In silico unwinding of Caenorhabditis elegans microRNA duplexes to evaluate thermodynamic end stabilities improves predictions of microRNA strand selection

ABSTRACT

microRNAs (miRNAs) are endogenous ~22 nucleotide long, non-coding RNAs that post-transcriptionally regulate gene expression. During miRNA biogenesis, stem-loop-containing miRNA precursors are enzymatically cleaved to form a small RNA duplex. Cleavage positions are determined based on the position of structural motifs and junctions on the stem-loop precursor. The duplex end containing a favourable 5’ nucleotide and lower thermodynamic stability is subsequently loading into an Argonaute protein. Typically, one duplex (guide) strand is retained in Argonaute and becomes functional whereas the other (passenger) strand is degraded. Therefore, accurate structural predictions of miRNA intermediates and quantification of duplex end stabilities are important towards understanding miRNA biogenesis. Here, we compiled predicted secondary structures for all Caenorhabditis elegans miRNA hairpins and duplexes at physiologically relevant temperatures. We developed a new approach to calculate the thermodynamic stability of miRNA duplex ends, which resulted in improved predictions of miRNA strand selection. Our approach introduces hard constraints to folding algorithms to restrict base-pairing of terminal nucleotides, which improves modelling of in vivo duplex end unwinding. We propose that constrained RNA folding can be used to evaluate local stabilities within an RNA secondary structure.

Introduction

While approximately two-thirds of the human genome is actively transcribed into RNA, less than 3% of the human genome appears to encode for proteins [1,2]. A large fraction of transcribed loci corresponds to non-coding RNAs, which can be broadly categorized into two major classes according to their length: long non-coding RNAs (typically defined as ≥ 200 nucleotides long) and small non-coding RNAs (≤200 nucleotides) [3]. The secondary structure of non-coding RNA molecules often defines their biological activities [4]. Dysregulation of non-coding RNAs, including mutations expected to impact secondary structure, have been reported in several human diseases including cancer [5–8]. Long non-coding RNAs can have extensive secondary or tertiary structures that contribute to diverse biological functions including regulation of gene expression, control of protein activity, or serving as scaffolds for the assembly of biomolecular complexes [9,10]. Some small non-coding RNAs, such as ribosomal RNAs (rRNAs) or transfer RNAs (tRNAs), have a secondary structure that is critical for their biological activities. Other classes of small non-coding RNAs, including microRNAs (miRNAs), play repressive roles in gene expression by guiding the RNA-induced silencing complex (RISC) to specific mRNA targets [11,12].

As miRNAs rely on intermolecular base pairing to identify mRNA targets, an extensive secondary structure would likely be prohibitive towards target recognition. However, during miRNA biogenesis, the secondary structure of miRNA precursors plays critical roles in generating the correct mature miRNA sequence. Primary miRNAs (pri-miRNAs) are genomically encoded as single-stranded RNA species that contain a self-complementary stem-loop segment [13,14]. During canonical miRNA biogenesis, the microprocessor complex cleaves pri-miRNAs at a defined position by measuring the length of the base-paired stem relative to unpaired structural junctions on the pri-miRNA to produce an ~ 70 nucleotide hairpin precursor miRNA (pre-miRNA) [15–23]. Additional sequence features and motifs have also been identified that influence the accuracy of microprocessor cleavage [24–27]. Dicer subsequently cleaves the pre-miRNA to remove the terminal loop from the hairpin and generate an ~ 22 nucleotide miRNA duplex [28–32]. In addition to structural elements of pre-miRNAs, Dicer recognizes sequence features including the GYM or YCR motifs to identify the correct cleavage position [33–41]. The miRNA duplex is then loaded into an Argonaute protein, and a single, mature strand (guide/miR strand) is retained in the Argonaute, while the other strand (passenger/miR* strand) is degraded [29,42–46]. The MID domain of Argonaute prefers to load the duplex end containing a more favourable 5’ nucleotide (U > A>G>C) and lower thermodynamic end stability [47–50]. However, these guidelines are not sufficient to describe the strand preference of ~20–25% of miRNA duplexes where the guide strand originates from the duplex end containing a less favourable 5’ nucleotide and higher relative thermodynamic stability [51], suggesting that additional cis (structural, thermodynamic or sequence-based) features or trans-acting factors likely influence miRNA strand choice. Thus, accurate calculations of free energy within miRNA duplexes may facilitate identification of additional features impacting strand selection.

Understanding the various thermodynamic and structural elements influencing miRNA biogenesis relies on accurate predictions of RNA secondary structure and stability. As RNA folding is largely dependent on intramolecular base-pairing, the secondary structure of RNA molecules can be predicted from their primary nucleotide sequence [52,53]. Prediction of RNA secondary structure relies on identifying the minimal free energy (MFE) structure of the folded RNA, which can be calculated using experimentally derived parameters for neighbouring nucleobase-stacking effects on the thermodynamic stability of different RNA helices [52,54–58]. The nearest neighbour energy model has been used for algorithmic prediction of RNA secondary structure, including by the widely used ViennaRNA package [59]. The nearest neighbour model uses an additive sliding window approach to calculate free energy, where the thermodynamic stability of dinucleotide pairs is summed across the length of an RNA structure [55,56,59]. The free energy of RNA is also directly linked to temperature, and temperature-sensitive differences in RNA folding allow for alternative functions of ‘RNA thermometers’ in bacteria, yeast and plants [60–64]. The default parameters of RNA folding algorithms typically assume the RNA should be folded at 37°C but allow for user-defined changes in folding temperature [59]. As 37°C is outside the physiological range of many organisms including the nematode Caenorhabditis elegans, using default parameters for these algorithms may lead to inaccurate predictions of RNA secondary structure. Notably, the current versions of the miRNA repositories miRBase v22.1 [65] and MirGeneDB v3.0 [66] often disagree with their predicted structures of C. elegans miRNA precursors (Figure S1).

