RNA switches with encoded folding paths depicted in Figures 1 and 2 were designed using Kinefold online server (45,46) (). A starting GG doublet was chosen to ensure efficient T7 transcription. The sequence of the ‘direct’ RNA switch (i.e. 5′-ABCD-3′, 73 nt, Figure 2) is: 5′-GGAACCGUCUCCCUCUGCCAAAAGGUAGAGGGAGAUGGAGCAUCUCUCUCUACGAAGCAGAGAGAGACGAAGG-3′. The ‘reverse’ RNA switch has exactly the opposite sequence (or reversed orientation), i.e. 5′-DCBA-3′. A ‘reverse’ sequence with a single mutation U₃₈/C₃₈ was also studied to unambiguously establish the correspondence between branched versus rod-like structures and the two migrating bands on non-denaturing polyacrylamide gels (Figure 3); i.e. 5′-GGAAGCAGAGAGAGACGAAGCAUCUCUCUCUACGA- GGC₃₈AGAGGGAGAUGGAAAACCGUCUCCCUCUGCCAAGG-3′. Complementary DNA oligonucleotides including T7 promoter and KpnI/StuI/BamHI restriction sites at sequence extremities were bought from IBA-Naps, Germany.

Sequences were inserted into pUC19 plasmid (between KpnI and BamH1 restriction sites) using enzyme removal kits (Qiagen) and cloned into calcium competent E.coli (DH5α strain) following standard cloning protocoles. Following plasmid extraction (Genomed kit), inserts were sequenced and cut at the StuI blunt restriction site located at the end of the desired DNA template. Run off transcription was performed in vitro using T7 RNA polymerase (New England Biolabs) for up to 4–5 h at 37 or 25°C. Heat renaturation was performed from 85°C to room temperature in ∼10 min (starting from 95°C gave the same results). Renatured and co-transcriptional native structures were then separated on 12% 19:1 acryl-bisacrylamide non-denaturing gels (1× TAE, temperature <10°C) and observed using ethidium bromide staining (0.1 μg/μl): Ethidium bromide slightly re-equilibrates the molecule equilibrium partition between branched and rod-like structures during heat renaturation but has no measurable effect at room temperature on the strongly biased partition between co-transcriptional native structures. Controls using denaturing gels (6%, 19:1 acryl-bisacrylamide, RNA in formamide and 8 M urea, 50°C) showed that >90% of transcripts had the expected run off transcription length. Virtually no other bands (i.e. <5% of total transcript) were observed on non-denaturing gels either (i.e. apart from the single or double bands shown in Figures 2–4).

We decided to investigate the basic mechanisms and coding requirements for efficient RNA folding paths with a stringent test case. Following the RNA switch design depicted on Figure 1, we set out to encode two (oppositely oriented) folding paths on the same RNA sequence. The proposed bistable RNA switch should form a branched structure (1) and a rod-like structure (2) with approximately the same free energy at equilibrium, and yet be guided into either one of these structures only, depending on the direction of synthesis (49), (Figure 1).

In practice, however, 5′–3′ versus 3′–5′ folding paths cannot be probed on the same RNA sequence, as there is no RNA polymerase known to perform transcription in ‘opposite’ (3′–5′) direction. Hence, instead of studying a single RNA sequence, we have actually used a pair of RNA switches with exactly opposite sequences, i.e. 5′-ABCD-3′ and 5′-DCBA-3′ (see Materials and Methods). It is important to note that, in general, such pairs of RNA molecules do not adopt related structures at equilibrium, due to the large asymmetry between free energies of stacking base pairs with reversed orientation (e.g. 5′-GC/GC-3′ ≃ −3.4 kcal/mol and 3′-GC/GC-5′ ≡ 5′-CG/CG-3′ ≃−2.4 kcal/mol). For this reason, the pair of direct (D) and reverse (R) RNA switches, we have designed (Figure 2), forms at equilibrium a branched structure (1) and a rod-like structure (2) constructed around symmetric helices, that are exactly conserved under sequence reversal (dashed lines on Figure 2 indicate mirror symmetry of Pa, Pb, Pc and Pd helices). Thus, comparing transcription products of the direct and reverse sequences probes the directionality of their folding paths while keeping the equilibrium structures of both switches essentially equivalent by symmetry; since structures 1D and 1R (respecively, 2D and 2R) of the direct and reverse switches are built on the same helices Pa and Pb (respectively, Pc and Pd), their sole free energy difference concerns the small sequence dependent contribution of single-stranded regions in the branched (respectively, rod-like) structure (GNRA tetra-loops and all other sequence-dependent tabulated loops have been avoided).

