acceptodds
Under review as a conference paper at ICLR 2027

Multimodal Diffusion over Sequence and SE(3) Manifolds for Protein-Protein Complex Design

Abstract

Designing high-affinity target-specific protein binders is critical to modern therapeutics. Existing protein generation methods are unsatisfactory on this task because many of their designs fail to achieve joint geometric and physicochemical complementarity at the binding interfaces. To address this, we propose PPFrame, a physicochemically enhanced multimodal diffusion method for protein-protein complex design. The core idea is to co-design a target-conditional protein sequence and structure by jointly diffusing over the discrete sequence space and the continuous structural manifold. We also introduce IPAFormer to serve as the core network inside the diffusion. IPAFormer's architectural insight is its interleaving of geometry-based invariant point attention and global sequence self-attention. We curate PPIBench, a dataset of 602,364 protein-protein complexes across diverse target scaffolds from the Protein Data Bank (PDB). Our PPFrame is first pretrained on PPIBench and then finetuned on two important real-world tasks: mini-binder design and antibody design. PPFrame consistently outperforms the best baselines across the three tasks, improving design success rates by 17.8%, 8.0%, and 1.9%, respectively. Notably, PPFrame generates a candidate broadly neutralizing antibody (bNAb) against HIV-1 with an ipTM of 0.81 and a binding energy of -52.1, highlighting its strong utility for real-world biotherapeutic discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein complex design, protein-binding protein design, protein complex sequence and structure co-design

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_1PA3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PPFrame, a multimodal diffusion framework for protein–protein complex design, aiming to address the insufficient sequence–structure consistency in existing binder design methods and the difficulty of simultaneously satisfying binding interface geometry and physicochemical complementarity.

Specifically, PPFrame combines discrete amino acid sequence diffusion with continuous SE(3) structure diffusion to enable joint sequence–structure generation conditioned on the target protein. The authors designed IPAFormer, which alternates Invariant Point Attention and Transformer Self-Attention to integrate 3D geometric information with sequence dependencies, and introduces residue charge and hydrophobicity features as well as a self-conditioning inference mechanism to improve generation quality.

In addition, the authors constructed PPIBench from the PDB, containing 602,364 oriented complex samples, for model pretraining, and further fine-tuned the model on Mini-binder and antibody design tasks. Experimental results show that PPFrame achieves overall better performance than the selected baseline methods across the three tasks in terms of computational success rate, predicted binding energy, and sequence–structure consistency. The paper also uses ablation experiments to analyze the roles of components such as physicochemical features, self-conditioning inference, joint generation, and auxiliary losses.

Critical strengths

The first strength is well-motivated and coherent methodological design.The paper addresses an important challenge in protein-protein complex design by jointly modeling protein sequences and structures. The integration of discrete sequence diffusion, SE(3) structural diffusion, and physicochemical features provides a reasonable framework for improving sequence-structure consistency and binding interface complementarity.

The second strength is consistent empirical improvements across multiple protein design tasks. The proposed method achieves competitive performance across general protein complex design, mini-binder design, and antibody design, outperforming the evaluated baselines in average success rates and predicted binding energies. Notably, PPFrame demonstrates strong sequence-structure consistency, achieving 99.39% co-designability in antibody design. These results provide empirical evidence for the effectiveness of the proposed framework across different protein design scenarios.

Critical weaknesses

The first weakness is lack of direct comparison with the most closely related work, PPDiff. PPFrame and PPDiff address highly similar research problems and share substantial similarities in their experimental settings. In particular, their mini-binder and antibody benchmarks have the same data sources and identical dataset split statistics. However, PPFrame does not include PPDiff as a baseline or provide a direct comparison under a unified evaluation protocol. Since the reported success rates and candidate generation settings are not fully consistent across the two papers, the results reported separately cannot establish the relative performance of the two methods. This makes it difficult to assess whether PPFrame provides a meaningful improvement over existing joint sequence-structure diffusion approaches. A direct comparison using consistent data splits, evaluation metrics, and candidate generation budgets would substantially strengthen the paper's claims regarding its methodological contributions.

