acceptodds
Under review as a conference paper at ICLR 2027

mdRNAcat: A Molecular Dynamics RNA Catalog of REST2 Conformational Ensembles

Abstract

Learning molecular models of RNA requires simulation data that capture conformational variability and provide consistent physical training signals. Here, we introduce mdRNAcat, a curated dataset of all-atom molecular dynamics (MD) simulations of 48 RNA systems selected to promote sequence diversity. Each system is simulated using replica-exchange with solute scaling (REST2), with 32 Hamiltonian states and production lengths of 200–300 ns per state. The dataset provides atomic coordinates, instantaneous forces, and forces reevaluated under the unscaled reference Hamiltonian, together with structural annotations and replica-exchange metadata. Coordinates and forces are available at 1 ps intervals for the reference state and 100 ps intervals across all states in a standardized HDF5 format. We characterize conformational sampling, exchange behavior, base pairing, and sugar-pucker distributions to document the properties of the simulated ensembles. mdRNAcat provides a reusable resource for developing and evaluating data-driven RNA models, including coarse-grained force fields.

Primary area:
datasets and benchmarks
Keywords:
dataset, molecular dynamics, biophysics, RNA

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_23L6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

mdRNAcat is a molecular dynamics dataset for RNA, which contains RNA systems simulated with REST2 enhanced sampling. Each system uses 32 REST2 states and 200-300 ns of simulation per state. The main feature is that it provides not only RNA coordinates, but also instantaneous atomic forces and forces recomputed under the unscaled reference Hamiltonian. The lowest REST2 state is stored every 1 ps, while configs from all REST2 states are stored every 100 ps. This dataset can be a resource for force matching, learned molecular potentials, coarse-grained RNA models, etc.. The dataset was validated through seq and structure-motif coverage of PDB RNA data, vs conventional MD, replica-exchange acceptance & energy overlap, qualitative RNA structural behavior, etc..

Critical strengths
  1. This dataset fills a gap in RNA machine learning. The RNA dynamics problem is important, especially bridging with the functional RNA design and interactions with other molecules. Most of the current ML models deal with static RNA structures, facing the challenge of data scarcity. This dataset may fill in the RNA dynamics dataset gap, facilitating the further development of dynamics-aware RNA ML models.
  2. The biophysics perspective of this paper is generally sound (to me). Firstly the dataset contains rich metadata (coordinates, forces, scaled/reference forces, exchange information, structure annotations, etc.). Secondly this dataset use REST2 to reach configs beyond normal 300K sampling and then recompute forces for those configs under the unscaled Hamiltonian, which is useful for training potentials on configs that are rare in ordinary MD. Thirdly, the authors make the effort to diagnose REST2 behavior, including the report of transition probability, energy overlaps, conformational landscapes at different temperature, per-system energy-overlap plots, etc.. The mean transition probability of 0.183 is close to the desired regime for RE. That is useful evidence that the simulations are not obviously broken.
Critical weaknesses

[1] ML utility is not demonstrated.

  1. There is no enough ML elements. Following my comments in the strengths part, although I think the biophysics is generally grounded and well motivated, there is no model trained on mdRNAcat. There is no baseline, no train/val/test split (even no such recommendation on the split), no seq generalization experiment, no family/structural-level OOD benchmark. As a result, the manuscript demonstrates the generated MD data, but doesn't yet show that the dataset works well for the ML problem that motivate the paper. As in this ML conference, I would expect to see some downstream ML-based/AI4Science style experiments. I think this would also reveal practical problems in the dataset that trajectory-level analysis may not expose. The current manuscript looks more suitable for biophysics or RNA journal like NAR.

[2] Physical validity & coverage of the simulated ensembles are insufficiently established