Figure 2 demonstrates that, in spite of these strong similarities between equilibrium structures, the two RNA switches are indeed guided towards two distinct native structures upon in vitro transcription (see Materials and Methods). The correspondence between branched versus rod-like structures and migrating bands on polyacrylamide gels was unambiguously established using a single mutation U₃₈/C₃₈ on the reverse sequence (Figure 2). This mutation stabilizes the branched structure (UG > CG) relative to the rod-like structure (AU < AC) at equilibrium, hence demonstrating the correspondence between branched structure and lower band, (Figure 3). Note, however, that this mutation also perturbs the co-transcriptional folding path by redirecting about half of the molecules into the branched structure, hence illustrating the difficulty to dissect independently folding paths from equilibrium structures with sequence mutations only (see Introduction).

These results strongly support the co-transcriptional folding principles depicted on Figure 1 which primarily rely on the difference in helix length (i.e. ∣Pa∣ > ∣Pc∣ for the direct switch and ∣Pc∣ > ∣Pb∣ for the reverse switch) to code for the co-transcriptional formation of structures 1D and 2R, respectively. This small asymmetry between successive overlapping helices in the direction of transcription induces a divergence in structural cascades between the two co-transcriptional folding paths; namely, the red helix Pc cannot displace and replace the longer (and stronger) helix Pa previously formed by the direct switch during transcription (∣Pa−Pc∣/∣Pc∣ ≃15%), while Pc does displace and replace the shorter (and weaker) helix Pb initially formed by the nascent reverse sequence (∣Pc−Pb∣/∣Pc∣ ≃30%). The efficacy of these folding paths using T7 polymerase appears maximum (≃100%) for the direct sequence while it is ∼90% for the reverse switch at 37°C, suggesting that the branch migration exchange between Pb and Pc is not always successful in these conditions (T7 transcription rate is ∼200–400 nt/s at 37°C); however, we found that the folding bifurcation is almost always achieved (≃100%, Figure 4) at lower temperature 25°C which decreases 3-to 4-folds T7 transcription rate (50, 51). This small improvement in bifurcation efficiencies suggests that the decreasing elongation rate indeed prevails over other opposite kinetic factors at lower temperature. In particular, nucleation of Pc (which probably involves the opening of one or two base pairs of Pb, as sketched in Figure 2) and the subsequent branch migration between Pb and Pc are probably both slowed down at lower temperature.

Overall, this demonstrates that the competition between native and non-native helices can lead to efficient co-transcriptional folding paths of RNA switches independently from their actual equilibrium structures.

Moreover, we found that the folding path of the reverse switch could be significantly redirected toward structure 1R (≃50%) through transient antisense interactions, (Figure 4 right panel). This is simply achieved using a 7 nt long antisense oligonucleotide designed to interfere through competing interactions, with the encoded exchange between helices Pb and Pc, (Figure 5). Note, in particular, that the hybridized antisense probe is eventually displaced by the longer (and stronger) downstream helix Pa of each nascent reverse sequence (no global shift of equilibrium bands is observed between transcriptional folds formed in the presence or absence of antisense probe; Figure 4). This shows that transient antisense interactions can, in principle, control multiple turnovers of redirected folding pathways.

Hence, transient intra- and inter-molecular base pair interactions can efficiently regulate the folding of nascent RNA molecules between alternative long-lived native structures, irrespective of their actual thermodynamic stability. Indeed, once formed, the co-transcriptional structures 1D and 2R remain trapped out-of-equilibrium for more than a day at room temperature (data not shown) demonstrating that these RNA switches can reliably store information on physiological time scales with their co-transcriptionally folded structures. In another context, the ability to control folding between distinct long-lived structures of nucleic acids using electrical (49) or thermal (52) stimuli, instead of transcription, could also lead to nanotechnology applications.

Although our conclusions are based on particular examples of synthetic RNA switches related by sequence reversal and helix symmetries, we want to stress that these strong symmetry constraints are solely instrumental in demonstrating the possible independence between encoded folding paths and low-free energy RNA structures. These symmetries are not directly used nor necessary to achieve efficient co-transcriptional folding. On the contrary, imposing such strong sequence symmetries greatly limits the additional ‘information content’ that can possibly be encoded on the sequence. In the next subsection, we discuss how this use of sequence symmetries can actually be formalized to provide quantitative estimates on the minimum coding requirement for selective folding paths of generic RNA swiches.

