acceptodds
Under review as a conference paper at ICLR 2027

EnzyShell: Dynamic Shell-Guided Enzyme Backbone Generation

Abstract

Enzyme design requires generating protein backbones that scaffold catalytic motifs and accommodate substrates. However, directly generating full backbones from these functional constraints involves two distinct requirements: inferring a suitable local protein environment around the catalytic motif and substrate, and generating the full backbone to support this environment. To address this challenge, we propose EnzyShell, a dynamic shell-guided framework for enzyme backbone generation. We explicitly model this local environment as a functional shell comprising the motif and residues surrounding either the motif or the substrate. We then train two complementary flow models for full-backbone generation, conditioned on the substrate together with either the motif or the functional shell. During sampling, we dynamically construct the functional shell from backbone predictions of the motif-conditioned model, provide it to the shell-conditioned model, and combine their predicted vector fields to guide enzyme backbone generation. Evaluated on the EnzyBind benchmark spanning six EC classes, EnzyShell improves backbone designability, EC match rate, and predicted over the evaluated baselines, while maintaining comparable catalytic-motif geometry.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
enzyme backbone design, generative models, flow matching

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_T621OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

EnzyShell is a framework for generating enzyme backbones conditioned on catalytic motifs and substrates. The main idea is to explicitly model the local structure env around the catalytic motif and substract (functional shell), and to guide full-backbone generation. The framework consists of two conditional flow models. The motif-conditioned model generates backbone predictions from the catalytic motif and substrate, while the shell-conditioned model uses the functional shall and substrate as conditions. During inference, the functional shell is dynamically extracted from the motif-conditioned model's intermediate predictions. The vector fields from both models are then combined to update a shared sampling trajectory.

Critical strengths
  1. EnzyShell is well motivated because residues surrounding catalytic sites can influence substrate recognition, active-site geometry, and catalysis. Representing this local environment explicitly as a functional shell is a reasonable design choice. I also appreciate that the shell includes residues near both the substrate and the catalytic motif, rather than defining the entire environment using substrate proximity alone.
  2. The method provides a clear way to connect local structural prediction with global backbone generation. In particular, dynamically constructing the shell from intermediate backbone predictions avoids requiring a predefined pocket structure at inference time. The combination of two conditional vector fields along a shared trajectory is also straightforward to implement within flow-matching models.
Critical weaknesses
  1. My main concern is whether the improvements can be attributed specifically to the proposed dynamic shell mechanism, rather than to the stronger underlying models, additional training objectives, or the combination of two networks. In Table 2, the motif-conditioned model alone already achieves a designability score of 0.689, which is substantially higher than the strongest external baseline (0.517). Dynamic shell guidance improves this score to 0.752, but the improvement over sequential two-model sampling is relatively small, from 0.726 to 0.752. Furthermore, Table 3 shows that using only the shell-conditioned vector field (α=1\alpha = 1), while still dynamically constructing the shell using the motif-conditioned model, already achieves 0.736. Combining the two vector fields increases this to 0.752. These results suggest that the different components contribute to performance, but their individual effects are not clearly separated. In particular, both models are trained with shell-weighted flow-matching losses and additional geometric supervision (Eqs. 9–13). Therefore, even the motif-conditioned model benefits from shell-specific training objectives.

  2. During training, the shell-conditioned model receives shells extracted from native structures. During inference, however, the shell is extracted from predictions of the motif-conditioned model. The authors address this mismatch by adding Gaussian noise to the shell coordinates during training. However, coordinate perturbation does not necessarily capture all the errors in predicted shells. In particular, the predicted shell may include incorrect residues or miss residues that belong to the native functional environment. The hard distance thresholds can also cause shell membership to change abruptly as the predicted backbone evolves. Additionally, at later sampling steps, the motif coordinates in the shell are taken from the motif-conditioned model's predictions instead of the prescribed motif coordinates. So I wonder how reliably the method preserves the original catalytic constraints during sampling.

