acceptodds
Under review as a conference paper at ICLR 2027

MechPath: Geometry-Preserving Flow Matching for Protein Mechanical Unfolding

Abstract

Understanding how proteins unfold under force is important for elucidating their mechanical functions and associated mechanopathologies, such as cardiomyopathy and muscular dystrophy. Steered molecular dynamics (SMD) simulations resolve unfolding structural dynamics under physiologically relevant pulling geometries, but their high computational cost limits extensive pathway sampling. Existing generative models for protein dynamics offer a scalable alternative, yet typically construct probability paths in ambient structural spaces that do not preserve the constraints imposed by mechanical pulling. Here, we formulate unfolding trajectory generation as probability transport directly on the path geometry induced by mechanical unfolding. We parameterize each unfolding pathway by separating monotonic unfolding progress along the pulling reaction coordinate from the remaining conformational degrees of freedom. Building on this formulation, we introduce MechPath, a flow-matching framework that learns the transport in unconstrained latent tokens and maps it back to structural trajectories through a geometry-preserving realization map. This construction makes the satisfaction of the mechanical constraints an invariant of the entire generative probability path. In the absence of a public benchmark, we further establish a large-scale dataset comprising 6,750 SMD trajectories from 1,350 proteins. On this benchmark, MechPath consistently outperforms general-purpose protein dynamics generators in structural accuracy, conformational-distribution recovery, whole-pathway similarity, and underlying contact-release dynamics. The results show that incorporating the geometry of externally driven molecular processes into generative probability paths substantially improves structural pathway generation, which provides a scalable route to study molecular mechanisms underlying mechanopathologies.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein unfolding, generative modeling, flow matching, molecular dynamics, protein mechanics

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_tvUHOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work the authors focus on the protein mechanical unfolding and propose a generative approach to reconstruct the unfolding trajectory using conditioned stochastic interpolation and directly enforcing the mechanical constraint of the problem and show how this customization of the solution to the problem at hand results in better generation compared to models built for generic thermodynamic motion instead. The paper contributes a new application space, as unfolding is not widely targeted by practitioners of generative MD despite its biological relevance, a new benchmark dataset, and a new approach to embed mechanical admissability into the model design.

Critical strengths
  • Novel approach, bringing attention to an application area of biological relevance that has not been widely targeted by the community yet.
  • Embedding the stationary state of the first residue, and the fact that protein monotonically extends into the design. Model shows success, the comparison to MD is sound.
  • Significant effort in defining a benchmark and setting a baseline (altho it is not clear if the data is available)
Critical weaknesses
  • The comparison to models that were not built for unfolding, especially for mechanical admissibility, does not make the most fair comparison although this may be acceptable given the limited baseline availability. However text should be crystal clear in the comparison.
  • The design embeds the well known property of the physical system, an inductive bias, into the model and shows the gains in consistency thanks to having to represent a simplified space. Practically makes sense altho may need to be defended for generalizability. One option is to propose that every driven process gets its own model design, another option is to find flexible, generalizable representation of external drivers to biomolecule motion.

EDIT: see discussion. rating increased accordingly.

Questions for authors

It would be great to understand the data availability statement.

3 comments

We thank the reviewer for recognizing the novelty of introducing generative modeling to a biologically important yet underexplored domain, the effectiveness of embedding mechanical constraints into the model design, and our significant effort in establishing an unfolding trajectory benchmark. We address your concerns below:

W1: Baseline Fairness & Clarity of Comparison: "The comparison to models that were not built for unfolding, especially for mechanical admissibility, does not make the most fair comparison although this may be acceptable given the limited baseline availability. However text should be crystal clear in the comparison."

Response: We thank the reviewer for this comment and appreciate the opportunity to clarify the extensive efforts we made to ensure a rigorous and fair evaluation. To our knowledge, no existing generative model targets structural trajectories of mechanical unfolding under a prescribed pulling geometry. We therefore adapted the strongest available general-purpose protein dynamics generators (MDGen, 4D-Diffusion, and TEMPO) under a strictly controlled protocol (Appendix F.4). To keep the comparison as fair as possible within this scope, all methods use the same dataset, protein-level splits, selected progress states, conditioning inputs, and evaluation pipeline. Furthermore, to eliminate any architectural differences across baselines and achieve a controlled comparison, we conducted a dedicated geometric representation ablation (Table 2) that holds the exact backbone architecture, training objective, and data fixed while varying only the generative coordinates.

