acceptodds
Under review as a conference paper at ICLR 2027

Learning Protein-Specific Elastic Interactions for Globally Coupled Protein Dynamics

Abstract

Proteins function through conformational flexibility and collective motion that a single reference structure does not specify. Molecular dynamics (MD) simulation captures these dynamics but requires long trajectories; learning-based methods that predict covariances from structure avoid this cost, but they couple each residue pair through a single scalar and provide no mechanism from which globally consistent collective motion arises. Elastic network models (ENMs) offer such a mechanism, describing global dynamics through explicit inter-residue interactions, yet they prescribe these interactions with fixed cutoffs and hand-designed rules. We introduce ENMulator, which combines the two: a network maps sequence and structure to protein-specific elastic interactions, which parameterize an elastic network whose Hessian induces a globally coupled, residue-level displacement covariance. A complementary branch predicts residue-wise anisotropic covariances, and a blockwise geodesic congruence calibrates the residue marginals of the elastic covariance toward these predictions while exactly preserving the canonical correlations between residues. On the ATLAS test set, in both covariance and sampled-ensemble evaluation, it improves recovery of covariance geometry, collective modes, and correlated motion over prescribed ENMs and direct covariance predictors, reducing DCCM error by 17% relative to the next-best method, while remaining competitive on residue-wise fluctuations. These results support learned elastic interactions as a principled bridge from a single reference structure to globally coupled protein dynamics without simulation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein dynamics, Learned elastic interactions, Covariance prediction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nKRMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ENMulator, a framework for predicting protein dynamics from sequence and a single reference structure. The method learns protein-specific elastic interactions to construct a Hessian-based, globally consistent displacement covariance, while a complementary branch predicts residue-wise anisotropic fluctuations. A geodesic congruence operation combines these predictions while preserving positive semidefiniteness and inter-residue canonical correlations. Experiments on the ATLAS benchmark demonstrate improvements over classical elastic network models and learning-based covariance predictors in covariance reconstruction, collective-mode recovery, and directional residue coupling. The authors further evaluate conformational ensembles generated from the predicted covariance.

Critical strengths
  1. Principled integration of physical modeling and machine learning. The paper presents a well-motivated framework that combines protein-specific learned interactions with the global mechanical structure of elastic network models. The Hessian-based covariance construction and geodesic congruence provide useful mathematical guarantees, particularly positive semidefiniteness and preservation of canonical correlations. This offers a meaningful methodological contribution beyond directly predicting pairwise correlations.

  2. Strong empirical validation of global protein dynamics. The paper evaluates multiple complementary aspects of protein dynamics, including covariance reconstruction, collective modes, directional coupling, and sampled ensembles. The consistent improvements over established baselines, supported by targeted ablations and theoretical analysis of directional coupling limitations, provide convincing evidence of the method's effectiveness.

Critical weaknesses
  1. Ambiguity between reference-centered covariance and genuine dynamical coupling. The MD supervision uses the reference-centered second moment, $\mathbb{E}[\Delta x\Delta x^\top] $, not the mean-centered covariance. Consequently, the predicted cross-residue correlations may partly reflect systematic displacements from the reference structure rather than genuine correlated fluctuations. This distinction is particularly consequential because recovering directional residue-pair coupling is a central claim of the paper. Although the authors acknowledge this difference, they do not establish whether their improvements persist under mean-centered covariance evaluation. Re-evaluating DCCM, canonical correlations, and collective modes using mean-centered MD statistics would substantially strengthen the physical interpretation of the results.

  2. Insufficient evidence that the elastic-network inductive bias drives the improvements. The experiments demonstrate improvements over classical ENMs and existing direct covariance predictors, but do not isolate whether these gains arise specifically from learned elastic interactions or more generally from enforcing a globally positive-semidefinite covariance. Since the direct predictors considered have restricted pairwise coupling representations, they do not provide a sufficiently controlled comparison. A matched baseline that directly predicts a globally consistent covariance (e.g., through a low-rank PSD parameterization), using the same encoder and supervision, would help establish whether the elastic-network construction provides a distinct advantage. Without such evidence, the paper's central methodological claim remains insufficiently validated.

Questions for authors
  1. Reference-centered versus mean-centered covariance. To what extent do the reported improvements in directional coupling and collective-mode recovery persist when evaluated using mean-centered MD covariance? Could the authors quantify how much of the predicted cross-residue coupling is attributable to mean displacement from the reference structure rather than correlated fluctuations? This would clarify the physical interpretation of the central results.

  2. Contribution of the elastic-network inductive bias. Can the authors clarify whether the improvements primarily arise from the elastic-network formulation or from enforcing a globally consistent PSD covariance? Would a controlled comparison against a direct PSD covariance predictor with a matched encoder and training objective yield similar performance? Such evidence would help establish the necessity of the proposed mechanical inductive bias.

3 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer nKRM

We thank the reviewer for recognizing both our methodological design and the breadth of our evaluation. In particular, the reviewer noted that combining learned protein-specific interactions with the mechanical structure of an elastic network, together with the guarantees of the Hessian construction and geodesic congruence, goes beyond direct covariance prediction, and that consistent results across covariance reconstruction, collective modes, directional coupling, and sampled ensembles provide complementary evidence. We address the reviewer’s remaining concerns and questions below.

[W1]

Ambiguity between reference-centered covariance and genuine dynamical coupling. … This distinction is particularly consequential because recovering directional residue-pair coupling is a central claim of the paper. … Re-evaluating DCCM, canonical correlations, and collective modes using mean-centered MD statistics would substantially strengthen the physical interpretation of the results.

The improvements are not driven primarily by systematic displacement from the reference: they persist when all methods are evaluated against the mean-centered MD covariance (Table R1; reference-centered / mean-centered). Directional coupling is particularly stable: ecce_{cc} remains 0.121 and is lower than BackFlip-2's on all 82 test proteins under both evaluations.

