acceptodds
Under review as a conference paper at ICLR 2027

DisAb: Self-Consistent Antibody Sequence–Structure Co-Design with Multi-Chain All-Atom Discrete Diffusion

Abstract

Antibody design requires amino-acid sequence and 3D structure to remain mutually compatible under antigen conditions. Existing generative methods often model sequence in a discrete space but structure in continuous geometric spaces, resulting in heterogeneous generation processes that are difficult to synchronize and reuse across different design tasks. We introduce DisAb (Discrete Diffusion for Antibody), a unified framework that represents antibody sequence and multi-chain all-atom structure as residue-aligned discrete tokens and jointly generates them through a single masked discrete diffusion process. To enable this formulation, we develop DisAb-VQ, to our knowledge the first residue-level VQ-VAE that tokenizes multi-chain atom14 protein geometry while preserving intra-chain structure and relative inter-chain placement. We further introduce Boltz-2-based sequence–structure self-consistency as a reference-free potential for sequential Monte Carlo (SMC) guidance at inference time. We evaluate DisAb in both bound-antigen design and the more challenging unbound setting, where the complete bound antigen–VH–VL complex is generated from an isolated antigen monomer. DisAb-VQ accurately reconstructs multi-chain all-atom structures, while DisAb achieves an Aligned DockQ of in the bound setting, improved to with SMC guidance. In the unbound setting, SMC improves raw DockQ from to while also improving VH–VL geometry, CDR sequence recovery, and sequence–structure self-consistency. These results demonstrate that unified discrete modeling provides a flexible framework for antibody design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody design, protein sequence–structure generation, discrete diffusion, vector-quantized autoencoder, sequential Monte Carlo

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_yRqoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DisAb, a conditional generation framework that places antibody sequence and multi-chain structure in a single discrete space for joint CDR sequence and VH-VL structure co-design given an antigen condition. Evaluation is conducted under two settings: bound, where the bound antigen is given, and unbound, where only an isolated antigen monomer is given and the model must generate the full bound complex including the docking pose. Neither setting uses native framework-region coordinates. At inference, optional SMC guidance uses Boltz-2 sequence-structure self-consistency as a reference-free potential. The main reported results include multi-chain reconstruction backbone RMSD of 0.68 to 1.09 Å, Aligned DockQ of 0.711 under the bound setting after SMC guidance, and DockQ of 0.149 under the unbound setting.

Critical strengths
  1. The multi-chain tokenizer fills a real gap. Existing structure tokenizers (FoldToken, GCP-VQ-VAE, ESM3, DPLM-2) are single-chain, backbone-only, or both. Under a single global fit, DisAb-VQ reconstructs full complexes to 0.68 Å backbone RMSD on CAMEO2024 and 1.09 Å on SAbDab (Table 1, Figure 2), showing that inter-chain geometry is indeed preserved through the discrete bottleneck.

  2. By the standards of the field, the evaluation is careful. The test set is temporally held out and filtered by antigen homology, and both tokenizer and generator training corpora are audited against the test set (Appendix B). The SMC analysis includes an equal Boltz-2 budget comparison against independent best-of-16 and honestly reports bootstrap intervals that include zero (Tables 5 and 6).

  3. The unbound task is well motivated and rarely attempted. Generating a full bound complex with a docking pose from an isolated antigen monomer without framework coordinates is harder than the bound setting and closer to practical use. The paper compares against both a same-condition baseline (Germinal) and a stronger-condition baseline (RFantibody).

Critical weaknesses
  1. No clear advantage is shown in the bound setting. DisAb is essentially on par with dyMEAN on the main metric, Aligned DockQ (0.692 vs. 0.691), and lags on all other metrics, most importantly CDR-AAR (0.544 vs. 0.738, and only 0.585 even after SMC, Table 2). Sequence recovery is a core deliverable of co-design methods, and a gap of roughly 19 percentage points is difficult to offset. Moreover, this tie is confounded. DisAb uses about 1.3M pretraining structures and a 2.6B-parameter model stack, whereas the competitor is a much smaller 2023 model. The paper does not report baseline model sizes or training data, as acknowledged in Appendix B.4. The current experimental design cannot separate the contribution of the discrete formulation from that of model and data scale, so the available evidence does not support the central claim that the discrete formulation itself provides an advantage.

  2. The unbound results are not yet usable in practice, and the SMC gain in this setting lacks statistical resolution. DockQ 0.149 and complex RMSD 10.4 Å (Table 3) indicate that poses are largely incorrect and far below acceptable quality. The improvement emphasized in the abstract has a bootstrap interval of [−0.011, 0.054] on the only benchmark for this setting (n=36, Table 6). Stating this improvement without qualification exceeds what the data can support.

  3. The tokenizer comparison is a feasibility demonstration, not a strict comparison. Other tokenizers cannot decode multi-chain structures, so the full-complex portion of Table 1 has no baselines. Where a comparison is possible, the results are mixed: DisAb-VQ lags GCP-VQ-VAE-large by a wide margin on the SAbDab antigen subset (1.15 vs. 0.66 Å backbone RMSD), and although its RMSD is competitive on the VH and VL chains, its codebook utilization there is much lower than GCP-VQ-VAE's (40% and 34%, versus 64% and 57%). The main text does not discuss either gap, stating only that the tokenizer performs particularly well on antibody structures (Section 5.2).

  4. The emphasis on the SMC component is disproportionate to its actual benefit, and computational cost is not reported. In the only controlled equal-budget test (n=30), guidance improves DockQ by only 0.008 with an interval that includes zero (Table 5), while inflating inference to roughly 128 Boltz-2 calls per target. The paper reports no wall-clock or GPU cost figures, which affects assessment of reproducibility and practicality.

