acceptodds
Under review as a conference paper at ICLR 2027

AbGaze: Attentive Geometric Representation Learning for End-to-End Antibody Design

Abstract

Computational antibody design requires representations that capture the geomet- ric patterns underlying antigen–antibody interactions, yet existing approaches of- ten rely on scalar distances or surface-intrinsic features, leaving cross-molecular geometry largely implicit. We present AbGaze, an end-to-end antibody design framework based on attentive geometric representation learning, which encodes distance, spatial direction, and surface-normal orientation of antigen surfaces rel- ative to antibody-residue local frames, and adaptively aggregates these geometric interactions according to their interfacial context. The learned interaction repre- sentation is shared across multi-CDR co-design, complex structure prediction, and affinity optimization, with local-frame geometric supervision further constraining the representation. AbGaze outperforms prior methods across all three tasks: rel- ative to the second-best method, it improves amino-acid recovery by 7.1% and reduces structural error by 14.9% on average over the six CDRs, improves inter- face docking quality (DockQ) by 6.6%, and raises the affinity improvement rate (IMP) by 32.5%.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody Design, Geometric Representation Learning, Molecular Interface Modeling, Generative Models

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

This paper introduces AbGaze, an antibody design framework that incorporates orientation-aware atom–surface representations into a dyMEAN-based architecture. The model encodes distances, approach directions, and surface-normal incidence in antibody-residue local frames, aggregates these features through attention, and uses geometric supervision to support sequence–structure design. Evaluation covers all-CDR design, CDR-H3 design, complex structure prediction, and affinity optimization.

Critical strengths
  1. Clear geometric motivation. Explicitly representing how antibody atoms approach an antigen surface is a reasonable approach to capturing interface geometry beyond scalar distances.

  2. Shared representation across tasks. The same representation supports multiple antibody modeling tasks, and the reported results suggest benefits for both sequence recovery and interface modeling.

  3. Consistent benchmarking protocol. The authors re-run baselines under a shared evaluation protocol and remove training clusters overlapping with the held-out benchmark.

Critical weaknesses
  1. Affinity optimization lacks independent validation. The same frozen ΔΔG regressor provides the optimization signal and evaluates the resulting candidates. Thus, the reported improvements primarily demonstrate successful optimization of this proxy and may reflect exploitation of the scorer. An independent evaluator or experimental measurements are needed to substantiate improved binding affinity.

  2. Design evaluation uses oracle selection. Selecting the highest-AAR candidate among five samples requires access to the native sequence, which is unavailable in practical design. Although applied consistently across methods, this protocol measures an oracle sampling outcome rather than deployable performance. Mean performance across candidates and top-1 performance under a ground-truth-free selection rule should be reported.

  3. Generalization evidence is limited. All four tasks are evaluated on the same 60-complex RAbD benchmark. CDR-H3-based clustering controls one source of similarity, but does not establish generalization to unseen antigen families. Additional benchmarks or antigen-disjoint evaluation would strengthen the conclusions.

  4. Ablations do not fully isolate the proposed contributions. Replacing the entire orientation-aware module with distance-based message passing changes both geometric features and aggregation. Separate ablations of approach direction, surface-normal information, and attention are needed to identify the source of improvement.

  5. Some gains are modest and require uncertainty estimates. For complex structure prediction, DockQ improves from 0.424 to 0.435 over the strongest baseline. Paired statistical analysis, confidence intervals, and per-target results would help establish the robustness of this improvement.

  6. Sequence recovery is insufficient to establish design utility. Higher AAR demonstrates improved native-sequence reconstruction, but does not necessarily imply better binding or developability. The lack of experimental validation limits the practical conclusions.

Questions for authors

Do the affinity improvements persist when evaluated by an independent scorer that was not used during optimization, and when mutation counts are matched across methods?

How do the design results change when reporting mean performance across all five candidates or selecting candidates using predicted confidence?

Can the direction features, surface-normal features, and attention mechanism be ablated separately?

Are the interface improvements statistically significant, and do they persist on antigen-disjoint or additional held-out benchmarks?

How much computational overhead does the surface representation introduce relative to dyMEAN and AbFlow?