Table R1. Comparison of protein dynamics metrics using reference-centered and mean-centered MD covariance.

ANMHinsenDynaProtBackFlip-2ENMulator
DCCM rr ↑0.779/0.7630.824/0.8030.579/0.5780.797/0.7950.856/0.841
Collective modes (top-10 overlap) ↑0.564/0.5640.636/0.6350.526/0.5330.625/0.6210.663/0.660
Canonical correlations (ecce_{cc}) ↓0.269/0.2700.185/0.1870.335/0.3340.280/0.2820.121/0.121

We use the reference-centered second moment because the reference is the only structure ENMulator receives, and alignment removes the same six rigid-body directions excluded by the elastic network. DynaProt and BackFlip-2 use the same convention. We clarified this choice and added the mean-centered evaluation in App. C.1, L904–911 and Table A3.

[W2]

Insufficient evidence that the elastic-network inductive bias drives the improvements. … the direct predictors considered have restricted pairwise coupling representations … A matched baseline that directly predicts a globally consistent covariance … using the same encoder and supervision, would help establish whether the elastic-network construction provides a distinct advantage.

The gains arise from the learned elastic interactions, not positive semidefiniteness alone. With the same encoder and supervision, fixed ANM springs remain at the ANM-level DCCM rr (0.778 vs. 0.779), whereas learned springs reach 0.851 without the marginal branch and 0.856 in the full model (Table A14). Because DynaProt and BackFlip-2 use one scalar per residue pair (Table 4), we also added the requested matched rank-64 PSD baseline with the same encoder and supervision. It underperforms both learned springs and full ENMulator across all four metrics (Table R2):

Table R2. Comparison of a matched low-rank PSD predictor, learned elastic springs, and the full ENMulator.

Low-rank PSDLearned springs onlyENMulator
Rel. Fro. ↓0.9170.9140.874
BW ↓0.8300.6900.666
Top-10 overlap ↑0.4750.6430.663
DCCM rr ↑0.7350.8510.856

We added this baseline and its analysis in App. G.3 and Table A15.

[Q1]

Reference-centered versus mean-centered covariance. To what extent do the reported improvements in directional coupling and collective-mode recovery persist when evaluated using mean-centered MD covariance?

They persist under mean-centered evaluation; see our response to W1 and Table R1.

Could the authors quantify how much of the predicted cross-residue coupling is attributable to mean displacement from the reference structure rather than correlated fluctuations? This would clarify the physical interpretation of the central results.

The effect is small. The reference-centered and mean-centered MD coupling maps are nearly identical, with a median Pearson correlation of 0.995 between their residue-pair coupling strengths s1s_1 (App. C.1).

[Q2]

Contribution of the elastic-network inductive bias. Can the authors clarify whether the improvements primarily arise from the elastic-network formulation or from enforcing a globally consistent PSD covariance?

They arise primarily from the elastic-network formulation; see our response to W2.

Would a controlled comparison against a direct PSD covariance predictor with a matched encoder and training objective yield similar performance? Such evidence would help establish the necessity of the proposed mechanical inductive bias.

No, the matched rank-64 PSD predictor performs substantially worse than learned springs (BW 0.830 vs. 0.690; DCCM rr 0.735 vs. 0.851); see our response to W2 and Table R2.

Dear Reviewer nKRM,

Thank you again for your thoughtful review and helpful comments on our paper.

We would like to briefly highlight the updates in our response. Following your comments, we re-evaluated DCCM, canonical correlations, and collective modes against mean-centered MD covariance (Table R1), and quantified the agreement between reference-centered and mean-centered coupling maps. We also added a matched rank-64 PSD covariance predictor using the same encoder and supervision (Table R2). These comparisons show that the improvements persist under mean-centered evaluation and provide further evidence for the contribution of the learned elastic-network formulation beyond enforcing a globally consistent PSD covariance.

We sincerely appreciate the time and effort you have already devoted to reviewing our paper. If time permits before the discussion period closes, we would be grateful if you could read through our response. We would welcome any further thoughts and would be happy to clarify any remaining questions.

Best regards, 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 positive evaluation.

Reviewer_5R4zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ENMulator, a hybrid approach that combines Elastic Network Models (ENMs) with deep learning to predict protein-specific Cα displacement covariance from a single structure. The method learns a residue-specific spring stiffness matrix to build a global Hessian, whose pseudoinverse yields a base covariance. To address the limitation that this global coupling restricts local anisotropic fluctuations, the authors introduce a marginal branch that predicts local covariances, which are then merged via a blockwise geodesic congruence (a matrix interpolation scheme).

Critical strengths
  1. Theoretical formalization of directional coupling. The paper formally shows that direct covariance predictors (which use one scalar per pair) are fundamentally limited in capturing direction-dependent coupling, and derives a lower bound for their error. This formalization is a valuable contribution to the community.

  2. Physically grounded inductive bias. Using the ENM Hessian to ensure a globally consistent, positive semidefinite covariance is a clever way to avoid the "black-box" nature of end-to-end deep learning models, providing a clear mechanical interpretation.

Critical weaknesses
  1. Marginal empirical gains do not justify the architectural complexity. While ENMulator outperforms classical ENMs, its performance against the strongest direct baseline, BackFlip-2, is mixed. In Table 2, ENMulator's RMSF correlation is 0.859, which is actually lower than BackFlip-2's 0.877, and its RMSF MAE is only marginally better. Given that the marginal branch was specifically designed to fix local anisotropic fluctuations, it is surprising that it does not clearly beat a pure direct predictor at this task. The complexity of the blockwise geodesic congruence seems unjustified by the empirical payoff.

