acceptodds
Under review as a conference paper at ICLR 2027

Scale-Recursive Generation of Long-Horizon Molecular Dynamics Trajectories

Abstract

Molecular dynamics (MD) simulations are a standard atomistic tool for studying protein folding, ligand binding, and other biomolecular processes, but their reliance on femtosecond-scale integration steps makes the simulation of millisecond-scale events prohibitively expensive. Generative surrogates offer a promising acceleration route, yet long-horizon trajectory generation poses a memory–accuracy trade-off: autoregressive models accumulate errors across successive predictions, while full-horizon temporal attention evaluates quadratically many frame pairs. Generating shorter segments reduces the jointly modeled sequence length but can affect agreement with reference trajectory distributions. We introduce ReScale, a scale-recursive factorization and sampling procedure for conditional molecular trajectories. ReScale jointly denoises separately maintained temporal states through recursive cross-scale conditioning, using a common generator architecture with level-specific parameters. This enables longer jointly generated sequences under practical memory constraints without sequential short-segment rollouts. We evaluate forward simulation on tetrapeptide and fast-folding protein systems, and transition path sampling on tetrapeptides. ReScale improves agreement with reference torsional distributions, populations along TICA coordinates, and MSM state occupations. Memory-capacity tests demonstrate 4.5× longer jointly generated sequences than MDGen at a 27.5 GiB peak-memory budget, while forward-simulation benchmarks show up to a 6× inference speedup over MDGen. Results on held-out proteins further support scale-recursive generation as a promising approach to scalable MD trajectory modeling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular dynamics, generative models, flow matching, trajectory generation, long-horizon generation, multi-scale modeling, protein dynamics, transition path sampling, AI for science

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_vuo7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents ReScale, a flow-based framework for joint trajectory generation. The claimed contribution is a coupled denoising procedure with bidirectional cross-scale conditioning for coarse-to-fine generation, enabling parallel fragment updates without sequential physical-time rollouts. The method is evaluated on the forward trajectory generation task on tetrapeptides and fast-folding proteins, and the transition path sampling task on Alanine-4, demonstrating strong distributional similarity with MD references and comparable inference speed across baselines.

Critical strengths
  • The experimental setup is fairly comprehensive, including sufficient baselines across different paradigms and molecular systems of varying sizes, ranging from small peptides to large proteins.
Critical weaknesses
  • The novelty of the paper is debatable. The main contribution of the proposed method appears to be coarse-to-fine joint trajectory generation across multiple temporal resolutions. However, to my understanding, MDGen [1] already supports trajectory upsampling, i.e., generating fine-grained frames conditioned on coarser frames. Moreover, TEMPO [2] further extends this idea to multi-resolution protein trajectory generation, which is conceptually quite similar to ReScale. However, the latter is not discussed or compared in the experiments.
  • The experimental results do not sufficiently support the superiority of the proposed framework. According to the ablation studies in Table 4, ATE, bidirectional conditioning, and RoPE provide only marginal improvements, and the observed differences may even be attributable to randomness. In contrast, only context-window conditioning and global attention appear to yield clear gains. In my view, however, these are rather essential components of a coarse-to-fine generation framework, and their effectiveness is largely expected. Under this setting, the relatively better performance reported in Tables 1–3 may be attributable more to stronger backbone models than to the proposed framework itself.

References

[1] Jing, B., Stärk, H., Jaakkola, T., & Berger, B. (2024). Generative modeling of molecular dynamics trajectories. Advances in Neural Information Processing Systems, 37, 40534-40564.

[2] Xu, Y., Wang, D., Zhou, Z., Yu, T., & Chen, M. (2026). Tempo: Temporal multi-scale autoregressive generation of protein conformational ensembles. Advances in Neural Information Processing Systems, 38, 104292-104324.

Questions for authors
  1. The notation should be defined more clearly. For example, Eqs. (4) and (6) use || and : to denote concatenation, respectively, which may cause confusion. In addition, the hyperparameters uku_k and vkv_k in Eq. (6) are not properly defined.
  2. The evaluation metrics for forward simulation primarily focus on distributional similarity, which may not be sufficient for evaluating the quality of a generated protein ensemble. In trajectory or ensemble generation, a primary requirement is that the generated conformations are physically valid, which is essential for ensuring the reliability of the predictions. Therefore, I suggest that the authors include additional metrics for conformational validity, such as the fraction of valid conformations based on protein CA geometry, as used in [1].
  3. The metrics used for transition path sampling are the same as those used for forward simulation, which may not be appropriate. Compared with distributional similarity, TPS evaluation should place greater emphasis on quantities such as the recovery of committors and transition states ([2-4]), the fraction of valid paths ([4]), transition negative log-likelihood ([4]), and etc. I suggest that the authors revise the evaluation metrics for the TPS task accordingly.
  4. The method appears to be inappropriately applied to the TPS task. The current pipeline generates a fixed-length trajectory conditioned on the two endpoint states, A and B. However, since consecutive states in the generated trajectory correspond to a fixed physical time interval (e.g., 1 ps), this formulation effectively imposes a predetermined transition time from A to B, which may not match the actual transition time of the physical process. As a result, the generated trajectories may distort the underlying kinetics, making their physical interpretation as transition pathways questionable.
  5. I could not find a clear description of the data preparation procedures for the ATLAS dataset. Could the authors provide more details on how the dataset was prepared?
  6. Although the 9/1/2 split on the fast-folding proteins does not introduce data leakage, the number of protein domains seen during training is still too limited, which reduces the practical value of the evaluation. This may also be reflected in Table 1, where the reported metrics are generally worse than those on the tetrapeptide systems. I would suggest using the fast-folding proteins only as a test set and evaluating on them with a model trained on ATLAS.
  7. The anonymous repository provided by the authors appears to be inaccessible. Could the authors please check the repository permissions or hosting configuration to ensure that reviewers can access it properly?