The second weakness is insufficient ablation studies to validate the key methodological contributions.The paper presents several ablation studies examining physicochemical features, self-conditioning, auxiliary losses, and joint sequence-structure generation. However, the two central methodological components emphasized by the authors, namely SE(3) structural modeling and the IPAFormer architecture, are not independently evaluated through controlled comparisons. More targeted ablation studies are needed to substantiate the claimed methodological advances.

The third weakness is insufficient code documentation and reproducibility support. The anonymous code repository does not appear to provide sufficient documentation and environment specifications to facilitate reproduction of the reported results. In particular, the lack of clear setup instructions, dependency specifications, and a well-documented workflow for data preparation, training, and evaluation makes it difficult to determine whether the proposed method and experimental results can be independently reproduced.

Questions for authors

Question1:Comparison with PPDiff. PPFrame and PPDiff address highly similar research problems and share substantial similarities in their experimental settings. In particular, their mini-binder and antibody benchmarks have the same data sources and identical reported dataset split statistics. However, PPFrame does not include PPDiff as a baseline or provide a direct comparison under a unified evaluation protocol. Since the reported success rates and candidate generation settings differ between the two papers, the results reported separately do not allow a fair assessment of their relative performance. Could the authors clarify whether the exact same sample-level training, validation, and test splits are used in both works? More importantly, could they provide a direct comparison with PPDiff using consistent evaluation metrics, candidate generation budgets, and selection strategies? Such a comparison would help establish whether PPFrame offers meaningful improvements over existing joint sequence-structure diffusion methods.

Question2: Ablation Studies on the Core Methodological Components. The paper emphasizes SE(3) structural modeling and the IPAFormer architecture as key methodological contributions. However, the current ablation studies do not independently evaluate their contributions. Specifically, I would like to ask the following questions: What is the performance gain from SE(3) frame-based diffusion compared with structure diffusion based solely on Cα coordinates? Can the authors quantify the contribution of this change under otherwise identical experimental settings? What is the independent contribution of IPAFormer compared with existing denoising architectures? Can the authors provide controlled comparisons to demonstrate the benefits of the proposed architecture? What are the individual contributions of IPA, Transformer self-attention, and ESM-2 embeddings? Can the authors conduct component-wise ablation studies to clarify how each component contributes to the overall performance improvements?

3 comments

Thanks for reviewing our paper and giving detailed feedback. Our responses are listed below:

Weakness1 and Q1: The first weakness is lack of direct comparison with the most closely related work, PPDiff. A direct comparison using consistent data splits, evaluation metrics, and candidate generation budgets would substantially strengthen the paper's claims regarding its methodological contributions. Could the authors clarify whether the exact same sample-level training, validation, and test splits are used in both works? More importantly, could they provide a direct comparison with PPDiff using consistent evaluation metrics, candidate generation budgets, and selection strategies?

Ans: We thank the reviewer for highlighting this connection and suggesting a direct comparison. We would like to first emphasize that PPFrame introduces substantive changes beyond replacing EGCL with IPA: a full-backbone representation (four atoms) based on residue frames, explicit physicochemical feature guidance, a more complex diffusion process, auxiliary structural supervision, and self-conditioned inference. Our ablations support the contributions of physicochemical features, auxiliary losses, and self-conditioned inference, providing evidence that the improvements extend beyond the choice of geometric attention layer.

Moreover, PPBench provides Cα-only coordinates, consistent with PPDiff’s Cα-only representation. In contrast, as described in Section 3.1 Problem Definition, PPFrame models backbone geometry involving N, Cα, C, and O atoms and therefore requires these coordinates for training and evaluation. Therefore, we curated PPIBench to support PPFrame training, leading to a different training/validation/test split between these two datasets.