  2. The evaluation is heavily reliant on post-hoc geometric relaxation and is narrowly scoped. First, the ensemble results (Table 3) apply a 60-step geometric relaxation to the Gaussian samples of all covariance-based methods. Gaussian sampling is intrinsically unphysical, so the final ranking might be dominated by the relaxation post-processing rather than the predicted covariances themselves. Second, the evaluation is entirely restricted to Cα-level dynamics on the ATLAS dataset. Claims about "globally coupled protein dynamics" feel overstated when side-chain dynamics and highly flexible/disordered regions are completely ignored.

  3. The "congruence" formulation may mask true biological ambiguity. By mathematically forcing the cross-residue canonical correlations to remain identical to the elastic base covariance, the model artificially decouples global coupling from local amplitude. In real proteins, a fluctuation in local residue flexibility often does alter the global coupling (e.g., local unfolding or hinge-bending motions). Forcing invariance here might impose an overly rigid constraint that limits the model's physical expressiveness.

Questions for authors
  1. The marginal branch underperforms BackFlip-2 on RMSF metrics. Could the authors provide a direct comparison where they replace BackFlip-2's marginal blocks with ENMulator's marginal outputs (keeping the same global coupling)? This would isolate whether the marginal branch itself is suboptimal or if the congruence constraint is restricting it.

  2. If the post-hoc geometric relaxation (Eq. 11) is ablated, how much of the ensemble-level improvement (Table 3) over BackFlip-2 remains? Is the relaxation masking the raw sampling differences?

  3. The lower bound for scalar coupling (0.182) assumes the whitened cross-covariance blocks have equal singular values. Does this assumption realistically represent the structural limits of DynaProt or BackFlip-2, or is it a mathematically convenient proxy?

3 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer 5R4z [1/2]

We appreciate the reviewer’s careful assessment and emphasis on two aspects central to our contribution: directional residue-pair coupling and the elastic-network inductive bias for globally consistent covariance prediction. We are encouraged that the reviewer found the mechanical interpretation and analysis of one-scalar-per-pair predictors valuable. We address the remaining concerns below.

[W1]

Marginal empirical gains do not justify the architectural complexity. … Given that the marginal branch was specifically designed to fix local anisotropic fluctuations, it is surprising that it does not clearly beat a pure direct predictor at this task. … The complexity of the blockwise geodesic congruence seems unjustified by the empirical payoff.

The added complexity is justified primarily by improved cross-residue coupling rather than RMSF: ENMulator is comparable to BackFlip-2 on RMSF, but consistently stronger on cross-residue covariance metrics (Table R1). Target-wise, the paired RMSF improvements are Δr=0.002\Delta r=0.002 [−0.008-0.008, 0.0130.013] and Δ\DeltaMAE =0.008=0.008 [−0.030-0.030, 0.0520.052], so neither metric differs consistently (Table A12). This is consistent with the role of the marginal branch, which calibrates the marginals of the elastic covariance rather than aiming to outperform a standalone direct predictor. Adding it to the learned-spring model improves RMSF rr from 0.819 to 0.859 (Table A14), while marginal calibration reduces RMSF MAE from 0.551 to 0.433 Å (Table A1). We clarified this role in Sec. 4.2, L237–240.

The cross-residue gains are consistent across targets:

Table R1. Target-wise improvements over BackFlip-2, with win rates and 95% CIs.

vs. BackFlip-2Win (%)Δ\Delta [95% CI]
Bures–Wasserstein ↓780.055 [0.039, 0.073]
Top-10 overlap ↑720.056 [0.038, 0.074]
DCCM rr ↑780.051 [0.036, 0.071]

The congruence itself improves BW from 0.755 to 0.666 and Rel. Fro. from 0.949 to 0.874 (Table A1). It is a closed-form 3×33\times3 transformation per residue with no additional learnable parameters and preserves canonical correlations exactly. With BackFlip-2's marginals, ENMulator's coupling still gives better BW (0.675 vs. 0.752) and DCCM rr (0.854 vs. 0.797; Table R3 in our response to Q1).

[W2]

The evaluation is heavily reliant on post-hoc geometric relaxation and is narrowly scoped. First, … the final ranking might be dominated by the relaxation post-processing rather than the predicted covariances themselves.

The performance differences are already present before geometric relaxation. Table 2 evaluates the covariances directly, and Table R2 shows that ENMulator retains its advantage over BackFlip-2 on both global and per-target DCCM before relaxation, with comparable RMSF.

Table R2. Comparison of protein dynamics metrics computed from generated conformations without geometric relaxation.

No relaxationANMHinsenDynaProtBackFlip-2ENMulator
Cα RMSF (=1.48)0.520.591.101.231.42
Per-target RMSF rr ↑0.750.800.860.870.86
Global DCCM rr ↑0.650.770.570.760.84
Per-target DCCM rr ↑0.690.810.650.800.86

Relaxation primarily repairs local geometry: the fraction of conformations with Cα–Cα breaks decreases from 100% to at most 3.2%, while the dynamics metrics change only modestly. We added the full before/after comparison in App. C.4 and Table A8.

Second, the evaluation is entirely restricted to Cα-level dynamics on the ATLAS dataset. Claims about “globally coupled protein dynamics” feel overstated when side-chain dynamics and highly flexible/disordered regions are completely ignored.

By "globally coupled," we refer specifically to the displacement covariance, where residue pairs are coupled through a shared Hessian. We narrowed the Abstract to "the globally coupled, Cα-level fluctuations of a protein around that structure" (L28–30), and moved the scope and limitations to Sec. 6 (L474–485), explicitly stating that ENMulator models harmonic fluctuations around a reference structure at Cα resolution and is trained primarily on ATLAS.

[W3]

The "congruence" formulation may mask true biological ambiguity. … In real proteins, a fluctuation in local residue flexibility often does alter the global coupling … Forcing invariance here might impose an overly rigid constraint that limits the model's physical expressiveness.

