acceptodds
Under review as a conference paper at ICLR 2027

RNADyn: A Benchmark for Generating and Understanding RNA Dynamics

Abstract

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.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
RNA dynamics, Molecular dynamics, Diffusion models, Molecular trajectory generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PpNkOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  • The paper introduces are large RNA-MD dataset, more than 2500 100ns RNA MD trajectories, including different very useful splits based on structure and flexibility.
  • One application for such a dataset is training MD surrogates, i.e. ML models generating trajectories that supposed can stand in for the MD. The authors validate this with a bond MAE, O3P max errors, chiral flips, clashes, an RMSF correlation coefficient and some PCA-W2 metric. The later two I don't understand, and their explanations, even though crucial for the dataset being suited as a benchmark are hidden in the appendix.
  • The authors train a model in a multi-task fashion, that beneath this MD surrogate function learns two supervised tasks from single conformers: residue covariance and motion coupling between residues. Mysteriously, this is called physical grounding.
  • Without any motivation, the authors restrict the training of the model to 12h, and apply the same constraint to all baselines. No analysis of the computational performance, the degree of optimization or other influential factors is given as part of a discussion in what sense this is a fair cutoff.
  • The model is compared to three other models regarding the previously mentioned MD surrogate metrics, where the model beats alternatives.
  • The model furthermore is compared to five models on predicting MD trajectory properties from single frames. Here I assume that the references are trained in a supervised fashion. Their model is better
  • In a third evaluation, the resulting representations are used for the downstream of RNA ligand binding prediction, where also the trained model beats the baselines in several cases, in 2/8 cases even by a fair margin.
Critical strengths
  • The paper provides the largest, most diverse MD dataset for RNA that I am aware of. This in itself is an invaluable resource to the community. I have criticism (see below) but still this is clearly worth publishing.
  • The paper sketches a way to build a powerful model on the basis of the dataset by means of multi-task learning. Such a MD surrogate is challenging to build, especially as even with the dataset, data is still scarce. The proposed multi-task learning approach seems to point in a good direction, though I don't think we see a solution.
Critical weaknesses
  • I have trouble understanding why RNADynBench is a benchmark. The dataset is a great contribution to the scarce world of RNA data, but where is are the real metrics with physical meaning? The considered metrics do not study dynamics at all, but only ensemble averages. They only look at the atom-level but not on the vastly important secondary structure not even speaking of the fold. In my opinion, the metrics are what makes a benchmark a benchmark. For example, the authors could have picked individual RNA molecules, split their conformational space into two regions and check the relative occupation in a trajectory - and hence measure the free energy difference.
  • RNADynNet uses a complex multi-task learning paradigm that I found harder to grasp than I expected. The alignment loss term I did not understand at all. While it is plausible that there is a certain amount of mutual benefit for the different learning tasks, I struggle with the situation that the only presented model is so complex. Why not work with a simpler baseline first and do hyperparameter optimization. Why such an unmotivated cap in compute?

The provided dataset is a clear step forward, but here the authors rightfully point out that 100ns are too short. In RNA, single temperature runs of even 10x longer are to short for molecules to explore conformation space. Unfortunately, the highly reduces the value of the dataset for the community and even poses the risks of non-expert uncritically relying on it, not knowing that it is not equilibrated. (this is the justification for my soundness rating).

Questions for authors
  • Why did you decide for the 12h cap in compute?
  • Why did you not provide longer trajectories, at least for a subset of the molecules?
  • How did you optimize hyperparameters of your models and your baselines.
8 comments

We sincerely thank the reviewer for the detailed and insightful review. We are encouraged that the reviewer regards RNADynBench as the largest, most diverse MD dataset for RNA and an invaluable resource to the community. We address each point below.

Summary 2: RMSF r and PCA-$W_2$ are unclear and explained only in the appendix.