To directly address the reviewer’s request, we trained and evaluated PPDiff on PPIBench. As shown below, PPFrame achieves a higher sequence design success rate (47.83% versus 28.26%) and more favorable binding energy (−67.80 versus −28.16). The largest differences concern structural fidelity: PPFrame achieves 65.22% on both designability and co-designability, compared with 0% for PPDiff under this evaluation. We also observed frequent steric clashes in PPDiff’s generated structures, consistent with its lower structural scores. The results demonstrate PPFrame’s advantage in producing designs that jointly satisfy sequence-quality and structural-consistency criteria. We have updated the results in Section5.9 and Table6, as well as the discussion with PPDiff in Appendix Section D in our manuscript.

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
PPDiff0.56881.3728.26%87.63%88.91%-28.160%0%
PPFrame0.59287.1847.83%33.22%57.49%-67.8065.22%65.22%

Weakness2: The second weakness is insufficient ablation studies to validate the key methodological contributions. However, the two central methodological components emphasized by the authors, namely SE(3) structural modeling and the IPAFormer architecture, are not independently evaluated through controlled comparisons. More targeted ablation studies are needed to substantiate the claimed methodological advances.

Ans: We thank the reviewer for emphasizing the need for targeted ablations. To examine the contributions of the SE(3) structural modeling in IPAFormer, we trained a variant that removes the IPA layers, and evaluated it on the general protein-protein complex design task.

Removing IPA layers degrades sequence design quality. The resulting designs exhibit lower interface and structural confidence, with ipTM decreasing from 0.592 to 0.326 and pLDDT from 87.18 to 72.18. The sequence design success rate decreases from 47.83% to 0%, supporting the importance of geometric information for sequence generation. We have updated the results in Appendix Section C.5 and Table25 in our manuscript.

ipTMpLDDTSR
— Without IPA layers0.32672.180%
PPFrame0.59287.1847.83%

Weakness3: The third weakness is insufficient code documentation and reproducibility support. In particular, the lack of clear setup instructions, dependency specifications, and a well-documented workflow for data preparation, training, and evaluation makes it difficult to determine whether the proposed method and experimental results can be independently reproduced.

Ans: We thank the reviewer for highlighting this issue and agree that clear documentation is essential for reproducibility. We have added a detailed README with environment setup instructions and step-by-step guidance for training and evaluation. These additions clarify how to run the code and support independent reproduction of the reported experiments.

Q2.1: What is the performance gain from SE(3) frame-based diffusion compared with structure diffusion based solely on Cα coordinates? Can the authors quantify the contribution of this change under otherwise identical experimental settings?

Ans: According to the reviewer’s suggestion, we constructed a Cα-only variant of PPFrame by replacing the IPA layers with equivariant graph convolutional layers and adapting the two auxiliary losses to the Cα representation. We trained and evaluated this variant on PPIBench.

Compared with the Cα-only variant, PPFrame improves sequence design success rate by 21.75 percentage points. The largest gains occur in structural designability and co-designability, which increase by 54.35 and 56.53 percentage points, respectively. These results support the benefits of PPFrame’s full-backbone modeling framework, particularly for structural fidelity and sequence-structure consistency. We have updated the results in Appendix Section C.5 and Table26 in our manuscript.

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
Cα-only variant0.53983.1426.08%50.39%71.86%-38.2710.87%8.69%
PPFrame0.59287.1847.83%33.22%57.49%-67.8065.22%65.22%

Q2.2: What are the individual contributions of IPA, Transformer self-attention, and ESM-2 embeddings?

Ans: We thank the reviewer for requesting a clearer breakdown. Our ablations assess the contributions of these components within PPFrame as follows:

IPA layers. As reported in our response to Weakness 2, removing IPA reduces ipTM from 0.592 to 0.326, pLDDT from 87.18 to 72.18, and sequence design success rate from 47.83% to 0%. These results support the importance of geometric information for generating sequences with high interface confidence.

Transformer self-attention. We have run the experiments in Section 5.4 (Fig. 2(d)). Removing sequence self-attention reduces structural designability from 65.22% to 41.30%. This demonstrates that sequence-context modeling contributes to structural design quality, supporting the complementary roles of sequence attention and IPA.