Questions for authors
  1. Why does CDR-AAR lag dyMEAN by roughly 19 percentage points? An oracle-structure experiment (native structure tokens, generated sequences) could help separate the tokenizer bottleneck from the effects of the loss weights (λa=0.5, λz=1.5).
  2. Which baseline numbers in Tables 2 and 3 were rerun by the authors? Please specify the protocol, model sizes, and training corpora.
  3. Given that the bootstrap interval for unbound DockQ is [−0.011, 0.054], how is the improvement statement in the abstract statistically justified by this interval?
  4. What are the wall-clock time and GPU cost per design under no guidance and under SMC (N=8, K=16)?
  5. Why is codebook utilization on SAbDab VH and VL much lower than that of GCP-VQ-VAE? Does this limit the diversity reported in Figure 4?
9 comments

Summary. Thank you for recognizing the multi-chain tokenizer, careful conditioning distinctions and isolated-antigen task. Your sequence-recovery question motivated an oracle experiment and a unified Chothia baseline comparison. We also expanded matched-budget and multi-candidate evaluations and measured inference costs.

W1: Bound performance and model/data scale

W1: No clear advantage is shown in the bound setting. DisAb is essentially on par with dyMEAN on the main metric, Aligned DockQ (0.692 vs. 0.691), and lags on all other metrics, most importantly CDR-AAR (0.544 vs. 0.738, and only 0.585 even after SMC, Table 2). Sequence recovery is a core deliverable of co-design methods, and a gap of roughly 19 percentage points is difficult to offset. Moreover, this tie is confounded. DisAb uses about 1.3M pretraining structures and a 2.6B-parameter model stack, whereas the competitor is a much smaller 2023 model. The paper does not report baseline model sizes or training data, as acknowledged in Appendix B.4. The current experimental design cannot separate the contribution of the discrete formulation from that of model and data scale, so the available evidence does not support the central claim that the discrete formulation itself provides an advantage.

Thank you for requesting stronger comparisons and clearer accounting of model/data scale. Antibody discovery typically tests several proposals and advances the strongest candidates. We use best-of-20 to characterize this candidate-pool capability, while preserving the original single-candidate results. The evaluation is motivated by multi-candidate workflows in UniMoMo, ICML 2025; AbDPO, NeurIPS 2024; AnewOmni, bioRxiv 2026; all methods receive 20 draws, with each metric taking its best value.

Table S1. Best-of-20 bound antibody design (232 targets; AbX: 200).

MethodAll-atomNative FR xyzAligned DockQ ↑DockQ_BB ↑TM ↑Cα RMSD ↓H3 RMSD ↓AA RMSD ↓CDR-AAR ↑
DiffAbNoYes—0.81400.98460.81322.2421—0.5734
AbX (200 targets)NoYes—0.77990.97791.00262.3081—0.5892
----------
dyMEAN (retrained)YesNo0.7192‾\underline{0.7192}0.7078‾\underline{0.7078}0.95631.39783.0956‾\underline{3.0956}1.6464‾\underline{1.6464}0.5407
AbFlow (retrained)YesNo0.70190.70510.9598‾\underline{0.9598}1.3517‾\underline{1.3517}3.22011.62090.5722‾\underline{0.5722}
DisAb-boundYesNo0.76820.80250.96641.15642.27191.64890.6377

DisAb leads the full-atom Aligned DockQ comparison and improves H3 RMSD over the two retrained full-atom baselines. These results use the existing DisAb checkpoint. dyMEAN and AbFlow are retrained on our datasets same as DisAb, with shared Chothia definitions and validation-loss checkpoint selection.

The antibody supervision comes from public SAbDab, the same structural-data resource used by the comparison methods. General-protein pretraining includes approximately 1.316M AFDB structures and PDB-derived complexes; the final expanded stage samples 320,000 examples per logical epoch for five epochs. The synthetic-unbound antibody data comprise 2,992 training and 314 validation examples. These stages and corpora are described separately.

The model stack comprises a 325.5M trainable generator, a 650M frozen ESM-PPI encoder, and a 1.67B frozen tokenizer. Only the generator is optimized during antibody diffusion training. AbX uses a frozen ESM2-3B encoder. Capacity serves different roles across these systems; geometric representation, conditioning and task-specific supervision determine how that capacity translates into design quality. Paper update: The main table reports shared-condition retrained comparisons, and the training appendix separates component sizes and data stages.

W2: Unbound docking performance

W2: The unbound results are not yet usable in practice, and the SMC gain in this setting lacks statistical resolution. DockQ 0.149 and complex RMSD 10.4 Å (Table 3) indicate that poses are largely incorrect and far below acceptable quality. The improvement emphasized in the abstract has a bootstrap interval of [−0.011, 0.054] on the only benchmark for this setting (n=36, Table 6). Stating this improvement without qualification exceeds what the data can support.

Thank you for highlighting both practical docking quality and the statistical support for unbound guidance. The best-of-20 results show that DisAb-unbound can generate acceptable all-atom complex poses for 24/36 targets (66.7%) without native antibody framework coordinates. This candidate-generation capability supports the generate-and-screen setting introduced in W1, where multiple antibody proposals are considered for each antigen.

Table S2. Best-of-20 unbound design on 36 targets.