We agree that local residue flexibility can alter global coupling. In ENMulator, this effect is captured through the learned elastic interactions, so canonical-correlation invariance does not prevent changes in global coupling. Each residue-specific stiffness aia_i contributes to every spring involving residue ii, thereby modifying the Hessian and its coupling to other residues. Removing aia_i reduces DCCM rr from 0.856 to 0.824 and RMSF rr from 0.859 to 0.846 (Table A17). We clarified this role of aia_i in App. G.4 (L1499–1501).

AuthorsOpenReview ↗

Rebuttal for Reviewer 5R4z [2/2]

[Q1]

The marginal branch underperforms BackFlip-2 on RMSF metrics.

For the RMSF gap itself, see our response to W1.

Could the authors provide a direct comparison where they replace BackFlip-2's marginal blocks with ENMulator's marginal outputs (keeping the same global coupling)?

Yes, we swapped the marginal covariances and the cross-residue coupling between BackFlip-2 and ENMulator in both directions. The results are shown in Table R3. We added this swap experiment and its analysis in App. G.3 and Table A16.

Table R3. Effect of swapping marginal covariance predictions and cross-residue coupling between ENMulator and BackFlip-2.

CouplingMarginalBW ↓DCCM rr ↑RMSF rr ↑
BackFlip-2BackFlip-20.7520.7970.877
BackFlip-2ENMulator0.7610.7970.859
ENMulatorBackFlip-20.6750.8540.877
ENMulatorENMulator0.6660.8560.859

This would isolate whether the marginal branch itself is suboptimal or if the congruence constraint is restricting it.

The results do not indicate that either is the main limitation. The congruence does not restrict the marginals: with BackFlip-2’s marginals, ENMulator’s coupling recovers BackFlip-2’s RMSF rr (0.877). Nor is the marginal branch systematically suboptimal: the gap from 0.859 to 0.877 is between test-set medians, and target by target ENMulator’s RMSF rr is not lower (see W1).

In contrast, the global covariance metrics follow the coupling. With either set of marginals, ENMulator's coupling gives BW 0.666–0.675 and DCCM rr 0.854–0.856, compared with BW 0.752–0.761 and DCCM rr 0.797 for BackFlip-2's coupling. Thus, the observed performance gap is driven primarily by the learned cross-residue coupling rather than by the marginal predictions.

[Q2]

If the post-hoc geometric relaxation (Eq. 11) is ablated, how much of the ensemble-level improvement (Table 3) over BackFlip-2 remains?

ENMulator retains its advantage over BackFlip-2 on every dynamics metric reported in Table R4 without relaxation. We scored the same samples before and after relaxation.

Table R4. Effect of geometric relaxation on ensemble quality and protein dynamics metrics for ENMulator and BackFlip-2.

BackFlip-2 (no relax.)ENMulator (no relax.)BackFlip-2 (relax.)ENMulator (relax.)
Conf. with Cα–Cα breaks (%) ↓1001003.20.0
Conf. with clashes (%) ↓10032.40.00.0
Global DCCM rr ↑0.7560.8420.7610.845
Per-target DCCM rr ↑0.7970.8610.8040.863
Pairwise RMSD rr ↑0.8170.8310.8120.828
PC sim > 0.5 (%) ↑29.335.429.337.8
Transient contacts JJ ↑0.3050.3410.3070.353

Without relaxation, the DCCM gaps are nearly unchanged: the global DCCM advantage is +0.086 before relaxation versus +0.084 after relaxation. The improvements in PC similarity and transient contacts also remain substantial, retaining about three quarters of their relaxed gaps (+6.1 vs. +8.5 percentage points and +0.036 vs. +0.046, respectively). We added the full before/after comparison in App. C.4 and Table A8.

Is the relaxation masking the raw sampling differences?

No, the differences are already present in the raw samples. Relaxation primarily repairs local geometry: the fraction of conformations with Cα–Cα breaks decreases from 100% to at most 3.2%, and clashes are reduced to 0%. In contrast, the dynamics metrics change only slightly; for the metrics in Table R4, the change is at most 0.012, except for PC similarity, which changes by 2.4 percentage points. Thus, the improvements over BackFlip-2 are not introduced by the relaxation.

[Q3]

The lower bound for scalar coupling (0.182) assumes the whitened cross-covariance blocks have equal singular values. Does this assumption realistically represent the structural limits of DynaProt or BackFlip-2, or is it a mathematically convenient proxy?

It is a structural limit, not a proxy. For DynaProt and BackFlip-2, this is an exact structural property of their covariance reconstruction rather than a proxy. Both models predict one scalar ρij\rho_{ij} per residue pair and construct the cross-covariance as Σij=ρijLiLj⊤\Sigma_{ij}=\rho_{ij}L_iL_j^\top. After whitening by the marginal covariances, this gives s1=s2=s3=∣ρij∣s_1=s_2=s_3=|\rho_{ij}|, so the three singular values are necessarily equal.

Their actual predictions confirm this up to numerical round-off: over all residue pairs in the 82 test proteins, s1−s3s_1-s_3 is at most 1.3×10−61.3\times10^{-6} for BackFlip-2 and 3.1×10−73.1\times10^{-7} for DynaProt. Consistently, both models have a direction preference of 0.000 in Table 4. We added this derivation and numerical verification in App. F (L1338–1345).

Dear Reviewer 5R4z,

Thank you again for your thoughtful review and helpful comments on our paper.

We would like to briefly highlight the updates in our response. Following your comments, we quantified target-wise improvements with confidence intervals (Table R1), performed the requested bidirectional swaps of marginal covariances and cross-residue coupling (Table R3), and evaluated the sampled ensembles before and after geometric relaxation (Tables R2 and R4). These analyses help isolate the contribution of the learned coupling and assess whether the improvements depend on relaxation. We also clarified how residue-specific stiffness affects global coupling, verified the scalar-coupling restriction analytically and numerically (Appendix F), and narrowed our claims to harmonic, Cα-level fluctuations.