ESM-2 initialization. We additionally trained and evaluated a variant without ESM-2 initialization on PPIBench. As shown below, incorporating ESM-2 initialization improves sequence design success rate, alongside improvements in binding energy, structural designability and co-designability. These results suggest that pretrained sequence knowledge contributes primarily to sequence design quality, with additional benefits for structural quality and consistency. We have updated the results in Appendix Section C.5 and Table25 in our manuscript.

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
Without ESM2 embedding0.56882.1634.78%46.15%57.21%-60.7963.04%60.86%
PPFrame0.59287.1847.83%33.22%57.49%-67.8065.22%65.22%

We welcome any further questions or comments and would be happy to provide additional clarification.

I appreciate the clarifications regarding PPDiff. However, the differences described mostly amount to recombining existing components—structural representations, architectural modules, training objectives, and inference strategies. While the overall framework performs well, the evidence presented is not sufficient to establish a clear and independent methodological contribution. For this reason, I will maintain my rating of 2 (Weak Rejection).

Reviewer_Txs8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

PPFrame is a sequence and structure co-design method that leverages a novel IPAFormer architecture to enable state of the art performance on PPIBench.

The authors also introduce PPIBench, a new benchmark for complex design curated from the PDB, using AF3's preprocessing pipeline.

On PPIBench, PPFrame generally outperforms, RFDiffusion+ProteinMPNN, APM, and Proteina-Complexa.

Critical strengths

PPFrame appears to be a strong method, consistently outperforming the 3 other baselines on PPIBench. In addition, the model appears to scale well to more parameters.

Critical weaknesses

I am concerned about the methods being compared against. The intro and abstract mention strong baselines; however, two popular methods, BoltzGen and BindCraft, are mentioned but not compared against. I would want stronger baseline comparisons before recommending acceptance.

Could the authors elaborate on what makes IPAFormer different from the layers/blocks used in APM? APM also mentions using ESM alongside IPA and a transformer to interleave structure and sequence. What causes the difference in performance between the two, as they appear to me to have similar architectures. In addition, the IPAFormer seems to be very similar to the SSINC layers from the PPDiff paper (Song et al), with the exception that IPA replaces the EGCL layers (both have sequence based self attention layers). Is the IPA layer alone contributing to PPFrame's strong performance? How does PPFrame compare against PPDiff?

Questions for authors

Why are BoltzGen and BindCraft not compared against? Both are cited in the introduction but not mentioned afterwards.

How is PPIBench different from PPBench (Song et al)? PPBench also follows the same curation technique as PPIBench, but with different numbers of proteins after all the filtration steps. How are the two different? How come PPFrame is not benchmarked on PPBench? Why curate PPIBench and not use PPBench?

3 comments

We sincerely thank the reviewer for the careful evaluation and insightful comments. We have addressed each concern, conducted additional experiments, and revised the manuscript accordingly. Below, we provide a point-by-point response:

Q1: I am concerned about the methods being compared against. BoltzGen and BindCraft, are mentioned but not compared against. I would want stronger baseline comparisons before recommending acceptance.

Ans: We thank the reviewer for this valuable suggestion. We have expanded our evaluation to include both BoltzGen and BindCraft on the general protein–protein complex design task. As shown in the table below, PPFrame achieves the highest ipTM (0.592), pLDDT (87.18), and sequence design success rate (47.83%), along with the most favorable complex structure binding energy (ΔG = −67.80), among the evaluated methods. These results provide additional evidence of PPFrame’s competitive performance against these strongest baselines. Though BoltzGen achieves the highest designability, its lower ipTM indicates lower binding interface confidence and is accompanied by a lower sequence design success rate. Additionally, PPFrame achieves higher sequence diversity and novelty than BoltzGen, which might be due to the ubiquitin memorization issue mentioned in their paper. More importantly, we follow the suggestion from reviewer Cm7b and calculate the unique successes by taking ipTM ≥0.8, pLDDT ≥80 and scRMSD <2Å. The results show that overall, PPFrame is better at generating diverse, successful designs.