References

[1] Yu, Z., Huang, W., & Liu, Y. (2026). Unified biomolecular trajectory generation via pretrained variational bridge. In The Fourteenth International Conference on Learning Representations.

[2] Seong, K., Park, S., Kim, S., Kim, W. Y., & Ahn, S. (2025, May). Transition path sampling with improved off-policy training of diffusion path samplers. In International Conference on Learning Representations (Vol. 2025, pp. 93040-93062).

[3] Tang, C., Pandey, M. P., Chen, C. G., Megías, A., Dehez, F., & Chipot, C. (2026). Breaking timescales with generative sampling of conformational transitions. Nature, 1-6.

[4] Raja, S., Sipka, M., Psenka, M., Kreiman, T., Pavelka, M., & Krishnapriyan, A. S. (2025). Action-minimization meets generative modeling: Efficient transition path sampling with the Onsager–Machlup functional. In ICLR 2025 Workshop on Generative and Experimental Perspectives for Biomolecular Design.

6 comments

Q1. Relation to MDGen and TEMPO

Thank you for pointing us to TEMPO. MDGen upsampling fills frames between supplied sparse observations. TEMPO first generates a coarse backbone trajectory autoregressively, then uses those coarse predictions to anchor finer sampling. In the ReScale-L3 implementation used for our new forward experiments, each solver call builds coarse and middle features from the current noisy descriptor trajectory. The finest network uses those features to predict all frame-level velocities, and the solver advances a single descriptor-space trajectory. This distinction is about when the coarse information is computed and how it conditions the fine prediction.

Q2. What the ablations establish

The largest changes in the submitted 30-epoch study come from removing the context window or replacing global attention with window attention. The directionality and RoPE results are smaller. The ATE comparison is mixed across metrics, including a slightly lower total torsion JSD when ATE is removed.

Submitted Table 4 settingTotal torsion JSD ↓Joint TIC-0/1 JSD ↓
Full model0.1450.351
Without context-window conditioning0.2810.484
Window attention in place of global attention0.2200.481
Unidirectional conditioning0.1580.359
Without RoPE0.1560.378
Without absolute time embedding0.1420.359

These results measure the contribution of the tested components within ReScale. Figure 3 separately holds the 10×1,000-frame rollout schedule fixed and gives total torsion JSD of 0.173 for ReScale versus 0.183 for the tested attention-based MDGen. Joint TICA JSD is 0.356 versus 0.380 under original evaluation protocol. Table 8 matches parameter count between two weight-sharing designs within ReScale. These comparisons support a role for context and attention in the tested architecture, with inference capacity and accuracy measured under the specific controls above.

We also replayed released MDGen and evaluated a longer-trained local native-window checkpoint on all 100 4AA test peptides. With the same starts, 10,000 scored frames at 10 ps, and common scoring, total torsion JSD is 0.1007 for ReScale, 0.1109 for released MDGen, and 0.1084 for local MDGen e2919. Joint TICA JSD is 0.2975, 0.3255, and 0.3335, respectively. At their native schedules these metrics favor ReScale against both checkpoints.

Please refer to Q5 response for reviewer r78c for details. Released MDGen weights · Protocol-aligned ReScale e225 EMA weights

Q3. Concatenation and parent offsets

In Eq. (4), concatenation appends the absolute time embedding along the feature axis. In Eq. (6), the window selects parent features along the temporal axis, with uₖ preceding tokens and vₖ following tokens. It includes offset zero. For the evaluated 10×10×100 4AA configuration, the implemented offsets are −5 through +4, giving ten context tokens. At the second level the parent grid is the root sequence. At the third level it is the middle-token grid in the relevant outer block. Out-of-range positions are zero-padded.

Q4. Physical validity of generated conformations

We applied the physical validity checks described in our response to Reviewer r78c (Q2) to the same uncorrected outputs used for the distribution comparisons on the 100 held-out 4AA peptides. A frame passes the Cα check if all pairwise Cα distances are at least 3.0 Å and adjacent Cα distances are at most 4.19 Å. For the heavy-atom check, we scored common atoms and flagged frames with any van der Waals overlap greater than 0.60 Å, excluding atom pairs separated by at most three covalent bonds. Cα validity criterion

Geometry metricReference MDReScaleReleased MDGen
4AA, Cα-valid frames ↑100.00%99.90%99.98%
4AA, frames with heavy-atom overlap greater than 0.60 Å ↓0.011%18.61%15.54%

Cα validity is near the ceiling for both models on 4AA. The heavy-atom check flags 18.61% of ReScale frames and 15.54% of released MDGen frames, compared with 0.011% of reference MD frames.

More experimental results about temporal fidelity can be found under the response Q1 for reviewer MS83.

Q5. What the Alanine-4 metrics measure

The Alanine-4 comparison in Table 3 measures agreement between conformational populations in endpoint-conditioned samples and the MD reference. The reported metrics comprise torsion and TICA distributions and MSM occupation. The table below summarizes representative metrics from the tested configurations. The manuscript specifies a common interpolation horizon and the settings use different output resolutions.