MethodAll-atomNative FR xyzDockQ ↑DockQ_BB ↑SR ↑Cα RMSD ↓H3 RMSD ↓Complex BB ↓Complex AA ↓CDR-AAR ↑
RFantibodyNoYes—0.4979—0.78340.92432.3978—0.4693
-----------
GerminalYesNo0.1257‾\underline{0.1257}0.1028‾\underline{0.1028}0.1111‾\underline{0.1111}2.0516‾\underline{2.0516}4.3682‾\underline{4.3682}7.7951‾\underline{7.7951}8.1203‾\underline{8.1203}0.2447‾\underline{0.2447}
DisAb-unboundYesNo0.27460.27020.66671.40502.14545.76146.09400.6085

At this shared candidate budget, DisAb-unbound achieves higher DockQ (0.2746 vs. 0.1257) and lower complex backbone RMSD (5.7614 vs. 7.7951 Å) than Germinal, while increasing acceptable-pose coverage from 4/36 to 24/36 targets. Its H3 RMSD is also lower (2.1454 vs. 4.3682 Å). These results demonstrate useful coverage of native-like complexes from isolated-antigen conditions. RFantibody remains an auxiliary comparison with stronger native-FR-coordinate conditioning. The reported best-of-20 structural metrics measure native-reference oracle coverage within the candidate pool.

Regarding unbound SMC, the original paired experiment increases acceptable-pose counts from 4/36 to 8/36, with a DockQ improvement of 0.0202 and a 95% interval of [−0.0113, 0.0543]. The small 36-case benchmark limits statistical precision; resolving this interval requires a larger paired unbound evaluation datasets, which is unavailable to us.

We therefore supplemented the evidence with a 232-target bound experiment at an equal budget of 16 Boltz-2 evaluations per target. SMC improves Aligned DockQ from 0.7020 to 0.7113, with a paired PDB-cluster 95% improvement interval of [0.0035, 0.0155], excluding zero. Cα RMSD, CDR-AAR and all-atom self-consistency also improve significantly. This provides complementary evidence for the effectiveness of the shared SMC guidance mechanism, while the unbound-specific estimate retains its reported interval.

Paper update: Table 3 and the folding appendix now report multi-candidate unbound performance. The revised abstract attributes the statistically significant SMC improvement to the matched-budget bound experiment, and the appendix explicitly reports the uncertainty of the 36-case unbound comparison.

We hope these additions address your concerns about candidate quality and the scope of the statistical claims.

W3: Tokenizer comparison and utilization

W3: The tokenizer comparison is a feasibility demonstration, not a strict comparison. Other tokenizers cannot decode multi-chain structures, so the full-complex portion of Table 1 has no baselines. Where a comparison is possible, the results are mixed: DisAb-VQ lags GCP-VQ-VAE-large by a wide margin on the SAbDab antigen subset (1.15 vs. 0.66 Å backbone RMSD), and although its RMSD is competitive on the VH and VL chains, its codebook utilization there is much lower than GCP-VQ-VAE's (40% and 34%, versus 64% and 57%). The main text does not discuss either gap, stating only that the tokenizer performs particularly well on antibody structures (Section 5.2).

Thank you for asking us to clarify capability and code usage. To our knowledge, DisAb-VQ is the first residue-level protein VQ-VAE that jointly represents multi-chain atom14 geometry with one structural token per residue. GCP-VQ-VAE, ESM3 and DPLM-2 provide the single-chain backbone comparisons; our whole-complex evaluation measures the additional representation capability.

Table S3. DisAb-VQ reconstruction across chain types and complex datasets.

CohortNDisAb-VQ BB RMSD ↓GCP-VQ-VAE BB RMSD ↓DisAb-VQ AA RMSD ↓
VH2320.40730.41040.9909
VL2320.37040.38200.9246
Antigen991.15380.65751.5365
CAMEO chains4000.55610.41281.1519
CAMEO complexes1830.6825—1.2354
SAbDab complexes1801.0860—1.5170

A multi-chain all-atom code compresses more information per residue, including relative chain placement, which intra-chain torsions alone cannot determine. This broader geometric scope provides a plausible explanation for the reconstruction trade-off on general proteins, while maintaining comparable VH/VL backbone precision.

Table S4. Structural-code utilization by dataset and chain type.

ScopeDisAb-VQ active / totalUsageGCP active / totalUsage
VH3290 / 819240.16%2634 / 409664.31%
VL2750 / 819233.57%2342 / 409657.18%
CAMEO chains8128 / 819299.22%——
CAMEO complexes8167 / 819299.69%——

The larger vocabulary activates more distinct codes on VH and VL. Broad AFDB/PDB training covers many protein folds, while antibodies occupy a narrower subset. Near-complete usage on CAMEO and specialized usage on antibodies are consistent with this domain coverage. Paper update: Section 5.2 discusses geometric scope, absolute active-code counts and vocabulary fractions.

Reviewer_kEqCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

DisAb jointly generates antibody CDR sequences and multi-chain atom14 structures through masked discrete diffusion. Both tasks use antigen information, epitope hotspots, and antibody framework sequences. The unbound variant takes an isolated-antigen structure and generates the bound complex. A learned tokenizer represents local geometry and relative chain placement, while optional Boltz-2-guided sequential Monte Carlo searches for sequence-structure agreement.

Critical strengths
  • The multi-chain atom14 tokenizer extends backbone-based discrete generation such as DPLM-2. Whole-complex backbone RMSD is 0.683 Angstrom on 183 CAMEO complexes and 1.086 Angstrom on 180 SAbDab complexes, using one global alignment (Table 1).
  • The two task formulations support bound-context design and isolated-antigen conditioning without directly supplying native antibody framework coordinates (Section 3).
  • The authors distinguish SMC-guided search from exact sampling and provide an equal-score-budget control, bootstrap intervals, and a ProteinMPNN scoring comparison.
Critical weaknesses
  • Bound-antigen token provenance needs clarification: Section 3 defines these tokens through target-complex encoding, and Section 4.1 mixes information across chains.
  • Bound AlignedDockQ deliberately removes global antibody pose, while guided raw DockQ averages 0.149 on the 36-case unbound benchmark. Success rates and failure counts are needed to interpret this result.
  • SMC's unbound DockQ intervals span zero.
  • Existing ablations establish benefits from pretraining and monomer conditioning, but do not isolate the design benefit of all-atom tokens. Unguided DisAb nearly ties dyMEAN's AlignedDockQ, with worse H3 RMSD and CDR recovery.
Questions for authors
  • Are bound-antigen tokens encoded without antibody atoms? Please document their provenance and test antibody removal if necessary.
  • What fraction of the 36 unbound cases reaches DockQ > 0.23?
  • Can you provide a matched backbone-token ablation and release the preprocessing, exact splits, and per-target evaluation identities?
6 comments

Summary. Thank you for emphasizing the tokenizer’s geometry, practical conditions and statistical interpretation. We clarified antigen-only encoding, measured external docking and candidate coverage, and strengthened matched-budget SMC evidence.

