acceptodds
Under review as a conference paper at ICLR 2027

Ribomotion: An Atomistic Molecular Dynamics Dataset for RNA Modeling

Abstract

Understanding RNA dynamics is essential for connecting structure to function, yet existing RNA resources primarily capture sequences and static structures. We introduce RiboMotion, an all-atom molecular dynamics (MD) dataset comprising nearly 100 μs of trajectories across 1,039 systems, including isolated RNAs and RNA–ligand complexes. We derive six benchmark tasks covering flexibility, geometry, and stability through nucleotide and pocket fluctuations, distance distributions, and persistent disruption of stacking and hydrogen-bond interactions. Benchmarking shows that trajectory-derived observables contain learnable information beyond initial geometry. As a demonstration of broader utility, multi-frame MD supervision improves a pretrained RNA conformation generator, reducing centered distance-distribution error by 25.4% and flexibility-amplitude error by 35.9% relative to the original model. These results suggest that MD-supervised generators could offer an alternative to direct molecular dynamics simulations for sampling RNA conformational ensembles. All data and benchmarks will be freely available to accelerate the development of AI models that capture RNA dynamics.

Primary area:
datasets and benchmarks
Keywords:
RNA,Molecular Dynamics,Dataset,Benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tEW5OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Modern RNA track is dominated by sequence/static-structure supervision, while functional RNA behavior is inherently ensemble- and dynamics-dependent. RiboMotion addresses a real gap by providing roughly 100$\mu$s of explicit-solvent MD across 1,039 RNA systems, with 6 downstream tasks, and the demo that multi-frame MD supervision can improve an RNA conformation generator.

Critical strengths
  1. The RNA dynamics problem is important to bridge to the functional design, and RiboMotion directly addresses that issue.
  2. RiboMotion have tried broader scales including different simulation regimes, different distinct systems, etc.; The data splitting well noted the similarity issues & temporal orders; the Protenix experiment is compelling with careful experiments controls.
Critical weaknesses
  1. The scientific meaning of the benchmark depends strongly on the reliability and sampling quality of the MD trajectories, but this is not validated sufficiently for a benchmark of RNA dynamics. Most systems are represented by only one approximately 80–100 ns trajectory. Independent simulations are available for only nine Core RNAs, and there is no independent-replica panel for the 309 RNA–ligand systems. Experimental validation is also very limited: it mainly consists of one UUCG construct, for which the source NMR structures agree better with the eNOE measurements than the MD ensemble. The chosen RNA force field is widely used, but RNA force fields, including ff99+bsc0+χOL3, have known limitations for some conformations and interactions. The paper itself also acknowledges that the simulated ensembles depend on the force field and simulation conditions. This issue is consequential because these simulations define the supervision and the ground-truth targets. Many biologically relevant RNA transitions can occur on timescales longer than the simulated windows. Therefore, it is unclear whether the models are learning transferable RNA conformational dynamics or mainly reproducing finite-window observables generated by this particular force field and simulation protocol. In my view, this distinction is fundamental for a benchmark whose main claim is to enable learning of RNA dynamics.

  2. Several benchmark design choices make the reported generalization difficult to interpret. For the Core distance task, entry into the evaluation population depends on diagnostics computed from the target trajectory itself, including half-window RMSF agreement, contact-occupancy agreement, and pair-distance consistency. Thus, the primary benchmark evaluates models on a population that has been retrospectively selected using the answer trajectories. I understand the motivation of removing poorly sampled targets, but this creates a target-conditioned evaluation population rather than a fully prospective benchmark. This is especially difficult to justify because Appendix F.5 already shows results on all 97 technically valid test RNAs. The RNA–ligand split also does not enforce ligand-scaffold separation. Therefore, the reported test performance does not clearly establish generalization to unseen ligand chemotypes. In addition, the primary hydrogen-bond MacroAP is unconventional: systems without positive events receive AP = 0, giving a maximum score of only 20/39 on the test set, while false positives inside these systems do not affect their AP. This makes the absolute value and model ranking difficult to interpret, especially because the supplementary system-weighted AP and Brier metrics give a different view of performance. Together, these design choices make it difficult to determine how well the benchmark measures the intended out-of-distribution prediction of RNA dynamics.