  1. The schema defines NN as the number of RNA atoms, and coordinates and forces are stored as F×N×3F \times N \times 3. But the simulations use explicit water and ions. If only RNA coordinates are stored, then two identical RNA conformations can experience different forces because the surrounding water and ions are different.
  2. The biochemical env may be unrealistic for many RNAs. The source dataset of the simulation is not balanced. Many riboswitches are included (22 out of 48 systems). However, the simulation protocol describes RNA in TIP4P-D water with KCl and doesn't describe retrained Mg2+Mg^{2+}, bound ligands, and other biologically important cofactors. This is important because many RNA 3D structures (especially riboswitches, which have different states) strongly depend on Mg2+Mg^{2+}, specific ions, bound metabolites, cofactors, etc.. If those components were removed from experimental structures, the resulting ensemble may not correspond to the native biological ensemble. The authors used structural behavior of the trajectories as evidence for dataset validity, and in my opinion the authors should report: 1) which heteroatoms were present in the PDB; 2) which were retained/removed; 3) whether the structure is apo or ligand-bound; 4) whether known structural Mg2+Mg^{2+} ions were retained. Otherwise, conformation diversity may come from destabilizing RNA by removing the native env.
  3. The enhanced-sampling validation is too weak. The main REST2-versus-MD comparison uses one system, 6qn3, and one conventional MD trajectory (with the report that conventional MD samples only 47.6% of the area by that of the REST2 in a 2-dim TICA projection). The problems are: 1) it's only one RNA; 2) only one conventional MD run; 3) “area covered in TICA space” depends strongly on the projection, binning, and definition of occupied area; 4) broader exploration does not demonstrate convergence and that the newly sampled states are physically correct.
  4. The coverage metric used 3-mer TF-IDF followed by truncated SVD + Euclidean distance, which mainly captures local nucleotide composition, not obviously a strong measure of biological/structural diversity. Similarly, the structural coverage is based on 13 hand-crafted motif descriptors.
Questions for authors
  1. Any ML elements planned to be included into the manuscript?
  2. Which atoms are included in the stored coordinates; whether the forces include solvent/ion interactions; whether solvent and ion configurations can be recovered?
  3. Why is 3-mer TF-IDF an appropriate metric for RNA sequence-space coverage?
  4. I'm confused on the dataset selection. Why the initial procedure produces 105 candidate structures then only 43 are selected? Why the others are removed? What do the "outside computational scope", "system size out of scope", "difficult to parameterize" mean? Why a lot of data is simply "not prioritized"? Any clarification?
  5. How do you deal with the Mg2+Mg^{2+}, bound metabolites, ions, other heteroatoms, etc. before simulation?
  6. For the 22 riboswitch systems, were their cognate ligands retained? If not, should these trajectories be interpreted as apo RNA simulations?
  7. How should users correctly reweight higher-state samples to the reference Hamiltonian?
3 comments
AuthorsOpenReview ↗

Response to Question 1

We thank the reviewer for the indepth critique and suggestions on our paper and want to answer the questions provided to us, as well as commenenting on the raised weaknesses. The below numbering will reflect the numbering used in the questions and reference the critical weaknesses associated which the question.

  1. Yes. In response to the reviewers' suggestions, we have added a downstream ML application in Section 5.4 and Appendix A.10 of the revised manuscript. We also thank the reviewer for the suggestions raised under Cirtical Weaknesses 1. We agree that the original manuscript would have benefited from an explicit downstream learning experiment with appropriate baselines and an assessment of generalization. We have now addressed these points in Section 5.4 and Appendix A.10 of the revised manuscript. As detailed in our response to Reviewer 1, we evaluated RNA all-atom reconstruction using three established approaches (ABC2A, RAFA and RFdiffusion3NA), supplemented by geometric and training controls. Importantly, we also fine-tuned a pretrained model on mdRNAcat and evaluated its generalization to previously unseen RNA systems, thereby directly addressing the reviewer's concerns about the absence of model training and empirical evidence of dataset utility. To ensure a meaningful evaluation, we selected holdout systems absent from both the fine-tuning data and the model's original training corpus, and fixed the evaluation protocol before examining the holdout results. Of six selected systems, four satisfied the predefined evaluation criteria, while two were reported as attrition rather than replaced. On the four evaluable systems, fine-tuning reduced the reconstruction error to 0.344 Å, compared with 0.555 Å for the original pretrained checkpoint and 0.558 Å for an EMA-matched control without additional training. Regarding train/validation/test splits, mdRNAcat is deliberately released as a general-purpose resource rather than a task-specific benchmark. Different learning objectives require different splitting strategies: for example, within-system conformational learning, generalization to unseen sequences, and transfer across structural families cannot meaningfully be evaluated using the same universal split. We therefore advocate splits appropriate to the intended application, with system-level separation for generalization experiments and explicit precautions against leakage between correlated trajectory frames. We acknowledge that our limited holdout experiment does not establish comprehensive family- or structural-level generalization. Nevertheless, the revised manuscript now provides a concrete ML application with multiple baselines, controlled fine-tuning and an independent holdout evaluation, demonstrating the practical utility of mdRNAcat without conflating a foundational dataset contribution with the development of a comprehensive RNA ML benchmark.

