acceptodds
Under review as a conference paper at ICLR 2027

RayMorph: Learning Protein Redesign from Structural Pairs

Abstract

Redesigning a protein to change its geometry while preserving its function is a central challenge in protein engineering. For instance, when a biosensor needs to be inserted into an internal loop of a target protein, its N- and C-termini need to be brought close together. In such cases, the desired change can be easier to illustrate with examples than explicit residue-level guidance. To address this challenge, we introduce RayMorph, a framework that learns protein redesign from pairs of re- lated structures exhibiting a desired property change. The paired structural data specifies the transformation by example, allowing it to be learned and applied to new structures. RayMorph jointly modifies geometry, sequence, and length while preserving the source fold. We establish correspondence between unequal-length structures through geometry-guided dynamic programming and construct bridge trajectories that combine residue additions and deletions with changes in coordi- nates and amino-acid identities. We introduce a curated AlphaFoldDB dataset of structurally matched protein pairs with varying N–C distances and terminal ac- cessibility. Among 488 proteins drawn from 50 Foldseek clusters excluded from training, 25.4% yield at least one candidate satisfying designability, source-fold similarity, and terminal-remodeling criteria after inverse folding and refolding.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, generative modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_SJT4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RayMorph, a novel framework trained on transition trajectories between protein pairs with undesirable and target properties. RayMorph can therefore redesign proteins without requiring the hand-crafted constraints and conditions used in previous protein design models. The authors evaluate their framework on redesigning proteins to have shorter N–C distances.

Critical strengths
  1. The idea is novel, and supporting diverse redesign objectives is an important problem. In practical protein design, we often encounter requirements that are difficult to incorporate into a design model. RayMorph offers a potential solution to this problem.
  2. The authors demonstrate their idea by redesigning proteins with shorter N–C distances, showing the effectiveness of RayMorph.
Critical weaknesses
  1. One concern is the validity of the intermediate states generated by the bridge between S₀ and S₁. It seems that their validity is not guaranteed. Could this affect training performance?
  2. The performance of RayMorph appears somewhat limited. The paper presents a sophisticated data pipeline for co-design, but the consistency between its generated sequences and backbones is substantially lower than that obtained with ProteinMPNN-designed sequences for the generated backbones. Would a structure-only backbone design approach followed by ProteinMPNN sequence design achieve better performance?
Questions for authors

See weakness, the paper overall is novel and interesting, my major concern is the performance of RayMorph, and I look forward to the authors’ response to this concern.

4 comments

Weakness #1: “One concern is the validity of the intermediate states generated by the bridge between S₀ and S₁.” “Could this affect training performance?”

Thank you for raising the concern that the invalid intermediate states generated by the bridge could affect training performance. The intermediates are not guaranteed to be physically valid proteins. As is standard in diffusion/flow protein structure design models, the intermediates are normally invalid but the final target the model is trained towards is a valid protein. To mitigate the risk of poor training, the intermediate coordinates are supervised by the conditional drift toward assigned endpoint coordinates x₁, and amino-acid labels are trained to match a₁ (see Section 3.3 of the paper). We also incorporated auxiliary geometric losses that promote structural validity. (see Appendix B.3)

Weakness #2: “Would a structure-only backbone design approach followed by ProteinMPNN sequence design achieve better performance?”

We agree that RayMorph’s generated sequences do not match the quality of ProteinMPNN-designed sequences. Please see §2 of the Overall Response for the structure-only ablation addressing this concern.

 

If we have addressed your concerns, we hope you will consider raising your score.

Thank you for the detailed response and for the additional experiments. The new baselines and the structure-only ablation help clarify the contribution of RayMorph, and I appreciate the planned revisions to the functional-preservation claims.

On Weakness #1, I remain only partly convinced. Unlike standard diffusion/flow models that start from pure noise, RayMorph starts from a valid protein, and its bridge states also involve discrete split/delete edits. For example, a split produces a child that inherits its parent's coordinates. The response describes how the supervision is constructed, but it doesn't quantify how invalid intermediate states affect training. It doesn't show geometric statistics of the bridge states, or an ablation on the bridge design. I would encourage the authors to add such an analysis in the revision.

Overall, I appreciate the authors' efforts. After considering the rebuttal, I have decided to maintain my score.

Thanks for your question. Below we show a new analysis suggesting that the specific concern you outlined—that the child inherits the parent’s attributes—is not a driving factor of training quality.

