Source-linked AI summary
RNADyn: A Benchmark for Generating and Understanding RNA Dynamics
Yiming Huang, Lennart Bastian, Hanqun Cao, Luis Vollmers, Tolga Birdal
TL;DR
RNA conformational dynamics are not captured by static structures, while standardized RNA dynamics data remain limited and generation and understanding are often treated separately. RNADynBench provides a standardized MD benchmark, and RNADynNet jointly learns trajectory generation and single-conformer dynamics representations. Physical grounding improves both generated dynamics and the physical information recoverable from fingerprints.
Problem
RNA function involves conformational changes that static structures do not capture, while large-scale standardized RNA dynamics data remain limited.
Source: §1 · p. 1Method
RNADynBench supplies RNA MD trajectories and derived statistics for RNADynNet, which jointly learns trajectory generation and single-conformer dynamics representations with physical supervision.
Source: §1 · p. 2, §2 · p. 2, §3.2 · p. 6Results
Takeaways & Limitations
RNADynBench and RNADynNet provide a foundation for RNA models that learn conformational dynamics beyond static structure.
Source: §6 · p. 9The 100-ns simulations may miss slower dynamics, and the study focuses primarily on isolated RNA rather than complete complexes.
Source: §6 · p. 9
Abstract
from arXiv · showhide
Ribonucleic acid (RNA) functions through conformational changes that are not fully captured by static structures. However, large-scale standardized RNA dynamics data remain limited, and existing approaches typically treat trajectory generation and dynamics understanding as separate objectives. Here, we introduce RNADynBench, a standardized RNA molecular dynamics (MD) benchmark with 2585 quality-controlled 100-ns all-atom trajectories and leakage-controlled splits. Building on RNADynBench, we develop RNADynNet, a unified model for RNA dynamics learning that uses a shared backbone for both trajectory generation and dynamics fingerprint extraction from a single conformer. It combines coordinate denoising, single-frame-to-trajectory alignment, and physical grounding to connect all-atom trajectory generation with dynamics representation learning. Physical grounding improves both generated dynamics and the physical information recoverable from these fingerprints. Across both test sets, including the high-flexibility challenge set, the generated trajectories achieve RMSF correlations of 0.875 and 0.766, while single-conformer predictions show comparable agreement with MD-derived dynamics. RNADynBench and RNADynNet together establish a benchmark and unified modeling framework for generating and understanding RNA dynamics.
1 INTRODUCTION
RNA function depends on conformational dynamics that static structures do not capture, yet RNA-specific data and dynamics-aware modeling remain limited. The paper introduces a standardized RNA MD benchmark and a unified model, using physical grounding to improve generated dynamics and learned representations.
1 INTRODUCTION
- RNA function involves conformational changes beyond static structures, making dynamics central to modeling its regulatory, recognition, and catalytic roles.RNA conformations can also be reshaped by ions, ligands, and proteins.Source: §1 · p. 1
- RNA structural data are less abundant and diverse than protein data, and current approaches leave challenges in conformational coverage and dynamics-aware modeling.RNA multiple-sequence alignments also pose distinct challenges, motivating RNA-specific data and evaluation.Source: §1 · p. 1
- In size-matched comparisons, RNA trajectories show greater lag-dependent displacement and generally larger local fluctuations and pairwise RMSD than protein trajectories.These observations motivate conditioning RNA modeling on molecular motion rather than static structure alone.Source: §1 · pp. 1–2
- MD trajectories quantify fluctuations and correlated motion, but standardized RNA MD datasets remain scarce and costly to construct.Trajectory generation consumes substantial GPU resources, while system preparation and simulation protocols require domain expertise.Source: §1 · p. 2
- RNADynBench supplies data for RNADynNet, whose physical grounding trains representations to predict MD-derived covariance and inter-residue motion coupling.The authors report that physical grounding improves both generated dynamics and learned dynamics representations; fingerprints also transfer to RNA–ligand binding prediction.Source: §1 · p. 2
2 RNADYNBENCH: AN RNA MOLECULAR DYNAMICS BENCHMARK
RNADynBench standardizes RNA MD data and evaluation for trajectory generation and single-conformer dynamics prediction. Its trajectories show greater conformational motion than size-matched protein trajectories, while benchmark construction includes curation, physical validation, and leakage-controlled splits.