  2. Originally the dataset stored only the coordinates and forces of the RNA atoms in the system. In response to the Reviewers question, we now also added the ions and the solvent as well, so all can be recovered, albeit at a lower frame rate due to storage limitations. Expanding on the question with the points raised in Critical Weaknesses 1: While it is true, that two identical RNA conformations can be subjected to differenct forces with explicit solvent, we disagree with the Reviewer that it is something to avoid. Contrary, we think that it is an important physical property of all-atom explicit water simulations. Any Machine Learning task learning the forces of such a simulation should also learn these properties, no single solute conformation could give the definite forces on the solute. Nonetheless to allow deeper analysis of the forces and wider applications, we added the ion and water coordinates and forces to the replica h5 files and adjusted the script to reflect that change as mentioned above.

AuthorsOpenReview ↗

Response to Questions 3 to 6

  1. We thank the reviewer for this observation. We agree that 3-mer TF-IDF representations and the selected structural motif descriptors cannot comprehensively characterize the biological and structural diversity of RNA. However, we would like to clarify that these metrics were primarily used to guide the selection and prioritization of additional RNA systems, rather than to establish an exhaustive or optimal measure of RNA sequence and structural coverage. Our objective was to use simple, reproducible descriptors to identify underrepresented regions of sequence-composition and structural-motif space, thereby promoting diversity within the available computational budget. We do not claim that these metrics are uniquely optimal, nor that they capture all biologically relevant relationships between RNA systems. Indeed, mdRNAcat is an evolving resource that has already been expanded since the original submission. Its utility is not dependent on any particular coverage metric: the publicly available structures and trajectories allow users to independently assess the dataset's diversity and suitability for their applications using alternative sequence, structural, or functional representations. We have clarified the intended purpose and limitations of these descriptors in the revised manuscript.

  2. We agree with the Reviewer that the original manuscript did not explain the selection sufficiently. We have revised the text to better explain our reasoning. We additionally simulated six further systems and added them to the dataset. Some systems were entirely removed for either of two reasons: too large sections of the RNA not captured in PDB to accurately reconstruct and residue modification that were integral to the structure not parameterized in the used force fields of the project. While it is possible in both cases to reconstruct some similar structure, we are limited in how many systems can be simulated by the amount of available compute resources, which is why those systems were deprioritized. The same is true for larger systems which require more than 32 replicas to simulate with sufficient exchange rates. While it is possible to simulate them, we chose to focus on systems with less compute requirements to sample more different systems.

  3. We thank the reviewer for highlighting the importance of biochemical context and experimental design when interpreting the conformational ensembles of any biomolecule. We agree that ligand binding, structural divalent ions, and other cofactors can substantially affect RNA stability and conformational preferences. In particular, some riboswitches adopt different conformational states depending on the presence of their cognate ligands. Our dataset was designed to provide a homogeneous collection of RNA conformational ensembles und a standardized biochemical environment, rather than to reproduce the specific experimental conditions associated with every deposited structure. To this end, the simulations were performed without bound ligands or retained structural divalent ions, using a consistent explicit-solvent KCl solution environement. This choice ensures the maximum consistency across systems for downstream learning tasks, but necessarily limits the interpretation of the resulting ensembles. In particular, the riboswitch trajectories should be interpreted as simulations of the corresponding apo RNA systems under the specified simulation conditions, rather than a representation of their complete native ligand-bound environments. We acknowledge that removing the stabilizing ions or cofactors distort conformational preferences, likely destabilizing experimentally observed structures. However, the correct treatment of structural divalent ions in particular is not trivial and continues to be a subject of frontier research in this area. We opted to build a dataset that would facilitate models learning the intramolecular interactions and non-divalent ions solvent effects as a reasonable first scientific first objective, leaving these important effects for subsequent special treatment.