We define a metric of an intermediate state’s validity as the average of the following two scores, already introduced in the paper:

  • Bonds %: the fraction of consecutive Cα pairs (i, i+1) between 3.0 and 4.5 Å apart.
  • Clash-free %: the fraction of residues with no other Cα within 3.5 Å, counting only residues at least 3 positions apart in the chain (|i − j| ≥ 3).

Our starting observation (see table below) is that the metric is near 100% at t=0 and t=1 but drops to approximately 82% for t=[0.4,0.6]. This supports the point you mentioned about intermediate states being invalid but also indicates that even at their worst, the intermediate states are only moderately invalid.

To address the concern of child-inheriting-the-parent driving this invalidity, we next sought to attribute intermediate-state invalidity to three causes:

  1. Linear interpolation between S0S_0 and S1S_1
  2. Sigma knob of the Brownian bridge marginal
  3. Child placement on top of parent (reviewer’s concern)

We argue that #1 and #2 are validity compromises any plausible architecture will have to make. Specifically, #1 is the standard interpolation framework used very commonly in diffusion/flow models; e.g., FrameFlow[1] and FoldFlow[2] use it for coordinate changes. As for #2, it is a necessary component for stochastic inference to enable sampling diversity (when prior is a protein not noise) and one-to-many learning in a generative approach. Thus it is only with #3 that a RayMorph-specific design choice comes into play.

As an alternative to child-on-parent placement, we introduce ‘triangular insertion’: placing a newborn residue 3.8 Å (one Cα–Cα bond) from both its parent and the parent's right neighbour. This directly eliminates the penalty on both bonds and clashfree scores.

The table below illustrates an ablation on how the 3 factors affect the validity of the intermediate states. Across 400 subsampled protein-pairs in a training run, each cell reports the average of bonds% and clash-free% on the entire protein chain across time in the bridge interpolation between S0S_0 and S1S_1. σ is the noise scale of the Brownian bridge; σ = 0 reduces it to linear interpolation. The current RayMorph setting is shown in bold.

Child PlacementInherit ParentInherit ParentTriangularTriangular
tσ = 0.1σ = 0σ = 0.1σ = 0
0.099.899.899.899.8
0.1–0.387.393.487.693.7
0.4–0.682.086.782.687.4
0.7–0.989.993.590.393.9
1.099.999.999.999.9

Based on this analysis, we see that the two necessary design choices—straight line linear interpolation and sigma noise in the construction of brownian bridges—are the major drivers of intermediate state invalidity. We argue that the current design choice of having the child inherit the parent is not the main driver of input invalidity and will most likely not affect training performance. This is likely due to sigma noise already perturbs the coordinate locations such that the child rarely overlaps the parent directly.

FactorEffect
Straight-line pathAbout 13 percentage points (100 → 87.4)
σ = 0.1 noiseAbout 5 percentage points (87.4 → 82.6)
Child inherit parent vs. triangularAbout 1 percentage point (82.6 → 82)

References

[1] Yim et al. (2023). Fast protein backbone generation with SE(3) flow matching.

[2] Bose et al. (2024). SE(3)-Stochastic Flow Matching for Protein Backbone Generation.

Dear reviewer, thank you for engaging with us constructively. In our response above, we show through a new geometric analysis that the invalid intermediates are moderately invalid and their invalidity are mainly the byproduct of the bridge construction which is crucial for our stochastic sampling objective, one to many learning, and a structure to structure generative modelling. If you have further concerns, please let us know. We would appreciate for you to raise the score if we addressed your concern.

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

RayMorph learns protein redesign from paired structures, jointly modifying coordinates, amino-acid identities, and sequence length through geometry-guided correspondence and a staged stochastic bridge. Experiments on N–C remodeling demonstrate transfer to held-out structural clusters, with designability evaluated through inverse folding and refolding.

Critical strengths

As a reviewer with a background in machine learning theory who approaches protein problems from an application perspective, I find the proposed methodology meaningful for tackling a challenging learning problem: learning structural redesign from paired endpoints while jointly changing coordinates, amino-acid identities, and sequence length. Establishing correspondence between unequal-length structures and coordinating continuous geometric updates with discrete edits require more than a straightforward application of existing generative models.

The geometry-guided, order-preserving correspondence and stochastic bridge provide a coherent approach to constructing supervision when intermediate transformation trajectories are unavailable.