RNADYNBENCH: AN RNA MOLECULAR DYNAMICS BENCHMARK
- RNADynBench contains 2,585 trajectories across 1,469 RNAs, totaling 258.5 µs, with leakage-controlled splits and shared protocols for two benchmark tasks.The tasks are generating 100-ns trajectories from an initial conformer and predicting MD-derived statistics from a single conformer.Source: §2 · pp. 2–3
Data Curation
- Curation retains the target RNA chains and only qualifying RNA partners, while excluding structures with unresolved backbone gaps, unsupported chemistry, or untraceable topology changes.Canonical mappings link source structures, simulation topologies, and released coordinates.Source: §2 · p. 3
Molecular dynamics simulation
- Each curated structure undergoes a 100-ns all-atom production simulation in explicit OPC water with the Amber RNA OL3 force field at 300 K and 1 bar.Coordinates are saved every 100 ps, and the final 5-ns NPT equilibration is excluded from released trajectories.Source: §2 · p. 3
Quality control and benchmark splits
- Quality control screens for simulation failures and artifacts, and related systems are assigned together to limit evaluation leakage.The release includes separate structural-generalization and high-flexibility test sets, plus homology-guard trajectories for auditing.Source: §2 · p. 3
Dataset characteristics
- RDF patterns for phosphate oxygens and water are consistent with prior nucleic-acid simulations, supporting the physical fidelity of RNADynBench.The median first peak was 2.675 Å for both O1P and O2P; the median first minima were 3.225 Å and 3.275 Å, respectively.Source: §2 · pp. 3–4
- Across two references, RNA showed consistently higher lag-RMSD and broader local RMSF and pairwise RMSD distributions than proteins over 1–20 ns.RNA and protein systems were reweighted to a shared heavy-atom-size distribution before comparison.Source: §2 · p. 4
3 RNADYNNET: UNIFYING RNA MD UNDERSTANDING AND GENERATION
RNADynNet uses one shared spatiotemporal diffusion backbone for trajectory generation and single-conformer dynamics fingerprints, coupling the paths through alignment and MD-derived physical supervision.
3 RNADYNNET: UNIFYING RNA MD UNDERSTANDING AND GENERATION
- Trajectory generation and single-conformer fingerprint extraction share a spatiotemporal diffusion backbone, with training coupling the paths through alignment and MD-derived physical supervision.The single-conformer path encodes a conformer into a dynamics fingerprint.Source: §3 · p. 4
3.1 TRAJECTORY AND SINGLE-FRAME TRAINING
- Heterogeneous coordinate noise encourages the denoiser to use cleaner spatial and temporal context to reconstruct noisier regions across residues and frames.The generation objective also supervises local structure, fluctuation measures, and short-lag temporal correlations.Source: §3.1 · p. 5
- Single-frame-to-trajectory alignment trains the encoder to predict residue features from a trajectory pass, transferring trajectory-level motion information to a single conformer.The trajectory target incorporates the full window through temporal attention; fingerprint extraction requires neither H nor Δt.Source: §3.1 · p. 5
3.2 PHYSICAL GROUNDING
- Physical grounding uses MD-derived residue covariance and inter-residue motion coupling to supervise intermediate representations from trajectory-anchor and single-conformer views.These statistics serve as supervision, not encoder inputs.Source: §3.2 · p. 6
4 EXPERIMENTS
The experiments evaluate trajectory generation, single-conformer dynamics prediction, and RNA–ligand binding transfer. Across these tasks, RNADynNet is compared with adapted trajectory-generation methods, dynamics-prediction baselines, and binding models using its frozen fingerprints.
4.1 RNA TRAJECTORY GENERATION UNDER DYNAMICS-AWARE EVALUATION
- RNADynNet improves all four reported geometric metrics over BioKinema on both test sets, with comparable or better RMSF correlation, lower PCA-W2, and faster sampling.Relative to BioKinema, maximum O3′–P deviation and clash rate fall by roughly half on both sets.Source: §4.1 · p. 7
- Physical grounding improves RMSF correlation, O3′–P continuity, and PCA-W2 on both test sets, while bond MAE remains similar.Clash rates decrease on test struct and change modestly on test flex.Source: §4.1 · p. 7
4.2 LEARNING RNA DYNAMICS FROM SINGLE CONFORMERS
- RNADynNet’s native heads outperform BackFlip and DynaProt on all seven reported metrics in both test panels, with RMSF correlations of 0.8665 and 0.7827.NMC correlations are 0.8115 and 0.7614 on test struct and test flex, respectively; frozen-probe results show alignment improves RMSF and covariance metrics, while physical grounding improves all seven metrics.Source: §4.2 · p. 8
- At later checkpoints, larger training subsets generally reduce NMC MAE and increase covariance-ellipsoid IoU.The evaluation spans 10–100% of training data and 1k–6k steps.Source: §4.2 · p. 8, Figure 4 · p. 8
4.3 RNA–LIGAND BINDING
- Frozen RNADynNet fingerprints improve RNA–ligand binding AUROC in five of eight settings, with gains concentrated in settings involving RNA distribution shift.On combined homology-and-fingerprint splits, AUROC rises from 0.7279 to 0.7673 on Biosensor and from 0.6176 to 0.6619 on Robin.Source: §4.3 · p. 8
5 RELATED WORK
Related work spans biomolecular MD resources, ensemble and trajectory generation, structure-based dynamics prediction, and methods that guide generative or representation learning. RNADynBench adds a larger RNA trajectory collection, while RNADynNet combines trajectory generation with single-conformer dynamics learning.