The revised Section 4.1 defines both metrics, and Appendix C.4 gives the full definitions and explanation. RMSF r is the Pearson correlation between the per-residue C1' fluctuation profiles of generated and reference trajectories. It measures agreement in the relative fluctuation pattern across residues. PCA-W2W_2 is the 2-Wasserstein distance between the distributions of generated and reference frames in the plane of the first two principal components of reference MD.

Summary 3: Why is supervising residue covariance and motion coupling called physical grounding?

We supervise intermediate representations with MD-derived residue covariance and inter-residue motion coupling. Both are physical observables of the MD simulations, which is why we called this supervision physical grounding (PG). Because both targets describe fluctuations and correlated motions, we now use the more specific name dynamics grounding (DG) in the revised version.

W1a: The metrics capture only ensemble averages, not dynamics or secondary structure.

Our model builds on BioKinema and follows its evaluation protocol for comparability, replacing protein-specific metrics with measures of O3'-P linkage geometry and sugar chirality. In the revised Table 2, we add metrics for secondary structure and dynamics. Base-pair occupancy error (Pair occ. MAE) compares how often each residue pair is base-paired in generated and MD trajectories. Lag-dependent RMSD measures the displacement between frames 1 to 20 ns apart, so unlike an ensemble average it depends on frame order. We report its error against MD (LAG-RMSD MAE) and its correlation with MD across RNAs (LAG-RMSD r). Backbone continuity (O3P cont.) is the percentage of generated trajectories whose backbone stays connected in every frame. At the global level, the existing PCA-W2W_2 metric compares the conformational distributions of generated and MD trajectories.

W1b: Split conformational space into two regions and measure the relative occupation and free-energy difference.

We thank the reviewer for this suggestion and have run the analysis (new Appendix D.4). We use the fraction of formed base pairs in a stem, QstemQ_{stem}, to divide conformational space into an open region (Qstem<0.5Q_{stem}<0.5) and a closed region (Qstem≥0.5Q_{stem}\geq0.5), with occupancies popenp_{open} and pclosedp_{closed}. Occupancies are meaningful for stems that visit both regions repeatedly, so we analyze the 38 stems, one per target, that open and close at least twice in the reference MD. The apparent free-energy difference is ΔG/kBT=−ln⁡(pclosed/popen)\Delta G/k_{\mathrm B}T=-\ln(p_{closed}/p_{open}). Compared with BioKinema, RNADynNet has lower apparent free-energy error, transition counts closer to MD, and higher region-occupancy error.

Table A. Two-region analysis of stem opening and closing.

MethodRegion-occupancy MAEApparent ΔG\Delta G MAE (kT)Transitions per 100 ns
Reference MD--8.82
BioKinema0.2612.1292.89
RNADynNet0.3001.6306.49
W2: The multi-task design is complex and the alignment loss is unclear. Why not start from a simpler baseline?

The revised Section 3 now thoroughly describes the architecture before the losses. In summary, our architecture is a single backbone trained with three objectives. With the generation objective, we train it to denoise MD trajectory windows. With alignment, we match the features the backbone computes from a single frame to those it computes from the trajectory window containing that frame. This is self-distillation: the trajectory pass acts as teacher and the single-frame pass as student, with no separate teacher network, and it transfers trajectory information to the single-frame features. With dynamics grounding (DG), we supervise these single-frame features with MD-derived residue covariance and motion coupling. Alignment and DG add three prediction heads and one truncated single-frame forward pass per training step. At inference, trajectory generation and descriptor prediction each take a single conformer as input.

We verify each component by ablation. Adding alignment to generation-only training raises the frozen-probe RMSF r on test_flex from 0.62 to 0.69, and adding DG raises it to 0.78 and improves all seven descriptor metrics on both test sets (Table 3). Adding DG also raises trajectory-level backbone continuity on both test sets (Table 2). As the simpler baseline the reviewer asks for, we trained the same backbone with descriptor supervision alone (DG-only, Table 3): RNADynNet obtains better performance on five of seven metrics on each test set.

