acceptodds
Desk rejected at ICLR 2027

ProteoAA: Residue-Aware Side-Chain Co-Evolution for All-Atom Protein Design

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

All-atom protein design requires coordination between backbone geometry, amino-acid identity, and side-chain packing. Yet staged design pipelines typically pass information only from backbone generation to sequence and side-chain design, leaving atomic interactions unable to guide the evolving backbone. We introduce ProteoAA, a residue-aware framework that enables bidirectional refinement through trainable adapters connecting frozen backbone and sequence–side-chain models. Backbone residue features inform sequence prediction and side-chain packing, while the resulting all-atom structure provides residue- and pair-level feedback for backbone refinement. A feedback event during diffusion incorporates this atomic information into generation, with side-chain atoms modeled conditionally on residue identity. Across unconditional protein generation, target-conditioned binder design, and side-chain packing benchmarks, Proteo-AA improves sequence–structure consistency, computational binder designability, and atomic geometry. On a standard ten-target binder benchmark, it achieves a mean designability of 31.48%, compared with 21.19% for PXDesign, and ablations support the contributions of both directions of information exchange. These results demonstrate the potential of coupling specialized pretrained models through explicit backbone–side-chain feedback for all-atom protein design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
all-atom protein design, diffusion generative model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_xcmjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • The paper proposes Proteo-AA, a bidirectional refinement framework that connects pretrained backbone generation and sequence–side-chain modeling through lightweight adapters. Side-chain atomic information is fed back into the backbone denoiser to improve structural consistency, while keeping both pretrained models frozen.

  • The framework introduces two complementary feedback pathways: B2C injects backbone residue features into FAMPNN to guide sequence and side-chain generation, while C2B uses residue-level and pairwise atomic information to refine backbone predictions. A single feedback event is performed during diffusion to balance structural refinement and computational cost.

  • Proteo-AA is evaluated on unconditional protein generation, target-conditioned binder design, and side-chain packing. It achieves a mean binder designability of 31.48% across ten targets, compared with 21.19% for PXDesign, and improves several side-chain packing metrics. Ablation studies support the contributions of both feedback directions, although comparisons with other iterative refinement methods would further clarify its methodological novelty.

Critical strengths
  • Consistent performance improvements over strong baselines. Proteo-AA achieves promising results across unconditional protein generation, target-conditioned binder design, and side-chain packing. In Table 1, it reaches 82.2% all-atom codesignability and 92.6% designability with ProteinMPNN redesign (PMPNN@1). In Table 2, it achieves a mean binder designability of 31.48% across ten targets, outperforming PXDesign (21.19%) and A-CODE with ProteinMPNN redesign (20.98%). These results demonstrate the potential of incorporating bidirectional backbone–side-chain feedback to improve protein design quality.

  • Effective bidirectional feedback supported by ablation studies. The ablation results in Table 2 show that removing the C2B adapter reduces mean binder designability from 31.48% to 25.14%, while removing both adapters reduces it to 24.19%. In Table 3, Proteo-AA also achieves the best results on six of seven reported side-chain packing metrics when evaluated on PXDesign backbones. These findings support the contribution of atomic-level feedback to backbone refinement and suggest that the proposed coupling improves both backbone–side-chain compatibility and downstream design performance.

Critical weaknesses
  • Lack of analysis on the number of feedback iterations. Proteo-AA performs a single feedback event during backbone diffusion, where FAMPNN generates an intermediate all-atom structure that is used to refine the backbone. However, the paper does not investigate how the number of feedback iterations affects performance. Although the ablation studies examine the contributions of the B2C and C2B adapters, they do not compare different numbers of feedback iterations. Evaluating multiple iterations would help clarify whether additional refinement leads to further improvements or introduces diminishing returns, while also quantifying the associated computational cost.

  • Lack of direct comparisons with existing iterative refinement methods. Although the paper compares Proteo-AA with multiple strong baselines across unconditional protein generation, binder design, and side-chain packing, it does not directly compare the proposed feedback mechanism with prior protein design approaches that explicitly employ iterative structural refinement, such as FAIR. Such comparisons would help establish whether the performance gains arise specifically from the proposed bidirectional adapter coupling, rather than from the general benefit of iterative refinement. They would also provide a clearer assessment of the method's novelty and practical advantages over existing approaches.