Biomolecular MD Data
- RNA trajectory resources remain smaller and more specialized than protein MD resources; RNADynBench contributes a larger all-atom collection with RNA-group partitions and shared protocols.Earlier resources include BIGNASim, initially with 14 RNA simulations, CASP15 MD refinement evaluation, and simulations assembled for DeepRMSF.Source: §5 · pp. 8–9
Ensemble and Trajectory Generation
- Existing RNA ensemble generators sample conformations without temporal order, while time-resolved trajectory generators were developed for proteins rather than RNA.The cited trajectory approaches include ConfRover, MDGen, BioMD, and BioKinema.Source: §5 · p. 9
Dynamics Prediction from Structure
- RNADynNet extends single-structure RNA dynamics prediction beyond scalar flexibility by learning MD-derived residue covariance and inter-residue motion coupling.Prior examples include RNA elastic networks and DeepRMSF for flexibility, and DynaProt and Ensembits for protein dynamics.Source: §5 · p. 9
Guided Generative and Representation Learning
- Where prior methods add auxiliary signals to generation or representation learning, RNADynNet jointly uses MD trajectories to supervise trajectory generation and single-conformer dynamics representations.The cited approaches include REPA, HoG-D, and PSG-JEPA.Source: §5 · p. 9
6 CONCLUSION
RNADynBench and RNADynNet provide a foundation for jointly learning RNA trajectory generation and single-conformer dynamics prediction. The study notes that its simulations and molecular scope constrain which dynamics and interactions are represented.
6 CONCLUSION
- Physical grounding improves generated dynamics and the physical information recoverable from single-conformer fingerprints; an initial RNA–ligand binding evaluation illustrates downstream transfer.The conclusion frames RNADynBench as a foundation for moving RNA modeling beyond static structure toward dynamics linking molecular structure, recognition, and function.Source: §6 · p. 9
- The 100-ns simulations may miss slower conformational dynamics, and the study focuses primarily on isolated RNA rather than complete RNA–protein or RNA–small-molecule complexes.Future work could use the benchmark to study mutations, partners, and environmental conditions, or integrate experimental observables.Source: §6 · p. 9
- RNA dynamics models may support understanding function, recognition, and therapeutics, but RNADynNet neither performs de novo RNA design nor establishes biological activity.The authors advise caution when applying such models in molecular-design or therapeutic pipelines.Source: ETHICS AND REPRODUCIBILITY STATEMENT · p. 10
A.1 STRUCTURE PREPARATION AND MD SIMULATION
- The released files contain RNA-only postprocessed coordinates; systems with periodic-image contacts during equilibration were rebuilt using 18- or 24-Å padding instead of the default 12 Å.The default solvent box was a rhombic dodecahedron.Source: §A.1 · p. 15
- Accepted trajectories required approximately 23,000 H200-equivalent GPU hours, including equilibration and production.Generation used NVIDIA H200, GH200, and A100 GPUs.Source: §A.1 · p. 16
A.2 QUALITY CONTROL AND DATASET SPLITS
RNADynBench applies structure, trajectory, and release checks, then uses leakage-controlled partitions and size-matched RNA–protein comparisons. The supplied figure caption identifies the plotted measures and their encoding but does not state comparative outcomes.
