acceptodds
Under review as a conference paper at ICLR 2027

Steric Hamiltonian Flows for Proteins

21 views

Abstract

Proteins are central to nearly all biological processes, and their diverse functions arise from their complex 3D structures. Recent advances in deep learning have transformed this field, with diffusion-based generative models revolutionizing protein design by enabling the creation of novel protein structures. However, these methods lack physical inductive biases and neglect the intrinsic physical realism of proteins, driven by noising dynamics that lack grounding in physical principles. To address this, we first introduce a physically motivated structured noising process, grounded in classical physics, that unfolds proteins into secondary structures (e.g., -helices, linear -sheets) while maintaining bonds and preventing collisions. We integrate this process with flow-matching to model protein backbone distributions while incorporating sequence information for conditional folding. Our method outperforms baselines trained on the same dataset and matches those trained on substantially larger ones across short- and long-chain generation, and motif-scaffolding tasks.

Primary area:
generative models
Keywords:
flow matching, generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_366VOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose a new flow based generative model for protein structures called StericFlow. The method introduces a physically structured noising process that destroys the protein structure into secondary structures like beta sheets and alpha helices and argues that these physically meaningful noising processes are beneficial. StericFlow is evaluated extensively against existing works on canonical protein structure generation benchmarks and performs comparably.

Critical strengths
  1. Paper is very well written and reads like a textbook.

  2. The idea of physically structured noising to secondary structures is elegant.

Critical weaknesses
  1. Hasn’t this idea of physically constrained residue connectivity already been proposed in previous methods like EigenFold [1] (which the authors compare against in the experiments) and other notions of a “harmonic” diffusion/prior [2,3]? I don’t see any real distinction in terms of methodology (eq. 5 and 6 are just specific choices of E(x) in eq. 1 and 2 of [1]), which is the primary reasoning for my weak reject rating, but am happy to raise my score if I am missing something.

  2. The practical benefits of such a physically constrained noising process used in StericFlow is not obvious (roughly the same performance as best baselines in Table 2).

  3. There’s a disconnect between the pitch of the work and the actual method. StericFlow states that it aims to noise/unfold protein structures into secondary structures. This pitch leads the reader to believe that the final, most noised structure should have sensical secondary structures (a mix of both alpha helices and beta sheets). However, from eq. 6 and 7, there’s no clear reason why we would arrive at such meaningful secondary structures outside of just a linear “harmonic” like prior, similar if not identical to [1]. The authors note of a alpha helical variant in Line 265, but I can’t seem to find the details of this in the main section or the Appendix.

[1] Jing, et al. “EigenFold: Generative Protein Structure Prediction with Diffusion Models.” ICLR MLDD (2023).

[2] Stark, et al. “Harmonic Prior Self-conditioned Flow Matching for Multi-Ligand Docking and Binding Site Design.” ICML (2024).

[3] Jing, et al. “AlphaFold Meets Flow Matching for Generating Protein Ensembles.” ICML (2024).

Questions for authors
  1. Can the authors provide more intuition for the Look-Ahead loss in Section 5? Is this not a loss to enable fewer step generation? Why is this helpful at all for the purposes of generation quality or physically meaningful learning?
2 comments

Thanks for your thoughtful review. We have carefully addressed your concerns and have outlined our responses to the key points raised in your review below.

Regarding the Distinction from EigenFold and Harmonic Priors

Thank you for this suggestion. We have carried out a detailed comparison with EigenFold and with the harmonic-prior formulations of [2] and [3], and we will add this comparison, to Section 4.2 and the related work. The differences are shown in Table below.