Questions for authors
  • It would be interesting to investigate whether organizing PXDesign and FAMPNN into multiple alternating blocks within a unified iterative framework could further improve performance. In particular, comparing different numbers of refinement rounds would help determine whether repeated backbone–side-chain interactions yield additional gains over the single feedback event used in Proteo-AA, and how these gains trade off against computational cost.

  • Could the authors compare Proteo-AA with at least one existing protein design method that employs iterative structural refinement or coupled optimization, such as FAIR? This would help establish whether the proposed bidirectional adapter mechanism offers advantages over existing iterative approaches in terms of design quality, structural consistency, and computational efficiency.

2 comments

[Weakness 1, Question 1]: Effect of multiple feedback iterations on designability and computational cost

Thank you for this suggestion on investigating how the number of feedbacks affect performance. We conducted additional experiments varying the number and timing of feedback events during backbone diffusion, evaluating whether repeated backbone–side-chain interactions improve binder designability. Our results show that feedback effectiveness depends more on timing than frequency. The best single-event configuration at step 350 achieves 31.48% mean designability, compared with 25.14% without C2B feedback. Adding an earlier feedback event (step 303) or a later event (step 395) to the default configuration reduces designability to 25.46% and 28.89%, respectively. In contrast, adding an event near the default step (steps 347 and 350) yields 31.18%, nearly identical to the single-event result. Overall, multiple feedback events provide little additional benefit, suggesting that appropriately timed feedback is more important than increasing the number of interactions.

Feedback configurationEventsMean designability (%)
B2C only, without C2B025.14
Step 303126.07
Step 350 (Default)131.48
Step 395127.62
Steps 303 + 350225.46
Steps 347 + 350231.18
Steps 350 + 395228.89
Steps 347 + 350 + 396328.78

Feedback effectiveness depends on backbone maturity more than feedback events. Earlier diffusion stages produce less resolved backbones, leading FaMPNN to generate potentially unreliable side-chain configurations. Feeding these premature structures back into backbone refinement may introduce misleading structural information that subsequent feedback cannot fully correct. A very late feedback left very few steps for the backbone module to correct results.

We additionally measured complete generation runtime, including backbone generation and all feedback operations:

Feedback roundsDecode (s)Full generation (s)Relative runtime
110.9619.551.00x
221.1629.761.52x
332.3040.672.08x

Increasing feedback from one to three rounds approximately doubles total generation time without corresponding improvements in designability. Together, these findings suggest that carefully selecting feedback timing is more effective than increasing feedback frequency. We will incorporate those results in our final version together with the full Proteo-AA table in the final revision.

[Weakness 2, Question 2]: Comparison with existing iterative refinement methods

Thank you for this suggestion in comparison with FAIR.

To assess iterative refinement baselines, we evaluated the official pretrained FAIR checkpoint as a downstream refinement stage following PXDesign and FaMPNN across all ten AlphaProteo10 targets, conditioning on local target-interface atoms. In total, we generated 1,000 runs by refining 20 candidate structures per target using five independent random seeds. Post-refinement mean designability reached only 0.3%, yielding just a single successful design for IR and two for PDL1. Further geometric inspection revealed that 90.87% of design-region bonds diverged by over 0.25 Å from standard ideal geometry. Because FAIR was originally trained for small-molecule-conditioned pocket generation, these experiments reflect its zero-shot transfer performance on binder design without task-specific fine-tuning, rather than an upper bound on its potential after retraining. We will include these FAIR benchmarks and structural metrics in the final revision, noting the boundaries of zero-shot transfer while expanding our discussion of related iterative optimization frameworks.

Dear Reviewer,

Thank you again for your thoughtful feedback and constructive suggestions. We would greatly appreciate it if you could take another look at our responses and let us know whether they adequately address your concerns. We would be happy to provide further clarification or discussion if needed.

Thank you for your time!