W3+Q2: 100-ns trajectories are too short and not equilibrated. Why not provide longer trajectories for a subset of the molecules?

With RNADynBench, we target near-native, short-timescale dynamics around experimental structures, the regime in which trajectory surrogates and single-conformer dynamics predictors are trained and evaluated. These trajectories sample local fluctuations, correlated motions, and base-pair opening around each experimental structure, though not the full conformational distribution. We prioritized structural diversity: the release covers 1,469 RNAs (258.5 μ\mus), more than existing RNA trajectory resources (Table 1). Tenfold longer runs for even 10% of the systems would roughly double its cost. RNA conformational exchange can occur on microsecond-to-millisecond timescales [1], and converged sampling remains challenging even for small RNA motifs [2]. Equilibrium sampling of RNA is therefore an open research problem that a tenfold longer run would not solve, and we do not consider it a criterion that a dataset of 1,469 RNAs can currently meet.

Within this scope, we ran a block-averaging analysis [3] on all 2,585 trajectories (new Appendix A.4): 94% of the trajectories with canonical base pairs contain at least 20 uncorrelated samples of base-pair retention (Table B). Base pairing is therefore well sampled within 100 ns. C1' fluctuations, which underlie our RMSF metric, pass the same test in 65% of trajectories, and the radius of gyration in 46%.

Longer trajectories (Q2). We extended trajectories to 500 ns for three systems (two with low and one with high flexibility). We compare each later 100 ns window with 0-100 ns and use the two halves of 0-100 ns as a control (Table C). Every later frame lies within 2.4 Å C1' RMSD of a frame from 0-100 ns. On these three systems, replacing the 100-ns reference with the 500-ns reference changes the RMSF r of BioKinema, MDGen, and RNADynNet by at most 0.04 and leaves their ranking unchanged.

We state this scope in the revised Limitations paragraph and in the dataset card. We will release the extended trajectories and per-trajectory sampling diagnostics with the dataset.

Table B (Table 7). Block-averaging analysis after discarding the first 30 ns.

ObservableTrajectories with ≥20\geq 20 uncorrelated samples
Base-pair retention (trajectories with canonical pairs)93.9% (2055/2188)
C1' mean-square fluctuation64.7% (1673/2585)
RMSD to trajectory average60.9% (1574/2585)
Rg46.3% (1198/2585)

Table C (Table 8). 500-ns extensions (ranges over the four later windows).

SystemRMSF r, later windows vs. 0-100 nsRMSF r, first vs. second halfBase-pair occupancy MAE, later windowsOccupancy MAE, first vs. second halfMax. C1' RMSD to nearest 0-100 ns frame
1F270.67-0.720.730.03-0.040.061.6
1L3D0.83-0.970.970.02-0.070.071.8
402D0.93-0.970.940.02-0.070.022.4

[1] Invisible RNA state dynamically couples distant motifs. PNAS, 2014.

[2] Toward Convergence in Folding Simulations of RNA Tetraloops. JCTC, 2022.

[3] Quantifying uncertainty and sampling quality in biomolecular simulations. Annual Reports in Computational Chemistry, 2009.

Q1:Why did you decide for the 12h cap in compute?

We would like to clarify that 12 hours was an approximate training budget and not a cap on every method. On two H200 GPUs, RNADynNet trains in about 9 hours and BioKinema in about 12.6 hours. We only cut off ConfRover at 12 hours because its official schedule takes over 120 hours. We have since trained it for 48 hours, selected its checkpoint by validation loss, and report this model in the revised Table 2. Compared with the 12h model, some metrics improve while others degrade, and it remains worse than RNADynNet on most metrics.

Q3: How did you optimize hyperparameters of your models and your baselines.

We use one policy for all methods: published settings where a single training run takes 9 hours or more (see Q1), and a grid search where training is cheap. We therefore retrain the baselines on RNADynBench with their released hyperparameters and take most base settings of RNADynNet from BioKinema. We set the maximum alignment weight (20) by gradient-ratio calibration, so that its gradient contribution is comparable to that of the generation objective, and we set the covariance and motion-coupling weights to 1, following BackFlip. We use these settings in all ablations. For the frozen probes and simple descriptor baselines in Table 3, we search the same grid of six learning-rate and weight-decay settings.