  4. Yes, they are apo RNA simulations, as explained in more detail in the answer to question 5.

AuthorsOpenReview ↗

Response to Question 7 and remaining Critical Weaknesses

  1. We thank the reviewer for this question. We would like to distinguish between statistical reweighting of conformational ensembles and the evaluation of molecular forces under the reference Hamiltonian. REST2 is designed to sample the correct equilibrium distribution at the reference thermodynamic state (λ = 1) through replica exchange. Consequently, no additional reweighting is required for equilibrium observables calculated from this state, subject to adequate convergence. Importantly, mdRNAcat already provides two sets of force labels for the sampled configurations: the forces corresponding to their respective REST2-scaled Hamiltonians, and forces independently re-evaluated under the reference Hamiltonian (λ = 1). Thus, configurations from higher thermodynamic states can be used directly with reference-Hamiltonian force labels, without any further force rescaling. We deliberately retain these additional configurations and force evaluations to facilitate applications such as force-based learning, conformational exploration, and out-of-distribution evaluation. They are not, however, automatically distributed according to the reference equilibrium ensemble. If users wish to incorporate higher-state configurations into equilibrium averages, an additional statistical reweighting procedure, such as MBAR, would be required, using the appropriate Hamiltonian energies and sufficient overlap between states. We have clarified this distinction and the interpretation of the provided force labels in the dataset documentation.

Additionally, we thank the reviewer for the observations on critical weakness 4. As discussed in our response to the fist two points of reviewer 1, REST2 is a well-established enhanced-sampling method whose methodological advantages have been extensively investigated and validated in the literature (Wang et al., 2011). While conventional MD efficiently samples local fluctuations within metastable conformational minima, it frequently fails to overcome substantial free-energy barriers, limiting exploration of the global conformational landscape. REST2 is specifically designed to alleviate these limitations. Our comparison was therefore intended to illustrate the enhanced conformational exploration accessible through mdRNAcat, not to independently revalidate REST2 or establish its computational efficiency. We agree that the occupied area in a two-dimensional TICA projection depends on the projection, binning and definition of occupied space. The reported comparison for 6qn3 should therefore be understood as an illustrative example rather than a universal measure of sampling efficiency. To provide a broader characterization beyond this single example, we have additionally included per-system RMSD plots of the reference thermodynamic state (λ = 1) for all RNA systems contained in the dataset in the Appendix. These allow readers to examine the extent of structural deviations sampled across the entire dataset, independently of the particular TICA projection and discretization used in the original comparison. Finally, we emphasize that enhanced conformational exploration, thermodynamic convergence and physical accuracy are distinct questions. Neither the TICA comparison nor the additional RMSD analyses are presented as proof of convergence or experimental validity, which require separate assessments. We have clarified these distinctions in the revised manuscript.

Reviewer_XwkPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

Authors propose a dataset of 48 MD trajectories on diverse RNA monomers in the 40-80 nt length range. They focus on long sampling with replica exchange in a consistent simulation setup. Then, they validate its soundness on RNA-relevant features, such as base pairing or sugar puckering.

Critical strengths

S1. Length and quality of the simulation, totalling 316 micro seconds.

Critical weaknesses

W1. Why is this useful for the ML community? 48 systems seems too narrow for training dynamic-aware models, maybe this could be useful as a benchmark for dynamic predictors (like BioEmu) or learnt potentials (like Allegro). But then, this definitely lacks a demonstration of these capabilities. Analyze sota models on your trajectory and that will represent a meaningful contribution.

Questions for authors

See W1. I am open to increasing my score if you use the dataset yourself and prove that its features can translate into machine learning gains.

1 comment

We thank the reviewer for these suggestions. While 54 RNA systems do not constitute exhaustive coverage of RNA sequence or structural space, such coverage is not the objective of mdRNAcat. Its primary contribution is a standardized collection of extensively sampled, force-labelled RNA conformational ensembles, rather than independent static structures. Generating these data is computationally extremely demanding: approximately 54 million CPU core-hours were allocated to the project over three years, almost entirely consumed by molecular simulations on 128-core AMD EPYC compute nodes. The resulting dataset comprises 355 μs of aggregate simulation time across 54 RNA systems. Its value therefore lies not merely in the number of distinct sequences, but in the extensive conformational sampling and corresponding atomistic force information available for each system. We would also like to clarify the reviewer's suggested applications. Allegro is an architecture for training equivariant machine-learned interatomic potentials, rather than an existing, pretrained RNA force field. To our knowledge, no publicly released RNA-specific Allegro potential is currently available. Developing and validating such a potential would constitute a substantial independent research contribution. Indeed, providing the force-labelled training data necessary to enable precisely this type of development is one of mdRNAcat's principal objectives. Similarly, the original BioEmu is protein-specific. Interestingly, BioEmu-2, has just been released in October 2026, after the ICLR submission deadline. It now extends its claimed capabilities to nucleic acids. This provides an excellent example of the potential utility of mdRNAcat: its publicly available conformational ensembles offer a resource for independently investigating how well emerging generative models such as BioEmu-2 reproduce biomolecular conformational diversity beyond the protein domain. Furthermore, as detailed in our response to Reviewer 1, we have added an explicit downstream ML experiment demonstrating that fine-tuning on mdRNAcat improves RNA reconstruction performance on previously unseen systems. We therefore agree that the suggested applications are scientifically valuable, but emphasize that enabling their development and independent evaluation is the purpose of releasing mdRNAcat, not a prerequisite that the dataset authors must themselves fulfill.