A notable strength is the example-driven formulation, in which desired changes are specified through related structural pairs rather than a supplied target structure or residue-level edit plan.

The curated dataset is also a useful contribution, combining fold similarity with terminal distance and accessibility to capture requirements relevant to N–C remodeling. Evaluation on held-out structural clusters provides encouraging evidence of transfer beyond the training pairs.

Critical weaknesses

My main concerns relate to the experimental validation.

  • The headline success rates are obtained by using ProteinMPNN to design sequences for RayMorph-generated backbones. RayMorph’s own sequences achieve substantially lower success rates, so these results provide stronger evidence for backbone redesign than for joint structure–sequence generation.
  • Comparisons with adapted existing methods or simpler redesign pipelines are needed to establish the added value of the proposed framework.
  • Fold similarity and structure-prediction confidence do not directly demonstrate preserved biological function or successful internal tagging; these claims require experimental validation.
Questions for authors

Given the substantially lower success rates of RayMorph-generated sequences, how much does joint sequence modeling contribute to backbone redesign compared with a structure-only variant followed by ProteinMPNN?

3 comments

Question #1 and Weakness #1: “These results provide stronger evidence for backbone redesign than for joint structure–sequence generation.”

We agree. Please see §2 of the Overall Response for the structure-only ablation and §3 for the corresponding claim revisions.

Weakness #2: “Comparisons with adapted existing methods or simpler redesign pipelines are needed to establish the added value of the proposed framework.”

Thank you for pointing out the need for baselines to establish the added value of RayMorph. Please see §1 of the Overall Response for the three baselines and results.

Weakness #3: “Fold similarity and structure-prediction confidence do not directly demonstrate preserved biological function or successful internal tagging.”

We agree. Please see §3 of the Overall Response for the proposed revisions to these claims.

 

If we have addressed your concerns, we hope you will consider raising your score.

Dear reviewer, we would appreciate your feedback on our response to your concerns about our paper. We thank you for your thoughtful comments and kindly ask you to review our extensive new experiments on baselines, the structure-only ablation, and the bridge-state ablations addressing Reviewer SJT4’s query. We hope these additions address your concerns and support raising or maintaining your score. If you believe our work’s strengths merit publication, we would appreciate your consideration of increasing our score.

 

Regarding the concern about functional preservation claims, we have further revised the manuscript to clarify our claims as follows:

“using fold similarity as a proxy for preserving function. We note that structural similarity suggests but does not guarantee functional similarity. Validating function preservation requires wet-lab experiments that are outside the scope of this work.”

“Limitations and future directions. One of the future direction is to validate Raymorph’s in-silico designs functional preservation and suitability for internal insertion experimentally.”

 

If there are specific passages in the manuscript that you believe overstate our findings, we would appreciate your guidance so that we can revise the claims accordingly.

Thank you to the authors for conducting extensive additional experiments to address the reviewers’ concerns. In particular, the results in Table 2 of Section 4.4 of the revised manuscript address my concerns about the reliance on ProteinMPNN. I also appreciate the practical concern raised by reviewer AydV about the scarcity of experimentally characterized redesign pairs. As I mentioned in my initial review, my primary expertise is in machine learning, and I have some experience applying it to protein structure prediction. I am therefore cautious about making a definitive judgment about the practical utility of this formulation. Still, I believe that learning from paired redesign examples could be useful for protein redesign. I will maintain my original score.

Reviewer_AydVOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RayMorph, an example-driven framework for protein structure redesign. Traditional protein design methods typically rely on explicitly specified structural constraints or objectives; however, complex structural modifications often involve multiple interrelated geometric features that are difficult to describe precisely through such explicit constraints. To address this, the authors propose learning the principles of structural modification by analyzing pairs of proteins that share a similar overall fold but possess different structural properties. Specifically, RayMorph first establishes residue correspondences between proteins of varying lengths using a dynamic programming approach based on geometric information. It then constructs a variable-length stochastic bridge that jointly models residue insertions, deletions, changes in amino acid identity, and shifts in 3D coordinates, enabling the co-redesign of both protein sequences and structures. The experiments focus primarily on N–C terminal remodeling—aiming to reduce the distance between the N- and C-termini and increase the solvent accessibility of the terminal connection regions, while preserving the original protein fold as much as possible.