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designabilityUnique Successes
BindCraft0.56885.2145.65%56.19%76.38%-52.3752.17%56.52%1.71
BoltzGen0.30986.4423.91%25.03%36.82%-54.5184.78%76.08%1.93
Proteina-Complexa0.52184.6030.43%53.77%82.69%-56.5756.52%54.34%3
PPFrame0.59287.1847.83%33.22%57.49%-67.8065.22%65.22%3.8

Q2: The IPAFormer seems to be very similar to the SSINC layers from the PPDiff paper (Song et al), with the exception that IPA replaces the EGCL layers (both have sequence based self attention layers). Is the IPA layer alone contributing to PPFrame's strong performance? How does PPFrame compare against PPDiff?

Ans: We thank the reviewer for highlighting this connection and suggesting a direct comparison. We agree that both methods interleave sequence attention with geometric modeling. However, PPFrame introduces significant changes to the structural representation, generative formulation, supervision, and inference procedure that extend beyond replacing EGCL with IPA.

First, the structural representation and corresponding network updates differ substantially. PPDiff generates Cα-only coordinates, whereas PPFrame represents backbone geometry through residue rotations, translations, and backbone torsions, enabling reconstruction of the N, Cα, C, and O atoms. Accordingly, IPAFormer explicitly updates residue frames and pair representations alongside sequence features. This changes the geometric information modeled throughout generation.

Second, PPFrame incorporates explicit physicochemical information and additional structural supervision. Hydrophobicity and charge features inform interface modeling, while auxiliary losses supervise backbone atomic positions and pairwise distances. Our ablations in Fig. 2(b) and 2(d) support the contributions of these components, providing evidence that the performance gains cannot be attributed to IPA alone.

Third, the diffusion and inference procedures differ. PPDiff combines uniform categorical sequence corruption with diffusion over Cα-only coordinates. PPFrame instead combines absorbing-state sequence diffusion with diffusion over residue rotations and translations. The learning objective also involves two auxiliary losses over the structures. PPFrame also employs self-conditioned inference, whose contribution to structural designability and co-designability is supported by Fig. 2(c).

Continue to reply to Q2:

To address the reviewer’s request directly, we trained and evaluated PPDiff on PPIBench. As shown below, PPFrame achieves a higher sequence design success rate (47.83% versus 28.26%) and more favorable binding energy (−67.80 versus −28.16). The largest differences concern structural fidelity: PPFrame achieves 65.22% on both designability and co-designability, compared with 0% for PPDiff under this evaluation. We also observed frequent steric clashes in PPDiff’s generated structures, consistent with its lower structural scores. The results demonstrate PPFrame’s advantage in producing designs that jointly satisfy sequence-quality and structural-consistency criteria. We have updated the results in Section5.9 and Table6, as well as the discussion with PPDiff in Appendix Section D in our manuscript.

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
PPDiff0.56881.3728.26%87.63%88.91%-28.160%0%
PPFrame0.59287.1847.83%33.22%57.49%-67.8065.22%65.22%

Q3: Could the authors elaborate on what makes IPAFormer different from the layers/blocks used in APM? What causes the difference in performance between the two, as they appear to me to have similar architectures.

Ans: We thank the reviewer for raising this question. We agree that IPAFormer and APM share the general combination of IPA and Transformer layers, a design previously explored in FrameDiff [1]. The key differences lie in how these components incorporate pretrained sequence knowledge and physicochemical information, as well as in the training objectives and inference procedures.

First, PPFrame explicitly incorporates physicochemical features, including hydrophobicity and charge, to inform the modeling of binding-interface complementarity. Our ablation in Fig. 2(b) shows that removing these features degrades performance, providing direct evidence of their contribution. Second, IPAFormer initializes its sequence-attention layers from pretrained ESM2-150M parameters, whereas APM uses aggregated ESM2-650M embeddings as residue representations. Third, PPFrame uses diffusion and score-matching objectives complemented by auxiliary structural losses, whereas APM’s Seq&BB module is trained using flow matching. Finally, PPFrame employs a self-conditioned inference procedure that improves structural designability and sequence–structure consistency, as supported by the ablation in Fig. 2(c).