In this subsection, we discuss how sequence symmetries can be utilized to estimate necessary base pairing conditions to encode efficient co-transcriptional folding paths. This requires, however, to reformulate base pairing conditions from an information content perspective, following the approach developped for biomolecular sequences (47,48).

In the following, we first establish a simple conservation law for information content. We then argue that upperbounds for the coding requirement of selective folding paths (or other molecular features) can be estimated by restricting the available coding space with strong sequence symmetries. Ultimately, upperbounds on coding requirements are related to the likelihood that a particular feature might arise from natural or in vitro selection.

Let us first recall what the information content of a biomolecule is, before showing how it can actually be estimated for designed RNA switches using sequence symmetries.

The information content I of a functional biomolecule corresponds to the number of sequence constraints that have to be conserved to maintain its function under random mutations (47,48). Expressed in nucleotide unit, the maximum information content that can be encoded on an N-nucleotide-long RNA sequence is precisely I_max = N nucleotides, which define a unique sequence amongst all D^N different RNA sequences with N nucleotides, where D is the size of the coding alphabet (D = 4 for nucleic acids; D = 20 for proteins). The fact that neutral mutations can accumulate on an RNA sequence without altering its function implies I < I_max = N and can be simply translated into a conservation law, I + J = N, where J, the sequence entropy, corresponds to the number of unconstrained nucleotides which generate Ω = D^J RNA sequences with the same function. Hence J = log_D(Ω) and I = N − J = log_D (D^N) − log_D (Ω) (47,48). Although J and I can be inferred by sampling sequence space as demonstrated in (48), their contributions to N do not usually correspond to a simple partition between J ‘meaningless’ and I ‘meaningful’ bases since many sequence constraints actually arise from non-local base–base correlations, as shown, in particular, with base pair covariations between homologous RNA sequences. For this reason, it is usually instructive to quantify information content separately within paired and unpaired regions, when considering RNA structures. Since ∼70% of all bases are usually paired in low-energy RNA structures, these base pairs typically contribute the most to total information content and thereby to the minimum coding requirement for a given RNA function. Hence, in the following, we will focus, for simplicity, on short RNA sequences (e.g. N < 100 nt) and consider, at first, only base paired regions, ignoring both wobble base pairs and unpaired regions.

With these crude initial assumptions, the information content of a short RNA sequence adopting a unique stable secondary structure can be estimated as I ≃ J ≃ N/2, since the first base of each Watson–Crick base pair can be chosen arbitrarily. Hence, overall, short RNA sequences adopting a unique stable secondary structure present a large sequence entropy J = J_u ≃ N/2 that can, in principle, be used to encode additional features such as alternative, low-energy structures (53,54) or possibly other molecular properties like co-transcriptional folding pathways, as shown in the previous subsection. For example, encoding the simple bistable RNA of Figure 1A but with exactly overlapping helices (∣Pa∣ = ∣Pb∣ = ∣Pc∣ = ∣Pd∣) requires that I ≃ 3N/4 nt be fixed once the initial J = J_b ≃ N/4 bases are chosen arbitrarily (e.g. in the first pairing region). Including also as stable structure the pseudoknot constructed around the same four complementary regions (so as to obtain a tristable RNA molecule) then implies that each pairing region is self-complementary and that only J ≃ N/8 bases can be chosen arbitrarily (e.g. in the first half of the first pairing region). Similarly, designing the same bistable RNA of Figure 1A but with symmetrical helices (conserved under 5′–3′ sequence inversion) implies that only around J = J_bs ≃ N/8 bases can be chosen arbitrarily (e.g. in the first half of the first pairing region).