Reviewer_RVPJOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ProteoAA, a framework that couples a pretrained backbone generator, PXDesign, with a sequence–side-chain model, FAMPNN, through trainable bidirectional adapters while keeping the base models frozen. The backbone-to-side-chain adapter transfers denoiser features to sequence prediction and side-chain generation. The side-chain-to-backbone adapter converts the resulting packed structure into residue- and pair-level feedback, enabling a corrective backbone evaluation at the same noisy state and noise level. Feedback is applied once during sampling, followed by final sequence and side-chain generation. Experiments cover unconditional protein generation, target-conditioned binder design, and side-chain reconstruction. The strongest result is an increase in mean computational binder designability from 21.19% for PXDesign to 31.48%.

Critical strengths
  1. A well-motivated and practical mechanism for incorporating atomistic information into backbone generation. The paper addresses a meaningful limitation of staged design pipelines: side-chain compatibility cannot influence an already generated backbone. The proposed adapters provide an explicit reverse information pathway without retraining the base models.

  2. Promising binder-design results with useful internal controls. The mean computational designability improves from 24.19% without either adapter to 31.48% with both adapters, supporting a meaningful contribution beyond the underlying pipeline.

Critical weaknesses
  1. The ablations do not sufficiently establish that side-chain geometry drives the improvement. C2B receives backbone coordinates, residue identities, side-chain geometry, confidence information, and, in its late branch, learned MPNN representations. Removing the entire adapter establishes the usefulness of this combined feedback, but does not isolate the contribution of explicit side-chain interactions, which is central to the paper’s motivation.

  2. The unconditional generation results lack a matched internal baseline. Table 1 compares Proteo-AA against external methods, but does not include the same PXDesign–FAMPNN pipeline with both adapters disabled under matched sampling and decoding budgets. Consequently, the results establish the competitiveness of the complete system but do not isolate the new interfaces’ contribution to unconditional generation. This is particularly relevant because the all-atom codesignability improvement over A-CODE is modest (82.2% versus 81.2%) .

  3. The evidence for improved atomic geometry needs reconciliation and clearer scope. In Table 3, Proteo-AA’s Bad Bond fraction increases from 0.0009 on PXDesign backbones to 0.0502 on corrected backbones, despite improvements in reconstruction metrics. This substantial change requires explanation before drawing broad conclusions about atomic quality. Moreover, the packing benchmark uses native sequences and denoised, artificially corrupted native backbones; it therefore supports reconstruction under backbone uncertainty more directly than atomic quality in de novo binder generation.

Questions for authors
  1. Contribution of side-chain geometry. Could the authors provide a matched C2B ablation that retains sequence and backbone conditioning while excluding explicit side-chain geometry and all representations derived from it? This comparison would help establish whether atomistic information provides additional benefits beyond the other feedback signals.

  2. Unconditional generation baseline. How does Proteo-AA compare with the same PXDesign–FAMPNN pipeline with both adapters disabled, under matched backbone-evaluation and sequence-decoding budgets? This would help isolate the adapters’ contribution to unconditional generation.

  3. Packing evaluation: What explains the increase in Proteo-AA’s Bad Bond fraction from 0.0009 on PXDesign backbones to 0.0502 on Proteo-AA–corrected backbones in Table 3? Please clarify whether the same metric definition and post-processing procedures are used in both settings. Is the supplied backbone held fixed for all packing methods, including Proteo-AA, with no additional C2B-based backbone correction?

  4. Scope of the ablation claims. Does the statement that the full model “consistently outperforms” its ablations refer to mean designability across targets? Table 2 contains individual targets on which ablated variants perform better, such as IR and TNFa. Please qualify the statement accordingly.

7 comments

[Weakness 1, Question 1]: Contribution of side-chain geometry to C2B feedback

Thank you for highlighting this distinction separating out the effect of side-chain geometry

We conducted an additional geometry-free C2B ablation that retains sequence and backbone conditioning while excluding explicit side-chain geometry and representations derived from it.

Our results show that side-chain geometry provides additional benefits beyond sequence and backbone conditioning. The geometry-free C2B variant achieves 27.09% mean binder designability, compared with 25.14% for B2C only and 31.48% for full Proteo-AA. While geometry-free feedback improves designability by 1.95 percentage points, incorporating side-chain geometry further improves performance by 4.39 percentage points.

