From scattering curves to structural descriptors: interpretable machine learning maps SAXS/WAXS signals to RNA structural details in defined RNA motifs

Solution X-ray scattering offers an ensemble-averaged picture of biomolecular structure and dynamics under near-physiological conditions. Small-angle X-ray scattering (SAXS) reports on global architecture, while extending to the wide-angle regime (WAXS) captures sub-nanometer-level details. However, SAXS/WAXS (SWAXS) interpretation is complicated due to the overlap of multiple correlation lengths. Here, we present a machine-learning (ML) framework that infers structural properties for a given RNA fold. Using molecular simulations to generate ensembles, LightGBM (light gradient-boosting machine) successfully learns mappings from SWAXS features to key geometric properties. We apply this approach to double-stranded RNA (dsRNA) duplexes and a two-way junction construct. The models accurately predict structural changes from SWAXS data and highlight diagnostic q -windows that encode duplex and junction fingerprints. Our results demonstrate that interpretable ML trained on RNA structural ensembles provides a robust route to extract structural information from solution scattering measurements of RNAs.

Solution X-ray scattering provides a uniquely non-destructive, ensemble-averaged view of biomolecular structure under near-physiological conditions and is widely used across structural biology and soft-matter science. Owing to its flexibility and minimal sample requirements, solution scattering has been applied to RNAs (Chen and Pollack, 2016; He et al., 2023; Kirmizialtin et al., 2012; Plumridge et al., 2019), DNAs (He et al., 2021; Baker et al., 2018; Wang and Bloomfield, 1991; Ober et al., 2022), proteins (Kikhney and Svergun, 2015; Rambo and Tainer, 2013), and multiscale complexes (e.g., polymers and aggregates) (Boldon et al., 2015; Da Vela and Svergun, 2020). At small angles (SAXS), measurements yield robust global parameters and low-resolution envelopes, whereas extending measurements into the wide-angle regime (WAXS) accesses sub-nanometer periodicities that can report on motif-specific fingerprints, particularly for well-defined molecular constructs (Au et al., 2015; Thiel et al., 2024; Zielinski et al., 2023a). Thus, in RNA structure modeling, scattering data have long been integrated with computational approaches to generate molecular envelopes (e.g., DAMMIN (Svergun, 1999) and GASBOR (Svergun et al., 2001)), evaluate or select atomistic models by comparing calculated and experimental profiles (e.g., CRYSOL (Svergun et al., 1995) and FoXS (Schneidman-Duhovny et al., 2010)), and provide restraints during model building (e.g., RNA Masonry (Chojnowski et al., 2023) and ERNWIN/SPQR (Thiel et al., 2024; Thiel et al., 2012)). A recent review by Mukherjee et al. covers the broader landscape of RNA 3D structure prediction and, in particular, how low-resolution SAXS data have been used for RNA 3D structure modeling and evaluation (Mukherjee et al., 2024).

Higher-angle SAXS data can, in principle, constrain conformational detail. For example, Kratky plots help distinguish compact from coil-like states (Kikhney and Svergun, 2015; Rambo and Tainer, 2013; Chen et al., 2019). A compact macromolecule typically exhibits a pronounced peak, whereas an unfolded random coil shows a plateau (He et al., 2022). More generally, however, much of the richer local-geometry information remains entangled in one-dimensional SAXS/WAXS (SWAXS) profiles and is difficult to extract directly.