In this study, we used RNAfold and RNAcofold from the ViennaRNA package [59,67,68] to predict the folding of C. elegans miRNA hairpins and duplexes at physiologically relevant temperatures. We show that more than 25% of C. elegans miRNA hairpins have increased predicted base-pairing at 20°C compared to the default temperature of 37°C. Most of the predicted changes in base-pairing affected the terminal ends or loop structure (75.9%, n = 60), although in some cases we observed altered positions of central bulges and mismatches at lower temperatures (24.1%, n = 19). We found that the predicted MFE structures of ~15% of miRNA duplexes also showed increased predicted base-pairing at 20°C. We also observed differences in the predicted folding pattern of miRNA hairpins and miRNA duplexes across the physiological range of C. elegans temperatures, including differences that would be expected to impact miRNA processing. Interestingly, the predicted MFE structures for ~30% of miRNA duplexes were partially unwound relative to the hairpin-derived miRNA duplex structure, suggesting that unwinding of many miRNA duplex ends is energetically favourable once the duplex is liberated from the hairpin structure. As miRNA strand selection is connected to thermodynamic asymmetry of duplex ends [48–50], energetically favourable unwinding may contribute to proper miRNA strand choice. To quantify the thermodynamic stability of miRNA duplex ends, we developed a new approach that calculates the stability of partially unwound duplex structures by introducing hard constraints into the folding algorithm [53]. We use these constraints to calculate the difference in the free energy of partially unwound duplexes and the fully wound duplex structures. We refer to the difference in the free energy as unwinding energy, which is proportional to the amount of energy that would be required to unwind that duplex end. Unwinding energy modelled thermodynamic asymmetry of miRNA duplex ends and out performed free energy values obtained from the nearest neighbour database [69] at predicting strand selection of C. elegans miRNAs.

Importantly, our analysis provides an updated catalogue of C. elegans miRNA structures and end stabilities at physiologically relevant temperatures, which may improve discovery of sequence features influencing miRNA biogenesis and strand selection. Furthermore, our in silico method of duplex unwinding allows for evaluation of thermodynamic asymmetries within miRNA duplex ends and may have broader use for assessing local stabilities within different RNA structures.

Results

Secondary structure predictions of C. elegans miRNA hairpins at different temperatures

To improve upon the current publicly available miRNA hairpin folds [65,66], we used RNAfold to predict structures of miRNA hairpins (Table S1) across the physiological range of C. elegans (15°C, 20°C and 25°C) compared to the traditional default temperature parameter (37°C). We acquired sequence information for all C. elegans miRNA hairpins from miRBase which included a mixture of pre-miRNA sequences and extended pri-miRNA hairpin sequences when available [65]. We found that more than 25% of C. elegans pre-miRNAs were predicted to fold differently at 20°C, a temperature commonly used for laboratory culture than at the non-physiological temperature 37°C (Table S1). Most of the temperature-dependent differences that we observed occurred at the terminal end or loop of the hairpin (75.9%, n = 60), although we also observed some examples where the position of central mismatches or bulges (24.1%, n = 19) were affected (Figure 1(A), Table S1). In some cases, the temperature-dependent changes in hairpin folding affected the locations of basal or apical junctions that are important for miRNA processing [70] (Figure 1(B,C), Table S1). We also found that some hairpins were predicted to fold differently within the physiological range of C. elegans (Figure 1(D–F), Table S1). Compared to 20°C, we observed alternative predicted folding for 3.5% of miRNA hairpins at 15°C and ~8% of hairpins at 25°C (Table S1). These findings show that temperature should be considered when performing algorithmic prediction of miRNA hairpin structure, with parameters assessed for physiological relevance.

Figure 1.

Differences in predicted miRNA hairpin folding across different temperatures. (A) Predicted folding of the mir-37 hairpin at 20°C (top) and 37°C (bottom) shows altered locations of central mismatches and bulges at different temperatures. (B) The position of the apical junction in mir-44 is shifted at 20°C (top) compared to 37°C (bottom). (C) The location of the mir-49 basal junction is altered at 20°C (top) relative to 37°C (bottom). (D) The predicted folding of the mir-356a hairpin suggests different folding is favored at the physiological temperatures 15°C (top) and to 20°C (bottom). (E) The mir-228 hairpin is expected to adopt a different structure at the physiological temperatures 20°C (top) and to 25°C (bottom). (A-E) Minimal free energy (MFE) structures of miRNA hairpins were acquired from RNAfold. Boxes indicate locations of temperature-dependent differences in predicted folding. miRNA hairpin sequences were obtained from miRbase v22.1 [65]. (F) Quantification of differences in predicted folding of miRNA hairpins using RNAfold relative to default (37°C) temperature conditions. Folding differences are categorized based on whether they affect terminal base pairing (including the loop structure) or central region of the miRNA hairpin. miRnas that have altered folding of terminal and central structures are categorized as ‘both’. The number of observed differences is given at the bottom of each bar (n = 253 miRNA hairpins), while the axis shows the % of total hairpins affected.