AspectHarmonic diffusion / prior [1,2,3]StericFlow (ours)
Order of the dynamicsFirst-order, dx=−∇E(x)dt+2dWdx = -\nabla E(x)dt + \sqrt{2}dWSecond-order damped Hamiltonian in (z,v)(z,v)
Markov propertyMarkov in xx; generator fixed by EE aloneNon-Markov in zz; z¨t≠0\ddot z_t \ne 0 carries the potential
Form of the energyQuadratic, required for a closed-form processHarmonic attraction plus a singular 1/r1/r steric barrier
CoordinatesCartesian residue positionsInternal backbone angles, with J(z)⊤J(z)^{\top} pullback
Terminal distributionGaussian N(0,Σ)\mathcal{N}(0,\Sigma), Σ−1\Sigma^{-1} the chain LaplacianAttractor at an extended, clash-free chain with variance
Non-local stericsNot represented; blind to sequence-distant contactsExplicit pairwise repulsion between all residue pairs

Two fundamental differences distinguish StericFlow from these formulations. First, its dynamics evolve in phase space, with z˙=v\dot z=v and v˙=−∇zU(z)−γv\dot v=-\nabla_z U(z)-\gamma v, making the structural trajectory non-Markovian in zz and not directly expressible through a first-order energy-based diffusion. Second, the singular, non-quadratic 1/r1/r repulsion precludes the closed-form Gaussian marginals available with quadratic potentials, necessitating simulated flow-matching targets rather than simulation-free training.

The key benefit is a provable minimum inter-residue distance: since H\mathcal{H} is a Lyapunov function, min⁡i≠j∥xi,s−xj,s∥≥k2/H0\min_{i\ne j}\|x_{i,s}-x_{j,s}\|\ge k_2/\mathcal{H}_0 for all s≥0s\ge0. Unlike harmonic priors, which only encourage chain connectivity and assign positive probability to clashes at every intermediate time, StericFlow guarantees collision-free trajectories by construction. This distinguishes a soft structural prior from a constraint on the process support.

We further compare against the k2=0k_2=0 variant, which removes steric repulsion while retaining the same architecture and evaluation protocol, isolating the contribution of the steric term.

Forward processDesignabilityNovelty (TM-Frac)avg.maxTMDiversity (Clusters)TM-Score
EigenFold0.710.130.470.33 (142)0.27
Linear OU0.930.140.440.35 (164)0.25
StericFlow (k2=0k_2=0)0.940.160.410.49 (210)0.25
StericFlow (β\beta)0.970.190.380.50 (220)0.22
StericFlow (α\alpha)0.980.170.400.52 (228)0.23

The steric term improves designability from 0.94 to 0.975 when pooling both variants (n=500n=500; one-sided p=0.017p=0.017), compared with Linear OU (p=0.005p=0.005). More importantly, it reduces clashing-pair rates to 0.03%, compared with 0.5% for FoldFlow and 1.3% for FrameDiff, a 17-fold reduction (Table 11), with fewer clashes at every reverse step (Figure 8).

Regarding the Practical Benefits of the Constrained Noising Process

We agree that designability margins are modest, partly because performance is already near saturation (0.93–0.98). However, StericFlow (β\beta) outperforms FoldFlow-OT across all five metrics simultaneously, including designability (0.97 vs. 0.94), novelty (0.19 vs. 0.18), and diversity (220 vs. 178 clusters), with a sign test of p=0.031p=0.031. Importantly, it achieves these gains without the usual trade-off between designability and diversity. StericFlow also produces substantially more β\beta-sheet content at comparable designability: 7.7% vs. 1.2–3.2% for the strongest baselines, approximately 2.4-fold higher. The benefits extend to more challenging settings, where StericFlow leads on all three long-chain metrics (Table 3), achieves 1404 motif-scaffolding successes versus 1099 and 889, and has the lowest clash and peptide-break rates (Table 4).

More broadly, our contribution is to establish the forward process as an additional design axis for protein generation. We provide the theoretical analysis and methodological components needed to use physically informed paths, including simulation-based flow-matching targets, a curvature-aware objective, and the J(z)⊤J(z)^\top pullback through reconstruction. The steric potential is one example; other differentiable potentials, including hydrophobic, hydrogen-bonding, and experimental restraints, can be incorporated within the same framework, at a one-time cost of under one second per 200-residue protein.