Table 3 settingReported framesTotal torsion JSD ↓Joint TIC-0/1 JSD ↓MSM occupation error ↓Reported time
ReScale10,0000.1920.4080.22811 s
MDGen5,0000.1940.4200.24362 s
MDGen2,0000.2970.4700.22862 s
MDGen1,0000.2560.5250.36760 s

Over the same interpolation horizon, ReScale produced 10,000 frames in 11 s, compared with 5,000 frames in 62 s for the closest MDGen configuration in total torsion JSD (0.192 vs. 0.194). ReScale also achieved a lower joint TICA JSD (0.408 vs. 0.420). These results support comparable agreement with the reference conformational populations at higher output resolution and shorter reported runtime under the tested settings.

Q6. What the fixed endpoint duration means

Both endpoint conformations and the observation horizon are supplied. The horizon fixes the time of the terminal observation. A sampled path may reach the final basin before that observation or revisit either basin along the way. Consequently, the experiment characterizes the distribution of configurations in a fixed-duration bridge. We will call it fixed-duration endpoint-conditioned interpolation and scope the claims to the reported population metrics and runtime.

Q7. ATLAS preparation and evaluation

We used the MDGen ATLAS split lists with 1,266 training, 39 validation, and 82 test proteins before length filtering. All replicas of a protein belong to the same partition. Each replica is processed separately. We remove hydrogens, superpose frames, and map the remaining heavy atoms to residue-specific atom14 slots in Å. We retain every 40th 10 ps snapshot to obtain a 400 ps frame interval. ReScale trains on 200-frame windows with a 256-residue crop. ATLAS data release

Held-out evaluation choiceSetting
CohortAll 49 of the 82 official test proteins with sequence length ≤256
CheckpointsReScale e170 EMA selected by backbone torsion JSD on four validation proteins, plus released MDGen
ConditioningSequence and first frame of MD replica 1 for both models
Reference and scoring gridReplicas 2 and 3. We score the first 200 frames of each output at 400 ps spacing.
AggregationShared reference features, paired comparison across proteins

Backbone and total torsion JSDs are 0.2922 and 0.3150 for ReScale versus 0.3091 and 0.3250 for released MDGen. The corresponding paired 95% intervals for ReScale minus MDGen are [−0.0218, −0.0115] and [−0.0130, −0.0069]. Joint TICA differences are within uncertainty. Released MDGen ATLAS checkpoint

Q8. Scope of the fast-folding protein results

The submitted DESRES experiment uses 12 fast-folding proteins with a 9/1/2 train/validation/test split. It evaluates within-DESRES performance on two held-out proteins using 10,000 frames at 200 ps, a nominal 2 μs horizon. We scope the claim to those systems. The 49-protein comparison above evaluates the ATLAS-trained model on its held-out ATLAS split.

Q9. Anonymous artifact access

Thank you for flagging this. We fixed the anonymous code link in the paper and verified that it opens without signing in.

We thank the reviewer again for pointing us to TEMPO and for asking us to clarify what our experiments establish. We have explained how ReScale uses coarse features during generation, clarified the scope of the ablations, and added geometry checks and a held-out ATLAS comparison. In the revision, we will describe the Alanine-4 task as fixed-duration endpoint-conditioned interpolation and make clear that the fidelity of individual transition paths will be further explored in our future works.

I thank the authors for their detailed rebuttal and the effort devoted to addressing my concerns. However, after carefully considering their responses, I find that most of my original concerns remain insufficiently addressed.

  1. Novelty (Q1). The authors argue that the distinction between ReScale, MDGen, and TEMPO lies in "when the coarse information is computed and how it conditions the fine prediction." However, I do not consider these differences sufficiently fundamental to establish a substantial methodological contribution, as all three approaches essentially follow a coarse-to-fine paradigm in which coarse-grained representations are generated before being refined into fine-grained trajectories. This significantly weakens the claimed novelty of ReScale, particularly given the absence of any discussion of TEMPO or experimental comparisons against it in the manuscript.
  2. Ablation studies (Q2). I find the authors' response to Q2 difficult to follow, as the newly provided table largely repeats the results already reported in Table 4 of the manuscript. Consequently, my original concern remains: the ablation studies provide limited additional insight into the proposed method. Since context-window conditioning and global attention are essential components of the coarse-to-fine modeling framework, their contributions are largely expected.
  3. Physical validity (Q4). First, I consider 4AA an overly simplistic benchmark for assessing the physical validity of generated molecular trajectories. A more convincing evaluation should involve larger and more challenging systems, such as those in ATLAS or established fast-folding protein benchmarks. Second, even on 4AA, ReScale performs substantially worse than MD under the relatively basic validity criterion based on heavy-atom overlaps. This raises serious concerns about the physical realism of the generated trajectories and the practical applicability of the proposed method.
  4. Transition path sampling (Q6). I am not convinced by the authors' attempt to characterize the task as "fixed-duration endpoint-conditioned interpolation" instead of TPS. Given fixed initial and final configurations, a generative model can naturally be trained to produce trajectories connecting these endpoints within a prescribed duration. However, this formulation does not establish whether the two states are dynamically reachable within that duration, how long the transition would physically take, or whether the generated trajectories capture the actual transition-state ensemble. These are fundamental aspects of transition path sampling that cannot be addressed merely by conditioning on two endpoints. Therefore, I maintain that the task formulation and evaluation metrics are not sufficiently aligned with the physical meaning of TPS, and the rebuttal does not adequately resolve this concern.
  5. Scalability and consistency of reported results (Q7). I noticed that the ATLAS evaluation of ReScale is restricted to proteins containing no more than 256 residues. Is this restriction motivated by out-of-memory issues or other scalability limitations when applying ReScale to larger protein systems? I would therefore appreciate a direct comparison of the inference efficiency of ReScale and MDGen. Additionally, the backbone and total torsion JSD values reported in the response to Q7 appear inconsistent with those presented in Table 2 of the manuscript. Could the authors clarify the reason for these discrepancies?