A.2 QUALITY CONTROL AND DATASET SPLITS
- Quality control checks source-to-trajectory residue and atom identities and excludes structures with ambiguous mappings, internal backbone gaps, unsupported modified residues, or failed topology reconstruction.Run-level and release checks also cover simulation stability, trajectory integrity, periodic boundaries, chain separation, metadata, and duplicate aliases.Source: §A.2 · p. 16
- The structural test set expands 113 entries mapped from the gRNAde Das test set to 117 trajectories, with sequence and structural guards used to control leakage.The passage specifies sequence identity and coverage thresholds and a TM-score threshold for direct structural screening.Source: §A.2 · p. 16
- Test-flex selection ranks grouped trajectories by median pairwise aligned C1′ RMSD, stratifies by length, chain multiplicity, source class, and quality warnings, and selects whole groups toward 100 trajectories.A seeded selection from remaining groups yields 102 validation trajectories; other groups form training.Source: §A.2 · p. 16
A.3 SIZE-MATCHED RNA–PROTEIN COMPARISONS
- RNA and protein systems are compared within five log heavy-atom-count bins, using common bin weights; ATLAS and dynamicPDB are matched separately.The caption specifies mean marker-atom RMSF, mean RMSD to the first sampled frame, and heavy-atom Rg coefficient of variation, with logarithmic vertical axes and split violins.Source: §A.3 · p. 16, Figure 5 · p. 17
B RNADYNNET ARCHITECTURE AND OBJECTIVES
RNADynNet shares a spatiotemporal diffusion backbone across trajectory generation and single-conformer representations, connecting them through alignment and physical supervision. The architecture combines coordinate and dynamics objectives with physical-statistics prediction.
B RNADYNNET ARCHITECTURE AND OBJECTIVES
- The section establishes notation for RNADynNet, including frame and residue indices, while directing readers to the binding-model section for quantities used only there.Bold symbols denote vectors, matrices, or tensors.Source: B RNADYNNET ARCHITECTURE AND OBJECTIVES · p. 17
B.1 BACKBONE AND TEMPORAL CONDITIONING
- A shared diffusion backbone supports generation and dynamics representations, with trajectory-informed alignment and physical supervision linking the two paths.The backbone uses spatial attention followed by temporal self-attention; coordinate prediction continues through the denoising path while latent objectives use an intermediate representation.Source: §B.1 · p. 17
B.2 TRAINING VIEWS AND FINGERPRINT EXTRACTION
- Training constructs aligned, centered trajectory and clean single-frame views, aligns their intermediate representations, and reports fingerprints from an EMA encoder on one conformer.The trajectory view uses one global rotation across frames, while the single-frame path receives a corresponding clean anchor; evaluation uses neither random rotation nor coordinate noise.Source: §B.2 · p. 18
B.3 TRAJECTORY GENERATION OBJECTIVE
- Trajectory generation combines coordinate denoising with local structural agreement, fluctuation matching, and short-lag autocorrelation losses.RMSF, relative/local RMSF, and autocorrelation terms apply only to windows with at least three frames; coordinate errors use a global-frame reconstruction convention without a new rigid alignment.Source: §B.3 · p. 19
B.4 PHYSICAL GROUNDING TARGETS AND PREDICTION HEADS
- Physical grounding predicts residue covariance and inter-residue motion coupling from intermediate representations, using MD-derived reference statistics and local coordinates.The covariance head predicts log-covariance and recovers covariance descriptors by matrix exponential at inference; the covariance and coupling heads are supervised separately, without an explicit cross-head consistency constraint.Source: §B.4 · pp. 19–20
C EXPERIMENTAL PROTOCOLS
- The code and configurations specify remaining masks, reductions, head dimensions, numerical constants, objective weights, and training, sampling, and downstream optimization settings.Source: C EXPERIMENTAL PROTOCOLS · p. 20
C.1 TRAINING CONFIGURATION
Training varies trajectory-window starts, spacing, and lengths, and gradually activates single-frame-to-trajectory alignment after a warm-up. Evaluation uses fixed test panels and a shared forecasting-and-interpolation sampling protocol.
C.1 TRAINING CONFIGURATION
- Training varies window starts, spacings, and lengths, converts stored-frame spacing to physical interval Δt, and mixes unconditional, initial-frame-conditioned, and endpoint-conditioned windows.Source: §C.1 · p. 20
- Physical grounding is active from the first update, while single-frame-to-trajectory alignment begins after warm-up and smoothly increases to its final weight.The alignment target comes from the online trajectory pass, while EMA weights are retained for generation.Source: §C.1 · p. 20
C.2 TRAJECTORY SAMPLING
- Evaluation uses test struct (117 trajectories from 62 RNAs) and test flex (100 trajectories from 65 RNAs); both panels use EMA denoising with coarse anchor generation followed by intermediate-frame interpolation.Each of the four samples is a complete trajectory on the 101-frame, 100-ns evaluation grid.Source: §C.1 · p. 20, §C.2 · p. 20
C.3 BASELINES AND ABLATION SETTINGS
The section specifies trajectory-generation and descriptor-prediction baselines, along with ablations that isolate alignment and physical grounding. The compared methods use distinct RNA inputs, representations, and training setups.
