Selection of the more complex 3state model for the trajectory also required that both the HMM lower bound value and Aikake information criteria, calculated as previously described, 51 favoured this model. The dwell times in the RP C and RP i states were compiled to separate dwell time distributions. The lifetime of the RP C state was determined by fitting the dwell time distribution to the mono-exponential decay function using Origin software (OriginLab Corporation, MA, USA).
We validated the above analysis procedure for its accuracy to detect the RP i state and dynamics RP O 's. To this end, we used DeepFRET software 52 to simulate FRET trajectories for each state RP C , RP i and RP O using FRET efficiency and FRET noise levels extracted from the experimental trajectories. Specifically, the FRET efficiency was 0.196 (noise 0.05), 0.318 (0.06) and 0.448 (0.05) for the simulated RP C , RP i and RP O state, respectively. The noise of the corresponding experimental FRET data was 20-28% smaller, i.e., 0.036, 0.047 and 0.040 for the RP C , RP i and RP O state, respectively. The complete trajectories for a RP C ? RP i ? RP O mechanism, were then assembled from the state-specific simulation by using a custom Python script and monoexponential distribution of state dwell lengths (as in experimental data). The trajectories to determine the detection efficiency of dynamic RP O complexes, i.e., complexes showing RP ISO $ RP O dynamics, were simulated by DeepFRET using FRET efficiency setting 0.279 (noise 0.06) and 0.462 (noise 0.06) for the RP ISO and RP O state, respectively. The mean lifetime of the RP ISO and RP O state complex was set as 0.085 s and 0.56 s, respectively. The trajectory length was 2-7 s (100-350 frames).
The histograms of E* values were fit in Origin software to one or two Gaussian distributions using Eq. (2) with n fixed as 1 or 2, respectively. The fit parameters E c *, w and A are the centre, width and area of the Gaussian distributions, respectively.