We thank the reviewer again for the time and detailed feedback. We believe that the new metrics and analyses, the 500ns extensions, and the clarified method description address the reviewer's concerns, and we kindly ask the reviewer to reconsider the assessment in light of the revised paper. We are fully available for discussion should any concerns remain.

Thank you for the extensive replies. I will consider your points thoroughly.

Reviewer_usdiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces RNADynBench, comprising 2,585 100-ns all-atom MD trajectories from 1,469 RNAs, with structure- and flexibility-oriented test splits. It also proposes RNADynNet, combining trajectory generation, single-conformer representation alignment, and supervision from MD-derived covariance and motion-coupling statistics. Experiments evaluate trajectory generation, dynamics-descriptor prediction, and transfer to RNA–small-molecule binding prediction.

Critical strengths
  1. The dataset substantially expands the available RNA MD resources. The standardized simulation and curation pipeline, together with grouped splitting and attention to sequence/structure overlap, could provide a useful foundation for future research.

  2. The experiments demonstrate improvements in geometric quality and MD-descriptor prediction, providing a useful baseline for the proposed resource. Evaluating both generation and single-conformer prediction also broadens the potential applications of the dataset.

Critical weaknesses

The main generation evaluation does not establish temporal fidelity. The geometry metrics, RMSF correlation, and PCA-W₂ in Table 2 remain unchanged when future frames are arbitrarily permuted. They therefore cannot distinguish an accurate trajectory from an unordered ensemble with similar conformational statistics. Although the model includes temporal attention and an autocorrelation training loss, held-out evaluation of temporal behavior is needed to support the trajectory-generation claim.

The existing ablations retain the generation objective while varying alignment and physical grounding. However, they do not compare against the same pretrained backbone trained solely to predict MD descriptors. Since physical grounding directly supervises the descriptor family subsequently evaluated, the probe improvements do not establish the incremental value of the unified formulation over a simpler supervised predictor.

Questions for authors

Can the authors report held-out lag-dependent displacement and autocorrelation, including time-shuffled controls?

What proportion of generated frames and complete trajectories satisfies backbone continuity and other chemical-validity criteria?

How does RNADynNet compare with the same Protenix backbone trained only for single-conformer descriptors, and with generation plus PG without alignment?

Which coordinates are supplied in each task: raw experimental coordinates, prepared/minimized coordinates, or the equilibrated production-time-zero conformer?

Can you provide auditable split definitions, nearest-training similarities, representative target-stability analyses, and paired RNA-group confidence intervals?

For Table 4, can you compare static, ungrounded, and grounded fingerprints under identical downstream training and clarify structure sources and checkpoint selection?

5 comments

We appreciate the reviewer's careful assessment, in particular the observation that the dataset substantially expands the available RNA MD resources and could provide a useful foundation for future research. We address each of their comments below.

W1 + Q1: The generation metrics are invariant to frame order. Can the authors report held-out lag-dependent displacement and autocorrelation with time-shuffled controls?

We sincerely appreciate this constructive suggestion, which strengthens the benchmark. The original metrics followed BioKinema's evaluation protocol and do not depend on frame order. In the revised manuscript, we add lag-dependent RMSD (LAG-RMSD MAE and LAG-RMSD r) and autocorrelation (ACF MAE), and we pair the two errors with time-shuffled controls built from the same generated frames (Table A; definitions in Appendix C.4). For RNADynNet, shuffling the frames increases both errors on both test sets, so the order of its frames carries temporal information.

Table A (Table 13). Temporal metrics with time-shuffled controls.