Predicted MFE structures of C. elegans miRNA duplexes at different temperatures

We next used RNAcofold to examine how temperature influences the MFE structures of C. elegans miRNA duplexes (Table S2). We obtained sequences for available miRNA guide and passenger strands from miRBase and generated MFE structures for 190 miRNA duplexes that have annotated guide and passenger strand sequences (Table S2). Compared to the physiologically relevant temperature 20°C, we found that 13.2% of miRNA duplexes have an altered MFE structure at the non-physiological temperature 37°C. In general, higher temperature was associated with decreased base-pairing of miRNA duplexes, especially at terminal duplex ends (73.5%, n = 25: Figure 2, Figure S2, Table S2). In some miRNA duplexes, we observed temperature-dependent (15°C to 37°C) changes in the positions of central mismatches/bulges (26.5%, n = 9: Figure 2(A), Table S2), which play a role in sorting small RNAs to specific Argonautes [71–73]. In other duplexes, increased temperature was predicted to result in partial unwinding of one duplex end (Figure 2(B–D), Table S2), leading to altered predictions of thermodynamic asymmetries important for small RNA strand selection [48,49]. We also observed cases where higher temperature affected the overall structure of mismatched nucleotides (Figure 2(E), Table S2) or resulted in rearrangements that affected both duplex ends and central mismatches (Figure 2(F), Figure S2, Table S2). Similar to miRNA hairpins, alternative folding of miRNA duplexes was also predicted to occur within the physiological temperature range of C. elegans (Figure 2(G–I), Table S2). Compared to 20°C, temperature-dependent differences in duplex structure were observed for 4.2% of miRNAs at 15°C and 4.7% of miRNAs at 25°C (Figure 2(G,H), Table S2). Temperature-dependent differences in hairpin or duplex folding did not appear to be associated with miRNA abundance (Figure S2), indicating that alternative folding was not limited to lowly expressed miRNAs. Collectively, these findings show that, similar to miRNA hairpins, temperature should be considered when performing algorithmic prediction of miRNA duplex structure, with parameters assessed for physiological relevance.

Figure 2.

Differences in minimal free energy miRNA duplex structures at different temperatures. (A) Minimal free energy (MFE) structures of the miR-2 duplex at 20°C (top) and 37°C (bottom). The expected positions of central mismatches and bulges are affected by different temperature conditions. (B) MFE structures of the miR-247 duplex at 20°C (top) and 37°C (bottom). The 3p end (guide end) of the miR-247 duplex is predicted to be partially unwound at higher temperature. (C) MFE structures of the miR-56 duplex at 20°C (top) and 37°C (bottom). The 5’ nucleotide of miR-56-5p (passenger strand) is predicted to become unpaired at higher temperature. (D) MFE structures of the miR-64 duplex at 20°C (top) and 37°C (bottom). The 3p end (passenger end) of the miR-64 duplex is predicted to be partially unwound at higher temperature. (E) MFE structures of the miR-784 duplex at 20°C (top) and 37°C (bottom). At lower temperature, a U:A base pair intervenes a centrally located loop structure to create two adjacent loops. At 37°C, the U:A base pair is disrupted to generate a single, larger loop structure. (F) MFE structures of the miR-8198 duplex at 20°C (top) and 37°C (bottom). The 3p end (guide end) of the miR-8198 duplex is predicted to be partially unwound at higher temperature and the location of central mismatches/bulges are shifted. (G) MFE structures of the miR-241 duplex at 15°C (top) and 20°C (bottom). At 20°C, the 3p end (passenger end) of the miR-241 duplex is predicted to be partially unwound while increased base-pairing is predicted at 15°C. (H) MFE structures of the miR-232 duplex at 20°C (top) and 25°C (bottom). The 5“ nucleotide of miR-232-5p (passenger strand) is predicted to be paired at 20°C and become unpaired at 25°C. (A-H) Minimal free energy (MFE) structures of miRNA duplexes were acquired from RNAcofold. Guide strands are shown on the top of each duplex structure and highlighted in maroon. Passenger strands are shown on the bottom of each duplex structure and highlighted in gray. Boxes indicate locations of temperature-dependent differences in predicted folding. miRNA 5p and 3p sequences and guide strand information were obtained from miRbase v22.1 [65]. (I) Quantification of differences in predicted folding of miRNA duplexes using RNAcofold relative to default (37°C) temperature conditions. Folding differences are categorized based on whether they affect the base pairing of terminal nucleotides (5” or 3’ duplex ends) or central region of the miRNA duplex. miRnas that have altered folding of terminal and central structures are categorized as ‘both’. The number of observed differences is given at the bottom of each bar (n = 190 miRNA duplexes), while the axis shows the % of total duplexes affected.

miRNA hairpin precursors may constrain miRNA duplexes into a suboptimal energy state

When we examined the MFE structures of C. elegans miRNA duplexes, we noticed that the base-pairing of the MFE structure often differed from the base-pairing of the equivalent nucleotides in the hairpin structure at the same temperature (Table S3). At 20°C, nearly one-third (31.1%) of miRNA MFE duplex structures were different from their predicted structure within the hairpin (Table S3). We observed differences in both the location of central mismatches and terminal nucleotide pairing (Figure 3, Table S3). We quantified the positions of mismatched or bulged nucleotides, which revealed decreased base-pairing of terminal nucleotides in the MFE duplex structure compared to the hairpin-derived duplexes (Figure S3, Table S3). We found that many of the duplex ends that became unpaired in the MFE structure corresponded to miRNA guide strands (Figure 3, Figure S3), which could contribute to thermodynamic asymmetry of duplex ends. In some cases, unpairing of terminal nucleotides appeared to result in more extensive unwinding of that duplex end (Figure 3). Our analysis suggested that miRNA hairpins may constrain duplexes into less energetically favourable states and that partial unwinding of some miRNA duplex ends could be energetically favourable once the duplex is liberated from the hairpin precursor. It is possible the difference in stability between the hairpin-constrained duplex and MFE structure may contribute to thermodynamic asymmetries of duplex ends.