More fundamentally, as analyzed in Remark 1, although unconstrained ambient-space models can in principle learn mechanical constraints from training trajectories that satisfy the constraints as interpolation noise vanishes (t→1t \to 1), doing so from finite data diverts modeling capacity away from capturing unfolding dynamics, which both fails to guarantee admissibility and degrades overall structural and distributional fidelity. Consequently, our central claims do NOT rest on admissibility metrics, but are also supported by constraint-independent metrics that evaluate the quality of generated unfolding trajectories without directly measuring mechanical constraint violations. On these metrics, MechPath outperforms the strongest baselines by 27% on PA-RMSD (13.0613.06 vs. 17.9517.95), 25% on FPD (21.3121.31 vs. 28.4028.40), and 52% on PCA-JSD (0.130.13 vs. 0.270.27) (Table 1). We have modified Section 4.1 (Baselines) and Section 4.2 to make this scope crystal clear.


W2. Inductive Bias & Generalizability to Driven Processes: "The design embeds the well known property of the physical system, an inductive bias, into the model... Practically makes sense altho may need to be defended for generalizability. One option is to propose that every driven process gets its own model design, another option is to find flexible, generalizable representation of external drivers to biomolecule motion."

Response: We appreciate this valuable perspective and agree it is the right question for the broader direction of this work. Our view sits between the two options the reviewer outlines. The construction is not specific to unfolding. It applies whenever an external driver imposes (i) linear constraints on the system (e.g., fixed sites) and (ii) a monotone scalar progress along a reaction coordinate. In MechPath the driver enters only through the pulling specification π, via the covector cπc_\pi, the carrier bπb_\pi and the projection PπP_\pi. The network architecture itself is driver-agnostic. As a result, changing the attachment residues or the pulling direction needs no architectural change, because these quantities are recomputed from π for each input. More broadly, our progress–conformation decomposition can potentially generalize to many other externally driven biomolecular processes where a driver imposes affine boundary constraints and ordered progress along a low-dimensional reaction coordinate. While our current benchmark focuses on a single constant-velocity terminal pulling setup to establish a clean baseline for protein unfolding, we have explicitly integrated this broader perspective into Section 5 Conclusion.


Q1: Data Availability Statement.

For the SMD benchmark, we will host the SMD datasets on Zenodo and Hugging Face Datasets under a CC BY 4.0 license immediately if the article is accepted and upon publication. We have added an explicit Data Availability Statement alongside the Reproducibility Statement in the revised manuscript to guarantee public access.

Thanks to authors for their work. I have also been following the responses to other reviewers as well and am happy to see expanded literature, increased rigor in checking data leak in train/test sets, and generalizability claims. The ability to reconstruct the fine-grain conformation is a relevant question as well, let me extend my thanks to reviewer wk2x for that. I think the manuscript quality increased significantly. I hope to see an explicit mention of the zero-shot test on DysN-R3 wrt wetlab SM-FM result on branching probability, either in the appendix or main text in the camera ready version of the work. I appreciate the added wording in the Conclusion about where this method can extend to (a driven process with affine boundary constraints and an ordered low dim reaction coord). I believe it will help the community to think more creatively about driven processes. The pulling experiment is perhaps the low hanging fruit here but given the impact and data generation effort, it is a good starting point.
I thank the authors also for adding the data availability statement. I increase my score accordingly to a clear accept.

We sincerely thank the reviewer for the careful follow-up, for reading our responses to the other reviewers, and for raising the score to a clear acceptance. Your comments throughout the discussion have substantially improved the manuscript, and we are grateful for your engagement. We hope our paper encourages the community to explore more driven molecular processes, and we are actively pursuing some of these directions ourselves. We also thank Reviewer wk2x for raising the question of fine-grained conformational reconstruction and generalization, which helped strengthen the paper.

Thank you again for your time, constructive feedback, and support of this work.

Reviewer_w2kxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

MechPath is a flow-matching model that jointly generates 15 subsequent Cα conformations from a protein’s sequence and initial structure. Its main methodological contribution separates unfolding progress from conformational variation, enforcing a fixed anchor and monotonically increasing projected extension throughout generation. The paper authors construct a benchmark of 6,750 coarse-grained steered-MD trajectories from 1,350 proteins, evaluating structural agreement, contact release and pathway distributions. The paper authors compare against adapted MDGen, 4D-Diffusion and TEMPO baselines, although the evaluation does not establish physical-time kinetics or detailed structural validity.

Critical strengths
  • Joint generation of the entire pathway is a useful design choice, although it is already present in prior work such as MDGen (Jing et al.).
  • The benchmark is a useful contribution for unfolding pathways, with evaluation at the structure, whole-pathway and contact-release levels.
  • The use of standard structural auxiliary losses near the clean endpoint is reasonable.

References

Critical weaknesses
  • The Cα-only representation limits assessment of backbone stereochemistry and side-chain interactions. It remains unclear whether generated pathways can be reconstructed into physically plausible higher-resolution structures, although the main neural-network components are not inherently restricted to Cα representations.