Critical strengths
  1. Protein design based on examples—rather than explicit rules or predefined property evaluators—represents a novel and practically significant direction.
  2. The authors have curated a dataset specifically for N–C terminal remodeling, and the experimental results demonstrate meaningful utility.
Critical weaknesses
  1. The proposed framework is evaluated only on a single specific problem; it requires demonstration across a broader range of application scenarios. Furthermore, the approach necessitates the construction of a dedicated dataset for each specific task, which entails high implementation costs. It remains unclear whether this method is applicable in scenarios where training data is scarce—a common occurrence in biological tasks—or where fewer than 10 examples exist.
  2. The paper lacks comparisons against several straightforward baselines. Given that a dataset has already been constructed, one could employ alternative training strategies—such as DPO (Direct Preference Optimization) or a two-stage approach involving the training of a reward model followed by fine-tuning. It is not clear whether the proposed method offers a significant improvement over these simpler baselines.
  3. According to the experimental analysis, designs generated by RayMorph achieve a success rate of only 4.1% when folded (using a structure prediction tool), whereas designs produced by ProteinMPNN achieve a success rate of 25.4%. The proposed method for the co-design of protein sequences and structures is problematic. The heuristic design approach employed alters the inherent relationship between a protein's sequence and its structure.
  4. While the paper demonstrates that the generated structures are capable of folding, it fails to provide clear evidence that the original functional properties are preserved; consequently, the claims made in the paper are, to some extent, overstated.
Questions for authors
  1. Require broader case studies and more baseline comparisons.
  2. The proposed method for the co-design of protein sequences and structures could be problematic.
  3. Overclaim regarding maintaining functional properties.
3 comments

Question #1 and Weakness #1: The proposed framework is evaluated only on a single specific problem; it requires demonstration across a broader range of application scenarios; require broader case studies.

We would like to highlight this as a preliminary demonstration of example-driven protein redesign, where we define a task such that its metric is meaningful and can easily be measured and observed (in contrast to other implicit metrics like thermostability, developability). In particular, since paired proteins can be of different lengths, the paper’s methodological contribution—a novel framework for sequence/structure-conditioned variable-length generation—is crucial. There are many tasks that could eventually be formulated as example pairs. Apart from the N–C remodeling task, there are other potential applications which necessitate changes in protein length, e.g., circular permutation protein design (where linkers have to be generated), miniaturisation or magnification of an existing protein, or even remodeling the length of antibody CDR loops binding to an antigen.

Weakness #1 (second part): Furthermore, the approach necessitates the construction of a dedicated dataset for each specific task, which entails high implementation costs. It remains unclear whether this method is applicable in scenarios where training data is scarce—a common occurrence in biological tasks—or where fewer than 10 examples exist.

We agree that the availability of examples is critical. In future work, we plan to explore a direction where a general-purpose redesign model is pretrained on synthetic pairs. User-provided redesign examples, even if there are only a handful of them, could serve as few-shot examples to tune on. We could also pose it as in-context learning, inspired by tabular foundation models such as TabPFN[1].

Question #1 and Weakness #2: The paper lacks comparisons against several straightforward baselines; more baseline comparisons.

Please see §1 of the Overall Response for the three simpler baselines. To your DPO suggestion: it was not immediately clear to us how DPO could be easily adapted here—it requires a variable-length generator that can handle structure-to-structure redesign across unequal lengths, and we could not find any such precedent work in the field. Our current reward-steered baseline uses a simpler, fixed-length ProteinMPNN–ESMFold loop. We appreciate the suggestion and will consider DPO and reward-based fine-tuning in future work. Also, DPO could potentially be used in conjunction with RayMorph. The latter provides variable-length generation while the former could steer towards specific objectives.

Question #2 and Weakness #3: “The proposed method for the co-design of protein sequences and structures is problematic.”

We agree that RayMorph’s generated sequences do not match the quality of ProteinMPNN-designed sequences. Please see §2 of the Overall Response for the structure-only ablation addressing the contribution of joint sequence modeling.

Question #3 and Weakness #4: “It fails to provide clear evidence that the original functional properties are preserved”; “Overclaim regarding maintaining functional properties.”

Thank you for pointing out this unintended implication. Structural preservation is relevant to retaining function, but does not guarantee it. Structural similarity and refolding consistency are commonly used for computational evaluation of protein designs, as in FrameDiff, Genie, and FoldFlow; these metrics assess structural quality rather than functional preservation. We are revising the manuscript to make this distinction explicit, as described in §3 of the Overall Response.