W1 and Q1: Antigen-only conditioning

W1: Bound-antigen token provenance needs clarification: Section 3 defines these tokens through target-complex encoding, and Section 4.1 mixes information across chains. Q1: Are bound-antigen tokens encoded without antibody atoms? Please document their provenance and test antibody removal if necessary.

Thank you for checking the conditioning boundary. The bound complex supplies the observed antigen conformation; preprocessing extracts its antigen chains before computing structure tokens. Encoding then operates exclusively on those antigen residues. Antibody target tokens are constructed in a separate VH/VL encoding, keeping the antigen condition independent of antibody coordinates.

Our cache check covers 232/232 inputs. On four cases containing single- and multi-chain antigens (2,526 residues), we varied or removed the antibody and obtained identical antigen tokens. Fresh encoding also matches the saved condition. Hotspot labels and framework sequence remain the declared design inputs throughout.

Paper update: The notation section explains this input construction in prose, including the scope of graph and attention operations.

W2 and Q2: Docking success and placement

W2: Bound AlignedDockQ deliberately removes global antibody pose, while guided raw DockQ averages 0.149 on the 36-case unbound benchmark. Success rates and failure counts are needed to interpret this result. Q2: What fraction of the 36 unbound cases reaches DockQ > 0.23?

Thank you for asking for a direct account of docking success. The original unbound experiment has 4/36 successful unguided designs and 8/36 SMC designs at DockQ ≥0.23\geq 0.23. We have now completed the multi-candidate comparison.

The practical objective is to find effective antibodies among multiple proposals. Following candidate generation and selection in UniMoMo, ICML 2025; AbDPO, NeurIPS 2024; AnewOmni, bioRxiv 2026, we supplement one-shot results with best-of-20. This evaluation measures whether the generated pool contains a high-quality solution; the structural metrics use their respective native-reference optima.

Table S1. Best-of-20 unbound design on 36 targets.

MethodAll-atomNative FR xyzDockQ ↑DockQ_BB ↑SR ↑Cα RMSD ↓H3 RMSD ↓Complex BB ↓Complex AA ↓CDR-AAR ↑
RFantibodyNoYes—0.4979—0.78340.92432.3978—0.4693
-----------
GerminalYesNo0.1257‾\underline{0.1257}0.1028‾\underline{0.1028}0.1111‾\underline{0.1111}2.0516‾\underline{2.0516}4.3682‾\underline{4.3682}7.7951‾\underline{7.7951}8.1203‾\underline{8.1203}0.2447‾\underline{0.2447}
DisAb-unboundYesNo0.27460.27020.66671.40502.14545.76146.09400.6085

DisAb-unbound achieves 24/36 acceptable poses, versus 4/36 for Germinal, and a lower best-of-20 complex backbone error (5.7614 versus 7.7951 Å).

For the bound model, antibody design and docking have distinct roles. The invariant tokenizer retains each independently encoded object's geometry; separate antigen and VH/VL codes leave the cross-object pose unspecified. We add HDOCK to place the generated antibody against the antigen and evaluate the resulting interface. Inputs are independently randomized in rotation and translation, and only the declared hotspot/CDR information is retained. The two oracle columns select native DockQ over one or ten HDOCK poses for each of 20 designs

Table S2. HDOCK redocking: 197 common targets, 20 designs per target.

MethodAll-atomNative FR xyzOracle 20×1 DockQ ↑Oracle 20×10 DockQ ↑SR (20×10) ↑
DiffAbNoYes0.18160.27030.4213
AbXNoYes0.18540.27960.4416
dyMEAN (retrained)YesNo0.2150‾\underline{0.2150}0.2971‾\underline{0.2971}0.4365‾\underline{0.4365}
AbFlow (retrained)YesNo0.12070.24440.3706
DisAb-boundYesNo0.23370.35000.4924

Paper update: The main unbound table includes the completed candidate-coverage results, and the redocking appendix explains how HDOCK evaluates binding-pose quality after design.

W3: Unbound SMC uncertainty

W3: SMC's unbound DockQ intervals span zero.

Thank you for identifying the uncertainty in the 36-case estimate. We retain its paired DockQ interval, +0.0202 [−0.0113,0.0543], and success counts. The full-cohort matched-budget experiment supplies significant evidence for the guidance mechanism:

Table S3. Matched-budget SMC versus independent sampling on 232 bound targets.

MetricIndependent best-of-16SMC 4×4Improvement [95% CI]
Aligned DockQ ↑0.70200.7113+0.0093 [0.0035, 0.0155]
VH+VL TM-score ↑0.95410.9589+0.0049 [0.0032, 0.0068]
Cα RMSD ↓ (Å)1.46471.3876+0.0771 [0.0413, 0.1178]
SC_AA ↓ (Å)1.91771.7797+0.1380 [0.0874, 0.1915]
CDR-AAR ↑0.50880.5665+0.0577 [0.0474, 0.0687]

These estimates use 232 targets grouped into 97 PDB clusters; Aligned DockQ retains 230 entries from 95 clusters. Resampling each PDB jointly 20,000 times gives strictly positive improvement intervals for every displayed metric. Both arms use the same checkpoint and 16 folding scores, comparing independent four-step trajectories with four particles and four stages.