  • The 16-state, progress-parameterized output does not establish physical-time kinetics, and calibration of unfolding-branch probabilities is not demonstrated.

  • Physical-validity evaluation is limited: no bond-distance validity rates or steric-clash assessment are reported. Although designed for ligand docking, PoseBusters (Buttenschoen et al.) illustrates the value of reporting geometric and clash checks alongside structural similarity; checks appropriate to the Cα representation are needed here.

    • The related-work discussion omits relevant generative models of molecular dynamics, including EquiJump (Costa et al.) and MarS-FM (Kapuśniak et al.). Their different modeling objectives and representations would help contextualize the contribution.
  • There is a minor presentation issue around the line break at lines 270–271.

References

Questions for authors
  • What fraction of generated pathways reaches the unfolded state?
  • Please quantify the physical validity of generated samples, for example through Cα bond-distance violation rates, non-neighbour clashes.
  • How much of PC1/PC2 reflects global extension rather than internal conformational differences? Why not consider TICA fitted to the reference MD which was deisgned to isolate low frequency movement (Pérez-Hernández et al.)?
  • Why were 16 total states chosen?
  • What is the similarity between the training and test proteins? Please report the distribution of each test protein’s maximum sequence identity to the training set.

References

8 comments
AuthorsOpenReview ↗

Response to Reviewer w2kx (1/3)

We thank the reviewer for recognizing the soundness of our approach, the value of our mechanical unfolding benchmark, and the effectiveness of joint pathway generation with structural regularization. Below, we address your concerns with new experiments and analyses.


W2: Physical-Time Kinetics & Branch Probabilities

The 16-state, progress-parameterized output does not establish physical-time kinetics, and calibration of unfolding-branch probabilities is not demonstrated.

Response: We appreciate the opportunity to clarify our modeling scope. MechPath is designed as a scalable complement to steered MD (SMD), not a replacement for physical-time kinetics. Its goal is to sample mechanically admissible pathways at ∼103×\sim 10^3\times lower computational cost (see lines 413-418 in our manuscript) to provide structural priors that initialize downstream umbrella or path-sampling simulations (Noé et al., 2019; Lu et al., 2024; Viguera Diez et al., 2024). Because experimental and SMD unfolding trajectories are scarce, enforcing pulling geometry as a hard generative invariant (Proposition 2) enables strong data efficiency and generalization to larger proteins (Figure 3).

To evaluate unfolding-branch probabilities against wet-lab experiments, we performed a zero-shot test on DysN-R3 (a 667-residue dystrophin fragment linked to Duchenne muscular dystrophy). While 100 SMD trajectories would require ∼103\sim 10^3 CPU-hours (approximately 10 CPU-hours per trajectory), MechPath generated 100 pathways in ∼10\sim 10 minutes. Comparing the domain-level unfolding order between spectrin-like repeats (SLRs) and the N-terminal actin-binding domain (NT) against experimental single-molecule force spectroscopy (SMFS; Hua et al., 2026), MechPath recovers both the dominant experimental branch (SLRs →\rightarrow NT) and the minor pathway (NT →\rightarrow SLRs) zero-shot:

MethodSLRs →\rightarrow NT (Dominant)NT →\rightarrow SLRs (Minor)
Experimental SMFS (Hua et al., 2026)0.750.25
MechPath (Zero-shot)0.870.13

While MechPath captures both pathway heterogeneity and the dominant unfolding mechanism zero-shot, the exact branch ratios (0.870.87 vs. 0.750.75) are not yet perfectly calibrated. Overall, this zero-shot agreement with wet-lab SMFS demonstrates that embedding mechanical geometry as a generative invariant enables large-scale exploration of unfolding pathways far beyond the practical sampling capacity of conventional MD, while providing physically reliable structural priors to accelerate downstream physics-based simulations. We have added the discussion of experimental calibration as an exciting direction in Section 5.


W1: CαC_\alpha Representation & All-Atom Backmapping Plausibility

The CαC_\alpha-only representation limits assessment of backbone stereochemistry and side-chain interactions. It remains unclear whether generated pathways can be reconstructed into physically plausible higher-resolution structures.

Response: We thank the reviewer for raising this important point. We adopted a one-bead-per-residue CαC_\alpha structure-based (Gō) model, a standard paradigm validated in mechanical unfolding studies (Cieplak & Marszalek, 2005; Sułkowska & Cieplak, 2008; Sikora et al., 2011), to make simulating a 6,750-trajectory public benchmark computationally tractable while preserving native-contact topology.