Questions for authors
  1. How much of the improvement comes specifically from dynamic shell construction?
  2. How accurate and stable are the dynamically predicted functional shells?
  3. How reliably the method preserves the original catalytic constraints during sampling?
  4. How reliable are the functional improvements across targets?
4 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the constructive feedback and suggestions. We have conducted additional experiments and analyses to address the reviewer’s concerns regarding component contributions, shell reliability, and functional improvements.

Weakness 1 & Question 1: Contribution of Dynamic Shell Construction and Shell-Specific Training Objectives

To address the reviewer's concern, we have conducted additional experiments to assess the respective contributions of dynamic shell construction and shell-specific training objectives, with newly added experiments marked as [New] below.

1. Dynamic shell construction

First, following the reviewer's suggestion, we added a fixed-shell shared-trajectory baseline using the same trained models and fusion settings as EnzyShell, except that dynamic shell updates were replaced with a fixed shell extracted from a backbone independently generated by the motif-conditioned model (as in fixed-shell sequential sampling).

As shown in the table below, dynamic shell updates improve designability (0.752 vs. 0.722), while also improving motif preservation and the number of successful structural clusters. Without vector-field fusion, dynamic shell-guided sampling also achieves slightly higher designability than fixed-shell sequential sampling.

2. Shell-specific training objectives

Second, we trained an ablated motif-conditioned model without the shell-specific training objectives, including the shell-weighted flow-matching term (Equation 9, Line 247), substrate loss (Equation 11, Line 260), and contact loss (Equation 11, Line 260).

As shown in the table below, removing these losses slightly improves standalone motif-model designability but reduces motif preservation, while leading to lower generation performance under dynamic shell-guided sampling.

We have incorporated the additional experiments into Section 4.3 (Table2, Lines 425–430, 444–445 and 457-463) and Appendix C.2 of the revised manuscript.

Sampling StrategyDesignability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
Motif-model0.6891.92167
Motif-model (w/o shell-specific objectives) [New]0.7012.04165
Fixed-shell sequential (w/o fusion)0.7261.94169
Fixed-shell shared-trajectory (w/ fusion) [New]0.7221.91166
Dynamic shell (w/o fusion)0.7361.91165
Dynamic shell (w/ fusion, w/o shell-specific objectives) [New]0.7312.01163
Dynamic shell (w/ fusion, EnzyShell)0.7521.82171
AuthorsOpenReview ↗

Official Comment by Authors

Weakness 2 & Questions 2–3: Shell Prediction Reliability and Catalytic Constraint Preservation

We agree with the reviewer's observation that Gaussian coordinate perturbation does not explicitly simulate shell membership errors. To address these concerns, we have conducted additional analyses of coordinate augmentation, shell membership accuracy and stability, and catalytic motif preservation.

1. Effect of Gaussian coordinate perturbation (W2)

We trained shell-conditioned models with different levels of Gaussian coordinate perturbation during training, while keeping the motif-conditioned model and all other experimental settings fixed.

As shown in the table below, moderate coordinate perturbation improves generation performance under dynamic shell-guided sampling. We have included this ablation in Appendix C.1(new Table 5) of the revised manuscript.

σmax⁡\sigma_{\max} (Å)Designability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
0.00.7281.94167
0.10.7371.90169
0.5 (default)0.7521.82171
1.00.7441.87163

2. Shell membership accuracy and stability (Q2)

We evaluated predicted shell membership against reference shells constructed from native enzyme structures using residue-index-based precision, recall, and intersection-over-union (IoU), excluding prescribed motif residues. Stability was measured using adjacent-step Jaccard similarity. We further evaluated geometric evolution by computing the RMSD between the motif model's clean predictions and the final generated backbone on fixed final-shell residue indices after motif-only alignment.

As shown in the table below, the predicted shells achieve high reference-shell agreement and rapidly stabilize across adjacent sampling steps, while dynamic shell construction continues to update geometric conditioning throughout sampling. We have included this analysis and the corresponding metric curves across sampling steps in Appendix C.3 (new Figure 3) of the revised manuscript.