SplitMetricConfRoverBioKinemaMDGenRNADynNet
test_structLAG-RMSD MAE ↓\downarrow28.12061.21150.83401.2572
test_structLAG-RMSD r ↑\uparrow0.88870.94530.09650.9441
test_structShuffled LAG-RMSD MAE29.25071.68291.10901.7317
test_structACF MAE ↓\downarrow0.24920.13920.09820.1382
test_structShuffled ACF MAE0.30940.30780.30660.3073
test_flexLAG-RMSD MAE ↓\downarrow56.90174.40743.80663.4863
test_flexLAG-RMSD r ↑\uparrow−0.0443-0.04430.90980.69320.9368
test_flexShuffled LAG-RMSD MAE57.31805.94443.26434.7319
test_flexACF MAE ↓\downarrow0.42600.25310.15590.2679
test_flexShuffled ACF MAE0.45590.45600.45540.4556
Q2: What proportion of generated frames and complete trajectories satisfies backbone continuity and other chemical-validity criteria?

With RNADynNet, more than 99% of generated frames pass the O3'-P continuity check, and its trajectory-level pass rates exceed those of BioKinema for both checks on both test sets (Table B). For example, 72.2% versus 22.0% of complete trajectories pass the continuity check on test_flex. A frame passes the continuity check if every within-chain O3'-P distance between consecutive residues is within 2.5 Å, well above the ideal O3'-P bond length of about 1.6 Å. It passes the chirality check if no sugar stereocenter flips relative to the initial conformer. The main results (revised Table 2) now also report trajectory-level continuity, alongside bond-length error and clash rate.

Table B (Table 12). Backbone-continuity and sugar-chirality pass rates.

SplitCheckReference MDBioKinemaRNADynNet
test_structFrames with O3'-P ≤2.5\leq 2.5 Å (%)100.0099.0299.90
test_structTrajectories with O3'-P ≤2.5\leq 2.5 Å throughout (%)100.0050.4390.81
test_structFrames without chirality flips (%)100.0094.1996.70
test_structTrajectories without chirality flips (%)100.0022.6189.50
test_flexFrames with O3'-P ≤2.5\leq 2.5 Å (%)100.0097.5099.05
test_flexTrajectories with O3'-P ≤2.5\leq 2.5 Å throughout (%)100.0022.0072.21
test_flexFrames without chirality flips (%)100.0094.0199.85
test_flexTrajectories without chirality flips (%)100.005.6087.94
W2+Q3: Compare with the same backbone trained only on descriptors, and with generation plus PG (now DG) without alignment.

Descriptor-only control. We trained the same backbone with descriptor supervision alone (DG-only) and evaluated it under the same protocol (Table C). RNADynNet obtains the better value on five of seven metrics on each test set, including all RMSF and motion-coupling metrics, while DG-only cannot generate trajectories.

Generation plus DG without alignment. We also trained this variant and evaluated it with frozen probes (Table C). RNADynNet is better on 9 of 14 metrics, including all five RMSF and covariance metrics on test_struct.

Table C (from Table 3). Descriptor-only and no-alignment controls. † denotes probes on frozen representations.

SplitMetricDG-onlyRNADynNetw/o align.†RNADynNet†
test_structRMSF r ↑\uparrow0.85490.86650.85070.8626
test_structRMSF MAE ↓\downarrow0.40450.39830.44210.4076
test_structCov. RMWD ↓\downarrow0.86380.86510.91890.8741
test_structCov. Sym. KL ↓\downarrow0.57020.57430.62130.5823
test_structCov. IoU ↑\uparrow0.52130.52600.50410.5276
test_structNMC r ↑\uparrow0.79710.81150.79920.7958
test_structNMC MAE ↓\downarrow0.16800.16290.16840.1694
test_flexRMSF r ↑\uparrow0.76700.78270.76690.7782
test_flexRMSF MAE ↓\downarrow1.60211.58001.54461.7560
test_flexCov. RMWD ↓\downarrow3.26603.27003.27553.3359
test_flexCov. Sym. KL ↓\downarrow0.83670.85330.87310.8775
test_flexCov. IoU ↑\uparrow0.45260.45280.44600.4522
test_flexNMC r ↑\uparrow0.73640.76140.75550.7577
test_flexNMC MAE ↓\downarrow0.20860.20270.20450.2035
Q4: Which coordinates are supplied in each task?

