MERGE-RNA: a physics-based model to predict RNA secondary structure ensembles with chemical probing

TL;DR

MERGE-RNA introduces a physics-based framework that integrates RNA thermodynamics, probe binding, and mutational profiling to infer secondary-structure ensembles from chemical probing data. By learning transferable parameters across sequences, concentrations, and replicates via a maximum-entropy approach, it predicts ensemble distributions and extrapolates to native conditions (zero probe) to reveal suboptimal states. The method demonstrates robust transferability across diverse RNAs, improves ensemble accuracy over traditional MFE-based predictions, and uncovers dynamic states such as strand displacement in bistable or thermosensitive RNAs. This approach enables quantitative characterization of conformational heterogeneity, offering a scalable route to understanding RNA function through ensemble dynamics and intermediate states with practical implications for RNA biology and design.

Abstract

The function of RNA molecules is deeply related to their secondary structure, which determines which nucleobases are accessible for pairing. Most RNA molecules however function through dynamic and heterogeneous structural ensembles. Chemical probing methods (e.g., DMS probing) rely on selective chemical modification of accessible RNA nucleotides to infer base-pairing status, yet the resulting nucleotide-resolution data represent ensemble averages over dynamic RNA conformations. We present MERGE-RNA, a unified, physics-based framework that explicitly models the full experimental pipeline, from the thermodynamics of probe binding to the mutational profiling readout. By integrating measurements across probe concentrations and replicates, our model learns a small set of transferable and interpretable parameters together with minimal sequence-specific soft constraints. This enables the prediction of secondary structure ensembles that best explain the data and the detection of suboptmal structures involved in dynamic processes. We validate MERGE-RNA on diverse RNAs, showing that it achieves strong structural accuracy while preserving essential conformational heterogeneity. In a designed RNA for which we report new DMS data, MERGE-RNA detects transient intermediate states associated with strand displacement, dynamics that remain invisible to traditional methods.

Figures (12)

• Figure 1: Schematic overview of the method. Our approach builds a physical model---illustrated on the left side of the figure---that represents the entire pipeline of chemical probing experiments. This includes RNA folding into an ensemble of secondary structures, probe binding, adduct formation, and the final mutational profiling readout. The model is trained on chemical probing data from multiple RNA sequences, probe concentrations, and experimental replicates. From these data, it learns a set of physically meaningful parameters shared across all experiments ($\mu$, $\Delta\mu_{\text{pairing}}$, $p_\mathrm{bind}(A)$, $p_\mathrm{bind}(C)$, $p_\mathrm{bind}(G)$, $p_\mathrm{bind}(U)$, $m_0$ and $m_1$, as defined in the main text). In addition, it estimates sequence-specific soft constraints $\lambda_i$ (one per nucleotide position). Once trained, the model can predict the structural ensemble of each RNA at any probe concentration. Importantly, it can also extrapolate to zero probe concentration, providing the unperturbed structural ensemble.
• Figure 2: (a) Cross-validation of physical parameters across five RNA systems. Each bar represents the per-datapoint loss obtained for the indicated system after training on three systems and testing on the remaining two. Yellow bars refer to results obtained when the corresponding system was excluded from the training set. (b) Normalized loss profiles for HCV IRES with soft constraints of varying magnitude applied. Red line: constraints from reference structure; light green line: constraints from minimum free energy (MFE) structure. Each curve represents an independent fit with physical parameters trained on different system triplets. The blue line corresponds to an improved fit achieved by optimizing the sequence-specific $\lambda_i$ parameters with MERGE-RNA. Stars mark are reused across panels to indicate the same ensemble. (c) Frobenius distance between base pairing probability (bpp) matrices, quantifying structural similarity between the HCV IRES reference structure and: MFE structure, bpp of the original thermodynamic ensemble, bpp obtained from the minimum of the red curve in panel (b), and bpp from the fully optimized model. (d) Arc plot visualization comparing the secondary structures of HCV IRES: reference experimental structure (red), ensemble obtained from the thermodynamics model (green), and ensemble obtained from MERGE-RNA (blue).
• Figure 3: Ensemble predictions for cspA 5$^\prime$ UTR at two different temperatures. (a,b) Arc plots corresponding to the DMS data collected at 10 and 37$^\circ$C, respectively. Predictions from MERGE-RNA (red and black) are compared with predictions from Ref. zhangStressResponseThat2018 (blue).
• Figure 4: Model accurately deconvolves mixed structural states from synthetic data. (a) Base-pairing probabilities inferred from fitting on synthetic data for a bistable RNA sequence. The data was generated from an RNA construct with two competing helices (helix 1 and helix 2), where the ground-truth population of helix 1 was systematically varied from $0$ to $1$ with steps of $0.2$. The model's predicted pairing probabilities are shown for each case. (b) Arc plots of the inferred structural ensemble for two limit cases: when helix 1 is dominant (gold) and when helix 2 is dominant (black). (c) The helix population inferred by the model is plotted against the ground-truth population. The inferred population is quantified as the median of the base-pairing probabilities over the respective helical regions, highlighted in panels (a) and (b).
• Figure 5: Ensemble inference on experimental data for a putative bistable RNA exhibits evidence of strand displacement. (a) Per-nucleotide DMS-MaP mutation profiles across 0--100 mM (20 mM steps; two replicates each) for a putative bistable RNA. We leave out of the plot the pairing binding sites for better visualization. Two regions (A1 on the left and A2 on the right, as defined in (\ref{['eq:seq_design']})) are competing to pair with the central region (B). They show a slope in the reactivity profile that is compatible with a strand displacement mechanism. (b) Pairing probabilities obtained from the ensemble predicted by MERGE-RNA (solid lines) after fitting on experimental data, with Turner 2004 mathewsIncorporatingChemicalModification2004 (crimson) or Andronescu 2007 andronescuEfficientParameterEstimation2007 (steel) thermodynamic models as baselines. MERGE-RNA is able to fix both the baseline predictions (dotted lines) and retrieve pairing probabilities consistent with the observed mutation rates. (c) Schematic depiction of the ensemble of structures as predicted by the baseline thermodynamic model (above) and MERGE-RNA fitted on the experimental data (below). While the thermodynamic model predicts only the two configurations with fully-formed helices, MERGE-RNA captures a wealth of intermediate states where strand displacement occurs.