ConfigurationMean designability (%)
Neither adapter24.19
B2C only, without C2B25.14
B2C + geometry-free C2B27.09
C2B only, without B2C29.59
Full Proteo-AA31.48

These findings support the contribution of explicit side-chain geometry to backbone refinement, beyond the improvements achievable through sequence and backbone conditioning alone. We will include this additional ablation in our revision and include the full AlphaProteo Table along with all other ablations to better isolate the role of atomistic information in C2B feedback.

[Weakness 2, Question 2]: Matched internal baseline for unconditional generation

Thank you for highlighting this missing comparison on unconditional generation.

We conducted additional unconditional generation experiments using the same PXDesign–FaMPNN pipeline with both B2C and C2B adapters disabled, generating 100 samples at each length L{100,200,300,400,500} under the same budget.

Our results show that our architecture substantially improves unconditional generation over the matched internal baseline. The adapter-free pipeline achieves 65.4% all-atom codesignability, 73.0% Cα codesignability, and 78.0% PMPNN@1 designability. In comparison, full Proteo-AA achieves 82.2%, 89.4%, and 92.6%.

ConfigurationAll-atom codesignability (%)Cα codesignability (%)PMPNN@1 (%)
PXDesign + FaMPNN, both adapters disabled65.473.078.0
Full Proteo-AA82.289.492.6

These results demonstrate that the improvements are not solely attributable to combining PXDesign and FaMPNN, but also reflect the benefits of the proposed bidirectional adapter architecture under the evaluated sampling setting.

We will include this matched internal baseline in Table 1 of our revised manuscript to better establish the contribution of our proposed interfaces.

[Weakness 3, Question 3]: Correction of Bad Bond Calculation

Thank you for identifying this inconsistency on bad bonds.

Upon investigation, we found an implementation mistake in the Bad Bond metric calculation for the Proteo-AA–corrected backbone evaluation, and we apologize for the incorrect reporting.

Bond lengths should be calculated within each predicted structure, independently of the native reference. However, our binder-scoring implementation incorrectly combined predicted side-chain coordinates with native backbone coordinates. This inadvertently counted backbone displacement relative to the native structure as bond-length violations, artificially inflating the reported Bad Bond fractions on Proteo-AA–corrected backbones.

We corrected the calculation to measure bond lengths consistently within each predicted structure. The corrected results are shown below:

MethodCorrected Bad Bond fraction ↓
FaMPNN0.00043
LigandMPNN0.00037
PackPPI0.00006
Proteo-AA0.00006

We separately reverified the Bad Bond results on the original PXDesign backbones and confirmed that those values are correct and remain unchanged. We also audited the remaining reported metrics to verify their correctness.

We further clarify that the supplied backbone is held fixed for all packing methods, including Proteo-AA, and no additional C2B-based backbone correction is performed during packing evaluation. "Proteo-AA–corrected backbones" refers to the backbone structures supplied as inputs to the packing methods.

The corrected results confirm that Proteo-AA achieves a low Bad Bond fraction, matching the best-performing baseline in this evaluation. We will update the manuscript with the corrected values and appropriately qualify our claims.

[Question 4]: Scope of the adapter ablation claims

Thank you for pointing out the ambiguous wording.

We clarify that our statement that the full model "consistently outperforms" its ablations refers to mean binder designability across the ten targets, not improvements on every individual target.

We will revise the statement to "Proteo-AA achieves higher mean designability across the ten targets than each adapter-removal variant" to accurately reflect the scope of our findings. We will incorporate this change in the final revision.

Dear Reviewer,

Thank you again for your thoughtful feedback and constructive suggestions. We would greatly appreciate it if you could take another look at our responses and let us know whether they adequately address your concerns. We would be happy to provide further clarification or discussion if needed.

Thank you for your time!

Reviewer_7ToWOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Proteo-AA, a framework that couples backbone diffusion model PXDesign with FAMPNN through lightweight bidirectional adapters. The method allows backbone representations to guide sequence and side-chain generation, while the resulting all-atom structure provides feedback to refine the backbone during sampling. Proteo-AA performs a single feedback event during backbone diffusion and then regenerates the final sequence and side chains on the optimized backbone. Experiments show that ProTeo-AA outperforms existing protein design method over several baselines.