Consequently, experimental interpretation commonly relies on theoretical modeling to generate structural ensembles. Similarity between a computed SAXS/WAXS profile and an experimental profile is then used to infer structural features (He et al., 2021; Bernetti et al., 2021). A widely used approach is forward modeling, in which ensembles are generated and refined against scattering data (Schneidman-Duhovny et al., 2012; Tria et al., 2015). Once agreement between the model and the measurement is established, the model provides a structural interpretation of the experiment. Recent RNA-focused advances have pushed this paradigm toward more integrated pipelines, including multiscale sampling and refinement frameworks that incorporate experimental SAXS restraints during model generation and hierarchical refinement (Thiel et al., 2024). Time-resolved SAXS further extends this paradigm by capturing dynamical behavior in real time, enabling structural inference along reaction coordinates and conformational transitions (Plumridge et al., 2018; Pollack, 2011; Zielinski et al., 2023b; Kirmizialtin et al., 2020). In parallel, emerging machine learning(ML)-assisted approaches use SAXS as an experimental filter or validator for predicted RNA models and explicitly confront ion- and heterogeneity-related challenges that complicate RNA scattering interpretation (Patt et al., 2025). Recent work has explored machine-learning analysis of scattering in materials and biomolecular systems (Patt et al., 2025; Chen and Pollack, 2020), yet an interpretable and general mapping from SWAXS signals to RNA global and local structural descriptors remains limited.

Here, we introduce a computational framework that links SWAXS signals to RNA structure using atomistic MD ensembles, a supervised learning model (LightGBM), and a SHAP-to-profile mapping that overlays feature attributions directly onto I (q) to provide q -resolved importance and directionality. In the present work, we do not aim to predict unknown RNA structures de novo or to train a single model that generalizes across arbitrary RNA sequences and architectures. In contrast to SAXS-guided structure-generation pipelines that primarily aim to produce 3D models consistent with experimental scattering, our focus is to decode which q -regions in SWAXS encode specific, physically meaningful structural descriptors within a defined RNA fold, thereby enabling direct, interpretable readout of structural changes from scattering profiles. In particular, we ask which regions of the SWAXS profile encode particular global and local descriptors for a given, already characterized fold.

Our results show that low- q primarily encodes global geometry, whereas wide-angle regions report on local order at shorter length scales, including groove-scale metrics, base-pair step parameters, and aspects of hydration. Trained on cluster-averaged MD snapshots and their computed SWAXS curves, the LightGBM model predicts both global and local descriptors, R g, helical radius, groove widths, and base-pair step parameters, for canonical duplexes (dsRNA) as well as a flexible helix-junction-helix (HJH) motif. The SHAP mapping yields quantitative rules connecting intensity features to structural variables, thereby translating SWAXS signals into mechanistic structural readouts for RNA motifs. Importantly, this provides complementary information to sequence-based or single-structure predictors by leveraging solution scattering to report ensemble-averaged, condition-dependent structural responses, including subtle changes that are directly encoded in wide-angle fingerprints. The framework is compatible with time-resolved and condition-dependent experiments and can be extended to heterogeneous RNAs, RNP assemblies, and ligand- or protein-stabilized states, where interpretable ML can help identify the q -regions most diagnostic of functional change. We demonstrate that for the duplex and helix–junction–helix motifs considered here, the methodology provides a route to infer structural changes within a known RNA fold from SWAXS data while retaining an explicit connection between scattering features and structural descriptors.

RNA samples and SWAXS calculations

RNA duplex and two-way junction motif, containing two 12 based-paired RNA duplexes tethered by a junction of rU5 were simulated with all-atom MD. Details of the MD simulation setup are briefly discussed in the next section, RNA MD Simulations, and full details can be found in Refs (He et al., 2021; He et al., 2022). We uniformly subsampled snapshots from the full MD trajectories for each sequence, yielding ∼ 10,000 MD conformations per system. For each conformational frame, the SWAXS profiles were calculated with CRYSOL (Svergun et al., 1995) on a linearly spaced grid (N = 301) up to q = 2.0 Å − 1. These profiles constitute the experimental observables on which the machine learning models are trained.