6 comments

We thank the reviewer for recognizing the geometric motivation, unified multi-task representation, and comprehensive evaluation of AbGaze. We provide the following clarifications and additional experimental results.

W1 & Q1. Affinity Optimization Protocol

We clarify an important distinction in our affinity optimization protocol: optimization and final evaluation use different ΔΔG\Delta\Delta G predictors.

Following dyMEAN and AbFlow, each method employs its own predictor to guide optimization. For AbGaze, this is a lightweight MLP trained on the generator's representations, whereas final evaluation uses the same pretrained MEAN GNN-based ΔΔG\Delta\Delta G evaluator across all methods.

The optimization predictors are supervised using scores from the GNN evaluator, so the two stages are not statistically independent. Nevertheless, they involve distinct architectures and parameters rather than optimizing and evaluating with the same predictor.

We have revised Section 4.5 and Appendix C to explicitly distinguish these two stages, correcting the ambiguous description in the original manuscript.

W2 & Q2. Native-Independent Evaluation

The best-of-five AAR selection provides a standardized sequence-recovery benchmark across methods. We additionally evaluate AbGaze without native-sequence-based selection in Appendix F (Table 11).

With a single deterministic prediction, AbGaze achieves 64.0% AAR and 0.407 DockQ, outperforming the reported results of dyMEAN (60.1%, 0.396) and AbFlow (59.7%, 0.379).

Moreover, Appendix E (Table 10) demonstrates the diversity of our generative decoding: five samples yield an average of 4.92 distinct CDR-H3 sequences, with a mean pairwise Hamming distance of 0.458.

These results demonstrate strong single-draw performance alongside diverse candidate generation.

W3. Generalization Beyond RAbD

Our evaluation extends beyond RAbD. Appendix D (Table 9) additionally evaluates AbGaze on the SAbDab test split used by IgGM.

AbGaze outperforms IgGM on five of six CDRs, including H2 (0.694 vs. 0.644), H3 (0.397 vs. 0.360), and L3 (0.677 vs. 0.635).

These results provide additional evidence across test sets, complementing the four-task evaluation on RAbD. Strict antigen-family-disjoint generalization remains a separate evaluation setting.

W4 & Q3. Component-wise Ablations

We have added Appendix G (Table 12), independently ablating directional encoding, surface-normal information, and adaptive attention.

Removing these components decreases AAR by 0.014, 0.018, and 0.021, respectively, and DockQ by 0.020, 0.011, and 0.023. Removing directional encoding and attention also increases RMSD by 0.897 and 0.894, respectively.

Together with the module-level ablation in Table 5 (AAR −4.1 percentage points; DockQ −0.040), these results demonstrate the contributions of individual geometric components to sequence recovery and interface modeling.

W5 & Q4. Statistical Reliability

We conducted paired bootstrap analysis with 10,000 resamples to quantify the uncertainty of DockQ improvements.

ComparisonMean Δ DockQ95% CIWin RateMedian Δ
vs HERN+0.011[-0.023, +0.044]46.5%+0.013
vs dyMEAN+0.027[-0.011, +0.065]58.3%+0.019
vs AbFlow+0.022[-0.016, +0.052]52.7%+0.021

Both comparisons exhibit positive mean and median improvements, with win rates above 50%.

W6. Practical Significance Beyond AAR

Our evaluation extends beyond sequence recovery to structural and interface metrics, including RMSD, lDDT, TM-score, and DockQ, which measures the accuracy of predicted antibody–antigen binding configurations.

We further introduce Epitope-F1, following the evaluation approach of ChiMERa-Bench. This metric measures agreement between the antigen residues contacted by the designed antibody and the reference epitope, combining contact precision and recall.

On all-CDR design, AbGaze achieves 0.449 Epitope-F1, outperforming AbFlow (0.420) and dyMEAN (0.408), corresponding to relative improvements of 6.9% and 10.0%, respectively.

Together with DockQ, these results demonstrate improvements in predicted interface geometry and epitope contact recovery, beyond native sequence reconstruction.

Q5. Computational Overhead

We additionally compare the computational costs of AbGaze with existing methods. AbFlow requires 4.56 s for training and 1.4 s per inference step (14s in total), while dyMEAN requires 2.21 s and 0.47 s, respectively.