C.3 BASELINES AND ABLATION SETTINGS
- ConfRover, MDGen, and BioKinema are adapted for RNA trajectory generation with distinct input representations, training procedures, and sampling settings.ConfRover generates one trajectory per target and is excluded from the K = 4 ranking.Source: §C.3 · p. 20
- Descriptor baselines use the experimental initial structure, with Geometry-GNN trained jointly and RNA-FM+2D GCN using frozen RNA-FM features.DynaProt combines validation-selected marginal and joint heads; daggered rows use probes trained on frozen representations.Source: §C.3 · p. 20
- Representation-learning ablations successively add trajectory alignment and physical grounding while retaining the generation losses.The variants use Lgen alone, Lgen plus Lalign, or the full model including PG; daggered rows use separate probes on frozen EMA encoders.Source: §C.3 · p. 21
C.4 EVALUATION METRICS AND MD REFERENCES
The evaluation defines trajectory-generation and single-conformer dynamics metrics against MD references, and specifies a downstream RNA–ligand binding protocol. The grounding study compares supervision targets and readout depth, with Block 17 leading generation RMSF correlation and Block 24 yielding lower predictive RMWD on both panels.
C.4 EVALUATION METRICS AND MD REFERENCES
- Generation evaluation measures geometry deviations and clashes alongside RMSF-profile correlation and PCA-W2 between generated and MD future-frame distributions.RMSF correlation compares residue-wise profiles across 101 frames; PCA-W2 compares equal-weight distributions of 100 future frames in an MD-fitted two-dimensional PCA space.Source: §C.4 · p. 21
- Single-conformer predictions are compared with full-trajectory MD descriptors, while a comparison between non-overlapping 50-ns blocks indicates finite-sampling variability.Covariance comparisons use centered Gaussians and positive-definite projected matrices; ellipsoid IoU reflects fluctuation scale, anisotropy, and orientation, but not mean-coordinate error.Source: §C.4 · p. 21
- RMSF and NMC scores report residue-wise fluctuation and pairwise coupling pattern correlations and absolute errors, with RMSF using the raw covariance trace.Scores are RNA-group macro means; RMWD and RMSF MAE are in Å, while other descriptor scores are dimensionless.Source: §C.4 · p. 21
- Reference intervals use cluster resampling for RNA-level geometry and conditional moving-block bootstrap samples for trajectory matching.These intervals describe resampling variability conditional on the observed trajectories.Source: §C.4 · p. 21
C.5 RNA–LIGAND BINDING PROTOCOL
- The binding protocol combines frozen RNADynNet residue fingerprints with frozen sequence and secondary-structure features, then scores RNA–ligand interactions without a complex pose or cross-molecule distances.The fingerprint and pretrained RNA and ligand encoders remain frozen while fusion layers, the interaction stack, and prediction head are optimized.Source: §C.5 · p. 22
D.1 GROUNDING TARGETS AND SUPERVISION DEPTH
- Block 17 has the highest generation RMSF correlation on both test panels, while Block 24 has lower predictive RMWD on both panels.The ablations compare covariance-only and motion-coupling grounding; Block 24 moves fingerprint readout and its alignment and PG supervision from Block 17.Source: §D.1 · p. 22, Table 8 · p. 22, Table 9 · p. 22
D.2 PREDICTIVE DATA SCALING
The predictive-scaling figure covers seven metrics across held-out splits, data fractions, and training checkpoints. Qualitative examples show stepwise generation for held-out RNAs spanning low, median, and high motion.
D.2 PREDICTIVE DATA SCALING
- Figure 6 presents seven metrics across both held-out splits, training-data fractions, and checkpoints.The figure’s orange curves darken from 1k to 6k updates across 10–100% of training data; shading gives pointwise 95% source-RNA-group bootstrap intervals for 6k only.Source: §D.2 · p. 23, Figure 6 · p. 23
D.3 QUALITATIVE TRAJECTORY GENERATION
- Figures 7–9 show stepwise denoising for held-out RNAs selected at low-, median-, and high-motion reference-MD C1′ RMSD percentiles.Each figure shows columns at 0, 25, 50, 75, and 100 ns, with the final rows comparing generated trajectories and reference MD.Source: §D.3 · p. 23, Figure 7 · p. 24, Figure 8 · p. 25
- Ribbon gaps in the qualitative trajectories mark O3′–P distances of at least 2.6 Å.The step rows are diagnostic x̂0 estimates at labeled noise scales.Source: Figure 7 · p. 24, Figure 8 · p. 25, Figure 9 · p. 26