These distinctions provide plausible explanations for PPFrame’s strong design performance, with the ablations directly supporting the contributions of physicochemical features and self-conditioned inference beyond the shared IPA-Transformer architecture. We have updated the discussion with APM in Appendix Section D in our manuscript.

[1] Yim et al. “SE(3) Diffusion Model with Application to Protein Backbone Generation.” ICML 2023.

Q4: How is PPIBench different from PPBench (Song et al)? PPBench also follows the same curation technique as PPIBench, but with different numbers of proteins after all the filtration steps. How are the two different? How come PPFrame is not benchmarked on PPBench? Why curate PPIBench and not use PPBench?

Ans: We thank the reviewer for requesting this clarification. PPBench and PPIBench follow the same general AlphaFold3-based curation procedure. Their primary difference is the structural representation and the corresponding coordinate-completeness requirements.

PPBench provides protein sequences and Cα-only coordinates, consistent with PPDiff’s Cα-only representation. In contrast, as described in Section 3.1 Problem Definition, PPFrame models backbone geometry involving N, Cα, C, and O atoms and therefore requires these coordinates for training and evaluation. PPIBench was curated to provide this additional structural information. This introduces a stricter completeness requirement: an example suitable for Cα-only modeling lacks other backbone atoms required by PPFrame.

Consequently, the released PPBench cannot be used directly with PPFrame without retrieving the additional coordinates from the source structures and applying the corresponding completeness checks. This motivated the construction of PPIBench, which is to support full-backbone co-design using the established curation framework. We have added the discussion between these two datasets in Appendix Section A.1 in our manuscript.

Q5: Some of the text in the PPFrame manuscript is nearly identical to the text in PPDiff paper (see Appendix C in the PPDiff paper and Appendix A in the PPFrame paper). Appendix C in PPDiff and Appendix A in PPFrame are nearly the same, apart from differences in the numbers. Appendix A in PPFrame does not cite Appendix C in PPDiff. In addition, the PPFrame and PPDiff models themselves are quite similar. The key difference appears to be Invariant Point Attention instead of EGCL (see 'Weaknesses' section); everything else appears to be very similar.

Ans: We thank the reviewer for identifying this issue. We acknowledge the substantial overlap between the descriptions of dataset curation in the two appendices and have added an explicit citation to PPDiff in Appendix A.1.

Regarding the methodological relationship, we want to emphasize that PPFrame introduces substantive changes beyond replacing EGCL with IPA: a full-backbone representation (four atoms) based on residue frames, explicit physicochemical feature guidance, a more complex diffusion process, auxiliary structural supervision, and self-conditioned inference. Our ablations support the contributions of physicochemical features, auxiliary losses, and self-conditioned inference, providing evidence that the improvements extend beyond the choice of geometric attention layer.

We also provide a direct comparison by training and evaluating PPDiff on PPIBench. Under this evaluation, PPFrame achieves a sequence design success rate of 47.83% versus 28.26% and co-designability of 65.22% versus 0%. These results support the practical value of PPFrame.

As clarified in our response to Q4, PPIBench follows the established curation framework but provides the complete backbone coordinates required by PPFrame. Its role is to support full-backbone co-design.

We welcome any further questions or comments and would be happy to provide additional clarification.

Reviewer_Cm7bOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PPFrame, a target-conditioned multimodal diffusion model that jointly generates binder sequences and backbone structures in discrete amino acid space and the continuous SE(3) manifold. Its core network, IPAFormer, interleaves invariant point attention with sequence self-attention and incorporates residue physicochemical features. The model is pretrained on the proposed PPIBench dataset and evaluated on general protein complexes, mini-binder design, and antibody design.