StepReference-shell IoU ↑Adjacent-step Jaccard ↑Shell RMSD to final backbone (Å) ↓
20.705-2.712
30.7600.8052.152
50.7940.9231.818
100.8020.9651.590
500.8120.9721.091
2000.8100.9870.575
3000.8080.9810.368

(The adjacent-step Jaccard at step 2 is omitted because the preceding shell contains only the prescribed motif, corresponding to an empty non-motif shell.)

3. Catalytic constraint preservation (W2, Q3)

We evaluated motif RMSD relative to the prescribed motif after motif-only Kabsch alignment, using both the sampling state xkx_k and the motif model's predicted clean coordinates x^1∣k(M)\hat{x}_{1|k}^{(M)}, which are used for dynamic shell construction.

As shown in the table below, the motif model's clean predictions preserve the prescribed catalytic motif geometry throughout sampling, while the fused sampling state progressively recovers this geometry. We have included this analysis and the corresponding motif RMSD curves in Appendix C.4 (new Figure 4) of the revised manuscript.

Stepxkx_k Motif RMSD (Å)x^1∣k(M)\hat{x}_{1\mid k}^{(M)} Motif RMSD (Å)
1017.320.14
5011.010.12
1006.070.08
2001.800.05
2500.930.08
3000.480.09
4000.100.10

Question 4: Reliability of Functional Improvements Across Targets

To address this question, we have conducted a paired target-level comparison between EnzyShell and RFdiffusion3 across all test targets.

As shown in the table below, computational functional performance varies across targets and metrics. We have included this analysis, the paired comparison table, and complete per-target comparisons across all five methods in Appendix C.5 (new Table 7, new Figures 5–7) of the revised manuscript.

MetricWinTieLoss
CLEAN-predicted EC match rate66 (34.7%)93 (48.9%)31 (16.3%)
UniKP-predicted kcatk_{\mathrm{cat}}97 (51.1%)–93 (48.9%)
Gnina docking score73 (38.4%)–117 (61.6%)

(Wins and losses indicate better and worse performance relative to RFdiffusion3, respectively.)

Thanks the authors for the rebuttal and additional experiments. I have no other issues and I'll maintain my positive scores.

AuthorsOpenReview ↗

Official Comment by Authors

Once again, we appreciate the reviewer's insightful and constructive comments, which have significantly enhanced the quality of our manuscript.

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

This paper proposes EnzyShell, a dynamic shell-guided framework for enzyme backbone generation that explicitly represents the local structural environment around catalytic motifs and substrates as a functional shell. The framework combines two complementary conditional flow models: a motif-conditioned model that infers the local environment while generating the full backbone, and a shell-conditioned model that uses the dynamically constructed functional shell as an explicit structural condition. During inference, their vector fields are combined along a shared sampling trajectory to guide backbone generation. On 190 enzyme–substrate pairs from the EnzyBind benchmark, EnzyShell achieves a designability of 0.752 compared with 0.517 for RFdiffusion3 while maintaining comparable motif RMSD (1.82 Å vs. 1.77 Å), and also reports improved predicted EC match rate and kcat.

Critical strengths
  1. The functional-shell formulation is intuitive and well motivated for enzyme backbone generation. Separating the inference of the local catalytic environment from full-backbone generation provides a reasonable decomposition of this challenging conditional generation problem.

  2. The ablation studies provide empirical support for the proposed dynamic coupling mechanism. Dynamic shell-guided sampling achieves a designability of 0.752, compared with 0.726 for sequential two-model sampling and 0.689 for motif-only sampling, while simply increasing the sampling budget of the standalone models yields only limited improvements. The analysis of the combination weight α\alpha further shows relatively stable performance over a range of intermediate values, suggesting that the observed gains are not dependent on a narrowly tuned setting.

Critical weaknesses
  1. The functional claims rely almost entirely on computational proxy models. EC match is obtained from CLEAN, kcat from UniKP, and binding from Gnina. The authors themselves state that these metrics do not establish catalytic activity and that UniKP values should not be interpreted as experimentally measured kinetics. Thus, claims about improved enzyme function should be interpreted much more cautiously than the structural-designability results.