We sincerely appreciate the time and effort you have already devoted to reviewing our paper. If time permits before the discussion period closes, we would be grateful if you could read through our response. We would welcome any further thoughts and would be happy to clarify any remaining questions.

Best regards, The Authors

Reviewer_eq4kOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a new look into a harmonic analysis of protein motions. The paper applies the harmonic approximation to protein trajectories, trains a model and compares the results to some physics-based approaches and also two recent DL models. The paper, however, does not discuss the limitations and inaccuracies of the harmonic approximation or the molecular dynamics and force fields used to generate the data. Technically, it combines a DynaProt-style local covariance prediction with also predicting stiffness and potential of springs. It provides some extensive validation, but uses only a single benchmark. More MD trajectories (e.g. MISATO), at least for validation, will strengthen the claims. Experimental conformations collected from the PDB will do even better validation.

Critical strengths

-- The paper provides an extensive list of physics-based baselines and shows that some of them (e.g., Hinsen's model) perform better than the previous DL-based SOTA methods, which questions rigour of assessment and problem formulation.

-- The paper presents a useful convex interpolation technique, referred here as "geodesic congruence", though without proper introduction or analysis.

Critical weaknesses

-- The evaluation is rather limited. More MD benchmark and some PDB samples will make the claims stronger. Generative techniques, like AlphaFlow or BBFlow, were not assessed on the main metrics.

-- The low-rank nature of the Hessian and covariance is not discussed. As a consequence, no mention of comparison is provided to the low-rank SE3-equivariant predictors, such as PETIMOT. The C-alpha approximation also limits the strength and capacity of the approach.

Questions for authors

-- One of the common uses of ENM models in structural bioinformatics is the fact that correlation matrices and motion components are essentially low-rank. This allows to efficiently compute low-rank pseudo-harmonic inverses, study PCA and NMA motions, and approximate motion covariances with just a few principal components. It also makes sense instead of learning local covariances and the full-size Hessian, as it is done in ENMulator, to learn only a few principal components of the covariance of the Hessian matrix, as it was done in the SeaMoon (https://scholar.google.com/scholar?cluster=8893814279311351016&hl=en&as_sdt=2007) and PETIMOT (https://scholar.google.com/scholar?cluster=16118281374142803448&hl=en&as_sdt=2007) methods. Please discuss the low-rank covariance approximation as objective and compare your results with those from PETIMOT.

-- Connection between eq 7 and eq 6 must be explained better. Probably, it makes sense to first introduce geodesic in eq 7, and then as a consequence you get this intermediate representation Di in eq6. Otherwise, eq 6 looks very superficial.

-- The actual question is why eq.7 represents a geodesic? Why do you call this alpha-interpolation a "GEODESIC CONGRUENCE"? how does it relate to the one from general relativity???

-- Please also include alpha=0 to the table A10.

-- I would also appreciate if the authors share the weights of the model, so we can confirm the presented results.

-- AlphaFlow and BBFlow seem to be very strong baselines. I would appreciate them to be included in Table 2... Maybe you can reduce the number of produced conformations?

-- Can you check the stability of the method wrt to the chosen reference conformation? What happens if you choose a different reference in the test?

-- Each ATLAS protein trajectory consists of 3 replicas. Please explain how these replicas were used. Please also compute metrics in Table 2 between 3 replicas in the test set - this will give you some expected standard deviation.

-- Please include limitations to the main text. You say "The model may not capture multimodal conformational landscapes or strongly anharmonic transitions" - it certainly cannot by design, please change the formulation.

-- Please provide a reference that explains how the harmonic approximation connects covariance and Hessian.

-- When using ANM and Hinsen analytical methods, how do you rescale to the MD temperature (the absolute amplitude)? Does your correlation metric is absolute amplitude (temperature) sensitive?

-- Why ANM covariances in Figure A4 are mostly white for a good half of proteins? Can you explain this behaviour?

-- "In an elastic network, how much and in which directions each residue fluctuates is not a free choice" - generally, this statement is not fully correct, as you have as many degrees of freedom in the Hessian, as the number of its principal components. In principle, one can learn some eigen-decomposition of the Hessian or the covariance, which guarantees independent fluctuations.

-- The Cholesky decomposition choice of the objective is not explained. It appears out of nowhere, and at least references to the previous works that used the Cholesky decomposition to learn fluctuations must be provided

-- eq 4 adds and subtracts the null-space with a unit eigenvalues. Is it optimal? Shall you try to add and subtract it with a smaller epsilon? Can you prove that the rigid-body variance is "negligible"??? In principle, it must be trivial to fully remove it, for example using an SVD/igenvector-decomposition-based pseudoinverse.

-- It will be very informative to actually analyse the learned stiffness constants and their distance behaviour. Does it resemble any physics-based models? is is similar to the one from Hinsen?

-- Finally, the Ca approximation used looks like a severe limitation that is not even fully compatible with the IPA geometric transformer and the relative geometry of one residue with respect to another using

8 comments
AuthorsOpenReview ↗

Rebuttal for Reviewer eq4k [1/4]

We thank the reviewer for the constructive review. We are encouraged that the reviewer found our evaluation against physics-based baselines informative and our geodesic congruence useful. We address each concern and question below.

[W1]

The evaluation is rather limited. More MD benchmark and some PDB samples will make the claims stronger.

We added both evaluations, reported in Tables R1 and R2, respectively.

MD benchmark. We evaluated 300 single-chain complexes from the MISATO validation/test splits, after fine-tuning ENMulator on 829 MISATO training complexes without homologs of the test set. ENMulator achieves the lowest BW and the highest DCCM and RMSF correlations, and its top-10 overlap is tied with Hinsen (0.626 vs. 0.625).

Table R1. Comparison of protein dynamics modeling methods on MISATO. BW uses per-target oracle scaling as in Table A20.