Under the corresponding evaluation settings, AbGaze requires 6.34 s for training and 1.8 s per inference step (18 s in total).

These measurements characterize the computational cost of incorporating our orientation-aware atom–surface representation and adaptive attention mechanism.

Thank you for the detailed response and additional experiments. The component-wise ablations directly address my concern about attributing improvements to the individual geometric features and attention mechanism. The deterministic single-prediction results also substantially alleviate the concern that the sequence-recovery gains depend on oracle selection. I acknowledge these improvements to the evidence.

I also accept the clarification that optimization and final affinity evaluation use different predictors, and revise my original criticism accordingly. However, since the optimization predictor is supervised using outputs from the final evaluator, the two remain coupled. The reported affinity gains therefore demonstrate optimization of an evaluator-derived objective, while independent evidence of improved binding remains limited.

My main remaining concern is the strength of the interface-modeling evidence. All three reported bootstrap confidence intervals for DockQ differences include zero. This does not establish the absence of a benefit, but it leaves the modest average improvements uncertain. The statement that win rates exceed 50% also needs qualification, as the comparison against HERN reports 46.5%.

The additional SAbDab evaluation and Epitope-F1 results are useful. Nevertheless, the former does not establish antigen-family-disjoint generalization, and the latter measures reference contact recovery rather than independently validating functional binding. Experimental validation is not essential for every methodological paper, but stronger independent computational evidence would help substantiate the practical claims here.

Overall, I view AbGaze as a well-motivated geometric representation improvement, and the rebuttal addresses several concerns constructively. However, uncertainty around the interface gains and the evaluator-coupled affinity assessment still limit the evidence for the claimed advances. I therefore maintain my Weak Reject recommendation.

We thank the reviewer for the constructive follow-up and for recognizing the value of our component-wise ablations and deterministic evaluation. While we agree that statistical uncertainty in structure prediction and evaluator coupling in affinity optimization require careful framing, we want to clarify why the interface-level evidence carries substantial weight for our core contribution.

1. Core Evidence for Antibody Design and Interface Fidelity

AbGaze is designed for antigen-conditioned antibody sequence–structure co-design, rather than standalone structure prediction or affinity optimization. We included the latter two solely to probe whether our geometric representation generalizes beyond its primary setting.

Within the core design task, AbGaze consistently improves sequence recovery across CDRs under the unified benchmark. Crucially, these gains do not rely on oracle selection or iterative sampling: under single deterministic inference, AbGaze achieves 64.0% AAR and 0.407 DockQ. Component-wise ablations confirm that directional encoding, surface normals, and adaptive atom–surface attention each directly drive this performance.

More importantly, Epitope-F1 is far more than a secondary metric—it is the foundational benchmark of spatial recognition in structure-based design. On all-CDR design, AbGaze achieves 0.449 Epitope-F1, outperforming AbFlow (0.420) and dyMEAN (0.408). While Epitope-F1 evaluates reference contact recovery rather than raw binding energy, capturing the true paratope–epitope interface is the ultimate litmus test for whether a model truly understands antigen recognition. Achieving superior Epitope-F1 proves that AbGaze’s orientation-aware geometric interaction mechanism genuinely steers the model toward the correct biological hotspots on the antigen surface, rather than relying on sequence memorization or arbitrary packing. This spatial fidelity is the strict prerequisite for any meaningful functional design.

2. Supplementary Structure-Prediction Results

We acknowledge that the DockQ confidence intervals include zero, meaning the gains do not reach statistical significance. We also thank the reviewer for catching the win-rate error; the original comparison against HERN was indeed 46.5%.

To further test structural refinement, we evaluated a task-specifically fine-tuned variant against HERN:

ComparisonMean ΔDockQ95% CIWin RateMedian ΔDockQ
Fine-tuned AbGaze vs HERN+0.0176[−0.0108, +0.0558]53.3%+0.0040

Although this fine-tuned variant yields a higher mean ΔDockQ than the unified model (+0.011), its confidence interval still crosses zero. We present this strictly as supplementary evidence rather than claiming statistical significance. Uncertainty in this auxiliary task, however, does not undermine the clear spatial and sequence gains established in our primary design benchmarks.