Similar estimates can be made including wobble base pairs (GU and UG) in addition to Watson–Crick base pairs (GC, CG, AU and UA). In that case, the available sequence entropy becomes J_u ≃ log₄ (6) · N/2 ≃ 0.65N for a molecule with a unique structure (i.e. with 6 possible base pairs), while we get for the previous bistable RNA, J_b ≃ log₄ (14) · N/4 ≃ 0.48N (i.e. with 14 possible quadruplet circuits including circular permutations: 2×CGGC, 2×UAAU, 2×UGGU, 4×CGGU and 4×UGAU) or J_bs ≃ log₄ (14) · N/8 ≃ 0.24N with additional symmetric helix restriction, as above. Hence, the possibility of wobble pairs tends to increase sequence entropy J, and to concomitantly decrease the number of sequence constraints I. On the other hand, including a significant fraction of wobble pairs (e.g. one-third) in designed structures tends also to facilitate the formation of unwanted, alternative low-energy structures (e.g. with fewer wobble pairs). Thus, including wobble and WC pairs on an equal footing effectively underestimates coding requirements, since preventing the formation of unwanted alternative structures then requires additional information constraints, especially for longer sequences (e.g. >150 nt). In practice, limiting the fraction of wobble pairs in designed structures can efficiently prevent the formation of alternative structures with limited additional sequence constraint. (e.g. we have used 3 wobble pairs out of a total of 50 bp in structures 1 and 2). This also justifies a posteriori the initial crude estimate we have made based on WC pairs only. Ignoring opposite effects of wobble base pairs and unwanted alternative structures is a reasonable first approximation of global information constraint requirement. More precised estimates are difficult to obtain, in general, and sequence candidates should always be tested for possible alternative structures with an appropriate RNA folding algorithm including base pair stacking free energies (we have used kinefold (46) which also includes pseudoknots and knots in RNA structures). Alternatively, it is usually possible, owing to the available sequence entropy, to implement highly constraining heuristics that prevent the formation of alternative structures. Typical heuristics are based on the limitation of short complementary substring occurrences in the sequence (52,55,56).

The previous coding requirement estimates demonstrate that structural information I and symmetry constraints I_s encoded on an RNA sequence can equivalently limit its available entropy J. In particular, combinations of structure and symmetry constraints can provide tighter upperbounds to possible coding increments I′ of any new feature encoded on the sequence, via the conservation law I + I_s + J = N, i.e. I′ < J = N − I − I_s.

This can be applied to estimate the minimum information that might be required to obtain two efficient opposite folding paths from the generic bistable RNA switch sequence of Figure 1A. From the previous estimate, we conclude that the two efficient folding paths for the direct and reverse sequences do not require to constrain more than ∼1/8th of the 42 bases that are paired in both low-energy structures (i.e. overlapping base pairs). It corresponds to assigning, for each folding path, a maximum of two or three overlapping base pairs not already constrained by the combination of branched and rod-like low-energy structures.

This limited coding requirement concerning overlapping base pairs reinforces, a posteriori, our intuitive design principles focussing, instead, on a few bases paired in only one of the two low-energy structures (i.e. non-overlapping base pairs), Figures 1 and 2. Thus, efficient co-transcriptional folding of the direct switch (Figure 2) primarily relies on the sole terminal GC base pair at the base of Pa to prevent the nucleation of Pc, while the exchange between Pb and Pc for the reverse switch hinges on the AC terminal mismatch at the base of Pb to facilitate the nucleation of Pc.

Hence, if all non-functional sequence symmetries are lifted, we expect that selective folding paths can indeed be readily achieved for a wide class of RNA sequences, as they require little encoded information beyond small asymmetries between alternative helices to guide or prevent their successive exchanges during transcription. Interestingly, this pivotal role of a few unpaired or transiently paired bases for efficient folding paths is also observed for other encoded molecular functions of RNAs. For instance, a few unpaired conserved bases usually prove essential for ribozyme functions or in vitro selected aptamers showing remarkable binding efficiency to specific target molecules (48).

Although many convincing reports have shown the importance of co-transcriptional folding (21,34–40), the underlying mechanisms behind efficient folding paths has remained elusive. As recalled in Introduction, this is mainly due to intrinsic difficulties to probe natural RNA folding paths independently from their equilibrium structures and possibly multiple functions.

In particular, studying the equilibrium folds of increasingly longer 3′-truncated transcripts has been argued to miss important out-of-equilibrium intermediates on the folding path of full-length molecules (57). This problem was, however, circumvented by using circularly permutated variants of wild-type sequences to study the folding pathways of E.coli RNaseP RNA (41,42) and Tetrahymena group I intron (43,44). In this approach depicted with a simple RNA switch on Figure 6, the natural 5′ and 3′ ends of the molecule are genetically connected, while two new ends are engineered from an alternative break point in the circularized sequence. This results in a circular permutation of the various domains on the linearized RNA. The resulting permutation in their transcription order indirectly probes folding paths of the full-length wild-type sequence. In particular, altering the connectivity of the primary sequence may lead to alternative co-transcriptional folds as illustrated in Figure 6. The rationale underlying such circular permutation scenario assumes that co-transcriptional folding primarily relies on sequential folding of the nascent chain into independent native domains with no specific coding for transient base pair interactions during transcription. This suggests, in particular, that co-transcriptional folding favors branched secondary structures (43) (as illustrated in Figure 6) and that the 5′–3′ directionality of RNA transcription plays little role as long as the different native domains can all successively fold during transcription.