To verify whether generated Cα pathways reconstruct into physically plausible all-atom structures, we backmapped all 16 states across 50 randomly selected test proteins into full-atom backbone and side-chain coordinates using PULCHRA (Rotkiewicz & Skolnick, 2008) and evaluated them with MolProbity (Chen et al., 2010):

MethodClashscore ↓\downarrowRamachandran Outliers (%) ↓\downarrow
SMD Reference (Ground Truth)9.809.83
TEMPO (Strongest Baseline)251.8937.88
MechPath (Ours)30.1120.98

MechPath reduces all-atom Clashscore by 88% (30.1130.11 vs. 251.89251.89) and Ramachandran outliers by 16.90 percentage points (20.98%20.98\% vs. 37.88%37.88\%) compared to TEMPO (the best-performing baseline in Table 1). Since even ground-truth SMD trajectories exhibit non-zero outliers (9.809.80 clashscore, 9.83%9.83\% outliers) when heuristic backmapping is applied to highly stretched chains under tension, these results confirm that preserving mechanical pulling geometry in the CαC_\alpha latent space yields substantially more stereochemically plausible backbones that can be reliably reconstructed into higher-resolution all-atom intermediates for downstream atomistic refinement. We have added the backmapping results to Appendix G.6.

AuthorsOpenReview ↗

Response to Reviewer w2kx (2/3)

W3, Q1 & Q2: CαC_\alpha Geometric Validity & Unfolded-State Completion

Quantify physical validity (e.g., CαC_\alpha bond-distance violations, non-neighbor clashes) following PoseBusters (Buttenschoen et al.), and report the fraction of pathways reaching the unfolded state.

Response: We thank the reviewer for this valuable suggestion. To evaluate physical plausibility alongside global unfolding completion across the test benchmark, we report three complementary metrics: (1) CαC_\alpha bond-distance pass rate (consecutive CαC_\alpha distances within 3.4–4.23.4\text{--}4.2 Å, following PoseBusters; Buttenschoen et al., 2024), (2) nonlocal clash-free rate (no distances <2.5<2.5 Å for sequence separation ≥3\ge 3), and (3) unfolded-state completion rate (Q≤0.10Q \le 0.10 and extension ≥80\ge 80 % of contour length). Because MechPath includes bond-length auxiliary loss Lbond\mathcal{L}_{\text{bond}} during training while baselines do not, we applied the same lightweight, reference-free geometric refinement (iteratively relaxing adjacent bonds and non-local overlaps while keeping anchor residues and projected extension fixed) across all methods for a strictly fair comparison:

ModelCαC_\alpha Bond Pass Rate (%) ↑\uparrowNonlocal Clash-Free Rate (%) ↑\uparrowUnfolded Completion (%) ↑\uparrow
MDGen57.80±12.9557.80 \pm 12.95100.00±0.01100.00 \pm 0.0146±1146 \pm 11
4D-Diffusion77.28±20.9177.28 \pm 20.9199.25±0.9499.25 \pm 0.9434±3534 \pm 35
TEMPO89.79±5.1089.79 \pm 5.1099.88±0.0899.88 \pm 0.0830±1.230 \pm 1.2
MechPath (Ours)99.41 ± 0.17100.00 ± 0.0078 ± 11

MechPath outperforms all baselines across both local physical validity and global unfolding completion, achieving 99.41%99.41\% bond-distance validity (+9.62+9.62% over TEMPO), 100.00%100.00\% clash-free frames, and 78%±11%78\% \pm 11\% unfolded-state completion (more than doubling TEMPO's 30%30\% and 4D-Diffusion's 34%34\%). Crucially, for MechPath, this geometric refinement requires only a minimal 0.80.8 Å mean CαC_\alpha coordinate adjustment and slightly improves pathway fidelity (PA-RMSD drops by 0.12–0.150.12\text{--}0.15 Å). We have added the complete validity metrics, refinement protocol, and paired evaluation results to Appendix G.7.


Q3: Decomposing PC1/PC2 Variance & Applicability of TICA

How much of PC1/PC2 reflects global extension vs. internal conformation? Why not consider TICA (Pérez-Hernández et al.)?

Response: Motivated by your insightful question, we decomposed the reference SMD principal components into global extension along the pulling axis vs. internal conformational variation. The results confirm that PC1 is dominated by global extension (97.7%97.7\% extension vs. 2.3%2.3\% internal), whereas PC2 almost purely captures internal conformational rearrangements (99.1%99.1\% internal vs. 0.9%0.9\% extension). Evaluating marginal Jensen–Shannon divergence (JSD) along each axis shows where MechPath's gains originate:

ModelPCA-JSD (PC1+PC2) ↓\downarrowGlobal Extension JSD (PC1) ↓\downarrowInternal Conformation JSD (PC2) ↓\downarrow
MDGen0.28±0.020.28 \pm 0.020.045±0.00330.045 \pm 0.00330.37±0.0330.37 \pm 0.033
4D-Diffusion0.27±0.090.27 \pm 0.090.096±0.03200.096 \pm 0.03200.32±0.0740.32 \pm 0.074
TEMPO0.30±0.030.30 \pm 0.030.060±0.00380.060 \pm 0.00380.34±0.0180.34 \pm 0.018
MechPath (Ours)0.13 ± 0.00410.027 ± 0.00340.18 ± 0.0048

MechPath's margin is most pronounced on PC2 (internal conformation), cutting JSD by 44% relative to the strongest baseline 4D-Diffusion (0.180.18 vs. 0.320.32). This directly supports our central argument: preserving mechanical admissibility throughout generation allows the model to devote its modeling capacity to internal conformational changes. We have incorporated the PC1/PC2 decomposition results into Appendix G.8.

Time-lagged independent component analysis (TICA; Pérez-Hernández et al., 2013) requires time-lagged covariance matrices computed over a constant physical lag time Δt\Delta t. Because our 16 states are subsampled uniformly in normalized mechanical progress p∈[0,1]p \in [0, 1] rather than uniform physical time (Appendix B.2), a fixed index lag Δk\Delta k corresponds to variable physical time intervals, making TICA ill-defined on progress-parameterized trajectories.

AuthorsOpenReview ↗

Response to Reviewer w2kx (3/3)

Q5: Sequence Identity Distribution & Generalization to Sequence-Dissimilar Proteins

Report the distribution of each test protein’s maximum sequence identity to the training set.

Response: Using full-length global pairwise alignments (BLOSUM62, gap open −11-11, extend −1-1) between all evaluated test proteins and training proteins, nearly half (45.88%) of test proteins share <50<50 % maximum sequence identity with any training protein (24.7124.71% have <30<30%, 21.1821.18% fall in [30[30%, 5050%), and 54.1254.12% have ≥50\ge 50%). Evaluating models on the strict <50<50% and <30<30% subsets confirms that MechPath does not rely on sequence memorization:

Max Seq. IdentityModelPA-RMSD (Å) ↓\downarrowCR-W1\text{CR-}W_1 ↓\downarrowPCA-JSD ↓\downarrowFPD (Å) ↓\downarrow
<50% (n=39)MDGen61.42±2.2561.42 \pm 2.250.335±0.0180.335 \pm 0.0180.356±0.0110.356 \pm 0.011119.86±3.83119.86 \pm 3.83
4D-Diffusion26.36±10.3026.36 \pm 10.300.223 ± 0.0150.264±0.0930.264 \pm 0.09339.97±15.4539.97 \pm 15.45
TEMPO18.15±0.2418.15 \pm 0.240.237±0.0030.237 \pm 0.0030.321±0.0060.321 \pm 0.00627.84±0.5127.84 \pm 0.51
MechPath15.44 ± 0.660.253±0.0090.253 \pm 0.0090.175 ± 0.01223.57 ± 0.34
<30% (n=21)MDGen70.57±1.3270.57 \pm 1.320.329±0.0240.329 \pm 0.0240.389±0.0230.389 \pm 0.023140.94±2.23140.94 \pm 2.23
4D-Diffusion29.90±11.9729.90 \pm 11.970.227 ± 0.0180.274±0.1080.274 \pm 0.10846.34±18.0346.34 \pm 18.03
TEMPO19.93±0.1819.93 \pm 0.180.243±0.0040.243 \pm 0.0040.294±0.0060.294 \pm 0.00631.33±0.2231.33 \pm 0.22
MechPath16.72 ± 0.640.257±0.0120.257 \pm 0.0120.155 ± 0.02026.36 ± 0.28

Even on the most challenging <30%<30\% identity subset, MechPath outperforms the strongest structural baseline TEMPO by 16% on PA-RMSD, 16% on FPD, and 47% on PCA-JSD, which provides additional evidence of MechPath's advantages in structural accuracy and pathway distribution modeling. We have added the full sequence-identity results and subgroup tables to Appendix G.9.


Q4: Choice of 16 States

Why were 16 total states chosen?

Response: We selected 16 progress-aligned states (X0X_0 plus K=15K=15) to capture major conformational rearrangements while filtering out picosecond thermal noise (∼435 ps\sim 435\text{ ps} average spacing across ∼6,500\sim 6,500 raw 1 ps1\text{ ps} SMD frames for a 100-residue protein). Our systematic resolution ablation across 4, 8, 16, and 32 states (detailed in our Response to Reviewer gyVx, Q4) shows that doubling from 16 to 32 states yields no gain in PA-RMSD (12.86→13.3612.86 \rightarrow 13.36 Å) or PCA-JSD (0.127→0.1230.127 \rightarrow 0.123) while increasing training cost by ∼4.44×\sim 4.44\times (180→800 minutes180 \rightarrow 800\text{ minutes}).