 

If we have addressed your concerns, we hope you will consider raising your score.

 

References

[1] Hollmann et al. (2025). Accurate predictions on small data with a tabular foundation model.

Thank you for the detailed response. While the rebuttal clarifies the scope of the contribution, my primary concerns regarding generalization, data requirements, and functional validation remain unresolved.

The N–C remodeling experiment provides an initial demonstration of example-driven, variable-length redesign. The other applications discussed in the rebuttal are plausible, but the response reports no results on a second redesign task. The evidence therefore remains limited to a single application, leaving it unclear how well the framework transfers to other settings.

The need to construct a dedicated paired dataset for each task is a practical limitation. Experimentally characterized redesign pairs are often scarce, so performance with fewer than ten examples would be particularly informative. Synthetic pretraining, few-shot adaptation, and in-context learning are reasonable directions for future work. Their feasibility in this setting remains untested.

My most consequential concern is functional preservation. The authors clarify that structural similarity and refolding consistency assess structural quality. The reported results provide no direct evidence that the redesigned proteins retain their original function. Revising the manuscript to state this distinction is appropriate. It also narrows what the paper can claim about the method’s practical utility.

I decided to lower my rating, as I do not think the paper can be accepted in the current version.

Dear reviewer, thank you for your response. We hope our reply helps clarify a potential difference in our understanding of the paper’s contribution. In light of these clarifications, we respectfully ask you to reconsider the score reduction and retain your previous rating of 2.

 

  • Functional validation: To address the most consequential concern, which is functional preservation of RayMorph’s generated candidates. As it is stated before, the paper made the best attempt of using an in-silico metric of structural similarity (TM_x0) between the RayMorph proposed structure and it's starting source protein as a reasonable proxy for functional preservation. This metric serves as an important in-silico screening criterion for prioritizing proposals for subsequent experimental validation. Experimental functional preservation claims requiring wet lab verification are completely out of scope for methods venues like ICLR and not sufficient grounds for rejection from this venue. As for the paper revision, we have clearly stated and make the distinction as follows:

“using fold similarity as a proxy for preserving function. We note that structural similarity suggests but does not guarantee functional similarity. Validating function preservation requires wet-lab experiments that are outside the scope of this work.”

“Limitations and future directions. One of the future direction is to validate Raymorph’s in-silico designs functional preservation and suitability for internal insertion experimentally.”

If there are specific passages in the manuscript that you believe overstate our findings, we would appreciate your guidance so that we can revise the claims accordingly.

 

  • Demonstrate applicability to 10-example settings: We would like to stress that the task to demonstrate performance on training with experimentally characterized redesign pairs with fewer than 10 examples is out of scope and infeasible in a machine learning setting. Training RayMorph does not require experimentally characterized redesign pairs, and its data is mainly sourced from AFDB. This curation of data can be easily extensible to other redesign settings using different prescribed transformation functions. More importantly, RayMorph has demonstrated its capability on limited data setting through the performance on NE+ unseen clusters and NE0 clusters, where these clusters have very little overlap with the training set. This demonstrates that RayMorph can still perform NC remodelling even on held out unseen protein clusters.

 

We thank you again for your time and thoughtful consideration, and hope these clarifications help resolve the remaining concerns about the scope of our contribution.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Overall response to shared concerns

We thank the reviewers for their thoughtful comments. Following our preliminary response, we address the three shared concerns below. Individual reviewer comments are addressed separately.

1. Baselines

RayMorph is trained only from example pairs, without knowledge on how it's constructed. We established 3 baselines with increasing information: the first uses the same information as RayMorph, the second additionally receives numeric objectives, and the third is given the numeric objective and uses an explicit restraint optimization. Each baseline is iterated 3 times, corresponding to the 3 proposals of RayMorph, and evaluated on NE+ using criterion (e) for structure-refolded quality. RayMorph achieves higher success than all three baselines on both NE+ sets.

  1. Paired examples, without explicit numerical objectives. This is the closest comparison in information supplied: like RayMorph, the reward model learns from source–target pairs. We employ a ProteinMPNN [1] → ESMFold [2] loop, steered by a reward function trained on aligned RMSD: f(source) = alignedRMSD(source, target), f(target) = 0. The reward model has no direct access to dNC or dExp.