Critical strengths
  1. The paper presents a clear and easy way to connect backbone diffusion model with sequence and side-chain design model. Both models are frozen and only two adapters are trained, making the method lightweight. And the idea of allowing side-chain packing information to feed back to backbone generation is intuitive and well motivated.
  2. The experiments are thorough and the model performs great across different tasks.
Critical weaknesses
  1. The method is motivated as bidirectional backbone–side-chain refinement and the authors use words like "Co-Evolution", which will make people think that the backbone design and sequence design might contact each other frequently and evolve at the same time. But only a single feedback event is performed during the 400-step diffusion process. It would be useful to evaluate whether repeated feedback provides additional benefits. Also the word "Co-Evolution" might be an established term in protein biology, maybe the authors can rephrase that.
  2. The feedback step is fixed at σt\sigma_t=0.429. The paper studies how B2C behaves at different noise levels, but does not provide an end-to-end ablation showing how the choice of feedback timing affects final design quality.
Questions for authors

I am curious about multiple feedbacks and different noise level feedbacks performance.

3 comments

[Weakness 1]: Clarification of "Co-Evolution" terminology

Thank you for this observation and pointing it out.

We agree that the term "Co-Evolution" may suggest more frequent backbone–side-chain interactions than our current method performs and may be confused with its established meaning in protein biology.

We will replace this terminology with "bidirectional refinement" referring to backbone-to-side-chain conditioning and side-chain-to-backbone feedback. We will also explicitly clarify that the default Proteo-AA sampler performs a single feedback event during the 400-step diffusion process. We will revise the corresponding descriptions to more accurately reflect our method.

[Weakness 1, Weakness 2, Question 1]: Effect of feedback timing and multiple feedback events

Thank you for these suggestions investigating feedback timing and events count.

We conducted additional end-to-end experiments varying the number and timing of feedback events during backbone diffusion to investigate how feedback timing and repeated backbone–side-chain interactions affect binder designability.

Our results show that feedback effectiveness depends more on timing than frequency. The best single-event configuration at step 350 achieves 31.48% mean designability, compared with 25.14% without C2B feedback. Adding an earlier feedback event (step 303) or a later event (step 395) to the default configuration reduces designability to 25.46% and 28.89%, respectively. In contrast, adding an event near the default step (steps 347 and 350) yields 31.18%, nearly identical to the single-event result. Overall, multiple feedback events provide little additional benefit, suggesting that appropriately timed feedback is more important than increasing the number of interactions.

Feedback configurationEventsMean designability (%)
B2C only, without C2B025.14
Step 303126.07
Step 350 (Default)131.48
Step 395127.62
Steps 303 + 350225.46
Steps 347 + 350231.18
Steps 350 + 395228.89
Steps 347 + 350 + 396328.78

We observe that feedback effectiveness depends on backbone maturity rather than the number of feedback events. At earlier diffusion stages, the backbone remains insufficiently resolved, potentially leading FaMPNN to generate unreliable side-chain configurations. Feeding these immature structures back into backbone refinement may introduce misleading structural information that subsequent feedback cannot fully correct. Conversely, feedback introduced too late leaves insufficient diffusion steps for the backbone module to incorporate the corrective information effectively.

Within each feedback event, the provisional and corrective backbone evaluations use identical noisy coordinates and noise levels. Therefore, varying the feedback step allows us to evaluate how feedback timing along the diffusion trajectory affects final designability.

We additionally measured complete generation runtime, including backbone generation and all feedback operations:

Feedback roundsDecode (s)Full generation (s)Relative runtime
110.9619.551.00x
221.1629.761.52x
332.3040.672.08x

Increasing feedback from one to three rounds approximately doubles total generation time without corresponding improvements in designability.

Together, these findings suggest that carefully selecting feedback timing is more effective than increasing feedback frequency. We will include these additional results in our revision to clarify the effects of feedback timing and repeated refinement, and to more accurately state the role of bidirectional interactions in Proteo-AA.

Dear Reviewer,

Thank you again for your thoughtful feedback and constructive suggestions. We would greatly appreciate it if you could take another look at our responses and let us know whether they adequately address your concerns. We would be happy to provide further clarification or discussion if needed.

Thank you for your time!

