RNADyn: A Benchmark for Generating
and Understanding RNA Dynamics

Yiming Huang¹Lennart Bastian¹Hanqun Cao²Luis Vollmers³Tolga Birdal¹

¹ Imperial College London² The Chinese University of Hong Kong³ Technical University of Munich

Overview

RNADyn brings together standardized all-atom RNA trajectories and a unified model for generating and understanding RNA dynamics. RNADynBench provides the data foundation; RNADynNet learns to generate trajectories and extract dynamics fingerprints from a single conformer.

Key Contributions

RNADynBench — A standardized data foundation. 2,585 quality-controlled, 100-ns all-atom RNA trajectories, with leakage-controlled splits and evaluation protocols for trajectory generation and dynamics prediction.
RNADynNet — A unified dynamics model. A shared backbone connects RNA trajectory generation and single-conformer dynamics representation learning through trajectory-informed alignment and physical grounding.
All-atom RNA trajectory generation. A benchmark evaluating molecular geometry, fluctuations, and conformational distributions across structural-generalization and high-flexibility test sets.
Transferable dynamics representations. Single-conformer fingerprints encode information about MD-derived fluctuations and motion coupling and support transfer to RNA–ligand binding prediction.

RNADynBench

Experimental structures, standardized simulations, and shared evaluation protocols.

2,585MD trajectories

1,469Unique RNAs

100 nsPer trajectory

258.5 μsTotal simulation time

Manuscript Figure 1: RNA size coverage, hydration and size-matched RNA–protein motion comparisons — **Structural coverage and physically plausible hydration.**
Panel **(a)** summarizes RNA length and size coverage, showing increasing mean radius of gyration with sequence length. Panel **(b)** shows phosphate hydration profiles with a first-shell peak near 2.7 Å, consistent with prior nucleic-acid simulations and supporting the physical fidelity of RNADynBench.
**RNA exhibits greater conformational motion than size-matched proteins.**
Compared with ATLAS and dynamicPDB, RNADynBench shows higher lag-RMSD **(c)**, local RMSF **(d)** and pairwise RMSD **(e)** within the sampled simulation window.

Explore RNADynBench →Browse trajectories →

RNADynNet

RNADynNet unifies RNA trajectory generation and single-conformer dynamics representation learning within a shared spatiotemporal diffusion backbone. Trajectory-informed alignment and physical grounding connect the two capabilities during training.

Trajectory generation

Generate RNA conformational trajectories from an initial conformer, conditioned on a temporal horizon and sampling interval. A hierarchical forecasting-and-interpolation sampler produces the trajectory.

Dynamics fingerprints

Extract a dynamics-aware representation for each residue directly from a single conformer, without sampling a trajectory. Physical prediction heads further estimate residue covariance and normalized motion coupling (NMC), with RMSF derived from the predicted covariance.

Trajectory denoising

Reconstruct RNA trajectories under heterogeneous coordinate noise, learning from spatial relationships between residues and temporal context across frames.

Single-frame-to-trajectory alignment

Train single-conformer representations to match trajectory-contextualized residue features, transferring information about motion into the single-frame encoder.

Physical grounding

Supervise both trajectory and single-conformer representations with MD-derived residue covariance and inter-residue motion coupling, tying learned features to measurable dynamics.

Results & Examples

Trajectory generation

Generate an RNA trajectory from an initial conformer. Compare geometric quality, motion statistics and runtime across methods.

Method	K	Bond MAE ↓	O3′–P max ↓	Chiral flip (%) ↓	Clash / 1k ↓	RMSF r ↑	PCA W₂ ↓	Time / trajectory
Reference	—	0.03308	0.1938	0	0	0.9819	0.7295	8.1 h
ConfRover	1	0.19315	120.4747	0.03103	581.4520	0.0666	9.3105	340.4s
MDGen	4	0.08024	15.1488	0.03103	95.9398	0.3737	2.4200	7.2s
BioKinema	4	0.02612	4.4682	0.03836	2.0746	0.8770	1.9276	40.2s
w/o PG	4	0.02418	2.4455	0.03097	1.1884	0.8594	1.9085	/
RNADynNet	4	0.02424	1.8858	0.03085	0.8225	0.8746	1.8739	24.5s

Method	K	Bond MAE ↓	O3′–P max ↓	Chiral flip (%) ↓	Clash / 1k ↓	RMSF r ↑	PCA W₂ ↓	Time / trajectory
Reference	—	0.03307	0.1966	0	0	0.9633	3.1137	13.9h
ConfRover	1	0.16506	133.0420	0	419.7848	0.2288	11.1922	501.5s
MDGen	4	0.10786	18.9945	<10⁻⁵	88.9058	0.2994	8.9282	11.0s
BioKinema	4	0.02950	8.3586	0.02012	5.8241	0.7536	5.9406	47.8s
w/o PG	4	0.02555	9.1171	0.00160	2.8416	0.7440	6.1744	/
RNADynNet	4	0.02567	4.2400	0.00093	3.0404	0.7658	5.9169	32.9s

RNA-group macro averages. Distances in Å; runtimes measured per trajectory on H200. K denotes generated samples per case: ConfRover uses K=1; other learned methods use K=4. Reference reports MD quantities; w/o PG omits physical grounding. “/” indicates an unreported runtime. Geometry uses the future 100 frames; RMSF uses all 101 frames. Reproducibility resources are pending release.

Trajectory examples

Reference MD and RNADynNet side by side: 101 saved frames over 0–100 ns. Each pair shares a fixed view and scale, after C1′ alignment to the initial MD structure.

3IVN_A · Low motion · 0–100 ns.

View denoising snapshots

Manuscript denoising figure for 3IVN_A — Intermediate rows are diagnostic clean-trajectory estimates at selected denoising steps; denoising steps are distinct from physical time.

Dynamics from a single conformer

Predict residue-level fluctuations and motion coupling directly from one RNA conformer, without sampling a trajectory.

RMSF profiles

Compare predicted nucleotide fluctuation amplitudes with the MD reference. RMSF is derived from the predicted residue covariance.

Motion coupling

Normalized motion coupling (NMC) describes how nucleotide motions are related. Each matrix shows RNADynNet below the diagonal and MD above.

Structure coloring

The RMSF color scale places fluctuation amplitudes in structural context, highlighting relatively stable and more mobile regions.

Test-set performance

Native prediction heads · RNA-group macro averages.

Metric	test_struct	test_flex
RMSF r ↑	0.8665	0.7827
RMSF MAE (Å) ↓	0.3983	1.5800
Covariance RMWD ↓	0.8651	3.2700
NMC r ↑	0.8115	0.7614
NMC MAE ↓	0.1629	0.2027

Paper

RNADyn: A Benchmark for Generating and Understanding RNA Dynamics

Yiming Huang, Lennart Bastian, Hanqun Cao, Luis Vollmers, Tolga Birdal

arXiv preprint · 2026 · arXiv:2610.03712

View PDF

BibTeX

@article{huang2026rnadyn,
  title={RNADyn: A Benchmark for Generating and Understanding RNA Dynamics},
  author={Huang, Yiming and Bastian, Lennart and Cao, Hanqun and Vollmers, Luis and Birdal, Tolga},
  journal={arXiv preprint arXiv:2610.03712},
  year={2026}
}