Overall, I still believe that the work falls below the acceptance standard of ICLR, and I therefore maintain my original recommendation of clear rejection.

Reviewer_r78cOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a molecular dynamics emulator built on established generative modelling techniques and architectures. Its main contribution is a representation with coarse and fine levels jointly denoised through cross-scale conditioning. Through this approach, the paper authors aim to reduce autoregressive error accumulation and improve long-horizon generation.

The paper primarily builds on and compares against MDGen (Jing et al., 2024), while including comparisons with other methods in several subsections. Evaluation uses similar metrics on tetrapeptides, DESRES fast-folding proteins, and ATLAS proteins.

References

Critical strengths
  • The paper addresses autoregressive error accumulation, an important challenge in molecular dynamics emulation.
  • The problem is well motivated and relevant to the conference.
  • The paper includes an ablation study examining the contributions of individual components.
Critical weaknesses
  • Uncertain fairness of the tetrapeptide comparison. The MDGen baseline is substantially worse than originally reported (backbone torsion JSD: 0.263 versus 0.130), raising concerns that the 400-epoch budget may leave the models undertrained and favour ReScale. Longer training and evidence of convergence are needed to substantiate the claimed improvements. If comparable baseline undertraining also affects ATLAS and DESRES, the validity and fairness of those comparisons would likewise be undermined. The reviewer considers this concern sufficiently serious that, unless it is addressed by the authors, the reviewer would maintain a clear reject recommendation, but happy to reevaluate once resolved.

  • Limited support from ATLAS for the long-timescale motivation. The original ATLAS paper explicitly acknowledges that “While three replicates of 100 ns simulations offer valuable insights into dynamics properties of proteins with a relatively stable structure, exploration of rare events or major conformational rearrangements of large proteins may require longer simulations” (Vander Meersche et al., 2024). MarS-FM (Kapuśniak et al.) uses mdCATH (Mirarchi et al.) to evaluate broader structural changes and generalisation. The mdCATH dataset contains trajectories that are mostly 500 ns long, simulated at several temperatures, with greater conformational changes observed at elevated temperatures, particularly 450 K.

  • Limited assessment of generalisation on fast-folding proteins. The use of 12 fast-folding proteins partially addresses the concern above. However, having only two proteins in the test set limits the assessment of generalisation to unseen proteins. The paper authors already acknowledge the limitations associated with the small training set.

  • Potential value of mdCATH. By combining diverse protein domains with trajectories exhibiting substantial conformational changes, mdCATH offers properties relevant to both the generalisation and conformational-exploration concerns above.

  • Missing discussion and comparison with MarS-FM. MarS-FM (Kapuśniak et al., 2025) addresses related challenges involving generalisation, autoregressive error accumulation, and long-timescale dynamics, and reports improved performance on unseen proteins relative to its evaluated baselines (MDGen). A comparison on mdCATH would therefore be relevant. MarS-FM also uses an AlphaFold2/MDGen-style residue representation. Given these methodological connections and overlapping objectives, the reviewer considers discussion of this work necessary.

  • Physical validity of generated samples. The reviewer considers atom clashes and other measures of physical validity important aspects of the evaluation. PoseBusters illustrates the importance of such checks in the related setting of protein–ligand docking (Buttenschoen et al., 2024).

References

Questions for authors
  • Could the paper authors evaluate the model on mdCATH (Mirarchi et al., 2024), as motivated above (and thus compare with Mars-FM Kapuśniak et al. (2025))? The reviewer would reconsider the concerns about benchmark coverage in light of results demonstrating improved performance on this dataset.

  • Could the paper authors provide quantitative measurements of the physical validity of generated samples, including steric clashes?

  • Could fixed temporal partitions become problematic when training on longer trajectories? A single coarse representation may encode a block containing conformations from different metastable states. Could this compression make transitions between those states harder to preserve during generation? Would incorporating a Markov-state-based approach, such as that used in MarS-FM (Kapuśniak et al., 2025), help address this potential issue?

  • Did the paper authors compare against the Hyena-operator variant included in the original MDGen paper (Jing et al., 2024)? That work demonstrated generation of 100,000-frame trajectories (proof-of-concept experiment on a single pentapeptide). How would comparison with this variant affect the memory-efficiency claims and the comparison presented in the abstract?

  • Could the paper authors explain why their MDGen baseline has a backbone torsion JSD of 0.263, compared with 0.130 in the original MDGen paper? This discrepancy raises concerns about the validity and fairness of the comparison, including whether the 400-epoch cutoff favours ReScale while leaving both models undertrained. The reviewer requests justification of this training budget and longer training of both methods. (MarS-FM’s published tetrapeptide results could also be included for context directly from the MarS-FM paper once first point clarified). Could the paper authors also clarify the baseline training protocols for ATLAS and DESRES and provide evidence of convergence? If baseline undertraining also affects these datasets, the corresponding comparisons would require re-evaluation.

References

6 comments

Q1. mdCATH and MarS-FM

Thank you for proposing mdCATH as a complement to ATLAS. We compared three checkpoints on 20 held-out domains at 450 K. For each domain we used the same starting conformation and scored 200 conformations per seed across three seeds against the other four MD replicas. Both 200-epoch runs used the same 96 training domains. We fit TICA and the reference MSM on the four reference replicas at a 4 ns lag. The MSM uses 20 microstates and four PCCA+ macrostates, with 99.31% mean reference-frame coverage.