Paper update: The abstract and conclusion draw on the significant matched-budget result; Appendix C reports paired estimates for each setting.

Reviewer_waj4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DisAb, a discrete diffusion model for antibody sequence–structure co-design. It first uses a multi-chain all-atom VQ-VAE to encode protein structures into residue-level discrete tokens. A shared masked diffusion Transformer then jointly generates CDR sequences and full VH/VL structures, conditioned on the antigen and antibody framework sequences. For unbound design, it also generates the bound antigen–antibody complex and docking pose from an isolated antigen structure. The authors further introduce Boltz-2-guided SMC to improve sampling. Experiments evaluate structure reconstruction on CAMEO2024 and SAbDab, and antibody design on 232 bound and 36 unbound cases, using structural accuracy, DockQ, sequence recovery, and self-consistency metrics.

Critical strengths
  1. The paper introduces a multi-chain all-atom structural tokenizer that represents protein complexes as residue-level discrete tokens while preserving inter-chain geometry. This extends existing protein structure tokenization methods, which mainly focus on single-chain backbone representations.

  2. The paper studies a challenging unbound antibody design setting. Given an isolated antigen structure and antibody framework sequence, the model jointly generates CDR sequences, full VH/VL structures, and the bound antigen–antibody complex, including the docking pose.

  3. The unified discrete diffusion framework provides a common modeling approach for antibody sequence and structure generation, and supports both bound and unbound design settings.

Critical weaknesses
  1. Incomplete comparison with recent baselines. The paper overlooks several relevant methods with publicly available code and pretrained weights, including AAMFM [1], AbX [2], BoltzGen [3], TiDE-Ab [4] and ABOPD [5]. The authors should include comparisons under compatible settings or better justify their omission.

  2. Limited experimental advantages and evaluation metrics. DisAb does not show a clear improvement over existing methods. For example, dyMEAN achieves better CDR-AAR (0.738 vs. 0.585) and all-atom RMSD (1.642 vs. 1.885) than DisAb with SMC. The evaluation would also benefit from independent AF3-based metrics, such as scRMSD, ipTM, and iPAE. Although H3 diversity is reported, the paper lacks a systematic analysis of the diversity of successful samples after AF3-score filtering

  3. Insufficient justification for discrete structural representations. The main motivation for using VQ-VAE and unified discrete diffusion is not well supported by experiments. Continuous autoencoders or VAEs could also encode full-atom geometry, such as atom14/atom37 coordinates in residue-local frames, into continuous latent representations. However, the paper does not compare VQ-VAE with continuous VAE or discrete diffusion with continuous latent diffusion or flow matching. Without these comparisons, it is unclear what advantages the proposed discrete formulation offers beyond demonstrating its feasibility.

  4. Unclear role of the external tFold encoder. The model uses a frozen 650M-parameter tFold sequence encoder, but no ablation is provided to measure its contribution. It is also unclear whether tFold features are recomputed from the intermediate sequence at every denoising step or cached from the conditioning sequence. The authors should clarify the exact training and inference pipeline, including which sequence tokens are visible to tFold. Comparisons with standard amino-acid embeddings and an analysis of the additional computational cost would help separate the benefits of external pretraining from those of the proposed DDiT architecture.

  5. Limited evidence for the effectiveness of SMC. The paper includes an equal-scoring-budget comparison against best-of-16 sampling, but the DockQ improvement is small (0.659 to 0.667) and not statistically significant. Moreover, the main SMC configuration uses 8 particles and 16 scoring steps, while the equal-budget comparison only evaluates 16 scoring calls. Figure 6 also shows that better self-consistency does not consistently lead to lower native RMSD. A more thorough comparison with best-of-N under matched total computational budgets is needed to justify the benefit of SMC.

References:

[1] Shi et al. Antigen-specific Antibody Multi-modal Foundation Model for Functional Antibody Design. arXiv:2607.20057, 2026. https://arxiv.org/abs/2607.20057

[2] Zhu et al. Antibody Design Using a Score-based Diffusion Model Guided by Evolutionary, Physical and Geometric Constraints. ICML 2024. https://proceedings.mlr.press/v235/zhu24j.html

[3] Stark et al. BoltzGen: Toward Universal Binder Design. bioRxiv, 2025. https://doi.org/10.1101/2025.11.20.689494

[4] Kim and Baek. De novo epitope-specific antibody design via time-dependent guidance. Bioinformatics, 2026. https://doi.org/10.1093/bioinformatics/btag455

[5] ABOPD. https://arxiv.org/abs/2607.18835

Questions for authors

All the questions are in weakness. My current recommendation would be Weak/Clear rejection unless the authors can adequately address these concerns with convincing explanations and additional evidence.

13 comments

Summary. Thank you for recognizing the tokenizer, unified generation framework and isolated-antigen setting. We added AbX, standardized the Chothia baseline comparison, expanded candidate-budget evaluation, and measured folding confidence and inference costs.

W1: Recent baselines

W1: Incomplete comparison with recent baselines. The paper overlooks several relevant methods with publicly available code and pretrained weights, including AAMFM [1], AbX [2], BoltzGen [3], TiDE-Ab [4] and ABOPD [5]. The authors should include comparisons under compatible settings or better justify their omission.

Thank you for recommending these comparisons. With the available compute, we prioritized AbX (ICML 2024) because its code and weights are released and its sequence–structure design task is relevant. We use the RAbD checkpoint with 100-step sampling. Its 200 eligible targets are deposited after the reported July-2023 training cutoff; that sample count is marked in the comparison. We will consider AAMFM, ABOPD, TiDE-Ab and BoltzGen as resources permit.