For trajectory generation and single-conformer dynamics prediction, we supply the production-time-zero conformer. For RNA-ligand binding prediction, we use the RhoFold-predicted RNA structures provided with GerNA-Bind.

Q5: Provide split definitions, nearest-training similarities, target-stability analyses, and paired RNA-group confidence intervals.

Data split. We split at the RNA-group level (Appendix A.2). test_struct contains 117 trajectories from 113 gRNAde test entries. To prevent leakage, we exclude 18 further trajectories that are too similar to a test_struct target from training, model selection, and evaluation. These have sequence identity and coverage of both sequences ≥80%\geq 80\%, a TM-score ≥0.45\geq 0.45 (normalized by the longer RNA), or the same sequence or structure cluster as a test target. We select test_flex by flexibility from the remaining groups.

Nearest-training similarity. Across all 217 test trajectories, the median nearest-training TM-score is 0.30 and the median sequence identity is 0.63. We will release the split assignment and the nearest-training similarities of every trajectory with the dataset.

Target stability and confidence intervals. In the MD reference row of Table 3, we compare descriptors computed from two non-overlapping 50-ns halves of each trajectory, and in Appendix A.4 we add per-trajectory block-averaging diagnostics. For RNADynNet, the ordered-minus-shuffled ACF MAE is -0.169 on test_struct (95% paired RNA-group bootstrap CI -0.184 to -0.152) and -0.188 on test_flex (-0.199 to -0.176).

Q6: Compare static, ungrounded, and grounded fingerprints under identical downstream training, and clarify structure sources and checkpoint selection.

We compare static (from the pretrained Protenix backbone), ungrounded (trained without DG), and grounded (RNADynNet) fingerprints using the same downstream model and training protocol (Table D). Grounded fingerprints achieve the highest AUROC in five of eight settings. All three use the RNA structures released with GerNA-Bind, and we select downstream checkpoints by validation AUROC.

Table D. Binding AUROC of static, ungrounded, and grounded fingerprints.

FingerprintBiosensor RandomRNA homologyLigand fingerprintHomol. & fingerpr.Robin RandomRNA homologyLigand fingerprintHomol. & fingerpr.
Static0.95920.90770.69510.75380.70120.68100.67700.6217
Ungrounded0.95790.91770.68230.75930.70920.65670.67600.6421
Grounded (RNADynNet)0.95820.90660.70410.76730.71190.68990.67600.6619

Once again, we sincerely thank the reviewer for their thoughtful engagement. We hope the added temporal metrics, chemical-validity checks, and controls address the concerns. We also hope that these efforts provide sufficient reason to raise the score.

Dear Reviewer usdi,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

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

DNADynBench proposed an RNA molecular dynamics (MD) benchmark with 2500+ all-atom trajectories, with leakage controlled splits, which is significantly larger than existing benchmarks, by first extracting RNA structures from RNA-Solo, then using MD simulation. They use quality control to remove artifacts and other issues. To construct leakage controlled splits they RNA identity, and various measures of structural or other similarity. Building on this dataset, they propose a model for RNA trajectory generation and dynamics representation learning from an initial conformer. The model is trained using a multi-objective losses for trajectory generation, single-frame-to-trajectory alignment, residue covariance, and inter-residue motion coupling. For trajectory learning, heterogeneous noise is added along with a coordinate denoising loss. The results show better performance compared to adaptations of protein molecular dynamics to the RNA scenario.