Reviewer_YMg1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This article introduces a molecular dynamics dataset named mdRNAcat, which contains 48 RNA systems. These systems were all simulated using the 32-state REST2 enhanced sampling method. The dataset is provided in a standardized HDF5 format and includes atomic coordinates, structural annotations, and other metadata. Experimental evaluations on the coverage of structural motifs and the efficiency of conformational sampling demonstrated that REST2 can explore a broader conformational space compared to traditional molecular dynamics simulations.

Critical strengths
  1. The dataset presented in this paper covers 48 RNA systems, all of which are studied through standardized simulation methods. By employing the enhanced sampling technique REST2, this research explores a broader range of potential RNA conformations and provides a set of RNA kinetic data sets; all the data in this dataset follow a unified generation process and format standard.
  2. This dataset is stored in the HDF5 format and comprehensively records various simulation conditions and processes. Additionally, this research elaborates on the simulation execution details, the exploration process of various conformations, and the structural characteristics of RNA, which helps to deeply understand the generation method of the data and facilitates other researchers to reproduce the related experiments.
Critical weaknesses
  1. The successful exchange of replicas and the overlap of energy distribution do not necessarily mean that the simulation has fully explored the conformational space of RNA, nor can they guarantee that the relative population distribution of different conformations has reached stability. It is recommended to verify the consistency of the sampling results over different time periods, report the round-trip statistics of the replicas, and estimate the number of effective independent samples to assess whether the sampling is sufficient and the results are reliable.
  2. Since REST2 uses 32 replicas while traditional molecular dynamics (MD) simulations only use a single trajectory, the same simulation duration is not equivalent to the same computational cost. The observed sampling advantage may partly be due to the investment of more computing resources. It is recommended to conduct comparisons under a fixed computational budget and include more types of RNA systems to determine whether these advantages are universal.
  3. Although analyses such as base pairing and ribose ring conformation indicate that the simulated structures are to some extent reasonable, this is not sufficient to determine to what extent these results reflect the behavior of real RNA. Through quantitative comparison with independent experimental data - while reporting the degree of agreement and the observed deviations - more powerful evidence can be provided for the reliability of the data.
  4. Although the article mentions that this dataset can be used to develop machine learning force fields and construct RNA conformation models, it has not yet been proven in specific tasks for its utility in practical applications. It is recommended to select representative learning tasks, compare the performance of the models before and after using this dataset, and verify whether it can improve prediction accuracy or enhance the generalization ability of the model. Currently, the practical value of this dataset in this aspect still needs to be further confirmed.
Questions for authors

Refer to the weaknesses

3 comments
AuthorsOpenReview ↗

Response to Critical Weakness 1 and 2

We thank the reviewer for the indepth critique and suggestions on our paper and want to address each of the points raised in turn.

  1. We thank the reviewer for raising this important point. We agree that favourable exchange acceptance rates and overlapping energy distribtuions demonstrate exchange feasibility but do not, by themselves, establish convergence of the sampled conformational ensembles We have therefore clarified the distinction between thermodynamic-state and replica trajectories in the dataset. In particular, the reference-state coordinate groups contain configurations contributed by different replicas following exchanges, rather than uninterrupted trajectories of individual replicas. The underlying replica identities and exchange records are retained, allowing users to reconstruct the exchange histories. In the revised version of the paper we added sections on convergence, effective samping size and replica round-trips. Especially for the convergence we see good results across all simulated systems with quick convergence of RMSD and Rg values. The median effective sample sizes of those two parameters are 42 and 196 respectively, which demonstrates ample sampling. While not all replicas complete round-trips within the 200 ns simulation time, we see a median of 6.5 round trips within a single REST2 simulation. We acknowledge that these diagnostics cannot provide complete convergence of every potentially slow conformational degree of freedom. Nevertheless, they provide substantially more informative characterization of the sampling quality and its limitations than exchange acceptance rates alone.

