acceptodds
Under review as a conference paper at ICLR 2027

DuET-MD: Training-Free Rare-Event Steering for Frozen Molecular Dynamics Surrogates

Abstract

Protein dynamics is central to understanding biological function, while atomistic molecular dynamics simulation, the gold-standard tool for characterizing protein dynamics, is too costly to observe rare transitions across high free-energy barriers. Recent MD surrogates accelerate trajectory generation by learning ps-to-ns finite-lag transitions, but faster propagation alone does not solve rare-event sampling: their default rollouts are concentrated around high-probability transitions learned from reference trajectories, leaving rare-event sampling unresolved. Applying conventional enhanced-sampling biases to frozen generative transition models is also non-trivial. We introduce DuET-MD, a training-free inference-time framework for steering frozen iterative MD surrogates. DuET-MD couples two particle processes across distinct clocks. Within each next-frame generation, inner particles use guidance and intermediate resampling to discover promising transitions before frame completion. Across physical time, an outer particle process uses the selected frame together with an estimated local normalizer to properly weight and resample competing trajectory prefixes. This enables path conditioning without retraining the surrogate. Across protein-folding and cryptic-pocket-opening tasks, DuET-MD increases the recovery of rare-transition candidates over default surrogate rollouts under matched finite sampling budgets. We further evaluate geometric validity and mechanistic ordering not encoded in the steering objective. DuET-MD thus extends accelerated MD surrogates from fast rollout to training-free rare-transition sampling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
molecular dynamics, protein dynamics, rare event sampling, inference-time steering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4UouOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3(+1)
Soundness
2
Read review
Paper summary

This paper presents DuET-MD, a training-free steering method for rare event sampling using frozen, pretrained generative MD surrogates. DuET-MD adopts a dual-clock particle inference scheme, which first generates candidates from a guided proposal using biasing potentials during the reverse diffusion process, and then resamples the predicted structures in the outer loop. This formulation enables sampling from low-density regions of the original proposal distribution with theoretical guarantees, while also preserving the dynamical behavior of the pretrained MD surrogate. Experiments on protein folding and cryptic pocket opening tasks show consistent improvement over the complete-frame Nested-SMC control method in rare-event sampling.

Critical strengths
  • Rare-event sampling is a critical challenge in generative molecular dynamics and cannot, in principle, be fully addressed by models that learn only finite-lag transition dynamics. This work presents a training-free method to address this problem and achieves consistent improvements across experiments, making it a valuable contribution to the field.
Critical weaknesses
  • The manuscript, particularly the Method section, could be substantially improved in terms of writing and clarity, as the current presentation is difficult to follow. For example, (i) The abstract and introduction both refer to the term "clocks", which is neither standard terminology in the field nor clearly defined before Eq. (2). This makes the presentation difficult to understand. (ii) The symbol YY in Eq. (2) is not explicitly defined, although it can be inferred from the context that it denotes the random variable associated with the reverse diffusion process. (iii) The term "checkpoint" first appears in Line 216 but is never explicitly defined. Since this concept is important for understanding the discussion in this paragraph, the authors should provide a clear definition when it is first introduced. (iv) Line 217 uses YiY_i, whereas Eq. (2) uses Yt,lY_{t,l}. My understanding is that the subscript ii denotes one of the internal particles, while t,lt,l denote the physical time and the diffusion time, respectively. Mixing these different subscripts without clarification can easily cause confusion, so the notation should be defined and used consistently.
Questions for authors
  1. My main concern is the issue of readability and clarity of presentation mentioned in the Weakness section. I believe the authors should carefully revise the manuscript, after which I would be willing to reconsider and potentially raise my score.
  2. Sections 4.2 and 4.3 are central to the proposed method, yet it is still difficult to accurately understand the overall inference procedure from the textual description alone. Given that the main text has not yet reached the page limit, I would suggest adding a workflow diagram or pseudocode in the Method section to help readers better understand the working pipeline.
  3. Line 701 is the first place where the guidance strength η\eta is introduced. Based on Section 4.2, it looks like the guided proposal is constructed using the gradient of the biasing potential during generation. However, I could not find a clear description elsewhere in the manuscript of how η\eta specifically enters the inference procedure. I believe the authors should provide a more explicit explanation and revise the manuscript to improve the clarity of the presentation.
  4. I noticed that Enhanced Diffusion Sampling [1] also performs rare-event sampling by applying a Feynman-Kac corrector to the diffusion model. Could the authors provide a comparison between this prior work and the proposed method in terms of methodology and experimental performance?
  5. In Table 2, for the PRMT6 case, the improvement in the fpocket DS metric is limited and even appears to be worse than that of the frozen model. Could the authors provide a case-specific analysis to explain why this occurs?