450 K mdCATH JSD ↓ReScale e200MarS-FM e200Released MarS-FM e1000
Backbone torsion0.30690.52420.2663
Side-chain torsion0.26510.45090.2537
Total torsion0.28660.48840.2603
TIC-00.52790.53610.4206
Joint TIC-0/10.73630.71410.7129
Reference-MSM occupation0.33810.45220.2155

Due to computational budget limits, currently we have re-trained results for both ReScale and MarS-FM at epoch 200. At 200 epochs, ReScale had lower JSD than MarS-FM on all three torsion metrics and MSM occupation. Every scored frame from MarS-FM e200 had at least one adjacent Cα pair more than 4.19 Å apart. Released MarS-FM, from the authors’ 4,304-domain training partition, had the lowest mean JSD on all six metrics.

Both methods use residue frames and torsion descriptors. MarS-FM learns to sample across MSM-defined states and expands samples as a tree. ReScale learns from ordered windows and generates a trajectory on a physical-time grid. The mdCATH comparison evaluates the resulting conformational ensembles on observables shared by both methods.

Under the 200 epoch training budget, ReScale showed better performance than MarS-FM. Also we are continuing to fine-tune ReScale beyond epoch 200 and will add the results once evaluation is complete.

References:

Q2. Physical validity and steric clashes

We measured physical validity on the same 100 held-out 4AA peptides per model. A frame passes the Cα check if all pairwise Cα distances are at least 3.0 Å and adjacent Cα distances are at most 4.19 Å. For the heavy-atom check, we excluded atom pairs separated by at most three covalent bonds and flagged frames with any van der Waals overlap exceeding 0.60 Å. Cα validity criterion

Cohort and metricReference MDReScale (225 epoch)Released MDGen (10k epoch)
4AA, Cα-valid frames ↑100.00%99.90%99.98%
4AA, frames with heavy-atom overlap ↓0.011%18.61%15.54%

Cα validity is near the ceiling for both models on 4AA. The heavy-atom check flags 18.61% of ReScale frames and 15.54% of released MDGen frames, compared with 0.011% of reference MD frames.

More experimental results about temporal fidelity can be found under the response Q1 for reviewer MS83.

Q3. Fixed partitions and state changes

A fixed block can contain conformations from different metastable states. In the evaluated ReScale-L3, the coarse levels compute conditioning features from contiguous blocks of the current noisy trajectory. The finest level also receives frame-wise descriptors, so a coarse token can summarize a block that contains multiple states.

We assigned frames to states using a reference-fitted MSM for the 99 4AA peptides with a usable fit. At the 1 ns grid, excluding the 10 ns grid, the probability of a different state assignment across a 100 ps interval was 28.8% at fragment boundaries and 15.4% at matched block midpoints for ReScale. The corresponding reference MD values were 9.6% and 10.3%. The excess boundary–midpoint contrast over MD was 14.1 percentage points, with a paired 95% interval of [12.4, 15.9]. This identifies a sensitivity to fixed partitions in the state-assignment diagnostic. The statistic measures differences in labels across sampled pairs. The excess in state-label changes at fixed fragment boundaries indicates a boundary effect in ReScale. Whether coarse compression causes this effect and how it affects the timing of physical transitions remain open.

We also compared uniform window sampling against enrichment for state-spanning windows on mdCATH. Both arms continued from the same fine-tuned checkpoint for 20 additional epochs. On 12 validation domains, mean ensemble JSD was 0.4135 with uniform sampling and 0.4088 with enrichment. This experiment indicates that it’s likely that MSM conditioning improves transition fidelity, and we will explore it more in future work. We thank the reviewer for this suggestion.

Q4. Hyena and the scope of the memory comparison

For a trajectory of T frames at fixed batch size, molecule size, and model width, MDGen’s Hyena mixer uses FFT-based convolution with O(TlogT) work and O(T) storage for the convolution tensors. The balanced three-level ReScale hierarchy in Section 2.8 has O(T^(4/3)) temporal attention-score entries per layer and denoising step. Doubling the horizon roughly doubles the size of Hyena’s convolution tensors and increases ReScale’s temporal score count by a factor of about 2.52 if the hierarchy is rebalanced. End-to-end peak GPU memory also depends on kernel implementations, activations, solver buffers, and FFT workspace.

Training for the direct comparison between Hyena-enabled MDGen and ReScale is underway on the same split, with matched training windows and updated budgets. We will report held-out torsion and TICA scores, training throughput, and peak training memory once the runs are complete. We will also report peak inference memory and throughput at matched physical stride and trajectory length, using the same hardware and precision.

At the 27.5 GiB peak inference memory budget reported in Figure 3, ReScale jointly generated 10,000 frames and attention-based MDGen generated 2,212. We will name the attention-based baseline explicitly in the abstract and Figure 3.

Q5. MDGen audit and longer training

We open-sourced the ReScale model weights here: https://anonymous-hf.com/a/72om7025lzz4/

Where the submitted number came from. The submitted 0.2633 backbone torsion JSD used a locally trained MDGen checkpoint with 5,000-frame windows, and the scoring aggregate mixed individual-angle and joint φ/ψ histograms. On the same archived trajectories, individual-angle scoring gives 0.2359. The fresh released-checkpoint run gives 0.1331, close to the published 0.130. The revised comparison uses the released weights and a separate longer-trained local MDGen with the native 1,000-frame windows. This replaces the original comparison based on a common epoch count. Released forward-simulation weights

What longer training changed. On four fixed validation peptides, backbone JSD for the native-window MDGen run improved from 0.1632 at epoch 399 to 0.1387 at epoch 2919. The released checkpoint scored 0.1502 on the same four peptides. For ReScale, backbone JSD on the full test split was 0.1216 at e225, and 0.1200 at e365. We use e225 EMA in the main comparison because it was selected by validation before full-test scoring. The later checkpoints show modest variation in this observable.