  2. We thank the reviewer for this observation. We would like to clarify that the REST2-versus-conventional-MD comparison is intended to illustrate the enhanced conformational sampling provided by mdRNAcat, rather than to benchmark the computational efficiency of REST2. Conventional MD effectively samples local conformational fluctuations within metastable minima but often fails to cross substantial free-energy barriers on accessible timescales, limiting exploration of the global conformational landscape. REST2 is a well-established enhanced-sampling method specifically designed to overcome this limitation, enabling exploration of conformational states that conventional MD may not reach within comparable computational budgets. Its methodological advantages have been extensively investigated and validated in the literature (Wang et al., 2011). mdRNAcat is an existing, substantial, and publicly available enhanced-sampling dataset, not a proposed sampling methodology. Its contribution is the provision of standardized, extensively sampled, force-labelled RNA conformational ensembles, for which, to our best no konwledge, no comparable public resources exists. Our comparison illustrates why the nominal reference-state trajectory length should not be equated with an equivalent length of conventional MD, and how enhanced sampling facilitates broader exploration of the conformational landscape. Re-establishing REST2's advantages over conventional MD is neither a claim nor an objective of this work, and would duplicate existing methodological studies rather than assess the scientific contribution of the released dataset. We have clarified this distinction in the revised manuscript.

AuthorsOpenReview ↗

Response to Critical Weankness 3

  1. We thank the reviewer for raising the important question of experimental validation. While we agree that comparison with experimental observations is valuable, we would like to emphasize that the majority of our reference structures were determined by X-ray crystallography, which provides experimentally constrained structures rather than direct measurements of solution-state conformational populations. Our simulations, by contrast, generate statistical conformational ensembles. Furthermore, to maintain a homogeneous simulation protocol suitable for downstream ML applications, we simulated isolated RNAs under standardized solution conditions. Consequently, the simulated ionic environment, concentrations, and absence of bound ligands or other cofactors may differ substantially from the experimental conditions. Such components can be essential for stabilizing particular RNA conformations (see our response to Reviewer 3, question 5). These differences complicate the interpretation of structural deviations from the experimental references. In particular, our newly added RMSD analyses characterize the extent of conformational deviations across the dataset, but cannot distinguish between effects arising from differences in biochemical conditions and those resulting from limitations of the simulations, including potential force-field inaccuracies. We therefore do not interpret these deviations alone as evidence of either successful sampling or simulation accuracy. For systems with suitable solution-state NMR measurements, comparison of ensemble-averaged experimental observables with corresponding quantities calculated from the simulations would provide a more direct assessment of conformational populations. Such an analysis would also require careful consideration of differences in experimental and simulated solution conditions. We consider this a valuable direction for further validation and inclusion in a camera-ready version of the manuscript.
AuthorsOpenReview ↗