  2. The baseline comparison is not fully condition matched. EnzyShell receives both catalytic motif and substrate, whereas the baseline methods receive whatever inputs their original formulations support. For example, Proteina receives the motif but no substrate. Therefore, part of the improvement over some baselines may reflect additional conditioning information rather than the proposed dynamic shell mechanism itself.

Questions for authors
  1. Can the authors provide a controlled baseline in which the same underlying architecture receives the same motif+substrate information but does not use dynamic shell guidance? The motif-only ablation helps, but a capacity- and conditioning-matched alternative would better isolate the contribution.

  2. Are any experimentally characterized enzyme designs or experimentally grounded activity measurements available to validate whether the gains in CLEAN/UniKP correlate with real catalytic function?

2 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the positive assessment and constructive suggestions. We address the reviewer's concerns regarding conditioning fairness and functional validation below.

Weakness 2 & Question 1: Conditioning Fairness and Controlled Baselines

We acknowledge that the external baselines differ in their conditioning inputs, which may partly contribute to the observed performance differences. To isolate the contribution of dynamic shell updates, we have added a new controlled baseline (fixed-shell shared-trajectory, marked as [New] below) using the same trained models, conditioning inputs, and fusion settings as EnzyShell. In this baseline, dynamic shell updates are replaced with a fixed shell extracted from a backbone independently generated by the motif-conditioned model (as in fixed-shell sequential sampling).

As shown in the table below, dynamic shell construction improves designability (0.752 vs. 0.722), while also improving motif preservation and the number of successful structural clusters. These results provide further evidence that the improvement is not solely attributable to additional conditioning information or model capacity. We have included this new baseline in Section 4.3 (Table2, Lines 425–430, 444–445 and 457-463) of the revised manuscript.

Sampling StrategyDesignability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
Fixed-shell shared-trajectory (w/ fusion) [New]0.7221.91166
Dynamic shell-guided (w/ fusion, EnzyShell)0.7521.82171

Weakness 1 & Question 2: Functional evaluation and experimental validation

We agree that computational functional predictions alone cannot establish actual catalytic activity. Previous studies have provided experimental evidence supporting the relevance of these predictors in their respective evaluation settings. Specifically, CLEAN has been evaluated for enzyme function identification, including experimental validation [1]. UniKP has demonstrated correlations with experimentally measured kinetic parameters and has been applied to experimentally validated enzyme mining and directed evolution [2].

However, we have not experimentally verified whether the observed improvements in CLEAN/UniKP predictions correlate with actual catalytic activity in our generated designs. We therefore interpret these metrics as supplementary indicators of potential functional relevance and have further clarified this distinction in the revised manuscript (Lines 24–26, 63–64, and 77–78).

[1] Yu et al., Science, 2023 (CLEAN).

[2] Yu et al., Nature Communications, 2023 (UniKP).

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you again for your constructive feedback. As the discussion period is approaching its end, we would appreciate it if you could take a look at our rebuttal and share any further thoughts, particularly on whether our additional experiments and analyses have addressed your concerns.

We would be happy to clarify any remaining questions.

Best regards,

The Authors

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

EnzyShell improves enzyme backbone generation by repeatedly predicting the local environment around an active site and using that prediction to guide the whole protein. Its desgnability is improved, but wether the designed enzyme can catalyze the chemical reaction is still a question.

Critical strengths

The method models the chemical environment. The shell captures neighboring structural context beyond the motif itself, which would be informative.

Critical weaknesses

Modeling issue:

The paper focus on enzyme backbone generation only. First, modeling the full-atom structure is important for enzyme-substrate interaction, at least for the active sites, and there are already a lot of full-atom enzyme design methods. Second, the method doesn't have any specific module that involves the special features for enzymes. Therefore, it can actually be used to design backbones for general protein families. The model should evaluate its performance in this way to see if the architecture is effective. Otherwise, special features from enzymes should be modeled so that the performance would be better.

Evaluation:

On several metrics, it performs worse than RFD2 and RFD3, maybe due to that it only models the backbone structure.

Additionally, the gain on some metrics specifically attributable to dynamic shell guidance is smaller than the headline comparison. The motif-conditioned model already achieves 68.9% designability. The full method adds 6.3 percentage points, or 2.6 points over sequential sampling. The larger external-baseline gap therefore should not be attributed entirely to dynamic shell updates.