Full 4AA test. All three arms use the same 100 test peptides, initial structures, 10,000 scored frames at 10 ps, and reference MD. MDGen uses 10×1,000-frame rollouts and ReScale uses 1×10,000. The metrics use one scorer and one reference fit per peptide.

4AA metric ↓ReScale e225 EMAReleased MDGenLocal MDGen e2919
Backbone torsion JSD0.12160.13310.1260
Side-chain torsion JSD0.08530.09460.0953
Total torsion JSD0.10070.11090.1084
TIC-0 population JSD0.20760.24260.2550
Joint TIC-0/1 population JSD0.29750.32550.3335
Reference-MSM occupation JSD0.21010.24990.2681

All metrics favor ReScale against both MDGen checkpoints.

ATLAS and DESRES. We compared ReScale e170 EMA with released MDGen on all 49 proteins in the official ATLAS test split with at most 256 residues. Both methods start from the first frame of replica 1. We score 200 frames at 400 ps spacing against replicas 2 and 3.

Held-out ATLAS metric ↓ReScale e170 EMAReleased MDGen
Backbone torsion JSD0.29220.3091
Side-chain torsion JSD0.33840.3411
Total torsion JSD0.31500.3250
TIC-0 JSD0.62400.6353
Joint TIC-0/1 JSD0.76910.7738

All metrics above favor ReScale, and analysis about reference-MSM occupation can be found below. The submitted DESRES experiment uses a 9/1/2 protein split and the stated 400-epoch local training schedule, with 10,000 frames at 200 ps on two held-out proteins. The released-weight comparisons above address 4AA and ATLAS.

Supplementary analysis · ATLAS reference-MSM occupation

For each of the 49 proteins, we fit a reference MSM on replicas 2 and 3 using the first two backbone TICs, 20 microstates, a 4 ns lag, and a four-state PCCA+ decomposition. The largest strongly connected set covered 51.5% of reference frames on average. We counted frames mapped outside this set in a shared unassigned bin. Reference-MSM occupation JSD was 0.4102 for ReScale e170 EMA and 0.3988 for released MDGen.

We thank the reviewer again for the thoughtful feedback.

Thank you to the authors for the detailed rebuttal and additional experiments. I appreciate the effort, particularly given the limited time and computational budget.

Q1. mdCATH and MarS-FM: The additional comparison is appreciated, but I remain concerned about the 200-epoch training budget, which suffers from the same convergence issues raised in the original review. Could the authors also clarify whether a sequence-similarity-based train/test split was enforced?

More importantly, I would argue that torsion JSD, TICA populations, and MSM occupations alone are insufficient to assess the quality of generated trajectories for larger proteins, particularly at 450 K. Structural observables such as RMSD, RMSF, radius of gyration, and secondary-structure content, as evaluated in the original mdCATH dataset paper, would be valuable here.

Q2. Physical validity: The Cα validity criterion is relatively permissive.

Q3. Fixed partitions: The explanation is helpful.

Q4. Hyena: Thank you for clarifying the scope of the memory comparison.

Q5. MDGen comparison: I appreciate the authors' transparency in identifying and correcting the original evaluation issues. However, the discrepancy between the originally reported MDGen backbone JSD (0.2633) and the corrected released-checkpoint result (0.1331) is substantial. This raises concerns about the reliability of the original evaluation and, consequently, confidence in the remaining comparisons.

Overall: I appreciate the authors' effort to address these concerns. However, the original MDGen comparison was not sufficiently fair, particularly regarding the number of training epochs. The fact that these issues became apparent following reviewer scrutiny raises concerns about the reliability of the original experimental comparisons. Therefore, I will maintain my current score and leave the final judgment on the significance of these issues to the ACs and SACs.

I welcome further discussion and clarification regarding the points highlighted above. I also believe that allowing sufficient time to train both ReScale and MDGen to convergence on the full training dataset, with appropriate and consistent splits and evaluation protocols, would substantially strengthen the work.

Reviewer_MS83OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

ReScale generates molecular trajectories through a hierarchy of temporal resolutions. Each level maintains a separate noisy latent state, conditioned on neighbouring coarse states during a shared denoising process. The architecture is repeated with separate parameters across levels, and fragments within a level can be updated in parallel. The paper evaluates forward generation on tetrapeptides and proteins, and endpoint-conditioned generation on tetrapeptides. It reports improved conformational distribution metrics and inference speed, as well as longer jointly generated sequences under a fixed memory budget.

Critical strengths
  1. The temporal hierarchy addresses an important computational bottleneck, with promising empirical gains. Figure 3 shows that the method generates 10,000 frames compared with 2,212 for the evaluated MDGen baseline. Lower torsion and joint TICA population JSD under a matched 10×1,000-frame rollout schedule further suggests that the improvements are not solely due to longer trajectories. The attention-cost analysis is appropriately qualified, and the protein results indicate potential beyond small systems.

  2. The ablations are informative, and the limitations are clearly acknowledged. Tables 4 and 8 examine temporal embeddings, context direction, parameter sharing, and attention structure, including a useful parameter-matched control. The authors also report MD-replica reference scores and acknowledge remaining protein-level errors. Notably, the tradeoff between improved local torsions and worse tetrapeptide state populations at greater hierarchy depths provides a balanced assessment of the method's limitations.