W4 & W5: Expanded Literature on Generative MD & Presentation

The related-work discussion omits relevant generative models of molecular dynamics, including EquiJump (Costa et al.) and MarS-FM (Kapuśniak et al.)... minor presentation issue at lines 270–271.

Response: Thank you for pointing out these relevant works. We have fixed the line break at Lines 270–271 in our original manuscript and expanded Section 1 to position MechPath against recent generative MD models—including multi-timestep/equivariant propagators (EquiJump, Costa et al., 2024; Timewarp, Klein et al., 2023), Markov-state flow matching (MarS-FM, Kapuśniak et al., 2026). Unlike these methods, which model unconstrained equilibrium dynamics or fixed-lag transitions, mechanical unfolding is an externally driven, non-equilibrium process requiring a fixed anchor and monotonically advancing extension, constraints that MechPath enforces as exact invariants throughout generation.

References

  • Buttenschoen et al. (2024). Chem. Sci.
  • Chen et al. (2010). Acta Crystallogr. D.
  • Cieplak & Marszalek (2005). J. Chem. Phys.
  • Costa et al. (2024). EquiJump, arXiv:2410.09667.
  • Hua et al. (2026). Resolving Heterogeneous Mechanical Domains via Physics-Aware Deep Clustering of Single-Molecule Force Spectroscopy Data. bioRxiv.
  • Kapuśniak et al. (2026). MarS-FM, ICLR.
  • Klein et al. (2023). Timewarp, NeurIPS.
  • Lu et al. (2024). Str2Str, ICLR.
  • Noé et al. (2019). Science.
  • Pérez-Hernández et al. (2013). J. Chem. Phys.
  • Rotkiewicz & Skolnick (2008). J. Comput. Chem.
  • Sikora et al. (2011). Nucleic Acids Res.
  • Sułkowska & Cieplak (2008). Biophys. J.
  • Viguera Diez et al. (2024). Mach. Learn.: Sci. Technol.
Reviewer_gyVxOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses the problem of modeling Steered Molecular Dynamic (SMD) simulation by proposing a new framework (MechPath), introducing a novel dataset and relevant metrics. MechPath is essentially a FlowMatching approach based on a system representation designed to satisfy the constrains typical on SMD simulations. An experimental analysis demonstrates the strong performance of the flow-matching MechPath method, setting strong foundations for future work.

Critical strengths
  • The paper is mostly clear, providing a sufficient motivation and guiding the reader in understanding the problems of Steered Molecular Dynamics
  • The paper designs a robust representation of the problem and clearly demonstrate its advantages compared to existing approaches in literature.
  • The paper provides a new dataset toghether with solid foundations for future research
Critical weaknesses

Main Concerns

I don't have any major concerns with the paper although I have to admit my limited familiarity with the field of SMD.

Other issues

  • Numerous sections in the papers are very dense in definitions and notation (e.g. lines 142-148, 205-223). In particular ζk\zeta_k and ζ~k\tilde\zeta_k, which are quite central for the proposed method are introduced very shortly. I recommend the author to provide additional intuition in these sections and simplify the notation where possible
  • The training objective described in section 3.2 introduces additional loss components that appear quite abrupt and not strongly motivated although the high-level intuition is clear.
Questions for authors
  1. The SMD problem is presented as a mechanical spring with stiffness ksk_s pulling on an end of a protein, which is fixed on the other end. Can the authors clarify why the set of admissible paths has only monotonically increasing extensions (Equation 1)? Why isn't the molecule allowed to "temporarily fight back" the spring?

  2. Can the authors clarify the rationale behind the hard threshold for the loss introduced in Equation 9? How does the proposed method perform without the additional regularization or with different regularization policies (e.g. regularization introduced earlier/later, soft regularization as a function of time)?

  3. In Table 1 the proposed MechPath model underperforms in terms of the RO-τ\tau metric, while in Table 2, the MechPath model augmented with local frame geometry outperforms the baselines. Is there any downside in including this additional information into the model input given that the mechanical feasibility constraint seem to be satisfied?

  4. The dataset introduced in this work reports only 15 states for each trajectory, which I guess is a severe reduction compared to the number of states produced during the ground-truth simulations. What is the total number of frames produced during the simulation? What is the temporal resolution of the reported sequences? Can the author clarify the rationale behind this specific choice?

3 comments
AuthorsOpenReview ↗

Response to Reviewer gyVx (1/2)

We sincerely thank the reviewer for recognizing our clear motivation, the robustness of the MechPath representation, and the value of our unfolding trajectory benchmark as a solid foundation for future research. Below, we address your questions on physical intuition, training objectives, and dataset design.