  2. Numerical objectives in a general iterative pipeline. We further asked how the baseline performs with direct access to dNC and dExp in the loop, which RayMorph is not explicitly informed of.

  3. Steering Boltz-2. We use Boltz-restr [3], an explicit steering approach adapted onto Boltz-2 [4], specifying target N–C distance reduction by 6 Å. This approach requires specifying the numerical objective and expressing it through supported restraints. We note that evaluating raw Boltz-restr proposals alone gives an overly optimistic view of remodeling success. For consistent evaluation, we feed the generated structure through the same ProteinMPNN–ESMFold pipeline.

MethodInformation givenNE+ seenNE+ unseen
RayMorphpairs only22.3%25.4%
RMSD-steeredpairs only2.9%3.3%
dNC/dExp-steered+ dNC, dExp4.3%8.2%
Boltz-restr (refolded)+ explicit restraint5.5%9.8%

2. Structure-only Ablation

We agree that the headline results provide stronger evidence for backbone redesign than for joint structure–sequence generation. To explore this further, we ran a structure-only RayMorph ablation by assigning “all alanine” to source and target sequences during training. Overall, on structure-refolded criterion (e), using sequence information modestly improves backbone redesign on NE+ seen while achieving comparable performance to the structure-only formulation on NE+ unseen.

The usefulness of sequence information may be task-dependent. By formulating RayMorph in a joint sequence–structure framework, we get the advantage of generality. If structure-only generation is more suitable for a specific task, it can be accommodated via hyperparameter settings that turn off sequence-relevant training. We will add this flexibility to the codebase so users can explore the optimal generative setting for their task. For N–C remodeling, we intend to retain the use of sequence information since it modestly improves or maintain performance.

MethodNE+ seenNE+ unseen
RayMorph22.3%25.4%
Structure-only RayMorph17%26.6%

3. Calibrating the Claims

We thank the reviewers for highlighting that fold similarity and structure-prediction confidence do not establish functional equivalence. We are revising the claims in the updated manuscript. Without biochemical assays to check protein function, which are outside the scope of this paper, our evaluation is limited to structural outcomes, including source-fold preservation and refolding designability. Computational evaluation through structural metrics is commonly used in protein design/modeling papers at ML conferences, such as FrameDiff [5], Genie [6], and FoldFlow [7]. We acknowledge the limitations of these metrics and will state them explicitly. Experimental validation will be left for future work. We will describe internal tagging as a motivating application requiring experimental validation.

References

[1] Dauparas et al. (2022). Robust deep learning–based protein sequence design using ProteinMPNN.

[2] Lin et al. (2023). Evolutionary-scale prediction of atomic-level protein structure with a language model.

[3] Hori et al. (2026). Distance-Restraint-Guided Diffusion Models for Sampling Protein Conformational Changes and Ligand Dissociation Pathways.

[4] Passaro et al. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction.

[5] Yim et al. (2023). SE(3) diffusion model with application to protein backbone generation.

[6] Lin and AlQuraishi (2023). Generating Novel, Designable, and Diverse Protein Structures by Equivariantly Diffusing Oriented Residue Clouds.

[7] Bose et al. (2024). SE(3)-Stochastic Flow Matching for Protein Backbone Generation.

AuthorsOpenReview ↗

Summary of revisions and responses

We would like to thank all reviewers and their overall feedback on improvements of the paper. Generally, the reviewers agree that our problem formulation is novel and interesting, both the introduction of example-guided protein redesign and its application to a new problem (N-C remodeling). They further appreciate the new dataset that can be a community resource and its construction pipeline. They also acknowledge the methodological contribution of variable length generative modelling approach to solve the problem.

We followed the reviewers’ guidance on improved performance analysis and attribution by a) extensive baseline comparisons, b) structure-only ablations, and c) geometric analysis on the intermediate states.

A remaining pushback is from reviewer AydV who ask to validate function preservation in generated candidates. In our humble opinion, truly validating functional preservation requires wet-lab experiments which are out of scope of this paper and not sufficient grounds for rejection from a ML venue. As we discuss below, our metric of structural similarity (TM_x0) between the RayMorph-proposed structure and the starting source protein should be viewed as an in-silico proxy for function preservation. This is consistent with the broad literature that structure generally precedes function.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.