MethodBW ↓DCCM rr ↑Top-10 overlap ↑RMSF rr ↑
ANM0.8810.7630.5550.765
Hinsen0.7590.8230.6250.808
ENMulator0.7210.8260.6260.827

We added this benchmark in Appendix H.1 and Table A26.

PDB samples. We evaluated 713 proteins from the PDB conformational collections used in PETIMOT, excluding homologs of the ATLAS training/validation sets, and fine-tuned ENMulator on the PETIMOT training set. ENMulator improves over classical NMA on every metric. PETIMOT, which is trained specifically for this benchmark, achieves the highest success rate and the lowest Min LS.

Table R2. Comparison of protein motion prediction methods on experimentally observed PDB conformational collections.

MethodSuccess rate ↑Min LS ↓Min magnitude ↓OLA LS ↓OLA magnitude ↓
NMA26.1%0.700.250.850.39
ENMulator31.3%0.660.210.830.36
BioEmu32.0%0.680.230.860.42
AlphaFlow32.5%0.680.240.860.44
PETIMOT42.6%0.620.220.840.41

We added this benchmark in Appendix H.2 and Table A27, and updated the Conclusion and Reproducibility Statement accordingly.

Generative techniques, like AlphaFlow or BBFlow, were not assessed on the main metrics.

We added AlphaFlow and BBFlow to Table 2, estimating the covariance from their generated conformations. Table 2 reports 250 conformations, and Table R3 also compares 50.

Table R3. Comparison of ENMulator with generative models on covariance-based protein dynamics metrics.

ConformationsBW ↓DCCM rr ↑Top-10 overlap ↑RMSF rr ↑
ENMulator— (one forward pass)0.6660.8560.6630.859
AlphaFlow500.6660.8560.7120.912
AlphaFlow2500.5790.8820.7560.921
BBFlow500.6860.8550.7100.893
BBFlow2500.6310.8610.7470.901

With 250 conformations, AlphaFlow and BBFlow are more accurate, but at far greater cost (5,149 s and 172 s vs. 7.8 ms for ENMulator). ENMulator matches AlphaFlow at 50 conformations and leads the covariance-based methods (see Q6).

[W2]

The low-rank nature of the Hessian and covariance is not discussed.

We added this discussion in Sec. 3, L189–190. Σ0=H+\Sigma_0=H^{+} is effectively low-rank because most of its variance is concentrated in a few soft modes. ENMulator has a participation ratio of 8.98, closer to MD (5.48) than BackFlip-2 (17.17) and DynaProt (40.69).

As a consequence, no mention of comparison is provided to the low-rank SE3-equivariant predictors, such as PETIMOT.

We added PETIMOT evaluation on the ATLAS test set with its own metrics (4 modes; Table R4). ENMulator is best on every metric:

Table R4. Comparison with low-rank protein motion predictors on the ATLAS test set using PETIMOT evaluation metrics.

MethodSuccess rate ↑Min LS ↓Min magnitude ↓OLA LS ↓OLA magnitude ↓Global SS ↓
Hinsen74.4%0.4930.1600.7300.3310.557
BackFlip-261.0%0.5330.1850.7410.3650.583
PETIMOT72.0%0.5140.1690.7250.3070.567
ENMulator75.6%0.4500.1270.6790.2630.480

On the 50 ATLAS test proteins with no homolog in the PETIMOT training set, ENMulator retains its advantage (Min LS: 0.465 vs. 0.532 for PETIMOT). We added these results in App. H.3, Table A28. We also revised Sec. 2 to include PETIMOT and other low-rank motion predictors.

The C-alpha approximation also limits the strength and capacity of the approach.

The Cα approximation applies only to the output. The IPA encoder takes N-Cα-C backbone frames, and the covariance is defined at Cα resolution. We removed the phrase “with Cα coordinates” from the input description in Sec. 4, L197 to avoid suggesting a Cα-only input, and state the Cα-resolution covariance as a limitation in Sec. 6.

AuthorsOpenReview ↗

Rebuttal for Reviewer eq4k [2/4]

[Q1]

One of the common uses of ENM models in structural bioinformatics is the fact that correlation matrices and motion components are essentially low-rank. This allows to efficiently compute low-rank pseudo-harmonic inverses, study PCA and NMA motions, and approximate motion covariances with just a few principal components.

ENMulator supports these. The eigendecomposition of its predicted covariance provides principal motion modes, which we evaluate through top-3/top-10 subspace overlap in Table 2 and against PETIMOT in Table R4 (see W2).

It also makes sense instead of learning local covariances and the full-size Hessian, as it is done in ENMulator, to learn only a few principal components of the covariance of the Hessian matrix, as it was done in the SeaMoon and PETIMOT methods.

ENMulator in fact predicts fewer numbers per protein than directly predicting a few principal components. The 3N×3N3N\times3N Hessian is assembled from one learned stiffness aia_i per residue (Eq. 3), while the local covariances add one 3×33\times3 matrix per residue (Eq. 5), corresponding to 7N7N predicted values in total compared with 12N12N for four 3D principal components. From this compact parameterization, Σ0=H+\Sigma_0=H^{+} yields the full covariance and its leading principal components.

Please discuss the low-rank covariance approximation as objective and compare your results with those from PETIMOT.

We additionally trained this alternative with the same encoder, directly predicting a rank-64 PSD covariance. As shown in Table R5, this alternative underperforms ENMulator across all four metrics. We added these results in Appendix G.3 and Table A15. For comparison with PETIMOT, see our response to W2.

Table R5. Comparison of rank-64 covariance prediction and ENMulator using the same encoder.

Rank-64 covarianceENMulator
Rel. Fro. ↓0.9170.874
BW ↓0.8300.666
Top-10 overlap ↑0.4750.663
DCCM rr ↑0.7350.856

[Q2]

Connection between eq 7 and eq 6 must be explained better. Probably, it makes sense to first introduce geodesic in eq 7, and then as a consequence you get this intermediate representation Di in eq6. Otherwise, eq 6 looks very superficial.