A design campaign produces and screens a collection of antibodies before choosing the most promising ones. Candidate-pool quality is consequently relevant to practical use. Motivated by UniMoMo, ICML 2025; AbDPO, NeurIPS 2024; AnewOmni, bioRxiv 2026, we add a shared 20-candidate budget and retain single-candidate evaluation. Each reported metric selects its optimum within that pool, quantifying the accessible quality of the conditional generator.

Table S1. Best-of-20 bound antibody design (232 targets; AbX: 200).

MethodAll-atomNative FR xyzAligned DockQ ↑DockQ_BB ↑TM ↑Cα RMSD ↓H3 RMSD ↓AA RMSD ↓CDR-AAR ↑
DiffAbNoYes—0.81400.98460.81322.2421—0.5734
AbX (200 targets)NoYes—0.77990.97791.00262.3081—0.5892
----------
dyMEAN (retrained)YesNo0.7192‾\underline{0.7192}0.7078‾\underline{0.7078}0.95631.39783.0956‾\underline{3.0956}1.6464‾\underline{1.6464}0.5407
AbFlow (retrained)YesNo0.70190.70510.9598‾\underline{0.9598}1.3517‾\underline{1.3517}3.22011.62090.5722‾\underline{0.5722}
DisAb-boundYesNo0.76820.80250.96641.15642.27191.64890.6377

DisAb supplies complete heavy-atom structures without native framework coordinates, achieving Aligned DockQ 0.7682 and H3 RMSD 2.2719 Å. Paper update: The main comparison now includes AbX and reports both candidate budgets in one table.

W2: Performance and additional metrics

W2: Limited experimental advantages and evaluation metrics. DisAb does not show a clear improvement over existing methods. For example, dyMEAN achieves better CDR-AAR (0.738 vs. 0.585) and all-atom RMSD (1.642 vs. 1.885) than DisAb with SMC. The evaluation would also benefit from independent AF3-based metrics, such as scRMSD, ipTM, and iPAE. Although H3 diversity is reported, the paper lacks a systematic analysis of the diversity of successful samples after AF3-score filtering

Thank you for recommending a closer comparison. We found that differences in CDR partition conventions exposed some benchmark CDR identities in the baseline inputs. We aligned the masks with Chothia, retrained dyMEAN and AbFlow on the shared split, and completed fresh generation and external evaluation. The updated tables report these results.

DisAb’s strongest gains are candidate-level interface coverage and H3 accuracy relative to the Chothia-retrained geometric baselines. The revised paper reports all methods under the same metric definitions.

For folding-based evaluation, we use Boltz-2 with its single-sequence capability, allowing uniform scoring of newly generated sequences without homolog searches. The open model makes this evaluation reproducible. Boltz-2 training includes complete-MSA dropout Boltz-2, Appendix C.1.1. Every method uses no MSA/templates, 200 sampling steps, ten recycles and five internal samples, retaining the confidence-ranked output.

Table S2. Best-of-20 Boltz-2 self-consistency for VH+VL (232 targets; AbX: 200).

MethodAll-atomNative FR xyzSC_BB ↓ (Å)SC_AA ↓ (Å)ipTM ↑iPAE ↓ (Å)
DiffAbNoYes0.9887—0.96202.4680
AbX (200 targets)NoYes1.1276—0.96342.4190
-------
dyMEAN (retrained)YesNo1.0473‾\underline{1.0473}1.41800.9604‾\underline{0.9604}2.5601‾\underline{2.5601}
AbFlow (retrained)YesNo1.11471.4961‾\underline{1.4961}0.96002.5658
DisAb-boundYesNo0.93321.58960.96272.4137

DisAb has the best iPAE in the no-FR group and improves backbone self-consistency over both retrained full-atom baselines. We appreciate the proposed quality-filtered diversity analysis: it would characterize variation among successful designs, and we plan it as follow-up work. Paper update: Main comparisons and the folding appendix present these results.

In the unbound setting, DisAb-unbound generates complete all-atom antigen–antibody complexes without native framework coordinates. Under best-of-20 evaluation, it achieves acceptable docking poses for 24/36 targets (66.7%), compared with 4/36 (11.1%) for Germinal, with higher DockQ (0.2746 vs. 0.1257) and lower complex backbone RMSD (5.7614 vs. 7.7951 Å). These results demonstrate strong candidate coverage and complex reconstruction under isolated-antigen conditioning, supporting antibody design when the native antibody structure is unavailable.

W3: Unified discrete representations and diffusion

W3: Insufficient justification for discrete structural representations. The main motivation for using VQ-VAE and unified discrete diffusion is not well supported by experiments. Continuous autoencoders or VAEs could also encode full-atom geometry, such as atom14/atom37 coordinates in residue-local frames, into continuous latent representations. However, the paper does not compare VQ-VAE with continuous VAE or discrete diffusion with continuous latent diffusion or flow matching. Without these comparisons, it is unclear what advantages the proposed discrete formulation offers beyond demonstrating its feasibility.

Thank you for asking us to clarify the modeling rationale. DisAb defines a joint conditional distribution over sequence and multi-chain structural tokens. With conditions cc and generated tokens y0=(a,z)y_0=(a,z),

pθ(a,z∣c)=∑y1:TpT(yT∣c)∏t=1Tpθ(yt−1∣yt,c),X=Dϕ(z,a).p_\theta(a,z\mid c)=\sum_{y_{1:T}}p_T(y_T\mid c)\prod_{t=1}^{T}p_\theta(y_{t-1}\mid y_t,c),\qquad X=D_\phi(z,a).

Normalized categorical transitions and fixed observed tokens define a valid conditional generative process. Masked training learns sequence–structure dependencies, and the decoder maps token samples to atom14 complexes. This common state space supports joint generation, structure-conditioned sequence design, and sequence-conditioned structure prediction through the same masking interface.