Q1: Physical Intuition Behind Monotonic Extension

Can the authors clarify why the set of admissible paths has only monotonically increasing extensions (Equation 1)? Why isn't the molecule allowed to "temporarily fight back" the spring?

Response: We appreciate the opportunity to clarify the distinction between raw thermal fluctuations and progress-parameterized structural pathways in our study. In raw SMD trajectories resolved at the femtosecond-to-picosecond scale, thermal agitation (kBTk_B T) indeed causes the protein to temporarily fluctuate against the harmonic spring. However, MechPath targets large-scale conformational transitions across the unfolding pathway rather than short-timescale thermal noise. Under sustained directional pulling, cumulative mechanical work over coarse-grained displacements substantially exceeds kBTk_B T, making progressive extension the dominant macroscopic trend, which is consistent with SMFS experiments where major unfolding events appear as discrete extension steps (Oberhauser et al., 2001; Bustamante et al., 2020).

Crucially, our preprocessing reparameterizes each trajectory along a monotonically normalized hybrid progress coordinate p∈[0,1]p \in [0, 1] and selects K=15K=15 states at strictly increasing progress levels (Appendix B.2). Thus, Equation 1 orders these state slices (0=e0<e1<⋯<eK0 = e_0 < e_1 < \dots < e_K) while allowing the protein chain to freely rearrange and resist tension internally within each equal-extension slice Tπ\mathcal{T}_\pi. We have added this clarification right after Equation 1.


W2 & Q2: Motivation of Auxiliary Losses & Hard Threshold (t>0.75t > 0.75)

Clarify the rationale behind the hard threshold in Equation 9, and show how the method performs without regularization or with alternative regularization policies.

Response: Thanks for this valuable suggestion. We introduce two auxiliary losses with a late-stage gate to refine local structural geometry near the clean data endpoint while avoiding gradient instability from noisy early-time extrapolations. Specifically, while the realization map guarantees global fixed-anchor and monotonic-extension constraints throughout generation, unconstrained flow matching does not automatically preserve local C-alpha bond geometry or supervise native-contact rupture timing. We therefore apply the two auxiliary losses to the decoded clean-pathway estimate. Because clean-endpoint extrapolation involves division by α(t)\alpha(t), where α(t)=sin⁡(πt/2)\alpha(t)=\sin(\pi t/2), prediction errors are amplified at early generative times. Enforcing structural constraints at these noisy stages can introduce unstable gradients. Following standard practice in protein diffusion (Yim et al., 2023), we activate structural supervision only when t>0.75t>0.75, where the estimated clean pathway is more reliable.

As shown in our additional ablation experiment below (within-run mean ±\pm std across test proteins), MechPath already achieves strong structural accuracy without any auxiliary loss (PA-RMSD 13.04\text{PA-RMSD } 13.04 Å), while combining both losses at t>0.75t > 0.75 yields the best overall performance (reducing PA-RMSD by 1.38%1.38\% and CR-W1\text{CR-}W_1 by 5.39%5.39\%):

MetricNo Aux. LossNo Contact LossNo Bond LossAll Losses (Default)
PA-RMSD (Å) ↓\downarrow13.04±8.0813.04 \pm 8.0812.97±8.0212.97 \pm 8.0213.32±8.1513.32 \pm 8.1512.86 ±\pm 8.12
CR-W1\text{CR-}W_1 ↓\downarrow0.2151±0.07970.2151 \pm 0.07970.2171±0.08240.2171 \pm 0.08240.2153±0.07650.2153 \pm 0.07650.2035 ±\pm 0.0814
RO-τ\tau ↑\uparrow0.3552±0.15730.3552 \pm 0.15730.3537±0.15670.3537 \pm 0.15670.3468±0.15170.3468 \pm 0.15170.3565 ±\pm 0.1594

Overall, these results confirm that MechPath's primary performance gains stem from enforcing global pulling geometry by construction, while late-stage auxiliary regularization provides a stable, complementary refinement to local bond geometry and contact-release dynamics. We have added this ablation table to Appendix G.4 and included a discussion of the rationale for the late-stage gate in Section 3.2 (training objectives).

AuthorsOpenReview ↗

Response to Reviewer gyVx (2/2)

Q3: Role of Local Frame Geometry

Is there any downside in including the additional local frame geometry information (Table 2) into the model input?