We reordered Sec. 4.2, L245–256 as suggested. In the revised numbering, the geodesic interpolation Pi#αMiP_i \#_\alpha M_i is defined in Eq. 6, and DiD_i is then defined in Eq. 7 to satisfy DiPiDi⊤=Pi#αMiD_i P_i D_i^\top = P_i \#_\alpha M_i.

[Q3]

The actual question is why eq.7 represents a geodesic? Why do you call this alpha-interpolation a "GEODESIC CONGRUENCE"? how does it relate to the one from general relativity???

Eq. 7 (Eq. 6 in the revision) is exactly the point at fraction α\alpha on the geodesic from PiP_i to MiM_i under the affine-invariant metric on SPD matrices, γ(t)=P1/2(P−1/2MP−1/2)tP1/2\gamma(t)=P^{1/2}(P^{-1/2}MP^{-1/2})^tP^{1/2} (Pennec et al., 2006; Bhatia, 2009). "Geodesic congruence" names the two steps: each marginal moves along this geodesic, and the full covariance follows by the congruence Σ↦DΣD⊤\Sigma\mapsto D\Sigma D^\top. It shares with general relativity only the notion of a geodesic as the shortest path under a metric; here the space is 3×33\times3 SPD matrices. We added these definitions and references in Sec. 4.2 (L245–260).

[Q4]

Please also include alpha=0 to the table A10.

We added α=0\alpha=0 (Table R6):

Table R6. Effect of the geodesic congruence interpolation parameter α\alpha on protein dynamics prediction.

αRel. Fro. ↓Trace ratio (→1)Global DCCM rr ↑Global RMSF rr ↑Per-target RMSF rr ↑Pairwise RMSD rr ↑
00.9491.0390.8410.3730.8300.387
0.750.8740.7610.8450.8250.8730.832

α=0.75\alpha=0.75 is better on five of six metrics. The median trace ratio of α=0\alpha=0 hides large per-protein errors (up to 375, vs. below 1.87 at α=0.75\alpha=0.75). We added this row to Table A10 (Table A13 in the revision).

[Q5]

I would also appreciate if the authors share the weights of the model, so we can confirm the presented results.

Yes, we will publicly release the trained model weights upon acceptance, as stated in App. A.

[Q6]

AlphaFlow and BBFlow seem to be very strong baselines. I would appreciate them to be included in Table 2... Maybe you can reduce the number of produced conformations?

We added both methods to Table 2 with 250 conformations, and Table R3 in W1 also reports 50. With 50, their BW rises from 0.579 to 0.666 (AlphaFlow) and from 0.631 to 0.686 (BBFlow). For 6kty_A, 250 conformations take 5,149 s (AlphaFlow) and 172 s (BBFlow), while ENMulator takes only 7.8 ms. We added the results to Table A23 and the runtimes to App. D.4 of the revision.

AuthorsOpenReview ↗

Rebuttal for Reviewer eq4k [3/4]

[Q7]

Can you check the stability of the method wrt to the chosen reference conformation? What happens if you choose a different reference in the test?

Yes, ENMulator is stable for references within 1.5 Å of the chosen reference conformation (Table R7).

Table R7. Effect of alternative reference conformations within 1.5 Å of the chosen reference conformation.

ReferenceRel. Fro. ↓BW ↓DCCM rr ↑Top-10 overlap ↑RMSF rr ↑
Original0.8720.6620.8580.6610.850
Alternative MD frame0.8730.6670.8570.6560.859

Accuracy decreases with distance from equilibrium: beyond 4 Å, DCCM rr drops by 0.071 and BW rises by 0.125, as expected for a harmonic model. We added this analysis to App. G.8 and Table A24, and the limitation to Sec. 6.

[Q8]

Each ATLAS protein trajectory consists of 3 replicas. Please explain how these replicas were used.

We pool the three replicas (3,000 frames) into one target covariance, following AlphaFlow and DynaProt. We clarified this in Sec. 5.1 (L346–349) and App. C.1 (L897–899).

Please also compute metrics in Table 2 between 3 replicas in the test set - this will give you some expected standard deviation.

We conducted this comparison, scoring each prediction against single MD replicas (Table R8):

Table R8. Comparison of model predictions and inter-replica agreement on the ATLAS test set using individual MD replicas as reference.

PredictionBW ↓DCCM rr ↑Top-10 overlap ↑RMSF rr ↑
Another MD replica0.7030.8390.7170.909
ENMulator0.7160.8530.6570.847
BackFlip-20.8240.7840.5970.859

Across target replicas, the standard deviation of the median is 0.011 (DCCM rr) and 0.006 (BW), well below the gaps to BackFlip-2. We added this analysis to App. G.9 and Table A25.

[Q9]

Please include limitations to the main text. You say "The model may not capture multimodal conformational landscapes or strongly anharmonic transitions" - it certainly cannot by design, please change the formulation.

We moved the limitations to Sec. 6, L474–485. Also, we revised the statement in Sec. 6 (L478–479) to: "By design, the model cannot capture multimodal conformational landscapes or strongly anharmonic transitions."

[Q10]

Please provide a reference that explains how the harmonic approximation connects covariance and Hessian.

Under the harmonic approximation, p(Δx)∝exp⁡(−Δx⊤HΔx/2kBT)p(\Delta x)\propto\exp(-\Delta x^\top H\Delta x/2k_BT) implies Cov(Δx)=kBT H+\mathrm{Cov}(\Delta x)=k_BT\,H^{+}. We added these references (Atilgan et al., 2001; Cui & Bahar 2005) right after Eq. 2 (L160–161).

[Q11]

When using ANM and Hinsen analytical methods, how do you rescale to the MD temperature (the absolute amplitude)?