Figure 3.

miRNA hairpins may constrain duplexes into less energetically favourable states (A) Differences in the miR-39 hairpin-derived duplex (top) and MFE duplex (bottom) structures at 20°C. The 5’ nucleotide of miR-39-3p (guide strand) is predicted to be paired in the hairpin, but is energetically favoured to unwind in the MFE structure. (B) Differences in the miR-54 hairpin-derived duplex (top) and MFE duplex (bottom) structures at 20°C. Two nucleotides that are paired in the miR-54 hairpin are energetically favored to unwind in the 3p end (guide end) of the MFE duplex leading to unwinding of the 3p duplex end. (C) Differences in the miR-64 hairpin-derived duplex (top) and MFE duplex (bottom) structures at 20°C. The positions of mismatches/bulges in the 3p end (passenger end) of the miR-64 are favored to rearrange in the MFE structure relative to how they are positioned in the hairpin structure. (D) Differences in the miR-70 hairpin-derived duplex (top) and MFE duplex (bottom) structures at 20°C. Three nucleotides that are paired in the miR-70 hairpin are energetically favored to unwind in the 3p end (guide end) of the MFE duplex leading to extensive unwinding of the 3p duplex end. (E) Differences in the miR-83 hairpin-derived duplex (top) and MFE duplex (bottom) structures at 20°C. Two nucleotides that are paired in the miR-83 hairpin are energetically favored to unwind in the 3p end (guide end) of the MFE duplex leading to unwinding of the 3p duplex end and rearrangement of a mismatched nucleotide pair to a bulged nucleotide. (F) Differences in the miR-5595 hairpin-derived duplex (top) and MFE duplex (bottom) structures at 20°C. The 3p duplex end (guide end) of the miR-5595 is favoured to rearrange so that the 5’ nucleotide in the MFE structure becomes unpaired. (A-F) Minimal free energy (MFE) structures of miRNA hairpins were acquired from RNAcofold. Hairpin structures were acquired by introducing folding constraints to mirror the duplex structure within the MFE hairpin that was acquired using RNAfold. Guide strands are shown on the top of each duplex structure and highlighted in maroon. Boxes indicate locations of temperature-dependent differences in predicted folding. Passenger strands are shown on the bottom of each duplex structure and highlighted in grey. miRNA 5p and 3p sequences and guide strand information were obtained from miRbase v22.1 [65].

In silico unwinding of miRNA duplexes addresses how the base-pairing status of terminal nucleotides impacts duplex stability

Previous methods to quantify duplex end stability have used the nearest neighbour model that uses energy parameters obtained at 37°C [48–51]. As 37°C is a non-physiological temperature for C. elegans, we developed an alternative method for quantifying duplex end stability that uses physiological temperature parameters for C. elegans miRNA duplexes. Our approach was to calculate the stability of partially unwound duplexes compared to the fully wound MFE or hairpin-derived duplex structures (Figure 4(A)). We refer to the difference in free energy (ΔΔG) of the partially unwound duplex and the fully wound state as the ‘unwinding energy’ for each duplex end (ΔG), which should be proportional to the amount of energy required to unwind that end of the duplex (Figure 4(A)). An advantage of this approach is that it considers how unwinding terminal nucleotides impacts overall duplex stability, which is not necessarily considered when using values from nearest neighbour databases. To calculate unwinding energies, we first used RNAcofold to quantify the stability of each C. elegans miRNA duplex at 20°C. We then introduced hard folding constraints to RNAcofold [53] that prevented base-pairing of the terminal 1–4 nucleotides on each duplex end and determined the free energy of the partially unwound duplex structures (Figure 4(A)). The constraint files also instructed RNAcofold how to enforce base-pairing throughout the rest of the duplex (summarized in Figure 4(A)), which allowed us to recapitulate in unwinding of MFE (Figure 4(B), Table S4), hairpin-derived (Figure 4(C), Table S5) duplex structures or partially constrained duplex structures (Figure 4(D), Table S6). Using this method, a ΔG value of zero was obtained when the corresponding nucleotides were already unpaired in the MFE duplex structure, supporting that additional energy would not be required to unwind those terminal nucleotides (Figure 4(B–D), Tables S4-S5). On the other hand, a negative ΔG indicated that it was energetically favourable for the duplex to unwind the corresponding nucleotides, which was often observed for the non-MFE, hairpin-derived duplex structures (Figure 4(C), Table S5). As expected, the average unwinding energy of duplex ends corresponding to 5’ end of guide strands was lower than the duplex ends corresponding to the 5’ end of passenger strands, indicating that less energy would be required to unwind the 5’ guide end of the duplex (Figure 4(B–D), Tables S4-S5). Therefore, our calculations of unwinding energy illustrate the intrinsic thermodynamic asymmetry of miRNA duplex ends.