Response: There is virtually no computational downside because adding virtual dihedral angles and residue frame orientation adds only 5 scalar channels per residue (Appendix C.1, Lines 1068–1071) with only a minor trade-off in global ensemble PCA-JSD (0.13→0.150.13 \rightarrow 0.15). In Table 1 (Lines 324–333), the base MechPath model (RO-τ=0.35\text{RO-}\tau = 0.35) already outperforms MDGen (0.290.29) and TEMPO (0.290.29), trailing only 4D-Diffusion (0.370.37), while adding local frame angles in Table 2 (Prog.-Conf. with Angles) boosts RO-τ\text{RO-}\tau to 0.380.38 (surpassing all baselines) and further lowers PA-RMSD (12.4212.42 Å) and FPD (19.9419.94 Å). We used the pure progress–conformation model as our default in Table 1 to cleanly isolate the contribution of our pulling-geometry decomposition, while offering the angular variant in Table 2 as a ready-to-use option when local rupture ordering is prioritized.


Q4: Simulation Frame Count, Temporal Resolution & Choice of K=15K=15 States

What is the total number of frames produced during simulation, what is the temporal resolution, and what is the rationale for selecting 15 states?

Response: Under our constant-velocity SMD setup (vpull=0.005 nm ps−1v_{\text{pull}} = 0.005\text{ nm ps}^{-1}, 1 fs1\text{ fs} integration step, coordinates saved every 1 ps1\text{ ps}; Appendix B.1), unfolding a representative 100-residue protein produces ∼6,500\sim 6,500 raw frames at 1 ps1\text{ ps} temporal resolution. Retaining K+1=16K+1 = 16 states (X0X_0 plus 15 unfolding states) corresponds to an average temporal spacing of ∼435 ps\sim 435\text{ ps}. Because adjacent 1 ps1\text{ ps} frames are dominated by redundant thermal vibrations, we subsample uniformly along our hybrid mechanical progress p∈[0,1]p \in [0, 1] (Appendix B.2) so every state captures a distinct structural state.

To quantitatively justify K+1=16K+1 = 16 states, we performed a resolution ablation across 4, 8, 16, and 32 states:

States (K+1K+1)PA-RMSD (Å) ↓\downarrowCR-W1\text{CR-}W_1 ↓\downarrowRO-τ\tau ↑\uparrowPCA-JSD ↓\downarrowFPD (Å) ↓\downarrowTrain Time (50 ep)Inference (600 trajs)
4 states10.93 ±\pm 7.220.17 ±\pm 0.050.45 ±\pm 0.160.15 ±\pm 0.1119.11 ±\pm 15.38——
8 states12.00 ±\pm 7.100.19 ±\pm 0.080.43 ±\pm 0.180.11 ±\pm 0.1020.18 ±\pm 12.98——
16 states (Default)12.86 ±\pm 8.120.20 ±\pm 0.080.36 ±\pm 0.160.13 ±\pm 0.1221.15 ±\pm 14.99180 min2 min
32 states13.36 ±\pm 7.920.22 ±\pm 0.080.31 ±\pm 0.140.12 ±\pm 0.1021.61 ±\pm 15.18800 min4 min

While coarser trajectories (4 or 8 states) yield lower per-state errors simply because interpolating fewer states is an easier task, they may miss fine-grained intermediate barriers. Conversely, doubling from 16 to 32 states yields no improvement in PA-RMSD ($12.86 \rightarrow 13.36 $ Å) or PCA-JSD (0.127→0.1230.127 \rightarrow 0.123) while increasing 50-epoch training time by 4.44×4.44\times (180→800 min180 \rightarrow 800\text{ min}) due to the O(KN2+NK2)\mathcal{O}(KN^2 + NK^2) spatiotemporal attention complexity (Appendix C.2). Thus, 1616 states provide the optimal accuracy–efficiency trade-off. We have added this ablation table to the appendix G.5 and included a discussion of the rationale for K=15K=15 in Section 4.1 (Dataset).


Additional Presentation Comments.

To ease notation density in Sections 2 and 3.1, we have added a notation lookup table in Appendix A.1 clarifying the two-stage latent-to-structure mapping.

The latent conformational variable is ζ~k\tilde{\zeta}_k, which preserves the fixed anchor but is not constrained to an equal-extension slice. It is projected to ζk\zeta_k which removes the component along the mechanical reaction coordinate and preserves the prescribed extension.

References

  • Bustamante et al. (2020). Single-molecule studies of protein folding with optical tweezers. Annu. Rev. Biochem.
  • Oberhauser et al. (2001). Stepwise unfolding of titin under force-clamp atomic force microscopy. Proc. Natl. Acad. Sci. U.S.A.
  • Yim et al. (2023). SE(3) diffusion model with application to protein backbone generation. ICML.
Reviewer_gyVxOpenReview ↗

Response to the Authors

I wish to thank the authors for their detailed answers and additional experiments that I believe can further strengthen the paper.

I, therefore, confirm my evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.