Training and inference use different kinds.

Training shells come from native structures. However, inference shells come from model predictions. Coordinate noise during training addresses some errors, but does not explicitly reproduce incorrect shell membership or larger structural mistakes. Missing ablations on shell radii, noise, and auxiliary losses leave their individual contributions unclear.

Questions for authors

See above limitations

3 comments
AuthorsOpenReview ↗

Official Comment by Authors

We thank the reviewer for the comments and suggestions to improve the manuscript. We address the concerns below and describe the corresponding revisions.

Weakness 1: Scope and Limitations of Backbone-Level Enzyme Design

We agree that full-atom modeling, particularly catalytic side-chain geometry and enzyme–substrate interactions, is important for enzyme design. Our work focuses on motif- and substrate-conditioned enzyme backbone generation, rather than full-atom enzyme design.

Our motivation is that preserving catalytic motif geometry alone is insufficient to fully characterize the structural environment required for substrate accommodation. To address this challenge, EnzyShell explicitly incorporates catalytic motif and substrate features and introduces a functional shell representation that captures their local structural environment. Specifically, the functional shell is constructed from residues surrounding both the catalytic motif and the substrate and is dynamically updated during generation. Shell-specific training objectives further emphasize the structural organization surrounding catalytic sites.

As shown in the table below (with newly added experiments marked as [New]), the full EnzyShell framework improves designability over motif-only generation (0.752 vs. 0.689). Moreover, removing the shell-specific training objectives, including the shell-weighted flow-matching term (Equation 9, Line 247), substrate loss (Equation 11, Line 260), and contact loss (Equation 11, Line 260), reduces both designability and motif preservation, providing further evidence for the contribution of these training objectives to enzyme backbone generation.

Although EnzyShell builds upon general neural network components, its conditioning, structural representation, and training and sampling mechanisms are specifically designed for enzyme backbone generation under catalytic motif and substrate constraints. As discussed in the Conclusion section of the original manuscript, extending EnzyShell toward all-atom modeling is an important direction for future work.

Sampling StrategyDesignability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
Motif-model0.6891.92167
EnzyShell w/o shell-specific objectives [New]0.7312.01163
EnzyShell0.7521.82171

Weakness 2: Performance Evaluation and Contribution of Dynamic Shell Guidance

1. Comparison with RFdiffusion2 and RFdiffusion3

We acknowledge that EnzyShell does not outperform all baselines across every evaluation metric. Specifically, RFdiffusion3 achieves slightly better motif RMSD, while RFdiffusion2/3 achieve better Gnina docking scores. Nevertheless, EnzyShell achieves the best results in designability, the number of successful structural clusters, CLEAN-predicted EC match rate, and UniKP-predicted kcatk_{\mathrm{cat}}. As discussed in Weakness 1, our framework focuses on backbone-level generation and does not explicitly model atomic-level enzyme–substrate interactions, which may limit performance on certain metrics.

2. Contribution of Dynamic Shell Guidance

We also agree that the overall improvements over external baselines should not be attributed solely to dynamic shell guidance. To better isolate its contribution, we have introduced a fixed-shell shared-trajectory baseline using the same trained models, conditioning inputs, and fusion settings as EnzyShell, but replacing dynamic shell updates with a fixed shell extracted from a backbone independently generated by the motif-conditioned model (as in fixed-shell sequential sampling).

As shown in the table below (with newly added experiments marked as [New]), dynamic shell updates improve designability (0.752 vs. 0.722), while also improving motif preservation and the number of successful structural clusters under matched settings. These results support the contribution of dynamic shell construction without attributing the full performance gap over external baselines to this component alone. We have included this controlled comparison in Section 4.3 (Table2, Lines 425–430, 444–445 and 457-463) of the revised manuscript.