I really appreciate the authors' response. It addresses my original questions.

However, the results also suggest that C2B feedback is quite sensitive to timing: feedback earlier or later substantially reduces the gain, while additional feedback events do not help and can even hurt. This raises some concern about whether the improvement is robust or depends on a narrowly selected operating point.

I will therefore keep my current rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers

We sincerely thank all reviewers for their constructive feedback and recognition of Proteo-AA's contributions. We summarize the additional experiments conducted to address their concerns.

Terminology, methodological clarification, and scope of claims (7ToW, RVPJ).

  • Terminology clarification (7ToW): We agree that "Co-Evolution" may imply simultaneous backbone–side-chain updates. We will replace it with "Bidirectional Refinement," referring to backbone-to-side-chain conditioning and side-chain-to-backbone feedback.
  • Side-chain packing evaluation (RVPJ): We clarify that the supplied backbones remain fixed during side-chain packing without additional C2B correction.
  • Scope of performance claims (RVPJ): We will revise "consistently outperforms" to refer specifically to higher mean designability across targets.
  • Atomic geometry claims (RVPJ): We will qualify claims regarding atomic geometry to reflect that the packing evaluation primarily measures reconstruction under backbone uncertainty, not de novo binder quality.

Feedback timing and frequency (7ToW, xcmj).

  • Feedback timing (7ToW): We conducted additional end-to-end experiments varying the feedback step. The default step 350 achieves 31.48% mean designability, compared with 26.07% at step 303 and 27.62% at step 395, suggesting that feedback effectiveness depends on backbone maturity.
  • Feedback frequency (7ToW, xcmj): We evaluated multiple feedback events. Adding feedback at steps 347 + 350 achieves 31.18%, nearly identical to the default single-event configuration, while other multi-event configurations perform worse. These results suggest that increasing feedback frequency provides little additional benefit.
  • Computational cost (xcmj): Increasing from one to three feedback events raises complete generation time from 19.55 s to 40.67 s (2.08×). But we do not observe corresponding improvements in mean designability.

Contribution of side-chain geometry (RVPJ).

We introduced a C2B control that retains sequence and backbone conditioning while excluding explicit side-chain geometry and its derived representations. Full Proteo-AA achieves 31.48% mean designability, compared with 27.09% for geometry-free C2B and 25.14% without C2B, supporting the additional contribution of atomistic information.

Matched internal baseline for unconditional generation (RVPJ).

We evaluated the same PXDesign–FaMPNN pipeline with both adapters disabled under the same budget. Full Proteo-AA improves all-atom codesignability from 65.4% to 82.2%, Cα codesignability from 73.0% to 89.4%, and PMPNN@1 from 78.0% to 92.6%, supporting the contribution of our method.

Correction of Bad Bond calculation (RVPJ).

  • Identified error: The original implementation incorrectly combined predicted side-chain coordinates with native backbone coordinates when evaluating Proteo-AA–corrected backbone inputs.
  • Corrected results: Computing bond lengths within each predicted structure yields Bad Bond fractions of 0.00043 (FaMPNN), 0.00037 (LigandMPNN), 0.00006 (PackPPI), and 0.00006 (Proteo-AA). The corrected results still hold the previous trend.
  • Consistency verification: We separately reverified the original PXDesign-backbone results and all other metric calculations in this section. All of them remain unchanged. We will correct the affected values in Table 3 and Appendix Table 5.

Comparison with existing iterative refinement methods (xcmj).

We evaluated the official pretrained FAIR checkpoint as a downstream refinement stage for binder design, obtaining 0.3% mean designability in zero-shot transfer. We acknowledge FAIR's original task setting and will clarify that this experiment evaluates zero-shot transfer.

We sincerely thank the reviewers again for their valuable feedback and suggestions.