Critical strengths
  • Dataset seems to be well constructed and is of larger size than existing RNA MD benchmarks, which provides a useful contribution.
  • Model seems well designed and shows very good results on a comprehensive range of tasks, metrics, and compared to a good number of baselines (particularly considering that there are few baselines developed specifically for RNA and the current baselines are adapted from protein based work).
Critical weaknesses
  • Some details of RNADynNet are mentioned in Figure 2 but the explanation of them in the main paper seems a bit brief making it hard to understand, particularly for readers less familiar with the area. For example, I am not sure if the residue covariance, and inter-residue motion coupling losses have been explained, as well as whether L_gen in equation (2) is the same as L_denoise in equation (3).
Questions for authors

See weaknesses above.

3 comments

We sincerely thank the reviewer for the valuable comments and for finding the dataset well constructed and the model well designed, with very good results on a comprehensive range of tasks. We address the reviewer's concerns below.

W1a: Some details of RNADynNet are mentioned in Figure 2 but the explanation of them in the main paper seems a bit brief making it hard to understand, particularly for readers less familiar with the area. For example, I am not sure if the residue covariance, and inter-residue motion coupling losses have been explained.

We thank the reviewer for pointing this out. The revised Section 3 now opens with an Architecture paragraph that explains each component of Figure 2: the encoder and decoder, the trajectory and single-frame passes, and where alignment and dynamics grounding act.

The two targets describe how each residue fluctuates and whether residues move together. With reference-MD C1' coordinates aligned to the initial conformer, let δri\delta\mathbf r_i be the displacement of residue ii from its time-averaged position. The residue covariance Σi=⟨δriδri⊤⟩\boldsymbol\Sigma_i=\langle\delta\mathbf r_i\delta\mathbf r_i^\top\rangle is a 3×33\times3 matrix that generalizes the scalar RMSF to directional fluctuations: its eigenvectors and eigenvalues give the directions and amplitudes of the motion, and the square root of its trace is the RMSF. Its loss is a squared error on the matrix logarithm in a local residue frame. Inter-residue motion coupling is Cij=⟨δri⊤δrj⟩C_{ij}=\langle\delta\mathbf r_i^\top\delta\mathbf r_j\rangle, which is positive when residues ii and jj move in the same direction and negative when they move in opposite directions. Its normalized form is ρij=Cij/CiiCjj∈[−1,1]\rho_{ij}=C_{ij}/\sqrt{C_{ii}C_{jj}}\in[-1,1]. Its loss matches the overall pattern of the matrix (cosine similarity), its magnitude, and the normalized coupling between different residues. These MD statistics are training labels, not model inputs. The revised Section 3.2 now gives these interpretations together with the definitions of both targets and their losses, and Appendix B.4 describes the prediction heads.

W1b: Is L_gen in Eq. (2) the same as L_denoise in Eq. (3)?

Lgen\mathcal L_{\mathrm{gen}} and Ldenoise\mathcal L_{\mathrm{denoise}} are not the same. The coordinate-denoising loss Ldenoise\mathcal L_{\mathrm{denoise}} is one of several terms in the generation objective Lgen\mathcal L_{\mathrm{gen}}: Lgen=Ldenoise+λsLDLsmooth-lDDT+λRMSFLRMSF+λrelLrel-RMSF+λlocalLlocal-RMSF+λACFLACF.\mathcal L_{\mathrm{gen}}=\mathcal L_{\mathrm{denoise}}+\lambda_{\mathrm{sLD}}\mathcal L_{\mathrm{smooth\text{-}lDDT}}+\lambda_{\mathrm{RMSF}}\mathcal L_{\mathrm{RMSF}}+\lambda_{\mathrm{rel}}\mathcal L_{\mathrm{rel\text{-}RMSF}}+\lambda_{\mathrm{local}}\mathcal L_{\mathrm{local\text{-}RMSF}}+\lambda_{\mathrm{ACF}}\mathcal L_{\mathrm{ACF}}. We take this objective from BioKinema, which extends the denoising and smooth-lDDT losses of AlphaFold 3 with fluctuation (RMSF) and short-lag autocorrelation terms. Alignment and DG are separate terms in the total objective. We state this relation in Section 3.1 of the revised paper and give the individual terms in Appendix B.3.