Sampling StrategyDesignability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
Fixed-shell shared-trajectory (w/ fusion) [New]0.7221.91166
Dynamic shell (w/ fusion, EnzyShell)0.7521.82171
AuthorsOpenReview ↗

Official Comment by Authors

Weakness 3: Training–Inference Shell Mismatch and Robustness

We agree that using native structure-derived shells during training and model-predicted shells during inference introduces a potential distribution mismatch. Gaussian coordinate perturbation can mitigate coordinate-level errors but does not explicitly simulate incorrect shell membership or larger structural errors. To address these concerns, we have conducted additional analyses of shell prediction reliability and ablations of coordinate perturbation, shell radii, and shell-specific training objectives.

1. Shell membership accuracy and stability

We evaluated predicted shell membership against reference shells constructed from native enzyme structures using residue-index-based precision, recall, and intersection-over-union (IoU), excluding prescribed motif residues. Stability was measured using adjacent-step Jaccard similarity. We further evaluated geometric evolution by computing the RMSD between the motif model's clean predictions and the final generated backbone on fixed final-shell residue indices after motif-only alignment.

As shown in the table below, the predicted shells achieve high reference-shell agreement and rapidly stabilize across adjacent sampling steps, while dynamic shell construction continues to update geometric conditioning throughout sampling. We have included this analysis and the corresponding metric curves across sampling steps in Appendix C.3 (new Figure 3) of the revised manuscript.

StepReference-shell IoU ↑Adjacent-step Jaccard ↑Shell RMSD to final backbone (Å) ↓
20.705-2.712
30.7600.8052.152
50.7940.9231.818
100.8020.9651.590
500.8120.9721.091
2000.8100.9870.575
3000.8080.9810.368

(The adjacent-step Jaccard at step 2 is omitted because the preceding shell contains only the prescribed motif, corresponding to an empty non-motif shell.)

2. Ablation of shell radii

We evaluated different shell radius settings during inference while keeping the trained models and all other sampling settings fixed. As shown in the table below, the default radii (6 Å for the motif and 8 Å for the substrate) achieve the best performance across all three metrics, while generation performance remains relatively stable across the tested settings. We have included this ablation in Appendix C.6 (new Table 8) of the revised manuscript.

Motif radius (Å)Substrate radius (Å)Designability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
4.06.00.7431.92168
6.0 (default)8.0 (default)0.7521.82171
8.010.00.7351.95166

3. Ablation of Gaussian coordinate perturbation

We trained shell-conditioned models with different levels of Gaussian coordinate perturbation during training, while keeping the motif-conditioned model and all other experimental settings fixed.

As shown in the table below, moderate coordinate perturbation improves generation performance under dynamic shell-guided sampling. We have included this ablation in Appendix C.1 (new Table 5) of the revised manuscript.

σmax⁡\sigma_{\max} (Å)Designability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
0.00.7281.94167
0.10.7371.90169
0.5 (default)0.7521.82171
1.00.7441.87163

4. Ablation of shell-specific training objectives

We trained an ablated motif-conditioned model without the shell-specific training objectives, including the shell-weighted flow-matching term (Equation 9, Line 247), substrate loss (Equation 11, Line 260), and contact loss (Equation 11, Line 260).

As shown in the table below (with newly added experiments marked as [New]), removing these objectives slightly improves standalone motif-model designability but reduces motif preservation, while leading to lower generation performance under dynamic shell-guided sampling. We have included this ablation in Appendix C.2 (new Table 6) of the revised manuscript.

Sampling StrategyDesignability ↑Motif RMSD (Å) ↓# Succ. Clusters ↑
Motif-model0.6891.92167
Motif-model (w/o shell-specific objectives) [New]0.7012.04165
Dynamic shell (w/ fusion, w/o shell-specific objectives) [New]0.7312.01163
Dynamic shell (w/ fusion, EnzyShell)0.7521.82171

Together, these experiments provide evidence for the accuracy and stability of predicted shell membership and the contributions of the evaluated training and conditioning choices.

AuthorsOpenReview ↗

Official Comment by Authors

Dear Reviewer,

Thank you again for your constructive feedback. As the discussion period is approaching its end, we would appreciate it if you could take a look at our rebuttal and share any further thoughts, particularly on whether our additional experiments and analyses have addressed your concerns.

We would be happy to clarify any remaining questions.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.