Critical weaknesses
  1. The latent generative procedure is insufficiently specified, making reproducibility a challenge. The manuscript does not explain whether the descriptor projections are fixed or learned, how the decoder is trained, or what ensures reconstruction of molecular configurations. Additional clarification around the projection, decoder, and reconstruction objectives is necessary to better understand the conditions under which the proposed model performs.

  2. The empirical comparisons do not yet establish a reliable advantage over existing methods. Several baselines use modified evaluation or sampling procedures without sufficient validation. Most notably, LaM-SLidE's reported tetrapeptide torsion/joint TICA JSD increases from 0.125/0.315 in the original paper to 0.563/0.713 here, alongside a change in trajectory block length. The evaluation also omits relevant recent methods, including MarS-FM, STAR-MD, and PLaTITO. While these methods differ in their sampling objectives and representations, comparisons on shared observables would better establish ReScale's competitiveness.

  3. The evaluation does not establish temporal fidelity. The distributional metrics are invariant to frame ordering meaning that they can't discern between realistic trajectories from independently generated conformations. This is particularly important given the paper's claims about long-horizon generation and reduced errors at rollout boundaries. Temporal diagnostics such as decorrelation, transition rates, and path validity, as originally reported in MDGen, are needed to demonstrate meaningful improvements in trajectory generation.

  4. The computational advantage is not established under matched budgets. The reported inference gains omit training costs, parameter counts, and consistent timing and memory measurements across baselines. The existing parameter-matched ablation addresses sharing within ReScale but does not establish an advantage over equally resourced single-level models. Comparisons with optimized MDGen variants and matched training budgets would clarify whether the temporal hierarchy provides a genuine quality–cost improvement.

Questions for authors
  1. Information preservation: A few additional metrics would be helpful: can descriptor reconstruction error, torsion distributions, and structural validity at each hierarchy depth be evaluated on held-out trajectories? If the projections are jointly learned, it would be helpful to compare them against fixed or independently pretrained representations and monitor variance (during training). This could help clarify whether the latent representations preserves information or allow for degenerate solutions.

  2. Baseline fidelity and competitiveness: It would be helpful to validate MDGen and LaM-SLidE under their original sampling settings, then compare them under matched block lengths and computational budgets? Including MarS-FM on the tetrapeptide benchmark (granted that they operate on Markov states and not actual time), evaluated using ensemble quality rather than temporal ordering, would further strengthen the comparison. Including torsion/TICA errors, state populations, structural validity, and inference costs across repeated runs would be very helpful.

  3. Temporal fidelity: Can you evaluate whether the hierarchy improves temporal consistency using lagged autocorrelations and rollout-boundary errors? Comparisons against a sequential coarse-to-fine cascade under matched budgets would help isolate the benefit of joint hierarchical generation. For endpoint-conditioned sampling, path validity or reference-MSM likelihood under matched physical durations would provide additional evidence beyond distributional metrics.

  4. Scalability and computational efficiency: It would be helpful to contrast quality-cost tradeoffs against optimized MDGen and ConfRover baselines on shared protein benchmarks, including STAR-MD on ATLAS if available? Inference time, memory, training GPU-hours, and parameter counts alongside ensemble quality, structural validity, and temporal accuracy would be helpful to observe. Comparisons with PLaTITO on shared C_$\alpha$ observables would also help assess broader protein-sampling performance. These experiments would establish whether ReScale offers a meaningful efficiency advantage at comparable accuracy.

2 comments

Q1. Representation and reconstruction

Thank you for your careful reading. We clarify that the implementation used for these experiments integrates a single descriptor-space trajectory. As an example for K=3K=3, at every evaluation of the velocity field, all three levels receive different folds of the same noisy tensor; the first two levels then produce hidden features that condition the next level, while the finest level produces the velocity for every frame.

OperationWhat it doesHow it is fitted
Temporal foldingConcatenates the descriptors of contiguous frames along the feature axis. The original entries can be recovered by reversing the reshape.Deterministic
Projection and cross-scale conditioningEach level projects its folded input. Child levels receive projected parent features as temporal context.Learned jointly through the finest-level flow-matching objective
Frame-level outputPredicts the velocity of the 21 descriptor channels per residue, including the relative rigid transform and seven torsion sine–cosine pairs.Learned finest-level output head
Coordinate reconstructionNormalizes the output quaternion and torsion pairs, composes the residue frames with the initial reference frames, and constructs atom14 coordinates using fixed residue geometry.Fixed geometric mapping

For the 10,000-frame 4AA model, each coarse token covers 1,000 frames, each middle token covers 100, and each fine token covers one. The finest level receives the frame-wise noisy descriptors directly alongside the compressed parent features.

For training, we sample descriptor-shaped Gaussian noise and interpolate to the MD descriptors with the GVP path. The loss regresses the derivative of this path, with gradients through the full hierarchy. At inference, the solver integrates the resulting velocity field and recomputes the parent features at every call. This is the procedure implemented in the evaluated checkpoint.

Q2. Baseline fidelity and ensemble quality

ReScale weights used in this table: https://anonymous-hf.com/a/72om7025lzz4/

MDGen. The submitted backbone torsion JSD of 0.263 came from a local 5,000-frame-window checkpoint and an aggregate that included joint φ/ψ histograms. Using individual-angle histograms gives 0.236 on the archived trajectories. We replayed the released MDGen weights at their native 10×1,000-frame rollout schedule and obtained 0.1331, compared with the published 0.130. We also evaluated a separately trained local MDGen checkpoint with native 1,000-frame windows through epoch 2919.

All three arms use the same 100 test peptides, initial structures, 10,000 scored frames at 10 ps, and MD references. Both MDGen arms use 10×1,000 frames and ReScale e225 (#epoch = 225) EMA uses 1×10,000. The table uses an aligned scorer in the MDGen paper and reports Jensen–Shannon distance.