The formulation follows multimodal tokenization in ESM3, sequence–structure discrete diffusion in DPLM-2, and masked diffusion in LLaDA/UniDisc. The categorical transition probabilities also support SMC reweighting and resampling with a positive self-consistency potential. The matched-budget experiment demonstrates its practical value: Aligned DockQ improves by 0.0093 [0.0035,0.0155] at 16 external scores per target.

Our baselines cover relevant alternatives: DiffAb combines continuous geometric and categorical sequence diffusion. DisAb extends this landscape with one residue-aligned categorical interface for backbone, side chains and relative chain placement. A controlled continuous-latent replacement would be a valuable further study when training resources permit.

References: ESM3, DPLM-2, LLaDA, UniDisc.

Paper update: Sections 3–4 specify the joint token state and conditional diffusion process; the matched-budget appendix supplies quantitative evidence for its guidance interface.

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

DisAb combines a multi-chain atom14 VQ-VAE with masked discrete diffusion over aligned sequence and structure tokens. It considers bound-antigen conditioning and a harder isolated-antigen setting. Boltz-2 sequence–structure self-consistency provides a reference-free potential for SMC-guided search. The evaluation includes tokenizer reconstruction, 232 bound-context antibody cases, and 36 paired unbound/bound cases.

Critical strengths
  1. A residue-level tokenizer that reconstructs multiple chains and their relative placement is a useful technical contribution. The whole-complex evaluation uses one global alignment, rather than independently aligning each chain, so the reconstruction results genuinely test more than monomer folding. The reported backbone RMSDs of approximately 0.683 Å on CAMEO complexes and 1.086 Å on SAbDab complexes are encouraging.

  2. The evaluation distinguishes several commonly conflated settings: framework sequence versus coordinates, aligned antibody quality versus raw docking, and structural self-consistency versus affinity. The appendix includes an equal-Boltz-2-budget comparison and paired uncertainty, as well as training and conditioning ablations. These disclosures make it possible to assess both the promise and the limitations of the approach.

Critical weaknesses
  1. The bound-conditioning information needs an explicit audit. Section 3 defines the observed antigen tokens as tokens encoded from the target bound complex, whereas Section 4.1 describes an encoder with inter-chain edges and complex-global attention. If the native antigen–antibody complex is encoded before the antigen tokens are selected, those observed tokens can carry information about the withheld antibody geometry. This would exceed conditioning on antigen conformation alone. The manuscript explicitly describes independent antigen-only encoding for the unbound track, but does not resolve the corresponding bound case clearly. I cannot establish from the manuscript whether this occurs in the implementation; this is a high-priority clarification, not a confirmed leakage finding. Separately, the high bound Aligned DockQ scores remove the antibody's placement relative to the fixed antigen using the native antibody, so they cannot establish docking performance even if the conditioning is clean.

  2. The evidence for the prospective docking and SMC gains is still limited. Raw mean DockQ on 36 unbound cases increases from 0.129 to 0.149. Table 6 gives a paired 95% interval of [-0.0113, 0.0543] for the +0.0202 difference. The equal-score-budget bound comparison improves DockQ by 0.0080, with interval [-0.0482, 0.0647]. These analyses do support some sequence-recovery and internal-structure benefits, but do not establish a reliable docking gain from sequential guidance. The larger 232-case comparison uses a substantially larger guided search budget than unguided generation, so it does not resolve the equal-budget question. Self-consistency is also the optimized selection score and cannot serve as independent confirmation of binding quality.

Questions for authors
  1. Are bound antigen conditioning tokens computed by encoding the antigen alone or the full native complex? Please provide the exact conditioning path. If they are full-complex tokens, what happens when the antigen alone is encoded at both training and evaluation?

  2. Can raw antigen-frame docking performance be reported for bound conditioning alongside Aligned DockQ, with an explicit statement of which coordinates are transformed and how? This would separate antibody reconstruction from recovery of the antigen–antibody arrangement.

  3. At a matched external-scoring budget, what is the distribution of acceptable docking outcomes for guided search versus independent sampling, especially on the unbound set? Please report per-target outcomes and uncertainty, and distinguish supported sequence/internal-geometry benefits from inconclusive docking gains.

3 comments

Summary. Thank you for highlighting the geometric representation and the importance of conditioning and docking protocols. We added direct conditioning checks, external redocking, and a full-cohort matched-score experiment.

W1 and Q1: Bound-antigen token provenance

W1: The bound-conditioning information needs an explicit audit. Section 3 defines the observed antigen tokens as tokens encoded from the target bound complex, whereas Section 4.1 describes an encoder with inter-chain edges and complex-global attention. If the native antigen–antibody complex is encoded before the antigen tokens are selected, those observed tokens can carry information about the withheld antibody geometry. This would exceed conditioning on antigen conformation alone. The manuscript explicitly describes independent antigen-only encoding for the unbound track, but does not resolve the corresponding bound case clearly. I cannot establish from the manuscript whether this occurs in the implementation; this is a high-priority clarification, not a confirmed leakage finding. Separately, the high bound Aligned DockQ scores remove the antibody's placement relative to the fixed antigen using the native antibody, so they cannot establish docking performance even if the conditioning is clean. Q1: Are bound antigen conditioning tokens computed by encoding the antigen alone or the full native complex? Please provide the exact conditioning path. If they are full-complex tokens, what happens when the antigen alone is encoded at both training and evaluation?

Thank you for requesting this clarification. The pipeline first extracts the antigen chains from the observed complex and encodes them alone. Antibody atoms never enter that graph or its attention. VH/VL target tokens are computed separately; the same encoding scopes are used for training caches and inference.