Response to Critical Weakness 4

  1. We thank the reviewer for emphasizing the importance of demonstrating the practical utility of mdRNAcat for downstream machine-learning applications. However, we would like to highlight an important distinction between the current state of RNA and protein molecular ML. While the latter benefits from an extensive ecosystem of datasets, pretrained models, established benchmarks and downstream applications, comparable resources remain considerably less developed for RNA, particularly for conformational ensembles with atomistic force labels. Addressing this lack of foundational training data is precisely the motivation for mdRNAcat. Requiring mature downstream models and benchmarks to establish the utility of the datasets needed to develop them creates a circular dependency that risks hindering progress in an already underserved research area. Furthermore, the utility of a new scientific dataset should not be measured exclusively by improvements to existing models, but also by the previously inaccessible learning objectives and research questions it enables. We therefore believe that the provision of a substantial, standardized, publicly available and force-labelled RNA simulation dataset constitutes a valuable scientific contribution in its own right. Developing the accompanying ecosystem of models, applications and benchmarks is necessarily a broader community effort, which we explicitly seek to enable rather than claim to accomplish within a single contribution. To demonstrate one use that we could test within the revision period, we asked whether fine-tuning an existing reconstruction model on mdRNAcat conformations improves heavy-atom reconstruction on held-out RNA systems (cf in Section 5.4 and Appendix A.10 of the revision). Specifically, we investigated the reconstruction of missing nucleotide heavy atoms from a sparse representation containing only three anchor atoms per nucleotide (P, C4′ and N9/N1). Unlike reconstruction from individual static experimental structures, our task evaluates distinct conformations of the same RNA loops sampled within mdRNAcat, allowing us to investigate whether models can recover their conformational variability. We evaluated six RNA systems with two conformations each, comparing ABC2A, RAFA and RFdiffusion3NA. RFdiffusion3NA achieved a mean reconstruction error of 0.511 Å, compared with 1.009 Å for ABC2A, and outperformed both alternative methods in the evaluated comparisons. We additionally introduced a local-template baseline to distinguish learned reconstruction from the geometric information already encoded in the conditioning atoms. More importantly, we investigated whether mdRNAcat could directly improve a pretrained reconstruction model. Following fine-tuning on eight mdRNAcat systems, the model achieved a mean reconstruction error of 0.344 Å on four evaluable holdout systems, compared with 0.555 Å for the original pretrained checkpoint and 0.558 Å for an EMA-matched control without additional training. This control accounts for differences in exponential-moving-average weight handling between the pretrained and fine-tuned checkpoints. These results provide direct empirical evidence that mdRNAcat can improve downstream learning performance, including on systems excluded from both fine-tuning and the original model's training corpus. We emphasize that this is a limited demonstration of utility rather than a comprehensive benchmark, and that it does not yet evaluate the additional potential of the force labels. Ultimately, the purpose of mdRNAcat is not to establish a new state-of-the-art result on one particular learning task, but to provide the scientific community with a reusable foundation for developing and evaluating a broad range of RNA molecular learning approaches. As a matter of fact, community interest can be measured on the public version of the dataset linked in the preprint of our paper: it has seen several thousand downloads in the last month on HuggingFace.

Comments on the paper

1 comment

We would like to thank the reviewers for their time and for their constructive comments. We take this opportunity to highlight the broader scientific motivation and intended contribution of mdRNAcat.

Molecular ML has achieved remarkable advances in protein modelling, supported by an extensive and continuously evolving ecosystem of structural datasets, molecular simulations, pretrained models and established benchmarks. However, comparable infrastructure remains substantially less developed for RNA and other non-protein biomolecules. This disparity is particularly striking, given the biological importance of the largely unexplored noncoding genome (the "dark genome"). Less than 2% of the human genome encodes proteins, while the remaining noncoding landscape contains diverse regulatory elements and functional RNAs, many of whose molecular roles remain poorly understood.

We believe that expanding the capabilities of molecular ML beyond proteins is an important scientific challenge, but one that cannot be addressed by model development alone. It presents a fundamental chicken and egg problem: developing capable models requires suitable datasets and benchmarks, yet establishing the utility of these foundational resources is often expected to depend on the prior existence of capable models and established benchmarks. Applying expectations developed around the mature protein ML ecosystem to emerging areas risks perpetuating this circular dependency.

mdRNAcat is our contribution towards breaking this cycle. We provide a publicly available, standardized, force-labelled RNA molecular simulation resource comprising 54 systems and 355.2 µs of aggregate conformational sampling, generated through approximately 54 million CPU core-hours of allocated computational resources. The resource is already attracting considerable public interest, with thousands of downloads on Hugging Face within the first weeks of its release.

Our objective is not to establish an entire RNA molecular ML ecosystem with a single publication, but to provide a foundation upon which datasets, models, benchmarks and applications can co-evolve through broader community participation. We believe such enabling contributions have scientific value independently of their immediate performance improvements on any individual downstream task.

Nevertheless, we have taken the reviewers' suggestions seriously and substantially strengthened the revised manuscript. In particular, we have added a controlled downstream ML demonstration showing that fine-tuning on mdRNAcat improves RNA atom reconstruction on previously unseen systems. We have also expanded the dataset and provided additional sampling diagnostics, documentation and clarification of limitations. Additionally, we have included solvent coordinates and forces upon the reviewers' suggestion.

We therefore respectfully ask that mdRNAcat be evaluated not only by the performance of individual downstream applications, but also by its potential to enable research that is currently constrained not only by the scarcity of standardized RNA conformational and force-labelled data, but also by the computational effort involved in producing such resources. We believe that supporting the development of foundational resources for underserved molecular domains is essential to extending the success of ML in structural biology beyond proteins.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.