3. Scope of Affinity and Generalization Claims

We agree that the optimization predictor and final evaluator remain coupled through shared supervision; thus, the reported affinity gains reflect optimization under the evaluator-derived objective rather than independent biochemical validation. Similarly, we recognize that our SAbDab evaluation does not guarantee strict antigen-family-disjoint generalization. We will explicitly state these boundaries in the revised manuscript.

Ultimately, these limitations in secondary tasks should be weighed separately from our primary contribution: an orientation-aware geometric interaction mechanism whose effectiveness is solidly backed by rigorous sequence–structure design benchmarks, deterministic inference, state-of-the-art interface recovery (Epitope-F1), and targeted ablations.

Reviewer_Aqc6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents AbGaze, an antibody design framework that represents antigen surfaces relative to antibody-residue local frames using distance, direction, and surface-normal information. It combines these features with attention and geometric supervision. The authors report improvements over several baselines in multiple task.

Critical strengths
  1. The method uses antibody-residue local frames to explicitly encode distance, spatial direction, and surface-normal information, enabling more fine-grained modeling of antigen–antibody interactions.
  2. Demonstrates consistent performance gains across multiple antibody design tasks.
Critical weaknesses
  1. The main ablation replaces the entire interaction module, so it does not show whether direction, surface normals, or attention drives the improvement.
  2. Validation relies on aggregate computational metrics, without examining representative successes or failures. This makes it difficult to assess whether the gains reflect meaningful improvements at specific antibody–antigen interfaces.
Questions for authors
  1. The manuscript is missing a required section, i.e., the AI USE STATEMENT.
  2. How are epitope residues identified at inference time? In a real-world antibody design setting, the binding epitope may be unknown.
  3. The paper selects the best of five candidates using native-sequence AAR, which is unavailable in real-world antibody design. Could the authors compare all methods using a native-independent selection strategy, such as random sampling or model-based ranking?
  4. Ablation replaces the entire interaction module, making it unclear how much direction, surface normals, and attention individually contribute to the performance gains.
  5. The paper lacks representative case studies and structural visualizations to demonstrate meaningful improvements at antibody–antigen interfaces.
  6. The paper does not provide publicly available code or implementation details.
3 comments

We thank the reviewer for recognizing the value of our antibody-local geometric representation and the consistent improvements across multiple tasks. We have revised the manuscript with additional component-wise ablations and structural visualization, and provide the following clarifications.

Q1. AI Use Statement

The original submission included an AI use disclosure on page 20, under the heading "The Use of Large Language Models (LLMs)." We have renamed this section "AI Use Statement" in the revised manuscript for clarity and consistency with the conference template.

Q2. Epitope Identification

AbGaze focuses on epitope-conditioned antibody design, consistent with established end-to-end antibody design settings. The target epitope is provided as conditioning information, rather than predicted by the model. Identifying unknown epitopes is an important but separate research problem. We have clarified this assumption in the manuscript.

Q3. Native-Independent Selection and Diversity

The best-of-five AAR selection serves as a standardized sequence-recovery benchmark across methods. To further assess performance without native-sequence-based selection, we evaluate a single deterministic prediction in Appendix F (Table 11).

With a single deterministic prediction, AbGaze achieves 64.0% AAR and 0.407 DockQ, outperforming the reported results of both dyMEAN (60.1%, 0.396) and AbFlow (59.7%, 0.379), even without best-of-five selection.

Moreover, our absorbing-state masked diffusion enables diverse designs for the same antigen. As shown in Appendix E (Table 10), five samples yield 4.92 distinct CDR-H3 sequences on average, with a mean pairwise Hamming distance of 0.458.

Thus, AbGaze maintains its performance advantage without diverse sampling, while additionally supporting diverse candidate generation, which is valuable for practical antibody discovery.

W1 & Q4. Component-wise Ablations

We have added Appendix G (Table 12), independently removing directional encoding, adaptive attention, and surface-normal information.

The corresponding decreases in AAR are 0.014, 0.021, and 0.018, respectively, while DockQ decreases by 0.020, 0.023, and 0.011. Removing directional encoding and adaptive attention also leads to substantial RMSD increases (+0.897 and +0.894). These results show that all three components contribute to sequence recovery and interface modeling, complementing the original module-level ablation in Table 5 (AAR −4.1 percentage points; DockQ −0.040).