RNA MD simulations

RNA duplex simulations. Initial A-form duplexes were built with Nucleic Acid Builder (NAB) (Case et al., 2005), solvated in TIP3P (Jorgensen et al., 1983) water, and neutralized with ions, yielding 400 mM KCl solution. Simulations were performed in GROMACS 5.0.5 with the AMBER14SB (Lindorff-Larsen et al., 2010) RNA force field. Periodic boundaries applied in all directions. Long-range electrostatics used particle mesh Ewald (PME) (Darden et al., 1993) with a grid spacing of 0.12 nm and an interpolation of order 4. The equations of motion were integrated using the Leap-Frog (Van Gunsteren and Berendsen, 1988) method with a time step of 2 fs. A 1.1 nm nonbonded cutoff were applied. Systems were equilibrated for 2 ns in the NPT ensemble at 300 K and 1 bar using a Berendsen thermostat (Berendsen et al., 1984) and a Parrinello-Rahman barostat (Parrinello and Rahman, 1981), followed by 100 ns restrained NVT with heavy-atom restraints (1000 kJ mol − 1 nm − 2) to relax solvent and ions, and then 300 ns of unrestrained production. Full details are provided in (He et al., 2021).

HJH RNA simulations. The HJH RNA was built with NAB. Simulations were run in GROMACS 5.0.5 with the HB-CUFIX force field, which incorporates hydrogen bonding and base stacking corrections (He et al., 2022) for RNA. Systems were solvated with TIP3P water and ions, resulting in a 200 mM KCl solution. Periodic boundaries are applied in all directions. A time step of 2 fs was used. Long-range electrostatics employed PME with 0.12 nm grid spacing and fourth-order interpolation. We used a 1.1 nm cutoff for nonbonded interactions. After 5000 steps of energy minimization, we performed 2 ns NPT at 300 K and 1 bar to set the volume, followed by 200 ns restrained NVT with heavy-atom restraints on RNA to relax solvent and ions. The final configuration seeded Well-Tempered Metadynamics simulations, implemented by Plumed (Tribello et al., 2014), to explore the HJH conformational landscape. Full protocols are provided in (He et al., 2022).

Data preparation

To reflect the ensemble nature of solution data, similar conformers were grouped by backbone-based clustering. This clustering is used to approximate the MD-derived conformational ensemble by a manageable number of representative subensembles. By averaging SWAXS intensities and structural descriptors within each cluster, we align with the ensemble-averaged nature of solution scattering and suppress frame-level noise. Clustering is performed separately for each RNA construct and is intended to capture conformational variability within a given fold, not variability across different sequences or motif classes. Frames were aligned on phosphate atoms, and k -means (Hartigan and Wong, 1979) was applied to flattened P-atom coordinates. Briefly, for each frame, the three-dimensional coordinates of all phosphate atoms were extracted, reshaped into a 3 N P -dimensional vector, and standardized (z-scored) prior to clustering. We then conducted standard k -means clustering with k = 1000 using the scikit-learn implementation (Pedregosa et al., 2011), which partitions the conformational ensemble into k clusters by minimizing the within-cluster sum of squared Euclidean distances between embedded phosphate coordinate vectors: where x i denotes the embedded phosphate coordinate vector for frame i, C c is the set of frames assigned to cluster c, and μ c is its centroid. With 10,000 MD frames, this choice of k = 1000 yields an effective structural resolution of ∼ 10 frames per cluster on average. In both duplex and HJH systems, ≥ 99% of clusters contained at least two conformers. For each cluster, SWAXS intensities and structural descriptors were averaged across member frames to obtain cluster-level observables.

The structural descriptors of the RNA constructs, including R g and helical/step parameters, were extracted and analyzed from the MD simulations using PLUMED 2.6.1 (Tribello et al., 2014) and x3dna-dssr (Lu and Olson, 2008; Lu et al., 2015), and were used as prediction labels in our supervised learning scheme. The cluster-averaged SWAXS profiles were used as inputs. We then trained a LightGBM (Ke et al., 2017) model to map SWAXS signals to global and local structural properties. Additionally, although CRYSOL treats solvent implicitly during SWAXS calculations, we also examined whether our LightGBM-based framework can recover information about coordinating water molecules from the scattering profiles. To this end, we quantified hydration by counting water molecules within 3.5 Å of the HJH RNA surface using the full MD simulation box for the frames employed in the SWAXS calculations.