Regarding the Terminal State and the α\alpha-Helical Variant

We apologize for the ambiguity in our terminology. The forward process produces a single extended, clash-free chain with either β\beta-like or canonical α\alpha-helical backbone dihedrals, rather than a mixture of folded secondary-structure elements. We will revise it accordingly. We will also add an appendix subsection specifying μα\mu_\alpha and the terminal configurations and forward clash profiles for both variants. The α\alpha variant differs only in the target mean μ\mu, with all other parameters unchanged. Notably, the two variants achieve similar designability (0.98 and 0.97), supporting the conclusion that the gains arise primarily from the probability path and steric barrier rather than the choice of terminal prior.

Regarding the Intuition for the Look-Ahead Loss

The motivation for the look-ahead loss is to account for the curvature of our probability path; robustness to inference steps is a consequence. For a linear path, z¨t=0\ddot z_t=0, so matching the velocity at each time is sufficient. Our path instead has z¨t=−∇U(zt)−γvt≠0\ddot z_t=-\nabla U(z_t)-\gamma v_t\ne0, meaning pointwise velocity matching captures only the local tangent and does not account for accumulated trajectory errors. The look-ahead loss addresses this by comparing predicted and target atom positions at future times s≥ts\ge t (Eq. 10), directly supervising the integrated trajectory rather than only its instantaneous velocity.

This also benefits generation quality: the loss supervises all four backbone atoms at intermediate times, transferring the clash-free path properties to the learned model, while coupling rotational and translational predictions through ζframe-to-atom\zeta_{\text{frame-to-atom}}. Consistent with this, designability remains stable across 50–200 inference steps (Table 9). The ablation without LLA\mathcal{L}_{\mathrm{LA}} is ongoing and will be reported during the discussion.

Reviewer_w55bOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces StericFlow, a protein backbone generative model with a physics-inspired forward process. The forward process unfolds proteins using a damped second-order ODE in backbone angle space, with attraction toward an extended structure and steric repulsion in Cartesian space. The simulated trajectories are then reversed using an SE(3) flow-matching model.

The method is evaluated on unconditional generation, sequence-conditioned folding, and motif scaffolding. The paper also provides theoretical results on invariance and convergence of the forward dynamics.

Critical strengths
  1. Using a structured, physics-inspired forward process instead of generic SE(3) noise is intuitive and well motivated.
  2. The paper implements the full pipeline and evaluates it on unconditional generation, sequence-conditioned folding, and motif scaffolding.
  3. Among PDB-only models, StericFlow shows high designability (0.94–0.98), good novelty and diversity, and slightly better long-chain generation.
Critical weaknesses
  1. The physics-based trajectory is used only for the R3\mathbb{R}^3 translation part. Therefore, the intermediate structures are not necessarily valid full backbones. The claim that the method “maintains bonds” is mainly true for the CαC_\alpha trace, not the full N–Cα–CN–C_\alpha–C backbone.

  2. The model does not use velocity vv as input. The second-order dynamics are only used to generate the training path, while the learned generative process is still first-order. This makes the “Hamiltonian” framing somewhat misleading.

  3. The benefit of using angular space is also unclear, since the repulsion, losses, and network mainly operate in Cartesian/frame space. Proposition 1 is straightforward and provides little support for the method.

  4. The harmonic potential treats angles as ordinary Euclidean variables and ignores their periodicity. This is especially problematic for ω\omega, whose target is set near +π+\pi. An angle near −π-\pi is physically very close to +π+\pi, but the current loss treats them as almost 2π2\pi apart.

  5. The results of CAMEO2022 results in Table 4 are weak for a folding task and do not support the claim that the method can "unify structure generation and folding", with TM-score=0.6 and RMSD=12.11A. A related concerns is joint training on conditional and unconditional objectives may hurt the conditional folding task. Maybe authors can test a conditional-only model first.

Questions for authors

The same as Weaknesses.

2 comments

Thanks for your thoughtful review. We have addressed your concerns and outlined our responses to the key points raised in your review below.