Figure 4.

In silico unwinding of C. elegans miRNA duplexes to evaluate thermodynamic stability of duplex ends (A) Example of in silico unwinding (UW) method for the let-7 miRNA. The stability of the MFE or hairpin duplex is calculated using RNAcofold (left) and folding constraints are introduced to calculate the stability of the partially unwound duplex (right). The stability of each duplex end (ΔG) is calculated by taking the difference in free energy (ΔΔG) of the fully wound duplex and the partially unwound duplex. The let-7 guide strand (let-7-5p) is shown on the top of duplexes in maroon highlight, and the let-7 passenger strand (let-7-3p) is shown on the bottom of duplexes in grey highlight. Box illustrates different folding methods to calculate free energy of duplex ends. (1) for MFE duplex constrained calculations, miRNA duplexes were forced to fold according to the duplex MFE structure. An additional set of constraints were then introduced to force progressive unwinding (1–4 nucleotides) of each duplex end. (2) for hairpin duplex constrained calculations, miRNA duplexes were forced to fold according to how they base-paired in the MFE hairpin structure. An additional set of constraints were then introduced to force progressive unwinding (1–4 nucleotides) of each duplex end. (3) for duplex ends constrained calculations, miRNA duplexes were not forced to fold according to any structure. Rather, duplexes were allowed to fold into the MFE structure after an additional set of constraints were introduced to force progressive unwinding (1–4 nucleotides) of each duplex end. (4) for nearest neighbor calculations, they were folded according to the MFE structure, and values for nearest neighbor energy were acquired from the nearest neighbor database [69]. Note that these calculations represent the stability of 1–4 nucleotides, rather than how unwinding of those nucleotides affects the stability of the resulting partially unwound duplex. (B) Quantification of thermodynamic end stability using in silico unwinding of constrained MFE duplexes. Duplexes were folded according to the MFE structure, or with the first one through four nucleotides of each miRNA duplex end forced to be unpaired. p-values for one nucleotide: p = 4.10x10⁻⁵, two nucleotides: p = 6.40x10⁻⁴, three nucleotides: p = 7.92x10⁻⁵ and four nucleotides: p = 0.002. (C) Quantification of thermodynamic end stability using in silico unwinding of hairpin-constrained duplexes. Duplexes were folded according to the structure embedded within the miRNA hairpin, or with the first one through four nucleotides of each miRNA duplex end forced to be unpaired. p-values for one nucleotide: p = 6.10x10⁻⁶, two nucleotides: p = 8.26x10⁻⁵, three nucleotides: p = 2.62x10⁻⁵ and four nucleotides: p = 4.86x10⁻⁴. (D) Quantification of thermodynamic end stability using in silico unwinding of unconstrained duplexes. Duplexes were allowed to refold into the MFE state following unwinding of the first one through four nucleotides of each miRNA duplex end. p-values for one nucleotide: p = 6.48x10⁻⁵, two nucleotides: p = 3.89x10⁻⁴, three nucleotides: p = 4.86x10⁻⁵ and four nucleotides: p = 0.001. (E) Quantification of thermodynamic end stability using nearest neighbor values. Duplexes were folded according to the MFE structure, and values for nearest neighbor energy were acquired from the nearest neighbor database [69]. Note that nearest neighbor values are provided at the 37°C, a non-physiological temperature for C. elegans. The stability of the rest of the duplex is not considered. p-values for one nucleotide: p = 0.012, two nucleotides: p = 0.298, three nucleotides: p = 0.032 and four nucleotides: p = 0.019. (B-E) the guide end of the duplex (miR) is shown in maroon color and the passenger end of the duplex (miR*) is shown in gray color. Each dot represents a miRNA duplex end (n = 190). Values for ΔG are given as kcal/mol. Boxes range from the first to third quartile of the data. The thick bar represents the statistical median. Lines extend to the minimum and maximum data point excluding outliers that were defined as 1.5 times the interquartile range. Statistical analysis: *p < 0.05, **p < 0.01, ***p < 0.001, n.S. (not significant) p > 0.05. (B-D) a value of zero indicates no difference in free energy between the wound and partially unwound duplexes and corresponds to duplex ends that are already unwound in the MFE or hairpin structures. A positive value indicates that the wound structure is less stable than the partially unwound structure.

We next examined whether partial unwinding of duplex ends may lead to unstable intermediate structures that are energetically favoured to further unwind or otherwise undergo structural rearrangements. To address this, we introduced new folding constraints that prevented base-pairing of terminal nucleotides but did not instruct RNAcofold to retain the predicted hairpin-derived nucleotide pairing, which allowed the partially unwound duplexes to reconfigure into MFE structures. We refer to these structures as ‘end-constrained’, as only the duplex ends are constrained during folding. Our quantification of unwinding energy revealed that the MFE state of some duplexes changes after partial unwinding, suggesting that unwinding of duplex ends could lead to progressive unwinding or structural rearrangements for some miRNAs (Figure 4(D), Table S6). When compared to energy values from the nearest neighbour tables (Figure 4(E), Table S7) [69], we found that our unwinding energy models outperformed nearest neighbour values at predicting the expected thermodynamic asymmetry of miRNA duplex ends (Figure 4(B–D)). In fact, the nearest neighbour model suggested that the terminal nucleotide was more stable for C. elegans duplex ends corresponding to the 5’ end of guide strands than the 5’ end of passenger ends (Figure 4(E), Table S7), which contradicts previous studies of small RNA asymmetry [48–50]. It is worth noting that the parameters of the nearest neighbour tables were determined at 37°C [69], which may lead to inaccurate representation of the thermodynamic asymmetry of C. elegans miRNA duplexes at physiological temperatures. We conclude that in silico unwinding of miRNA duplexes improves models of thermodynamic asymmetry of miRNA duplex ends and allows duplex end stability to be evaluated at a defined temperature condition. Overall, our annotation of miRNA duplex unwinding energies provides an improved resource quantifying C. elegans miRNA end stabilities.