4AA metric ↓ReScale e225 EMAReleased MDGenLocal MDGen e2919
Backbone torsion JSD0.12160.13310.1260
Side-chain torsion JSD0.08530.09460.0953
Total torsion JSD0.10070.11090.1084
TIC-0 population JSD0.20760.24260.2550
Joint TIC-0/1 population JSD0.29750.32550.3335
Reference-MSM occupation JSD0.21010.24990.2681

Against both MDGen checkpoints, ReScale yielded better performances. The backbone gap is smaller against e2919. For physical validity evaluation, Cα validity is near the ceiling for ReScale and MDGen (99.90% and 99.98%).

LaM-SLidE. Our 300 archived 4AA rollouts used ten 1,000-frame blocks at 10 ps across 100 peptides and three seeds. The 250-frame setting applies to DESRES. Common individual-angle scoring brings the local total torsion JSD from 0.563 to 0.539, and the joint TICA score remains 0.713. There is no public checkpoint for us to directly evaluate on. A frozen encoder–decoder check on actual MD inputs found geometry errors at reconstruction, before the latent rollout. We report the local reproduction and published results under their respective configurations.

STAR-MD. We omitted this baseline from our evaluation because its implementation was unavailable at the time.

MarS-FM on mdCATH. On 20 held-out 450 K domains, ATLAS-initialized ReScale fine-tuned for 200 epochs had lower torsion and MSM occupation JSD than MarS-FM trained from scratch for 200 epochs on the same subset. Released MarS-FM scored lowest on all six metrics. See our response to Reviewer r78c (Q1) for the full comparison.

Q3. Temporal fidelity

We measured backbone torsion and TIC-0 autocorrelations against reference MD on the same 100 test peptides, across lags from 0.1 to 5 ns. We used both the full trajectories and matched 10 ns scoring windows for ReScale and the released MDGen baseline. Shuffling the frame order raises autocorrelation MAE to about 0.248 for backbone and 0.723 for TIC-0 in both models.

Autocorrelation MAE ↓ReScale e225 EMAReleased MDGen
Backbone, full trajectory0.07970.0905
TIC-0, full trajectory0.24350.2569
Backbone, within 10 ns windows0.08070.0917
TIC-0, within 10 ns windows0.23030.2644

Both models retain time-order information. The ReScale point estimates are lower in the full and matched-window evaluations.

Inside fragments, the mean absolute one-frame backbone torsion step is 21.5° for ReScale, 21.2° for released MDGen, and 20.7° for reference MD. Both models are within 1° of the reference mean, with MDGen slightly closer.

Q4. Scalability and computational cost

4AA inference capacity and quality. At 27.5 GiB peak inference memory in Fig. 3, ReScale jointly generates 10,000 frames and the tested attention-based MDGen jointly generates 2,212. Under the shared 10×1,000 rollout schedule, ReScale achieves lower total torsion and joint TICA JSD.

Training-step cost with SDPA. We profiled compact ReScale-L3 and single-level MDGen on the same H200. The models used the same input within each benchmark, either a 256-residue ATLAS protein at 50 ps spacing or a 4AA peptide at 10 ps spacing. Both used batch size one, bf16, Adam, gradient checkpointing, and SDPA for temporal attention. We timed ten complete optimizer steps after three warm-up steps, with data loading and validation outside the timed region. Peak allocation includes parameters, gradients, and optimizer state. MDGen has 34.2M parameters in both profiles.

Benchmark and windowReScale paramsReScale frames/sMDGen frames/sReScale peak GiBMDGen peak GiB
ATLAS, 2,000 frames38.6M94261950.7339.90
4AA, 10,000 frames48.0M24,63114,2481.942.58

At 2,000 frames on ATLAS, ReScale achieved 1.52× throughput with 27% higher peak allocation. At 10,000 frames on 4AA, it achieved 1.73× throughput with 25% lower peak allocation (0.405 vs. 0.705 s per optimizer step). MDGen was faster at the shorter 4AA windows tested (1,000–5,000 frames). These profiles characterize optimizer-step cost at fixed input shapes.

Held-out protein quality. We separately evaluated the ReScale e170 EMA and released MDGen checkpoints on all 49 proteins of the official ATLAS test split that fit the 256-residue crop. Both start from the first frame of MD replica 1. We score 200 frames at 400 ps against replicas 2 and 3.

Held-out ATLAS metric ↓ReScale e170 EMAReleased MDGen
Backbone torsion JSD0.29220.3091
Side-chain torsion JSD0.33840.3411
Total torsion JSD0.31500.3250
TIC-0 JSD0.62400.6353
Joint TIC-0/1 JSD0.76910.7738
Reference-MSM occupation JSD0.41020.3988

On these proteins, ReScale’s advantage is clearest on 5/6 of these metrics.

Supplementary MSM diagnostic. Reference-MSM occupation JSD was 0.4102 for ReScale and 0.3988 for released MDGen. The paired 95% confidence interval for ReScale minus MDGen was [−0.0195, +0.0419]. The active set covered 51.5% of reference frames on average, with the remaining frames counted in a single unassigned bin.

We thank the reviewer again for the thoughtful feedback. It helped us clarify how ReScale generates and reconstructs trajectories, and sharpen the empirical comparisons. We re-evaluated released MDGen under its original sampling protocol, added temporal diagnostics, and profiled training steps on shared inputs. A matched-budget comparison with a sequential coarse-to-fine cascade and a full end-to-end quality–cost comparison remain open. We will make the scope of the current evidence clear in the revision.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.