Machine learning model development and testing

We employed a gradient boosting framework implemented via LightGBM (Ke et al., 2017) to predict a set of structural descriptors directly from SWAXS profiles (Equations 1–5). We trained each system separately. We applied a data-splitting strategy that splits the models into training (72%), validation (13%), and testing (15%) sets based on RNA conformations. LightGBM is a highly efficient tree-based learning algorithm, which builds an ensemble of weak learners (typically decision trees) in a stage-wise fashion to minimize a given loss function. We used the training set to fit the LightGBM model, the validation set is used for hyperparameter selection and early stopping, and the test set is used for final performance reporting. Briefly, for a dataset with n samples { (x i, y i) } i = 1 n, where x i represents the input features and y i are the corresponding labels, the model prediction y ̂ i is formed as a sum over the outputs K additive decision trees: where F denotes the function space of decision trees. At each boosting round t, the model is updated iteratively by adding a new tree f t to minimize the regularized loss: where l is the differentiable loss function that measures prediction error, and Ω (f) is a regularization term penalizing model complexity. LightGBM optimizes this loss using histogram binning and leaf-wise tree growth, as applied to high-dimensional datasets.

For all regression models, we used the LGBMRegressor implementation from the lightgbm Python library, interfaced through scikit-learn (Pedregosa et al., 2011), with a fixed set of hyperparameters across targets and systems. We did not perform an extensive hyperparameter search; a single LightGBM configuration (Table 1) was used for all targets and systems. Models were trained with up to 5,000 boosting iterations and early stopping based on the validation root-mean-square error (RMSE), with a patience of 100 rounds. The main LightGBM hyperparameters are summarized in Table 1.

SHAP analysis and q -to-structure mapping

We used SHAP (SHapley Additive exPlanations) (Shapley, 1953; Lundberg and Lee, 2017) with TreeExplainer on the LightGBM model to quantify each input SWAXS feature’s contribution to the prediction of a given structural descriptor. For an input feature i in the set F, providing both global and local interpretability of how SWAXS inputs influence predicted structural properties.

SWAXS q -to-structure mapping

Model inputs were SWAXS intensities on a common q -grid, so each input corresponds to I (q j). For a target structural descriptor, we summarized importance by the mean magnitude | ϕ j | ̄ and determined value → impact direction using a median split of x j = I (q j):

We then mapped SHAP onto the experimental curve I (q) by plotting points at (q j, I (q j)) with size ∝ | ϕ j | ̄ (importance) and color by Δ j (blue < 0, red > 0, zero-centered), yielding a direct visualization of which q -regions drive the prediction of that structural propertie and in which direction.

To infer RNA structural features from scattering profiles, we used the explicit-solvent MD simulations trajectories generated in (duplex RNA (He et al., 2021) and helix-junction-helix RNA (He et al., 2022)). The trajectories are clustered into representative conformational subfamilies, each approximating a basin. For each cluster, we compute the average SWAXS intensity I (q) and structural properties (e.g., R g, groove widths). Training on these cluster averages, rather than individual frames, suppresses frame-level noise, and improves identifiability. We train LightGBM regressors on SWAXS features, intensities sampled at defined q points, to predict the corresponding structural descriptors. Once training is achieved, we use SHapley Additive exPlanations (SHAP) analysis to map q -resolved contributions along the SWAXS curve to the predicted structural descriptors, directly highlighting the most informative regions of the profiles and the structural properties that they report. A schematic overview of the MD/ML framework is shown in Figure 1. Full details of the computations and model training are provided in the Materials and Methods section. Here, we report how this methodology allows us to decode structural features from RNA systems.