We will incorporate the corrections and wording clarifications into the main manuscript. All newly conducted experiments and detailed analyses are currently fit in appendix. Here are the specific changes we make (line numbers refer to the original submission):

  • L001–L002, Title: Replace "Co-Evolution" with "Bidirectional Refinement" to avoid implying repeated or simultaneous backbone–side-chain updates, addressing W1 raised by Reviewer 7ToW.
  • L015–L026, Abstract: Qualify claims regarding improvements in atomic geometry to reflect the scope of the packing evaluation, addressing W3 raised by Reviewer RVPJ.
  • L259–L267, Section 2.4: Reference the new end-to-end feedback-timing experiments in the appendix, addressing W2 raised by Reviewer 7ToW.
  • L349–L365, Section 3.1: Reference the adapter-disabled PXDesign–FaMPNN baseline in the appendix, addressing W2/Q2 raised by Reviewer RVPJ.
  • L394–L418, Section 3.2: Reference the additional feedback-timing, feedback-frequency, and geometry-free C2B ablations in the appendix, addressing W1/W2 raised by Reviewer 7ToW, W1/Q1 raised by Reviewer xcmj, and W1/Q1 raised by Reviewer RVPJ.
  • L415–L418, Section 3.2: Revise "consistently outperforms" to explicitly refer to higher mean designability across the ten targets rather than superiority on every individual target, addressing Q4 raised by Reviewer RVPJ.
  • L433–L437 and L450–L474, Table 3 caption and Section 3.3: Clarify that all packing methods use fixed input backbones without additional C2B correction and distinguish reconstruction under backbone uncertainty from de novo binder atomic quality, addressing W3/Q3 raised by Reviewer RVPJ.
  • L480–L485, Conclusion: Qualify claims regarding atomic geometry and performance consistency to reflect the evidence presented, addressing W3/Q4 raised by Reviewer RVPJ.
  • L433–L446, Table 3: Correct the Bad Bond fractions for Proteo-AA–corrected backbone inputs to 0.00043 (FaMPNN), 0.00037 (LigandMPNN), 0.00006 (PackPPI), and 0.00006 (Proteo-AA), addressing W3/Q3 raised by Reviewer RVPJ.
  • L648–L678, Appendix A (Related Work): We will add a discussion of FAIR's iterative, ligand-conditioned design and distinguish it from Proteo-AA's single-event backbone–side-chain feedback, addressing W2/Q2 raised by Reviewer xcmj.
  • L865–L879, Appendix C, Table 5: Update the corresponding Bad Bond values in the detailed binder-only packing results to maintain consistency with Table 3, addressing W3/Q3 raised by Reviewer RVPJ.
  • New Appendix E, Feedback timing: Add end-to-end designability comparisons across different feedback steps, addressing W2 raised by Reviewer 7ToW.
  • New Appendix E, Feedback frequency: Add single-, double-, and triple-event feedback experiments to evaluate the effectiveness of repeated refinement, addressing W1/Q1 raised by Reviewers 7ToW and xcmj.
  • New Appendix E, Computational cost: Report complete generation runtimes across feedback configurations to quantify the trade-off between refinement frequency and computational cost, addressing W1/Q1 raised by Reviewer xcmj.
  • New Appendix E, Geometry-free C2B: Include a matched ablation isolating the contribution of explicit side-chain geometry, addressing W1/Q1 raised by Reviewer RVPJ.
  • New Appendix E, Unconditional generation baseline: Include the adapter-disabled PXDesign–FaMPNN comparison under the evaluated sampling setup, addressing W2/Q2 raised by Reviewer RVPJ.
  • New Appendix E, Complete AlphaProteo-10 table: Provide the complete per-target results for the additional feedback and geometry-free configurations to support a more detailed analysis of the ablations, addressing concerns raised by Reviewers 7ToW, RVPJ, and xcmj.
  • New Appendix F, FAIR experimental setup: Describe the official pretrained FAIR checkpoint, downstream refinement protocol, and target-interface conditioning, addressing W2/Q2 raised by Reviewer xcmj.
  • New Appendix F, FAIR results: Report zero-shot binder designability and structural geometry measurements, addressing W2/Q2 raised by Reviewer xcmj.
  • New Appendix F, FAIR limitations: Discuss FAIR's original task setting, the zero-shot transfer mismatch, and the limitations of interpreting this experiment as a directly matched comparison, addressing W2/Q2 raised by Reviewer xcmj.

These revisions will strengthen the empirical support for our method and provide a more precise description of our work. We will await any further comments or suggestions and attach the final revised PDF.

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

Related papers

Discussion (0)

Sign in to comment.