Regarding the Informativeness of the Forward Process

We would first like to state a result that was not made explicit in the submission, as it bears directly on this assessment. The Hamiltonian H\mathcal{H} is a Lyapunov function satisfying H˙=−γ∥v∥2≤0\dot{\mathcal{H}} = -\gamma \|v\|^2 \leq 0. Since all terms in H\mathcal{H} are non-negative, each term remains bounded by its initial value H0=12∥v0∥2+U(z0),\mathcal{H}_0 = \frac{1}{2}\|v_0\|^2 + U(z_0), along the entire trajectory. This yields the following uniform bounds for all ≥0\geq 0, $ \min_{i \neq j}|x_{i,s}-x_{j,s}| \geq \frac{k_2}{\mathcal{H}_0}$, and

\leq \sqrt{\frac{\mathcal{H}_0}{k_1}}.$$ The first bound establishes a strictly positive lower bound on residue separation, while the second uniformly bounds bond-angle excursions over time. Thus, the forward dynamics are clash-free by construction, rather than merely clash-reduced. We emphasize that the distinction from Gaussian constructions concerns the support of the distribution, not parameter tuning. A Gaussian kernel assigns positive probability density to every configuration, including clashing configurations, at every intermediate time. In contrast, our dynamics exclude such configurations through the separation bound. Table 11 further shows that this property is largely preserved in the generated samples, with only 0.03% clashing pairs, compared with 0.5% for FoldFlow and 1.3% for FrameDiff. We have added this result to the main text as Proposition 3. **Regarding the Angular Parametrization and Proposition 1** The motivation for angular coordinates is constraint elimination rather than invariance. Bond lengths and chain connectivity are encoded by construction in the angular representation, eliminating the need to explicitly enforce these constraints during integration. In contrast, Cartesian dynamics require constraint solvers such as SHAKE or RATTLE, or stiff penalty potentials, which can introduce restrictive step-size requirements. Moreover, the resulting bound on angular deviations (the second inequality above) is directly expressed in the dynamical coordinates, whereas bond angles are nonlinear functions of Cartesian coordinates. Appendix D.2 derives the Cartesian variant, and we now clarify this motivation in the manuscript. Importantly, steric repulsion and losses are evaluated in Cartesian and frame space, where these quantities are naturally defined. Their gradients are pulled back to angular coordinates through $J(z)^\top$, so the dynamics evolve in angular space. The network operates on local frames for consistency with FrameFlow and FoldFlow-2. **Regarding the CAMEO2022 Results and the Folding Claim** The primary purpose of Table 4 is to establish a matched-data comparison under identical training conditions. Specifically, when trained on the same 24k split without pretraining, StericFlow consistently outperforms FoldFlow-2 across all four structural accuracy metrics: TM-score (0.60 vs. 0.55), RMSD (12.11 vs. 14.21 Å), GDT-TS (0.60 vs. 0.55), and lDDT (0.58 vs. 0.53). This advantage also holds on the held-out split across both protein-length categories. We acknowledge that comparisons with ESMFold and AlphaFold2 involve substantially different training data scales and have clarified this limitation in the revision. Nevertheless, StericFlow also achieves the best physical-validity metrics, including against AlphaFold2. We will revise the wording accordingly in the revised manuscript. **Regarding the Hamiltonian Framing** We follow Holderrieth et al. (2024) and adopt their reverse-learning formulation, in which the second-order Hamiltonian dynamics define the forward probability path, while the corresponding first-order marginal velocity field is learned through flow matching. Accordingly, Section 5 now clarifies this distinction between the second-order forward dynamics and the first-order generative model.

I thank the authors for their rebuttal and clarifications. However, I will maintain my score. It remains unclear whether the improvements are from the Hamiltonian formulation or from the angular coordinates. I also have concerns about the generalizability of the angular-coordinate process, including the dimension issues and the problem of periodic angles. Therefore, I will keep my score.