W2 & Q5. Structural Case Study

We have added a representative structural comparison in Appendix H (Figure 3), using the complex 1ic7. Compared with AbFlow, AbGaze generates a structure closer to the native antibody, with DockQ increasing from 0.615 to 0.935.

This case provides a qualitative illustration of interface improvements, complementing the quantitative results across the benchmark.

Q6. Code Availability and Reproducibility

Section 3 specifies the geometric representations, attention mechanisms, and optimization objectives, with further implementation and evaluation details provided in Appendices A–C and I. We plan to release the source code and pretrained checkpoints upon publication.

Thank you for the additional experiments and analyses. However, I remain concerned that requiring the target epitope to be known is a strong assumption in real-world antibody design, which substantially limits the practical applicability of the method. Furthermore, the structural analysis of a single complex is insufficient to demonstrate that the improvements are consistent across different antibody–antigen complexes. It also does not adequately establish the contribution of the proposed interaction module to the overall performance, leaving its effectiveness insufficiently interpretable. Therefore, these concerns remain unresolved, and I maintain my original rating.

We thank the reviewer for the follow-up. We would like to clarify three points regarding the task setting, the consistency of interface improvements, and the contribution of the proposed interaction mechanism.

1. Practical Relevance of Epitope-Conditioned Design

Specifying a target epitope is not merely a simplifying assumption, but an important objective in therapeutic antibody design, where antibodies are often required to engage particular functional regions rather than arbitrary antigen surfaces.

This setting is consistent with dyMEAN (ICML 2023) and RFantibody (Nature 2026). The latter experimentally demonstrates the feasibility of designing antibodies against user-specified epitopes.

AbGaze addresses how to design an antibody against a specified epitope, rather than which epitope to target in the first place. While identifying unknown epitopes is a valuable research problem, it is distinct from the design problem studied here.

2. Interface Improvements Beyond a Single Case Study

We agree that a single structural example cannot establish improvements across different complexes. Figure 3 is intended as a qualitative illustration, rather than the sole evidence for interface improvements.

To assess these improvements across complexes, we have conducted additional paired analyses of Epitope-F1 and FoldX-based energetic hotspot contact recovery under single-prediction evaluation, without native AAR-based selection.

MetricAbGazedyMEANMean Δ95% CI
Epitope-F10.4580.408+0.050[−0.000, +0.102]
Hotspot Recall0.2170.170+0.047[+0.011, +0.082]

Although the Epitope-F1 difference is not statistically significant, the hotspot recall improvement has a paired confidence interval excluding zero. This provides benchmark-level evidence of improved contact recovery for residues predicted by FoldX to be energetically important in native complexes, beyond the individual structural example.

3. Contribution of the Interaction Mechanism

The contribution of the interaction mechanism is evaluated through controlled ablations, rather than inferred from the case study.

In Table 5, removing the entire orientation-aware interaction module reduces AAR by 4.1 percentage points and DockQ by 0.040. In response to the reviewer's original concern, we have also added separate component-wise ablations in Appendix G (Table 12):

VariantΔ AARΔ RMSDΔ DockQ
w/o Direction−0.014+0.897−0.020
w/o Normal−0.018+0.118−0.011
w/o Attention−0.021+0.894−0.023

Removing any of the three components reduces both sequence recovery and interface accuracy. These results isolate the contributions of directional encoding, surface-normal information, and adaptive attention, directly addressing the effectiveness of the proposed interaction mechanism.

We believe the module-level and component-wise ablations, together with the new paired interface evaluations, should be considered alongside the qualitative case study when assessing the contribution of AbGaze.

References

[1] Kong et al. End-to-End Full-Atom Antibody Design. ICML, 2023.

[2] Bennett et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature, 2026.

Reviewer_S8ykOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes AbGaze, an end-to-end antibody design framework that introduces orientation-aware geometric representations and adaptive atom–surface attention to improve antigen–antibody interaction modeling. The method is well motivated and achieves promising results across multiple tasks.