In contrast, our results demonstrate that folding paths can efficiently guide RNA transcripts into distinct alternative structures even when competing branched-like conformations exist and could, in principle, form during transcription. This competition between local overlapping helices and even global alternative structures is, in fact, ubiquitous to the folding dynamics and thermodynamics of RNA molecules. For instance, co-transcriptional folding has long been known to induce structural rearrangements as the nascent RNA chain is being transcribed (16,17), demonstrating that transient helices almost inevitably participate in co-transcriptional folding paths. More recently, transient helices were also shown to affect force-induced unfolding paths of single RNA molecules in micromechanical experiments (58) (e.g. for E.coli, 1540 nt long 16S rRNA).

The present study, following earlier stochastic folding simulations reported in (45,46,57,58), demonstrates that this unavoidable competition between alternative base pairs can be exploited to precisely encoded co-transcriptional folding paths of RNA sequences. It is primarily achieved through branch migration exchanges between transient and native helices forming successively as the transcription proceeds from 5′–3′ ends of the sequence. This experimental finding is, in fact, corroborated by a recent statistical analysis of non-coding RNA sequences by Meyer and Miklós (59) who demonstrated the existence of a 5′-3′ versus 3′-5′ asymmetry in the relative positional correlation between native and non-native helices along primary sequences.

Non-coding RNAs typically tolerate a significant number of neutral mutations and co-variations in their sequence, which presumably facilitates their continuous adaptation to environmental changes. From an information content perspective, this tolerance to (concerted) mutations also suggests that (partly) unconstrainted nucleotides may be used to encode other alternative structures and functions on the same RNA sequence, a feature which might have favored the emergence of new functional RNAs and RNA switches in the course of evolution (60,61). In fact, much more functional, as well as non-functional information can be encoded on an RNA sequence. For instance, the pair of RNA switches we have designed (Figure 2) demonstrates that not only alternative structures but also selective folding pathways and strong sequence symmetries can all be encoded simultaneously on the same RNA sequence. Although such strong sequence symmetries are both non-functional and probably too stringent constraints to possibly emerge and adapt through natural or in vitro selection, they can be used to provide quantitative information on other encoded features of interest. For instance, equalling stacking contributions between alternative folds using helix symmetries also provides a powerful differential approach to uncover other free energy contributions from non-canonical tertiary structure motifs. In the present study, small but reproducible differences in band separation between direct and reverse switches at equilibrium (Figure 2) may possibly reflect a difference in tilt angle between helices Pc and Pd, due to differently structured interior loops in their respective rod-like structures.

In this study, we showed that co-transcriptional folding can efficiently guide RNA folding either toward branched structures (as for the ‘direct’ switch, Figure 2) or toward elongated rod-like structures (as for the ‘reverse’ switch, Figure 2) even though their helices are mutually identical. We also argued that only limited information is necessary to encode such selective folding paths for generic RNA switches: it essentially amounts to encoding the relative lengths of helices forming successively during transcription. Moreover, this strict hierarchy between successively exchanging helices can be somewhat alleviated by resorting to topological barriers based on ‘entangled’ helices (i.e. simple co-transcriptional knots) (46). Hence, we expect that the present findings concerning short RNA sequences (i.e. <100 bases) may be applied to design efficient folding paths for a wide class of larger RNA target structures. This could be achieved by encoding different series of local folding events leading to a succession of either rod-like or branched motifs at the 3′ end of the nascent RNA molecule during transcription. Such folding scheme also provides a theoretical frame to analyze selective folding paths of natural non-coding RNAs (57).

Finally, these results suggest that efficient folding pathways might have easily emerged and continuously adapted in the course of evolution the same way functional native structures have done so through mutation drift in sequence space; non-deleterious mutations are mostly neutral and conserve sequence folds and activity, while new functions may occasionally arise by rare hopping between intersecting networks of neutral mutations (neutral networks) (53,60,61). Furthermore, the fact that encoded folding paths may be redirected through transient antisense interactions (Figures 4 and 5) provides simple ‘all-RNA’ mechanisms to regulate the functional folding of RNAs in the absence of any elaborate control at the level of transcription initiation. From an ancestral ‘RNA World’ perspective, this constitutive coupling between RNA synthesis and RNA folding regulation may have also enabled the early emergence of autonomous RNA-based networks relying solely on intra- and inter-molecular base pair interactions. Indeed, RNA molecules cross-regulating their respective encoded folding paths could, in principle, be combined to perform essential regulation tasks, characteristic to all natural and engineered control networks (e.g. negative and positive feedback loops, feedforward loops, toggle switches, oscillators, etc.).