We fit one global spring constant per baseline on the training split, log⁡c∗=median[log⁡(tr Σγ/tr ΣMD)]\log c^*=\mathrm{median}[\log(\mathrm{tr}\,\Sigma_\gamma/\mathrm{tr}\,\Sigma_{\mathrm{MD}})], and fix it for all test proteins. We clarified this in App. D.1 (L1010–1021) and report the values in Table A9.

Does your correlation metric is absolute amplitude (temperature) sensitive?

No, our correlation metrics (DCCM rr, RMSF rr) and subspace overlap are invariant to global covariance scaling. Only the distance metrics (Rel. Fro., BW, RMSF MAE) are scale-sensitive. We clarified this distinction in App. E.2.

[Q12]

Why ANM covariances in Figure A4 are mostly white for a good half of proteins? Can you explain this behaviour?

The figure shows raw, unscaled covariances, and ANM's median trace is only 0.13× that of MD, a known underestimation of amplitudes by ANM (Koehl et al., 2021; Dubanevics & McLeish, 2022). We added this to the caption of Figure A4 (Fig. A5 in the revision; L1723–1725).

[Q13]

"In an elastic network, how much and in which directions each residue fluctuates is not a free choice" - generally, this statement is not fully correct, as you have as many degrees of freedom in the Hessian, as the number of its principal components. In principle, one can learn some eigen-decomposition of the Hessian or the covariance, which guarantees independent fluctuations.

A learned eigendecomposition can indeed parameterize the fluctuations independently. We revised the sentence in Sec. 4.2 (L237–240) to: "In an elastic network, one Hessian determines both the local fluctuation of each residue and the global coupling."

[Q14]

The Cholesky decomposition choice of the objective is not explained. It appears out of nowhere, and at least references to the previous works that used the Cholesky decomposition to learn fluctuations must be provided

We use a Cholesky factor to parameterize each marginal covariance, which guarantees a positive-definite covariance for arbitrary network outputs, following DynaProt (Bafna et al., 2026). We added this motivation and the reference in Sec. 4.1, L228–230 and App. B.5, L870–871.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the AC and all reviewers for their time and constructive feedback. Reviewers highlighted the physically grounded inductive bias and theoretical formalization of directional coupling (5R4z), the useful geodesic congruence and extensive physics-based baselines (eq4k), and strong empirical validation (nKRM). Below, we summarize the main concerns and how we addressed them.

  1. Contribution of the elastic-network inductive bias (nKRM, 5R4z, eq4k)

    • Concern: Whether the elastic-network inductive bias drives the improvements and whether the added complexity is justified.
    • Response: We added the matched rank-64 PSD baseline with the same encoder and supervision (App. G.3, Table A15). It underperforms both learned springs and full ENMulator across all four metrics. We also swapped the marginal covariances and the cross-residue coupling between BackFlip-2 and ENMulator in both directions. The observed performance gap is driven primarily by the learned cross-residue coupling rather than by the marginal predictions. With either set of marginals, ENMulator’s coupling gives lower BW and higher DCCM correlation than BackFlip-2’s coupling (Table A16).
  2. Reference-centered covariance and geometric relaxation (nKRM, 5R4z)

    • Concern: Whether the improvements persist under mean-centered covariance evaluation and without geometric relaxation.
    • Response: We added the mean-centered evaluation (App. C.1, Table A3). The improvements persist when all methods are evaluated against the mean-centered MD covariance. We also scored the same samples before and after relaxation (App. C.4, Table A8). The global DCCM advantage over BackFlip-2 is +0.086 before relaxation versus +0.084 after relaxation. Relaxation primarily repairs local geometry.
  3. Broader evaluation and missing baselines (eq4k, 5R4z)

    • Concern: More MD benchmarks, experimental PDB conformations, and comparisons with generative and low-rank motion predictors.
    • Response: We added evaluations on MISATO and experimental PDB conformational collections, after fine-tuning on the respective training sets (App. H.1–H.2, Tables A26–A27). We added PETIMOT evaluation on the ATLAS test set with its own metrics (App. H.3, Table A28). We also added AlphaFlow and BBFlow to Table 2, estimating the covariance from their generated conformations. We added the results to Table A23 and the runtimes to App. D.4.
  4. Reference sensitivity and MD replicas (eq4k)

    • Concern: Stability with respect to the chosen reference conformation and variability across the three MD replicas.
    • Response: We added both analyses (App. G.8–G.9, Tables A24–A25). We clarified that the three replicas (3,000 frames) are pooled into one target covariance (Sec. 5.1, App. C.1). Across target replicas, the standard deviation of the median is 0.011 (DCCM) and 0.006 (BW), well below the gaps to BackFlip-2.
  5. Mathematical formulation and physical interpretation (eq4k, 5R4z)

    • Concern: The geodesic formulation, scalar-coupling lower bound, and effect of local residue flexibility on global coupling.
    • Response: We added definitions and references for the geodesic under the affine-invariant metric on SPD matrices (Sec. 4.2), the α = 0 row (Table A13), and the justification and numerical analysis of the pseudoinverse (App. B.4, Table A2). For DynaProt and BackFlip-2, the scalar-coupling restriction is an exact structural property of their covariance reconstruction; we added the derivation and numerical verification in App. F. The effect of local residue flexibility on global coupling is captured through the learned elastic interactions. The learned distance kernel alone reaches the same DCCM as Hinsen (0.824), while adding the residue-specific terms increases it to 0.856 (App. G.4, Table A17).
  6. Scope of “globally coupled protein dynamics” (eq4k, 5R4z)

    • Concern: The scope of the harmonic approximation and Cα-level dynamics.
    • Response: We clarified in the Abstract and Sec. 6 that ENMulator models harmonic fluctuations around a reference structure at Cα resolution. By “globally coupled,” we refer specifically to the displacement covariance, where residue pairs are coupled through a shared Hessian. The IPA encoder takes N–Cα–C backbone frames (Sec. 4).

We thank the AC and reviewers again. The additional experiments and revisions have strengthened the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.