RNA duplex structural features from small and wide angle X-Ray scattering

We focus first on dsRNA duplexes. Duplexes exhibit well-defined periodicities in their structure, which generate rich scattering features at wide angles. However, these features are not independent. As shown in Figure 2A, the autocorrelations among SWAXS features are high. Many SWAXS features are coupled across multiple q regions. The correlation matrix for the geometric descriptors, on the other hand, shows a weaker correlation (Figure 2B). The correlations in structure are geometric restraints related to quantities such as helical inclination, base step roll, etc. To contextualize these relationships, we summarize the marginal distributions of each structural propensity, highlighting the accessible ranges and dispersion of global (e.g., R g, end-to-end distance) and local geometry (e.g., major-groove width) (Figure 2C). We construct the learning such that LightGBM models aim to predict structural properties using the SWAXS feature vector as input. We aim to find whether such a geometric property of interest is indeed measured by SWAXS and, if so, what unique SWAXS feature (q α, I (q α)) reports on the property.

The model accurately decodes multiple geometric descriptors from the SWAXS profiles. Figure 3 summarizes model performance across properties, reported separately as train/validation/test for these properties. Based on this ranking, the best descriptors confidently reported by the profile are R g, major groove width (PPmajor), minor groove width (PPminor), helical radius (helixradius_P), helical twist (helical_hTwist), and helical inclination (helical_Incl), respectively. Interestingly, all these properties come from the backbone, confirming that the phosphate backbone provides the strongest electron density contrast, and SAXS of duplex reports most confidently on the backbone. The nucleobase orientations show weaker correlation scores (0.6 ≤ R 2 ≤ 0.7), sugar geometry, and properties like step rise, tilt, and shift are invisible to SWAXS for duplex RNA.

The model achieves high accuracy in R g prediction, with train, validation, and test R 2 values approaching 1.0 (Figures 4A,B). SHAP analysis suggests the q -regions important for R g (Figure 4C). To show the dependence on real space projection of the structural feature and corresponding profile signal, we used beeswarm plots. Here, point color encodes the feature value; red points to the positive SHAP impact and blue corresponds to the opposite effect. Based on this, at q ∼ 0, the higher I (q) means the higher R g value, whereas blue color at q ≈ 0.05–0.06 Å -1 leads to higher R g when I (q) is lower. Mapping the SHAP impact onto the SWAXS profile (Figure 4D) shows that the model focuses on the Guinier regime of the curve, where it behaves consistently with ln ⁡ I (q) = ln ⁡ I (0) − R g 2 q 2 / 3.

In addition to R g structural properties with high confidence are shown in Figure 5. In each case, SHAP mapping reveals that multiple q features collectively encode the structural features. As shown in Figure 5A, SWAXS-informed ML resolves the signatures that encode the major groove. SHAP mapping identifies distinct q -windows with opposite contributions: negative at q ∼ 0.10, 0.60, and 0.90 Å -1 (higher I (q) associated with a narrower groove) and positive at q ∼ 0.30 and 0.70 Å -1 associated with a wider groove. This pattern indicates that multiple q -resolved signals act in concert to report major-groove variation. This is consistent with the SWAXS autocorrelation in Figure 2A and reflects the coupled nature of wide-angle signals. Minor groove width on the other hand, correlates strongly with wide angles (notably q ≳ 1.5 Å − 1, Figure 5B), consistent with the sensitivity of WAXS to fine radial and groove-scale order. Moreover, roll is coupled to SWAXS features across the full q -range, with a dominant negative correlation near q ∼ 0.25 Å -1 and broader correlations at higher q (Figure 5C). Furthermore, the helical radius shows positive correlations at around q ∼ 0.35 and 1.05 Å -1 (Figure 5D), with wider helices associated with enhanced intensity in these q regions. This reflects that these q regions correspond to real-space dimensions of duplex diameter. Scattering intensities near q ∼ 0.35–0.40 Å -1 are highly informative for predicting step slide (Figure 5E) and helical twist (Figure 5F). SWAXS signals near these q regions reflect the local arrangement of base pairs about the helical axis.