Reviewer_Q5b3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes StericFlow, which introduces damped Hamiltonian dynamics into the forward perturbation process of a protein generation model. Unlike conventional Gaussian diffusion or linear flow, the authors aim to gradually unfold proteins through attractive and repulsive potentials and a damping term in angle space, while preserving backbone continuity as much as possible and reducing steric clashes, and then use SE(3) flow matching to learn the reverse generation process. The experiments cover three tasks: unconditional protein backbone generation, sequence-conditioned structure generation, and motif scaffolding. The main results include: • The designability of unconditional generation reaches 0.97, an improvement over FoldFlow (OT), which achieves 0.94 under the same data settings. • Sequence-conditioned generation achieves a TM-score of 0.60 on CAMEO2022, surpassing the 0.55 achieved by the authors’ retrained FoldFlow-2. • Motif scaffolding reports 1,404 successful cases. These results suggest that the method is worth attention, especially the idea of incorporating protein geometric priors through a structured perturbation process. However, I believe the most critical issue at present is not whether the method has improved, but whether the authors have truly demonstrated that these improvements stem from the physical mechanisms they claim, and whether the mathematical theory in the paper correctly supports the corresponding conclusions.

Critical strengths
  1. Introducing physical constraints into the forward perturbation process is a valuable research direction. Most protein generation methods mainly learn geometric constraints through model architectures, loss functions, or final structure optimization. This paper attempts to directly modify the path of forward perturbation, allowing the generation model to learn the protein distribution from a certain structurally reasonable state with certain physical rationality. This idea of introducing inductive biases from probability path design rather than only from network architectures is worth exploring.
  2. The experiments are comprehensive and include a certain degree of fair comparison. The authors trained the model using approximately 24,000 PDB protein structures and retrained baselines such as FoldFlow-2 and MultiFlow, and tested on unconditional generation, sequence-conditioned folding, and motif scaffolding. Compared to only comparing with models using larger training data, this design is more persuasive.
  3. The analysis of the generation process itself is the highlight of the paper. Figures 2 and Appendix Figure 8 not only evaluate the final structure but also compare steric clashes in forward and reverse trajectories. Compared to only reporting designability, this type of analysis more directly corresponds to the core hypothesis proposed by the authors. I think these experiments currently mainly show that StericFlow can reduce collisions under a specific definition, but cannot fully prove that it has achieved strict physical constraints.
Critical weaknesses
  1. Theoretical soundness needs further clarification. There appear to be inconsistencies in the mathematical derivations. In particular, the SO(3) velocity direction in Section 4.1 seems inconsistent with the defined geodesic interpolation. Additionally, the proof of Proposition 2 appears to omit the chain-rule contribution when differentiating the repulsion potential in angular space. The authors should revisit these derivations and clarify the conditions under which the claimed convergence guarantees hold.
  2. The physical constraints are not sufficiently validated. Although the proposed repulsion potential reduces steric clashes, it does not explicitly guarantee physically valid residue distances. Moreover, Gaussian perturbations around simulated trajectories may violate the geometric constraints. It is also unclear whether the separately constructed rotational and translational flows jointly preserve valid backbone geometry. Further validation of bond geometry, peptide-bond continuity, and all-atom clashes is needed.
  3. The experimental evidence does not fully establish the contribution of the proposed method.The current ablations do not clearly distinguish the effects of Hamiltonian dynamics, steric repulsion, the secondary-structure prior, and the look-ahead loss. In particular, the improvement over the variant without repulsion is relatively modest. Furthermore, performance gains are not consistent across all tasks, and improvements in long-chain generation are relatively small. More controlled ablations, independent runs, and statistical analysis would strengthen the empirical claims.
Questions for authors
  1. Can the authors clarify the mathematical consistency of the proposed Hamiltonian flow and its convergence guarantees? In particular, the SO(3) velocity in Eq. (9) appears to have the opposite sign from the derivative of the geodesic interpolation defined in Section 4.1. Moreover, the proof of Proposition 2 seems to omit the Jacobian of the NeRF mapping when differentiating the Cartesian repulsion potential with respect to angular coordinates. Could the authors clarify these derivations and specify the precise conditions under which exponential convergence is guaranteed?
  2. How much of the improvement comes specifically from the Hamiltonian dynamics rather than the structured prior and auxiliary training objectives? The current ablation shows a relatively modest difference between StericFlow with and without the steric potential. Could the authors provide controlled ablations separating the contributions of the nonlinear trajectory, steric repulsion, secondary-structure prior and look-ahead loss, while keeping the architecture and training budget fixed?
  3. The paper reports substantially fewer Cα clashes, but the SO(3) and Cartesian flows appear to be constructed separately, and Gaussian perturbations are introduced around simulated trajectories. Could the authors evaluate bond lengths, bond angles, peptide-bond continuity and all-atom clashes at intermediate sampling steps, and clarify whether the stated physical guarantees apply to the final learned process?
  4. The improvements over matched baselines are encouraging but sometimes small, particularly in long-chain generation. Could the authors report confidence intervals across multiple random seeds, per-motif scaffolding results and details of the shared evaluation protocol? It would also be helpful to explain the discrepancy between the designability reported in Table 2 and Appendix Table 9.