My main concern is that the paper does not clearly distinguish its contributions from existing methods, particularly dyMEAN, which serves as its underlying architecture, and AbFlow, which shares a highly similar motivation and key technical ideas. I would consider raising my rating if the authors could better clarify these distinctions and provide targeted ablation studies and analysis to substantiate their contributions.

Critical strengths

1. Clear motivation and coherent methodology.

The motivation is clear and straightforward, and the proposed methods are well aligned with it. The overall framework is logically organized and easy to follow.

2. Comprehensive experimental evaluation.

The paper evaluates the proposed method across multiple tasks, including multi-CDR design, CDR-H3 design, complex structure prediction, and affinity optimization. The experimental results demonstrate promising improvements over existing baselines.

Critical weaknesses

W1. Although the motivation is clear, the underlying research problem is not particularly novel. Prior work, especially AbFlow, has already explored incorporating fine-grained antigen surface geometry into antibody design. The proposed geometric encoding and attention mechanisms appear to be relatively straightforward extensions of existing techniques, making the methodological novelty somewhat limited.

W2. The paper builds upon the general architecture of dyMEAN, but the relationship between the two frameworks is not sufficiently explained. In particular, it is unclear which components are directly inherited, which are modified, and which are newly introduced. A more explicit architectural comparison would help clarify the actual technical contributions of AbGaze.

W3.The ablation results suggest that the proposed components provide relatively limited and inconsistent improvements across evaluation metrics. Although the orientation-aware interaction module improves AAR, its impact on other structural metrics is relatively small. Similarly, local-frame geometric supervision improves structural accuracy but provides limited benefits to AAR and no improvement in DockQ. These results raise questions about the significance of the proposed components relative to their additional complexity.

Questions for authors

Please see Weakness

3 comments

We thank the reviewer for recognizing the clear motivation and comprehensive evaluation of AbGaze. We appreciate the opportunity to clarify its methodological distinctions from dyMEAN and AbFlow, particularly how the proposed representation changes the modeling of antibody–antigen interactions.

W1. Methodological Novelty Compared with AbFlow

The key distinction between AbGaze and AbFlow lies in how cross-molecular interactions are represented and used to guide structural refinement.

AbFlow introduces antigen surface geometry through its Surface Multi-channel Encoder (SME), which constructs geometric messages from atom–surface relationships and aggregates relative displacements for equivariant coordinate updates.

AbGaze addresses a more specific aspect of interface modeling: how an antibody atom approaches the antigen surface relative to its own local conformation, and which surface interactions should contribute to structural refinement.

Our method introduces a unified interaction mechanism:

  1. Antibody-local geometric representation. We express atom–surface displacements in antibody-residue local frames, explicitly encoding approach directions and surface-normal incidence. This makes local contact orientation directly accessible to the interaction model rather than leaving it implicit in geometric messages.

  2. Geometry-conditioned interaction selection. A two-level attention mechanism learns the relative contributions of antibody atoms and antigen surface vertices, identifying context-dependent interaction centers instead of aggregating surface displacements uniformly.

  3. Attention-guided geometric refinement. The selected interaction centers guide coordinate updates, with geometric descriptors and attention recomputed as the structural states evolve.

These components form a coherent mechanism connecting local contact orientation, adaptive surface interaction selection, and structural refinement.

The distinction is therefore not the use of antigen surfaces itself, but the explicit modeling of antibody-relative contact geometry and its integration into the interaction operator.

W2. Architectural Relationship with dyMEAN

AbGaze builds upon dyMEAN's full-atom modeling framework while introducing a redesigned antigen–antibody interaction mechanism.

Specifically, we retain dyMEAN's structural initialization, shadow-paratope docking, and multi-channel equivariant refinement backbone. The newly introduced components are the antibody-local atom–surface representation, hierarchical geometric attention, and attention-guided interaction updates.

We additionally incorporate local-frame geometric supervision and absorbing-state masked decoding to support structural regularization and diverse sequence generation.

Thus, dyMEAN provides the underlying full-atom modeling framework, while AbGaze contributes a new geometric interaction operator that explicitly couples antibody conformations with antigen surface orientation.

W3. Component Contributions and Empirical Evidence

The original ablations in Table 5 demonstrate the distinct roles of the interaction mechanism and geometric supervision.