Prediction of miRNA strand selection using unwinding energies of miRNA duplex ends

To investigate whether different methods of calculating the thermodynamic stability of miRNA duplex end affected predictions of miRNA strand selection, we used the twin-drive model [50] to calculate the predicted strand choice of C. elegans miRNAs. The twin-drive model uses a formula that weighs the relative effects of 5’ nucleotide identity (N-drive) and thermodynamic asymmetry (G-drive) to calculate the expected levels of each strand [50]. The twin-drive formula is given as ln(5p/3p) = kΔΔG_5p-3p + N_5p-N_3p, where k and N are constants affecting thermodynamic stability and nucleobase identity respectively [50]. We used the twin-drive formula, using previously described values for constants [50], to predict strand selection of C. elegans miRNAs using different methods of calculating thermodynamic stability for the first through fourth nucleotides on each end of the duplex (Tables S8-S11). Using the unwinding energy of MFE duplexes, we were able to correctly predict strand selection for 75.3% of miRNA duplexes when considering the thermodynamic stability of only the single terminal nucleotide on each duplex end (Figure 5(A,B), Table S8). When we calculated thermodynamic stability using additional nucleotides on each duplex end, we found that thermodynamic asymmetry was less predictive of miRNA strand selection (Figure 5(A,B), Figure S4A-B, Table S8). For example, when considering the unwinding energy of the three terminal nucleotides on each duplex end, we were able to correctly predict strand selection for 71.1% of miRNAs (Figure 5(A,B)). Predictions of miRNA strand selection were similar when using the unwinding energy of hairpin-derived duplexes (Figure 5(C), Figure S4C, Table S9), or end-constrained duplexes (Figure 5(D), Figure S4D, Table S10), although predictions using the unwinding energy of MFE duplexes slightly outperformed both alternative folding methods (Figure 5). Therefore, different methods of calculating unwinding energy appear to have similar predictive value for determining miRNA strand preference. As a comparison, we next used nearest neighbour energy values to predict strand selection of C. elegans miRNAs (Figure 5(E), Figure S4E, Table S11). We found that nearest neighbour stabilities predicted strand selection equally as unwinding energies of MFE duplexes when considering the stabilities for three or four terminal nucleotides on each duplex end (Figure 5(E), Figure S4E). However, nearest neighbour stabilities were less predictive of strand selection when considering one or two terminal nucleotide stabilities (Figure 5(E), Figure S4E). While the unwinding energy of the single terminal nucleotides for MFE duplexes correctly predicted strand selection of 75.3% of miRNAs, the nearest neighbour stabilities of terminal nucleotides were only successful at predicting strand selection of 58.4% of miRNAs (Figure 5, Figure S4). Thus, it appears that unwinding energies better model thermodynamic energies of duplex ends when considering few terminal nucleotides, possibly because our unwinding energies approach considers how unwinding of terminal ends affects the overall stability of the partially unwound duplex. Increasing the number of terminal nucleotides in our energy calculations also appeared to disproportionately reduce the predictive power for 5p dominant miRNAs compared to 3p dominant miRNAs (Tables S8-S11), which could indicate that additional factors promote selection of 5p miRNAs in C. elegans. Note that the constants we used to predict strand choice were not optimized for C. elegans miRNAs, as the twin-drive model was developed for best performance on murine miRNAs [50]. Thus, it is likely that future parameter optimization could further improve predictions of C. elegans miRNA strand choice. In general, our predictions of strand selection were strongly correlated to the calculated difference in free energy between the guide and passenger ends (Figure S5). For example, when considering the stability of terminal nucleotides using the unwinding energy model, we correctly predicted strand selection of 93.6% of miRNAs that had negative ΔΔG values (ΔG guide strand – ΔG passenger strand, Figure S5A), and only 39.6% of miRNAs that had non-negative ΔΔG values (ΔG guide strand - ΔG passenger strand, Figure S5A). This indicates that the current standing ‘twin-drive’ model [50] may underestimate the impact of factors other than miRNA duplex end stability in strand selection of those miRNAs. A comparison of different energy models revealed that strand selection of only 11 miRNAs (5.8%) was never predicted correctly across any nucleotide window, while strand selection of 61 miRNAs (32.1%) was always predicted correctly across all nucleotide windows (Supplemental Table S12). Thus, some form of thermodynamic asymmetry could potentially describe the observed strand choice of most C. elegans miRNAs. Collectively, our results support that unwinding energies model thermodynamic asymmetry of miRNA duplex ends and are predictive of miRNA strand selection in C. elegans.

Figure 5.