If you can solve my problem, I would be willing to increase my score.

3 comments

Many thanks for your thoughtful review. We have carefully addressed your concerns below.

Controlled Ablations Separating the Forward-Process Components

Thank you for this suggestion. We have separated these components at fixed architecture, training budget, data split and evaluation protocol. The results are shown in Table below.

Forward processDesignabilityNovelty (TM-Frac)avg.maxTMDiversity (Cluster)TM-Score
Linear OU (no Hamiltonian, no sterics)0.930.140.440.35 (164)0.25
++ damped Hamiltonian unfolding (k2=0k_2=0)0.940.160.410.49 (210)0.25
++ steric repulsion (β\beta-prior)0.970.190.380.50 (220)0.22
++ steric repulsion (α\alpha-prior)0.980.170.400.52 (228)0.23

The trajectory shape provides a modest designability gain (+0.01) but substantially increases diversity (164 to 210 clusters), while steric repulsion further improves designability (+0.03–0.04), novelty, and diversity, with a 17-fold reduction in clashing pairs; the negligible difference between terminal priors (α vs. β, +0.01) suggests that the gains come primarily from the probability path rather than prior-induced secondary-structure bias.

Regarding Confidence Intervals and the Evaluation Protocol

Thank you for this suggestion. We computed Wilson 95% confidence intervals and one-sided two-proportion tests across 250 samples per variant. Pooling both StericFlow variants (n=500), we find significant improvements over Linear-OU (+0.045, p = 0.005), the k2=0k_2=0 control (+0.035, p = 0.017), and FoldFlow-OT (+0.035, p = 0.017). StericFlow (β) also outperforms FoldFlow-OT across all five evaluation metrics (sign test, p = 0.031). All methods use the same 24,003-monomer split, length distribution, ProteinMPNN sampling settings, ESMFold refolding, and 2 Å scRMSD threshold.

Regarding Consistency of the Gains Across Tasks

Our central contribution is showing that the forward process itself is a useful design axis for protein generation, beyond architectures, data, and losses. Theoretically, Propositions 2 and 3 establish the path's well-posedness and guarantees, making StericFlow, to our knowledge, the first protein generative model with a clash-free forward process by construction. Methodologically, we introduce simulation-based flow-matching targets, an objective that accounts for non-zero path curvature, and the J(z)⊤J(z)^\top pullback through reconstruction. These components allow arbitrary differentiable potentials to be incorporated while retaining the theoretical guarantees, at an additional cost of under one second per 200-residue protein and no increase in model size.

Empirically, StericFlow is strongest among PDB-trained methods across all five Table 2 metrics, reduces clashing pairs 17-fold, leads on all three long-chain metrics, and achieves 1404 motif-scaffolding successes versus 1099 and 889 for baselines. It also achieves the lowest clash and peptide-break rates (below AlphaFold2) and 0.97 designability with substantially higher β-sheet content.

Regarding Validation of the Physical Constraints

We have evaluated the intermediate sampling steps, and this analysis is reported in Appendix C.4 and Fig. 8. Clashes are defined as pairwise residue contacts with d(xi,xj)<0.38d(x_i,x_j) < 0.38 and are tracked at every time step of the reverse process, from T=100T=100 to T=0T=0, averaged over 5 structures for each length in {100,150,200}\{100,150,200\}. As shown there, StericFlow produces fewer clashes than the vanilla diffusion and flow baselines at every intermediate step and across all tested lengths, and not only in the final generated structures; the property therefore holds for the learned process and not merely for the forward simulation.