Figure 6 showcases how SHAP feature attribution helps to elucidate structural details in experimental SWAXS profiles. Here, the same RNA duplex is measured under different ionic conditions, KCl and MgCl 2, leading to differences in the profiles. To interpret the differences in the structure of the RNA, we used the mapping we have discussed in Figure 4. To benchmark our predictions, we used the best-fit real-space ensembles previously reported He et al. (2021).

SHAP assigns importance for R g to the low- q Guinier region (Figures 4C,D). The ratio of I(q = 0.0)/I(q = 0.053) shows a higher value for MgCl 2 compared to KCl, suggesting R g of RNA in MgCl 2 is larger (orange band, Figure 6A). Consistent with this, the best fit structures in MgCl 2 condition show a larger R g than KCl (22.1 Å vs. 21.3 Å; Figure 6B), matching the low- q slope differences in the experimental curves. Second, the ML framework identifies q ∼ 0.3 Å − 1 as the dominant region for major-groove width information (Figure 5A), where increased I (q) is associated with groove widening. In the experimental profiles, this region shows a clear increase in KCl relative to MgCl 2 (green band, Figure 6A), suggesting that in KCl the duplex has a wider major groove. Consistent with the interpretation, the best fit structures show a wider major groove in KCl (9.3 Å vs. 8.5 Å; Figure 6D). (He et al., 2021) Third, SHAP highlights bands near q ∼ 0.35 and 1.1 Å − 1 informative regions for helical radius (Figure 5C), where higher I (q) is associated with an expanded helical radius. Consistent with this interpretation, the best fit structure shows a larger helical diameter in KCl (25.7 Å vs. 17.8 Å; Figure 6C).

Two-way junction RNA structural features from small and wide angle X-Ray scattering

Next, we examine whether this framework works for conformationally more flexible motifs, such as two-way juctions, which are ubiquitous in the RNA world (He et al., 2022; Bailor et al., 2011). The added flexibility through the junction leading to conformational heterogeneity is known to smear out some of the important periodic correlations. Despite the added flexibility, SWAXS still exhibits coherent autocorrelations across q (Figure 7A), structural properties, on the other hand, show weak correlation (Figure 7B). Most notable correlations are at helix-to-helix distance and global properties such as R g. A concise overview of these propensities, indicating variability across global (e.g., molecular R g) and local geometry (e.g., junction R g, coordination waters) are shown in Figure 7C.

LightGBM model performance across these properties reports an overall R 2 consistent with the expectation that conformational heterogeneity masks the fine details (Figure 8). Nevertheless, the top-performing targets include R g, end-to-end distance, and the water coordination around RNA; major groove width, and bending angle between the two helical segments. Some junction-scale descriptors (e.g., Junction_ R g) and base-pair step parameters (e.g., step_Slide) are modestly recovered; however, a robust interpretation is not recommended for properties possessing R 2 ≤ 0.5.

Based on the ML model, the q -resolved SWAXS features for two-way junction are bending/opening, intra-helix distance (Figure 9A), and R g for the RNA are reflected primarily at low q, with weak additional contributions around q ≈ 0.15 Å − 1 (Figure 9B). The coordination number of waters, which shows strong correlation with RNA global features, manifests itself in a broad q range (Figure 9C). Similar to the duplex, the major-groove width of the duplex regions is informed by bands near q ∼ 0.3, 0.6, 0.7, and 0.9 Å -1 (Figure 9C). The structural signature at q ∼ 0.3 Å -1 is diminished in the two-way junction, consistent with the junction interrupting helical periodicity and shortening the correlation length, leading to blurring the low- q signature of groove expansion. By contrast, higher- q features (≳ 0.6 Å -1) remain robust, reflecting the sensitivity of wide-angle scattering to local groove-scale geometry even in such a flexible RNA motif (Figures 9 D–E).