Predictions of miRNA strand selection. (A) Example let-7 duplexes illustrating nucleotides considered for thermodynamic stability calculations. The terminal nucleotides (left) or three terminal nucleotides (right) were considered. Ranked-order plots of predicted C. elegans miRNA strand asymmetry for (B) MFE duplex structures using unwinding energy, (C) hairpin-derived duplex structures using unwinding energy, (D) end-constrained duplexes using unwinding energy and (E) nearest neighbor energy values. (B-E) predicted strand ratios were determined using a formula from the twin-drive model [ln(5p/3p) = kΔΔG_5p-3p + N_5p-N_3p], using previously described values for constants [50]. miRnas are plotted in order of lowest to highest predicted ln(5p/3p) values. miRnas above dashed line are predicted 5p dominant and miRNAs below dashed line are predicted 3p dominant. Correct predictions (based on guide strands reported in miRbase) are indicated as black dots, and incorrect predictions are indicated as red dots.

Discussion

Structural predictions of C. elegans miRNA intermediates

Structural features of miRNA intermediates are critical for accurate miRNA biogenesis [70], yet structural repositories for miRNA structure information are often in disagreement (Figure S1). In this study, we explored different in silico folding methods to predict the structure of C. elegans miRNA hairpins and duplexes and calculate duplex end stabilities. Our analysis shows that temperature parameters of RNA folding algorithms have a substantial impact on predicted structures of miRNA intermediates. Previous studies, including our own [51], have used default temperature parameters (37°C) to predict the structure of C. elegans miRNA intermediates. However, as temperature is directly linked to free energy, folding C. elegans miRNAs at non-physiological temperatures may lead to inaccurate predictions of secondary structures. We show that performing RNA folding within the physiological temperature range of C. elegans (15–25°C) leads to increased predicted base-pairing of C. elegans miRNA intermediates. Decreased base-pairing observed at higher, non-physiological temperature could reflect partial ‘denaturation’ of the miRNA precursors, raising the possibility that miRNA evolution could involve temperature-sensitive adaptations. We also observed structural differences when miRNA intermediates were folded across different temperatures within the physiological range of C. elegans, including altered positions of structural junctions that are important for miRNA processing [70]. It is intriguing to speculate that temperature may regulate miRNA-dependent gene regulation in a manner akin to ‘RNA thermometers’ that have been described in other species [60–64]. However, additional studies will be important to uncover whether temperature-dependent differences in miRNA folding predictions accurately reflect physiological folding and are functionally relevant. Although currently lacking, testing temperature-dependent folding on experimentally validated RNA structures will be essential to fully test the accuracy of folding predictions.

During our analysis, we noticed that the base-pairing of MFE miRNA duplex structures was often different than the base-pairing of the equivalent nucleotides within the miRNA hairpin precursor. These differences likely arise from additional structural stability that is conferred from neighbouring base-pairs that are present in the hairpin precursor but are removed during miRNA processing events. This observation suggests that miRNA hairpins, at least temporarily, may constrain some miRNA duplexes into suboptimal energy states, although it is currently unclear whether this has functional significance. As thermodynamic asymmetries of miRNA duplex ends are associated with miRNA strand selection [48–50], it is possible that some miRNA hairpins ‘prime’ miRNA duplexes for asymmetric Argonaute loading and subsequent miRNA strand selection. Energetically favourable rearrangements of miRNA duplexes to reach their MFE states may promote unwinding of miRNA duplex ends and influence strand choice. Alternatively, miRNA duplexes may have sufficient time to rearrange into their MFE structures before miRNA strand selection occurs, in which case it is unlikely that hairpin-constrained duplex folding contributes towards thermodynamic asymmetries influencing strand choice. Additional work will be needed to clarify whether the base-pairing status of miRNA duplexes within their hairpin precursors influences miRNA strand biogenesis.

Evaluating thermodynamic asymmetry by in silico unwinding improves predictions of miRNA strand choice

We developed an in silico method to assess the thermodynamic stability of miRNA duplex ends using physiologically relevant temperature parameters. We introduced hard constraints to the RNA folding algorithm to restrict terminal nucleotide base pairing and calculate the stability of partially unwound duplex structures. By comparing the difference in stability of partially and fully wound MFE structures, we determined the unwinding energy of each duplex end. We reasoned that the unwinding energy should be proportional to the amount of energy required to unwind each duplex end. We used unwinding energy to predict miRNA strand asymmetry using the previously described twin-drive model [50]. We found that unwinding energy outperformed free energy values obtained from the nearest neighbour database [69], especially when considering the free energy one or two terminal nucleotides. In general, our energy models had the strongest predictive power when only considering the stability of the terminal nucleotide pair. Increasing the number of terminal nucleotide pairs considered in our energy calculations resulted in lower predictive power, especially for 5p dominant miRNAs. As miRNA strand selection is often disrupted in human disease [51], improved predictions of miRNA strand selection may help better understand the connection between miRNA strand choice and human diseases. Notably, we were able to correctly predict the preferred strand for ~75% of C. elegans miRNAs, suggesting that the twin-drive model is not sufficient to explain miRNA strand selection of all miRNAs. It seems likely that additional features of miRNA duplexes influence strand selection of C. elegans miRNAs in vivo.