Regarding the Hamiltonian Flow Derivations and Convergence Guarantees

We thank the reviewer for this careful reading. The apparent sign discrepancy arises because the two expressions use opposite time orientations. Section 4.1 describes the generative direction, $ r_t = \exp_{r_0}(t\log_{r_0}(r_1))$, which gives r˙t=log⁡rt(r1)1−t\dot r_t = \frac{\log_{r_t}(r_1)}{1-t}. In contrast, Eq. (9) reports the noising interpolant used in the FoldFlow and FrameFlow convention, where time runs from data to prior: r˙t=log⁡rt(r0)t.\dot r_t = \frac{\log_{r_t}(r_0)}{t}.

The two expressions are related by the time transformation t↦1−tt \mapsto 1-t, which reverses the tangent direction. Our implementation consistently regresses against log⁡rt(r1)/(1−t)\log_{r_t}(r_1)/(1-t), in agreement with the R3\mathbb{R}^3 component. We will revise the manuscript to clarify this distinction.

We agree that the Jacobian of the NeRF mapping was omitted from the written derivation. Let U(z)=k1Ut(z)+k2Ur(nerf(z)),U(z) = k_1 U_{\mathrm{t}}(z) + k_2 U_{\mathrm{r}}(\mathrm{nerf}(z)), and define J(z)=∂nerf(z)/∂zJ(z) = \partial \mathrm{nerf}(z)/\partial z. The correct gradient is

2k_1\Delta(z,z_{\mathrm{target}}) +k_2J^\top\nabla_x U_{\mathrm{r}}.$$ The corresponding Hessian contribution is \$\nabla_z^2 (U_r \circ f)=J^\top H_r J + \sum_{k=1}^{d_x} \frac{\partial U_r}{\partial x_k} \nabla_z^2 f_k $$The omission affects the constants in the proof, but not its overall structure. The implementation applies $J^\top$ exactly through automatic differentiation of the reconstruction; only the written analysis was incomplete. Step S1 relies only on the equations of motion and holds for any differentiable potential $U$. In S3, accounting for $J(z_{\mathrm{target}})^\top$ and assuming $\|J\|\leq C_J$ gives$$ \|\delta_i\| \leq \frac{C_1 C_J k_2}{2k_1}. $$In S2, the Hessian takes the form $2k_1 I+k_2 M$, where $\|M\|\leq C_M$. Consequently,$$ \omega_m^2 \geq 2k_1-k_2C_M, \$\$ which confirms the originally claimed scaling $\omega_m^2=2k_1+O(k_2)$. Regarding the precise conditions for exponential convergence, we will state the following assumptions explicitly in the revision: - **(A1)** $\gamma,k_1>0$. - **(A2)** The initial state is clash-free, so that $\mathcal{H}_0=\frac{1}{2}\|v_0\|^2+U(z_0)$ is finite. This holds for every PDB structure. - **(A3)** $U_{\mathrm{r}}$ acts as a barrier, ensuring that the sublevel set $\{\mathcal{H}\leq\mathcal{H}_0\}$ remains away from collisions and collinear configurations, and that $U$ is $C^2$ on this set. - **(A4)** $k_2<2k_1/C_M$. Under (A1)–(A3), Step S1, together with LaSalle's invariance principle, establishes convergence to the critical points of $U$. Assumption (A4) is additionally required to obtain the local exponential convergence rate established in S2. Accordingly, we restrict our global convergence claim to asymptotic convergence. Finally, (A4) is a checkable condition and is comfortably satisfied at our operating point, since $k_1/k_2=\ell_{\mathrm{prot}}\geq 60$ (Table 7). Again, thank you for your constructive feedback. We hope we have sufficiently addressed your concerns.

Thank you for your reply and additional explanations. I think this manuscript is interesting, and the research direction is also promising. However, the contribution and practicality, especially the improvement is not significant. So I will still keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.