Our goal is to introduce an ML framework that decodes RNA structural properties for a given fold directly from SWAXS over the full q -spectrum, explicitly leveraging the interpretation of solution data to access real-space projections. In this study, the models are trained separately for a duplex and a helix–junction–helix construct and are intended to interpret structural fluctuations and condition- or time-dependent perturbations within these known folds, rather than to perform de novo structure prediction across arbitrary RNA sequences and architectures. To achieve this, we computed SWAXS profiles from MD ensembles and trained LightGBM models with SHAP-based interpretation. For both rigid and flexible RNA systems, the approach recovers global and some local descriptors with high accuracy. We find that high- q features (q ≳ 1.0 Å − 1) carry substantial information about local structure. The SHAP-to-profile mapping that overlays feature importance directly onto I (q) links between SWAXS features and RNA structural readouts. We emphasize that SHAP attributions are interpreted here as statistical, model-based associations rather than direct evidence of physical causality. A q -region with high SHAP contribution is considered informative for predicting a given structural descriptor (Figure 6), but this does not imply that intensity variations at that q -value physically cause the corresponding structural change. Because a one-dimensional SWAXS profile reflects collective contributions from many interatomic distances, the SHAP-to-profile mapping should be viewed as a way to localize diagnostic q -regions and guide structural interpretation. This task is prohibitive manually, even at the maximum resolution, d min = 2 π / q max ∼ 3.0 Å. SWAXS alone cannot resolve atomistic detail without complementary atomistic models from MD simulations.

Specifically, our ML models confirm the importance of the Guinier regime for determining R g, consistent with experimental practice. The learned importance pattern recapitulates classical Guinier analysis (Blanchet and Svergun, 2013). Moreover, in duplexes, enhanced scattering near q ≈ 0.30 Å − 1 correlates with major-groove widening, consistent with prior observations (He et al., 2021). In the two-way junction, the attenuated q ≈ 0.30 Å − 1 signature reflects shorter correlation lengths introduced by the presence of a single-stranded junction, whereas high- q features remain robust reporters of local geometry. Notably, high- q features (q ≳ 1.0 Å − 1) also carry information about local hydration, enabling prediction of the water coordination number around the RNA two-way junction.

Looking ahead, this approach offers a practical tool to interpret experimental scattering for a given RNA fold. In its current implementation, the framework reliably learns how fluctuations within a specific RNA motif are encoded in its SWAXS profile and can thus be used to interpret condition- or time-dependent changes for that motif once its overall architecture is known. It enables readouts of structural geometry and hydration environments that shape RNA conformations and mediate molecular recognition. Immediate extensions include generalizing to junction-rich RNAs, RNP assemblies, and ligand- or protein-stabilized states, where interpretable ML can identify the q -regions most diagnostic of functional change. In addition, coupling the framework with time-resolved SWAXS may resolve transient intermediates and reconstruct dynamical pathways of RNA constructs, mapping evolving q -resolved signatures onto sequences of structural states in real time. Extending the approach to a more universal setting that spans different RNA lengths, base-pair counts, and secondary and tertiary architectures will require substantially larger, multi-system training sets and careful treatment of sequence and motif diversity, and we regard this as an important direction for future work.

In conclusion, this study establishes a framework for decoding RNA structural properties from SWAXS data. Using SHAP-enabled, interpretable machine learning trained on simulation-derived ensembles, we recover both global and fine local structural descriptors from q -resolved signatures. The approach enables rational, data-driven analysis of solution scattering and provides a practical tool to probe recognition, regulation, and perturbations in RNA structures. We envision its immediate use in interpreting time-resolved data as well as high-throughput screening of ligand-induced structural changes in RNA.