Questions for authors
  1. Clarification / benchmark design. Why are trajectory-consistency criteria used to determine entry into the primary Core distance benchmark rather than only to define diagnostic strata? Given the results in Appendix F.5, what is the reason for not using all 97 technically valid test RNAs as the primary evaluation population?
  2. Additional evaluation that would affect my assessment. Can the authors report a ligand-scaffold-disjoint evaluation while also preserving RNA-level separation? In particular, how much of the RNA–ligand performance remains when the test ligands have unseen chemical scaffolds?
  3. Sampling uncertainty. How much of the reported benchmark error can be attributed to finite-trajectory sampling variability? The independent-replica analysis on nine RNAs is useful but small, and no equivalent replica analysis is available for the ligand cohort. Can the authors provide a broader replica analysis, or otherwise quantify the sampling uncertainty relative to the model improvements?
  4. Event-label robustness. How sensitive are the stacking- and hydrogen-bond-disruption results to the chosen persistence threshold? For example, do the relative model rankings remain similar when the persistence criterion is approximately 0.5 ns, 1 ns, and 2 ns?
  5. Ligand trajectory integrity. How many RNA–ligand systems are affected by the whole-ligand translation correction described in Appendix A.3, and how much do the ligand-distance results change if these systems are excluded? The manuscript states that the available exports do not contain valid periodic boxes and that physical validation of these corrections would require the original boxed trajectories
Reviewer_Fvu1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors introduce RiboMotion, a molecular dynamics (MD) dataset for RNA and RNA-ligand complexes. The release of this dataset is timely and important and may shape the future of RNA generative models. Empirically, such data seems to improve existing biomolecular structure predictors (Protenix) in predicting RNA structural conformations.

Critical strengths
  1. RiboMotion fills a gap in the literature regarding the availability of all-atom RNA dynamics data for machine learning methods.
  2. RiboMotion's data seems to improve existing biomolecular structure predictors' ability to predict RNA structural conformations.
  3. The authors describe and motivate the composition of RiboMotion well.
Critical weaknesses
  1. The paper is written rather poorly, resembling verbatim LLM-produced text in several places (e.g., "RiboMotion cohorts", "Development-cohort", "All three arms", "diffusion-module", etc.). This makes it difficult for readers to easily grasp this paper's main findings/takeaways.
  2. The authors do not comprehensively utilize MD data for finetuning tasks. For instance, it would have been nice to see whether MD-derived metadata (e.g., energies or properties) could have made finetuning of machine learning methods more expressive.
  3. The authors do not provide any qualitative case studies of how MD finetuning improves prediction of particularly challenging RNA conformations to model. Visualizing some RNA trajectories predicted by a fine-tuned method (e.g., Protenix) would have helped in this regard.
Questions for authors
  1. Would the authors please use consistent spelling for RiboMotion? I noticed that the title of this paper on OpenReview differs from that referenced in the paper.
  2. Would the authors please rewrite portions of the manuscript (particularly Section 4) to improve readability? Specifically, rewriting any directly LLM-generated text would be very helpful.
Reviewer_PnfWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Authors introduce of a dataset of MD trajectories of RNAs and RNA-ligand complexes. The dataset covers 1,039 short systems concentrated in the 20-80 nt length, and a short simulation length per system of about 100ns. This kind of broad, short simulations is very close in spirit to an RNA MISATO dataset.

In addition, authors propose a benchmark with tasks assessing prediction of dynamic properties (where the ground truth is inferred from RiboMotion MD trajectories) from static structures. They retrain the diffusion module of Protenix by supervising it with multiple RiboMotion MD frames instead of a static structure, and show Protenix generates conformational ensembles closer to the MD ground truth on a heldout test set when supervised using multiple MD-frames than using a single structure.

The writing is too descriptive, lacking interpretation especially for the benchmark, and some points need to be clarified before accepting the paper regarding the methodology of the MD-supervised conformational ensemble generation where I suspect inflated results due to an overrepresentation of structurally similar tRNAs in train and test.

Critical strengths

S1. The dataset is well-motivated by the dynamic nature of RNA

S2. Novel benchmark for the prediction of RNA dynamic properties

S3. The proposed dataset seems to improve Protenix performance

Critical weaknesses

W1. Limited benchmark analysis. I'm guessing the submission caught you short and you did not have time to obtain exhaustive results before the deadline (Table3/4).

W2. Limited data analysis. For instance, there is no family breakdown. How much of your dataset is tRNA? Did you remove redundancy? Did you start from all RNA or the non redundant subset from BGSU?

W3. Possible overstatement: you fine-tuned protenix on dynamics, but you also fine-tuned it on RNA only. One could imagine that this is the main cause for RNA progress, not the inclusion of dynamic information. Could you include this (similar fine-tuning on RNA static structures) as a baseline ?

Questions for authors

Please adress the questions corresponding to the stated weaknesses. I am open to increasing my score if you address them well.

Below are additional smaller questions. Q1. Missing related works on benchmark for learning on RNA structure:

  • A Comprehensive Benchmark for RNA 3D Structure-Function Modeling (Wyss et al.)
  • Beyond Sequence: Impact of Geometric Context for RNA Property Prediction (Xu et al.)

Q2. What is the difference between what you call “first production frame” and “first exported MD complex”?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.