Removing the orientation-aware interaction module reduces AAR by 4.1 percentage points and DockQ by 0.040, indicating its contribution to sequence recovery and interface modeling. Removing local-frame supervision primarily affects structural accuracy, increasing RMSD by 0.147 Å and decreasing lDDT by 0.043.

We further provide component-wise ablations in Appendix G (Table 12) to examine the proposed interaction mechanism more directly.

Removing directional encoding, surface-normal information, and adaptive attention decreases AAR by 0.014, 0.018, and 0.021, respectively, and DockQ by 0.020, 0.011, and 0.023. These results demonstrate the contributions of the geometric representation and interaction selection components.

Importantly, the improvements are not dependent on the generative decoding strategy. Appendix F (Table 11) shows that even with a single deterministic prediction, AbGaze achieves 64.0% AAR and 0.407 DockQ, exceeding the reported results of dyMEAN (60.1%, 0.396) and AbFlow (59.7%, 0.379).

We additionally provide a structural comparison in Appendix H (Figure 3), where AbGaze achieves a DockQ of 0.935 compared with 0.615 for AbFlow on complex 1ic7.

Together, these results support the value of explicitly representing antibody-local contact orientation and adaptively aggregating surface interactions within the full-atom modeling framework.

Thanks for the detailed response and additional experiments. However, my concerns regarding the methodological novelty of AbGaze remain unresolved. Beyond reporting performance improvements, I believe adding more discussion and insights into what the model learns and why it works would help readers better understand its contributions. I therefore maintain my current rating.

We thank the reviewer for the follow-up. We would like to clarify three points regarding the task setting, the consistency of interface improvements, and the contribution of the proposed interaction mechanism.

1. Practical Relevance of Epitope-Conditioned Design

Specifying a target epitope is not merely a simplifying assumption, but an important objective in therapeutic antibody design, where antibodies are often required to engage particular functional regions rather than arbitrary antigen surfaces.

This setting is consistent with dyMEAN (ICML 2023) and RFantibody (Nature 2026). The latter experimentally demonstrates the feasibility of designing antibodies against user-specified epitopes.

AbGaze addresses how to design an antibody against a specified epitope, rather than which epitope to target in the first place. While identifying unknown epitopes is a valuable research problem, it is distinct from the design problem studied here.

2. Interface Improvements Beyond a Single Case Study

We agree that a single structural example cannot establish improvements across different complexes. Figure 3 is intended as a qualitative illustration, rather than the sole evidence for interface improvements.

To assess these improvements across complexes, we have conducted additional paired analyses of Epitope-F1 and FoldX-based energetic hotspot contact recovery under single-prediction evaluation, without native AAR-based selection.

MetricAbGazedyMEANMean Δ95% CI
Epitope-F10.4580.408+0.050[−0.000, +0.102]
Hotspot Recall0.2170.170+0.047[+0.011, +0.082]

Although the Epitope-F1 difference is not statistically significant, the hotspot recall improvement has a paired confidence interval excluding zero. This provides benchmark-level evidence of improved contact recovery for residues predicted by FoldX to be energetically important in native complexes, beyond the individual structural example.

3. Contribution of the Interaction Mechanism

The contribution of the interaction mechanism is evaluated through controlled ablations, rather than inferred from the case study.

In Table 5, removing the entire orientation-aware interaction module reduces AAR by 4.1 percentage points and DockQ by 0.040. In response to the reviewer's original concern, we have also added separate component-wise ablations in Appendix G (Table 12):

VariantΔ AARΔ RMSDΔ DockQ
w/o Direction−0.014+0.897−0.020
w/o Normal−0.018+0.118−0.011
w/o Attention−0.021+0.894−0.023

Removing any of the three components reduces both sequence recovery and interface accuracy. These results isolate the contributions of directional encoding, surface-normal information, and adaptive attention, directly addressing the effectiveness of the proposed interaction mechanism.

We believe the module-level and component-wise ablations, together with the new paired interface evaluations, should be considered alongside the qualitative case study when assessing the contribution of AbGaze.

References

[1] Kong et al. End-to-End Full-Atom Antibody Design. ICML, 2023.

[2] Bennett et al. Atomically accurate de novo design of antibodies with RFdiffusion. Nature, 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.