We thank the reviewer again for the time and feedback. We hope that these clarifications address the reviewer's concerns. We are fully available for discussion should any concerns remain.

Dear Reviewer zpLC,

Thank you again for your time and thoughtful review. As the discussion closes tomorrow, November 18 (AoE), we would appreciate hearing whether our rebuttal addresses your concerns. We would be happy to clarify any remaining points before the deadline.

Best, The authors

I thank the authors for the helpful reply, which has clarified my issues. I maintain my positive score.

Comments on the paper

1 comment

We thank all reviewers for their time and valuable comments. We are encouraged that the reviewers regard RNADynBench as the largest and most diverse MD dataset for RNA and an invaluable resource to the community (PpNk), find that it substantially expands the available RNA MD resources (usdi), and consider the dataset well constructed and the model well designed, with very good results on a comprehensive range of tasks (zpLC). At the same time, the reviewers asked for new evaluation metrics, evidence on how well the 100-ns trajectories are sampled, additional ablations, and a clearer presentation of the method. In response, we have run new experiments and revised the paper as follows:

  • New temporal, secondary-structure, and chemical-validity metrics (PpNk, usdi). The revised Table 2 adds lag-dependent RMSD (LAG-RMSD MAE and LAG-RMSD r), base-pair occupancy error, and trajectory-level backbone continuity. (i) Temporal behavior: Table 13 pairs lag-dependent RMSD and the autocorrelation error with time-shuffled controls. Shuffling RNADynNet's frames increases both errors on both test sets, showing that their order carries temporal information. (ii) Secondary structure: following the suggestion of PpNk, Appendix D.4 tests whether models reproduce stem dynamics on 38 stems that open and close at least twice. RNADynNet has a lower apparent free-energy error and transition counts closer to MD than BioKinema. (iii) Chemical validity: RNADynNet's backbones remain chemically valid (Table 12): more than 99% of its frames pass the continuity check, and its trajectory-level pass rates exceed those of BioKinema for both checks.

  • MD sampling diagnostics and 500-ns extensions (PpNk). Equilibrium sampling of RNA remains an open problem that tenfold longer runs would not solve, so RNADynBench targets near-native dynamics across 1,469 RNAs. Within this scope, we ran a block-averaging analysis on all 2,585 trajectories (new Appendix A.4, Table 7): 94% of the trajectories with canonical base pairs contain at least 20 uncorrelated samples of base-pair retention. The base-pair statistics, which serve as references for our secondary-structure metric, are therefore reliable within 100 ns. We also extended three systems to 500 ns (new Table 8). Later 100-ns windows agree with the first 100 ns in RMSF and base-pair occupancy as closely as the two halves of the first 100 ns agree with each other, and the longer reference leaves the ranking of the models by RMSF r unchanged.

  • Additional ablations (PpNk, usdi). We trained the same backbone with dynamics grounding (DG) alone, which we call DG-only, and a variant with generation and DG but without alignment (revised Table 3). RNADynNet is better than DG-only on 10 of 14 metrics, including all RMSF and motion-coupling metrics, while DG-only cannot generate trajectories. Alignment improves all five RMSF and covariance metrics on test_struct.

  • Clarified method presentation (PpNk, zpLC). We reorganized Section 3 to introduce the architecture before the objectives and clarified the dynamics-grounding targets. Section 4.1 now defines RMSF r and PCA-W2W_2, and we renamed physical grounding to dynamics grounding.

In addition to the point-by-point responses below, we used the additional page allowed in the revised manuscript to incorporate these changes. All significant changes are marked in blue in both the main paper and the Appendix for ease of reference.

We thank the reviewers again for their time and insightful feedback. We believe that these additions and revisions address the concerns raised in the reviews, and we kindly ask the reviewers to reconsider their assessments in light of this revised version. We are fully available for discussion should any concerns remain.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.