Critical strengths
  1. PPFrame jointly models sequence, structure, and physicochemical properties within a shared denoising network, providing a practical unified framework for improving sequence–structure consistency in target-conditioned binder design.
  2. The experimental evaluation is extensive, covering diverse scenarios with detailed ablations.
Critical weaknesses
  1. The evaluation protocol is insufficiently specified or inappropriate. See questions.
Questions for authors
  1. In section 4.2, what is the specific value of k when selecting the top-k candidates? Is this ranking procedure applied to each benchmark?
  2. The amount of evaluated candidates for each target is too small. Only 10 binder candidates are sampled for each target. The evidence for generalization is insufficient.
  3. The success rate only includes filters of confidence scores without considering self-consistency. In Table 2, on the H3 target, PPFrame achieves 97.36% on SR with only 31.57% on Consistency. Could the authors follow the report convention in Complexa [1] and evaluate "Unique Successes"?
  4. Could the authors provide more details regarding the baseline evaluation? Especially, whether supervised fine-tuning is also applied to each baseline?

[1] Didi K, Zhang Z, Zhou G, et al. Scaling atomistic protein binder design with generative pretraining and test-time compute[C]//International Conference on Learning Representations. 2026, 2026: 61056-61109.

1 comment

We sincerely thank the reviewer for the careful evaluation and insightful comments. We have addressed each concern, conducted additional experiments, and revised the manuscript accordingly. Below, we provide a point-by-point response:

Q1: In section 4.2, what is the specific value of k when selecting the top-k candidates? Is this ranking procedure applied to each benchmark?

Ans: We set k=1 and report top-1 results for all models. As specified in Lines 360–361 of the original submission, our method and all baselines follow the same ranking and evaluation procedures, ensuring a consistent and fair comparison.

Q2: The amount of evaluated candidates for each target is too small. Only 10 binder candidates are sampled for each target. The evidence for generalization is insufficient.

Ans: We thank the reviewer for this valuable suggestion. Within our available computational budget, we expanded the candidate pool to 30 and compared PPFrame with the strongest baseline, Proteina-Complexa, on general protein–protein complex design. We report the top-1, top-5 and top-10 results below, with candidates ranked by complex pLDDT for both methods. PPFrame continues to outperform Proteina-Complexa in sequence design, structure design, and sequence–structure consistency. We have updated the results in Section 5.7 and Table4 in our manuscript.

Top-1

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
Complexa0.70388.3256.52%51.38%79.18%-60.1767.39%60.86%
PPFrame0.75191.0665.22%38.42%59.61%-69.7476.08%71.74%

Top-5

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
Complexa0.59286.2151.30%42.78%53.19%-56.2861.73%58.69%
PPFrame0.66289.1761.30%36.12%42.17%-65.3271.74%70.86%

Top-10

ipTMpLDDTSRDiversityNoveltyDelta GDesignabilityCo-designability
Complexa0.56185.7945.43%36.14%48.61%-52.1959.56%57.39%
PPFrame0.62987.3157.39%34.91%39.16%-61.9769.13%66.08%

Q3: The success rate only includes filters of confidence scores without considering self-consistency. In Table 2, on the H3 target, PPFrame achieves 97.36% on SR with only 31.57% on Consistency. Could the authors follow the report convention in Complexa and evaluate "Unique Successes"?

Ans: We thank the reviewer for emphasizing the importance of self-consistency. We have evaluated Unique Successes on the general protein–protein complex design task, requiring successful designs to satisfy all three criteria: ipTM ≥ 0.8, pLDDT ≥ 80, and self-consistency RMSD ≤ 2. As shown below, PPFrame achieves the highest Unique Successes among all evaluated methods, supporting its ability to generate diverse designs that meet both confidence and self-consistency requirements. We have included these results into our manuscript in Section 5.8 and Table5 following the reviewer’s suggestions.

Unique successes
RFD+ProteinMPNN1.2
APM0.2
Proteina-Complexa3
PPFrame3.8

Q4: Could the authors provide more details regarding the baseline evaluation? Especially, whether supervised fine-tuning is also applied to each baseline?

Ans: We evaluated all baselines across the three tasks using their official implementations and released model weights, without additional supervised fine-tuning (SFT). Please note that for the general protein–protein complex design task, no method, including PPFrame, underwent additional SFT. In this setting, PPFrame still outperforms the baselines in sequence design, structure design, and sequence–structure co-design consistency. We have revised Appendix Section B line 783-787 of the manuscript to clarify the baseline evaluation protocol.

We welcome any further questions or comments and would be happy to provide additional clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.