Predictions of miRNA strand choice may be further complicated by the existence of isomiRs, which can result from alternative processing events during miRNA biogenesis or post-transcriptional modification of miRNAs [70]. While isomiRs that affect the sequence composition of miRNA duplex ends could affect predicted miRNA strand preference, it is currently unclear whether specific isomiRs are generated prior to duplex loading. Any isomiRs forming after the duplex is loaded into Argonaute, such as 3’ trimming, would not be expected to impact strand selection. As differences in both miRNA strand selection and isomiR profiles have been reported across different developmental and cellular contexts [74–83], it is intriguing to speculate that differences in miRNA strand preference could reflect altered isomiR duplex characteristics that favour alternative strand loading. Although we did not test how individual isomiRs may impact predictions of miRNA strand choice in this study, our analysis could be readily extended to duplexes comprising isomiR(s), which might be particularly interesting for miRNAs whose strand preference cannot be explained by current models.

Locally stable RNA structures can have important functions, and algorithms have been developed to identify locally structured regions of RNA molecules [84]. We used our in silico unwinding method to assess the local thermodynamic stability of miRNA duplex ends. In principle, this approach could be used to evaluate local stabilities of secondary structures within other RNA molecules.As some RNA binding proteins may recognize locally structured regions of their target RNA molecules [85], unwinding energies may help identify regions that may facilitate interactions with RNA binding proteins. Alternatively, it is possible that intermolecular interactions, between RNA and proteins or other nucleic acids, may locally restrict base-pairing across specific nucleotides of a given RNA molecule. In such case, constrained folding may improve predictions of how RNA secondary structure is altered upon intermolecular interactions and the energy requirement needed to facilitate structural rearrangements.

A current disadvantage of our approach is the time investment required to generate constraint files for folding algorithms based on careful manual curation of predicted structures. In this study, we visually inspected each miRNA fold and restricted terminal nucleotide pairing by individually marking the locations of terminal nucleotide pairs in dot-bracket notation (see Files S3-S5). This required annotation of both miRNA strands, as only restricting one nucleotide of a base pair would allow re-pairing of the other nucleotide in the resulting MFE structure. This time investment limited our ability to test our strand selection models on the miRNAs of other organisms. Towards that end, user-friendly automation of our in silico unwinding approach will substantially reduce the time required to perform such analyses and facilitate similar analyses across different RNA molecules and organisms. Collectively, our current analysis provides a careful manual curation of the structures and stabilities of C. elegans miRNA intermediates, a valuable resource towards understanding the mechanisms of miRNA biogenesis, strand selection, and, ultimately, miRNA-dependent gene regulation.

Materials and methods

Secondary structure predictions of C. elegans pre-miRNAs and miRNA duplexes

The sequences of C. elegans pre-miRNAs and mature miRNAs were retrieved from miRBase v22.1 [65]. RNAfold was used to predict pre-miRNA hairpin structures and RNAcofold was used to predict miRNA duplex structures [59,67,68]. We were unable to predict the duplex structure of miRNAs that lacked annotations for the passenger strand sequence. The parameter ‘−T 20’ was used to predict RNA folding at 20°C, or ‘−T X’ (where X is a specific temperature) was used to predict RNA folding at alternative temperatures. The following command was inputted into terminal for hairpin determination: RNAfold input.fa -T 20, and the following command was inputted into terminal for duplex determination: RNAcofold input.fa -T20, where input.fa is the name of the input files. The input files used for C. elegans pre-miRNAs (File S1) and miRNA duplexes (File S2) are provided in the supplementary material. miRNA duplex structures were recoloured using Adobe Illustrator such that the guide strand is maroon and the passenger strand is grey. Locations of mismatches and bulges were identified using dot-bracket nomenclature [59]. Terminal 3’ nucleotide overhangs were not included when scoring internal mismatch/bulge locations.

In silico unwinding of miRNA duplexes and predictions of strand selection

To calculate unwinding energies, folding constraints were introduced in RNAcofold to force unpaired nucleotides at the miRNA duplex ends: ‘-T 20 -C – enforceConstraint’, which mimics partial unwinding of duplex ends. The following command was inputted into terminal for constrained duplex folding: RNAcofold input.fa -T 20 -C – enforceConstraint, where input.fa is the name of the input files. The input files containing constraint information are provided in the supplementary material (Files S3-S5). File S3 was used for ‘MFE duplex constrained’ calculations, File S4 was used for ‘hairpin duplex constrained’ calculations and File S5 was used for ‘duplex ends constrained’ calculations. Unwinding energies for each duplex end were calculated by subtracting the energy of the fully wound (maximal free energy state) duplex from the partially unwound duplex (the first 1:4 nucleotides were unwound from each duplex end). Guide and passenger strand annotations were determined based on which strand had the most reported reads on miRBase [65]. For nearest neighbour stabilities, values were acquired from the nearest-neighbour database [69]. Predictions of miRNA strand selection were determined using the twin-drive model equation [ln(5p/3p) = kΔΔG_5p-3p + N_5p-N_3p] using constants previously described [50].

Statistical analysis

R statistical software (R Core Team, https://www.R-project.org/) was used to calculate all statistics. All p-values were calculated using two-tailed t-tests assuming equal variance among samples. All statistics are presented as mean ± one standard deviation unless otherwise specified. p-values are indicated by: *p < 0.05, **p < 0.01, ***p < 0.001, n.s. (not significant) p > 0.05.

Funding Statement

This work was supported by grant from the National Institutes of Health [R35GM124828 to A.Z.], the Kansas INBRE programme [P20 GM103418] and the Johnson Cancer Center at K.S.U.. J.C.M was supported by a fellowship from the NIH [F32GM148040].

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

All data supporting the findings of this study are available from the corresponding author upon reasonable request. Supplemental tables and files are available at figshare.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/15476286.2026.2649359