References

[1] Xie, Y., Winkler, L., Sun, L., Lewis, S., Foster, A., Jimenez-Luna, J., Hempel, T., Gastegger, M., Chen, Y., Zaporozhets, I., Clementi, C., Bishop, C. M., & Noé, F. (2026). Enhanced diffusion sampling: Efficient rare event sampling and free energy calculation with diffusion models. In ICML 2026 Workshop on Structured Probabilistic Inference & Generative Modeling.

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

This paper proposes Duet-MD, a training-free inference-time steering framework for frozen MD surrogate models. It consists of two processes, where the inner process uses gradient guidance and intermediate resampling, while the outer loop resamples the trajectory. The method is evaluated on two frozen MD surrogates with TICA-based bias potential steering, based on native contact recovery and $ C_\alpha$-based validity.

Critical strengths

1. Training-free rare-transition path steering of frozen MD surrogates

The paper addresses the gap where MD surrogates did not produce transition paths. The framework is almost training-free and can, in principle, be applied to surrogates.

Critical weaknesses

1. All baselines are the paper’s own ablations

The baselines compared against Duet-MD are Frozen and CN, which are close to ablations of the proposed method and therefore do not show how the method stands against any external approach. Being training-free is the central claim of the paper, but it is an advantage only if it costs little in performance, and no result shows that cost. On the folding tasks, conditioned path generation methods such as MDGen or ConfRover, both cited in the paper, could be added. Target hit rate would favor these methods since the endpoint is given, but reporting path level quantities such as physical plausibility of intermediates would still show what the CV-driven training-free settings gives up.

2. Validation of generated paths

Both backbones seem to generate heavy atom structures, yet the structural check reduces to CαC_\alpha distance criteria. Backbone torsions are not examined, and no energetic evaluation is reported at all, though force-field energies or the change under short unbiased relaxation are standard in transition-path work. This matters since steering pushes a frozen surrogate toward regions its training distribution covers poorly, and the MD surrogate does not constrain the output. A comparison against steered MD on the same systems would also be informative, since that is the physics-based analogue of what the paper does and would show whether the generated paths are comparable.

Questions for authors

1. Exclusions in benchmark

The folding benchmark covers three systems, where one is already reachable by frozen rollout and the other two show similar tendencies. Some smaller fast-folding proteins from the same reference paper is missing, such as Chignolin or Trp-cage, although the BBA TICA features were taken from the TPS-DPS evaluation setup. Additionally, the cryptic-pocket tasks use PVB alone, without ConfRover. Were these systems attempted, and if so what was observed?

Reviewer_4FDnOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper authors propose a training-free method for steering frozen MD surrogates toward rare conformational changes. It combines gradient guidance and particle selection within next-frame generation with weighting and selection across trajectory histories. Experiments with PVB and ConfRover show improved target reaching over unsteered generation.

Critical strengths
  • Evaluation covers different systems and two tasks, folding and pocket opening, with two frozen backbones overall.

  • The problem is well motivated: generating useful conformational changes without spending most of the computation on frequent local fluctuations.

Critical weaknesses
  • The reviewer finds the technical novelty limited: steering and particle methods are established, and their integration here needs stronger empirical justification.
  • The practical benefit relative to training-time resampling, as used in EquiJump, remains unclear without a comparison or discussion.
  • Quantitative diversity assessments and JSD comparisons of folding-related observables are missing. These require an appropriately conditioned reference or clearly justified reweighting.
  • The related-work discussion omits BioEmu and MarS-FM. The reviewer believe both works are relevant enough that should be included in prior work, particularly MarS-FM's which directly addresses the same problem of training being dominated by frequent, uninformative transitions.

References

Questions for authors
  1. Geometric validity does not establish a correct transition path. What further evidence can the paper authors provide about path quality?
  2. The folding setup uses reference-derived TICA coordinates. How does this dependence on reference MD limit applicability to unexplored systems?
  3. Could the paper authors report diversity and distributions of folding-related observables, such as radius of gyration and secondary-structure content? JSD against an appropriately conditioned MD reference would be useful, with any weighting explained.
  4. How would results compare with retraining the same backbones using TICA-cluster-based resampling of underrepresented states and their associated transitions?
  5. What is the elapsed-time cost relative to unsteered generation?

References

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.