We checked all 232 production inputs and repeated encoding on four single-/multi-chain antigen cases spanning 2,526 residues. Removing, moving or perturbing the antibody changes zero antigen tokens, and re-encoding reproduces the cache. These tests keep the declared antigen, hotspot and framework-sequence conditions fixed.

Paper update: Section 3 now describes the independent encoding procedure directly in words.

W1 and Q2: Docking beyond antibody alignment

W1: The bound-conditioning information needs an explicit audit. Section 3 defines the observed antigen tokens as tokens encoded from the target bound complex, whereas Section 4.1 describes an encoder with inter-chain edges and complex-global attention. If the native antigen–antibody complex is encoded before the antigen tokens are selected, those observed tokens can carry information about the withheld antibody geometry. This would exceed conditioning on antigen conformation alone. The manuscript explicitly describes independent antigen-only encoding for the unbound track, but does not resolve the corresponding bound case clearly. I cannot establish from the manuscript whether this occurs in the implementation; this is a high-priority clarification, not a confirmed leakage finding. Separately, the high bound Aligned DockQ scores remove the antibody's placement relative to the fixed antigen using the native antibody, so they cannot establish docking performance even if the conditioning is clean. Q2: Can raw antigen-frame docking performance be reported for bound conditioning alongside Aligned DockQ, with an explicit statement of which coordinates are transformed and how? This would separate antibody reconstruction from recovery of the antigen–antibody arrangement.

Thank you for proposing a docking-based assessment. Antibody design determines sequence and internal structure, while docking supplies the relative antigen–antibody pose. DisAb-VQ's geometric invariance preserves intrinsic conformation across rigid motions. In the bound setup, independent antigen and VH/VL encoding leaves their mutual pose unspecified. We therefore use HDOCK to assess interface quality after docking, alongside the joint-antibody-aligned metric.

Experimental antibody discovery screens multiple generated designs to identify effective candidates. We therefore measure the quality available within a fixed candidate budget, alongside single-candidate performance. This follows the multi-candidate design and selection paradigm of UniMoMo, ICML 2025; AbDPO, NeurIPS 2024; AnewOmni, bioRxiv 2026. Our best-of-20 is a per-metric oracle over the same 20 draws, assessing the coverage of DisAb’s learned joint distribution.

Both docking partners are independently rotated and translated before HDOCK, retaining the declared epitope/CDR information. We select by full-heavy-atom native DockQ over 20 designs and the first one or ten ranked poses per design; success means DockQ ≥0.23\geq0.23.

Table S1. HDOCK redocking: 197 common targets, 20 designs per target.

MethodAll-atomNative FR xyzOracle 20×1 DockQ ↑Oracle 20×10 DockQ ↑SR (20×10) ↑
DiffAbNoYes0.18160.27030.4213
AbXNoYes0.18540.27960.4416
dyMEAN (retrained)YesNo0.2150‾\underline{0.2150}0.2971‾\underline{0.2971}0.4365‾\underline{0.4365}
AbFlow (retrained)YesNo0.12070.24440.3706
DisAb-boundYesNo0.23370.35000.4924

DisAb reaches the highest oracle post-docking DockQ and success rate on this common cohort. Paper update: Section 5.3 and the redocking appendix state the design/docking distinction and the randomized-input protocol.

W2 and Q3: Matched-budget SMC and unbound outcomes

W2: The evidence for the prospective docking and SMC gains is still limited. Raw mean DockQ on 36 unbound cases increases from 0.129 to 0.149. Table 6 gives a paired 95% interval of [-0.0113, 0.0543] for the +0.0202 difference. The equal-score-budget bound comparison improves DockQ by 0.0080, with interval [-0.0482, 0.0647]. These analyses do support some sequence-recovery and internal-structure benefits, but do not establish a reliable docking gain from sequential guidance. The larger 232-case comparison uses a substantially larger guided search budget than unguided generation, so it does not resolve the equal-budget question. Self-consistency is also the optimized selection score and cannot serve as independent confirmation of binding quality. Q3: At a matched external-scoring budget, what is the distribution of acceptable docking outcomes for guided search versus independent sampling, especially on the unbound set? Please report per-target outcomes and uncertainty, and distinguish supported sequence/internal-geometry benefits from inconclusive docking gains.

Thank you for proposing the controlled comparison. We extended it to the full low-homology benchmark.

Table S2. Matched-budget SMC versus independent sampling on 232 bound targets.

MetricIndependent best-of-16SMC 4×4Improvement [95% CI]
Aligned DockQ ↑0.70200.7113+0.0093 [0.0035, 0.0155]
VH+VL TM-score ↑0.95410.9589+0.0049 [0.0032, 0.0068]
Cα RMSD ↓ (Å)1.46471.3876+0.0771 [0.0413, 0.1178]
SC_AA ↓ (Å)1.91771.7797+0.1380 [0.0874, 0.1915]
CDR-AAR ↑0.50880.5665+0.0577 [0.0474, 0.0687]

Both arms use 16 Boltz-2 candidate evaluations per target, with the same checkpoint and folding settings. The independent arm samples 16 complete four-step trajectories; SMC uses four particles and four scoring stages. Selection uses self-consistency in both arms. Intervals use 20,000 paired PDB-cluster bootstrap resamples: 232 entries from 97 PDBs, or 230 interface-valid entries from 95 PDBs for Aligned DockQ. All displayed intervals exclude zero. Sampling/scoring averages 312 s for SMC and 307 s for independent sampling.

These significant gains cover native-reference geometry and sequence recovery as well as all-atom self-consistency. For the original unbound SMC comparison, acceptable-pose counts are 4/36→8/36 and the DockQ difference is +0.0202 [−0.0113,0.0543]. The larger bound experiment provides the resolved matched-budget evidence for sequential guidance.

Table S3. Best-of-20 unbound design on 36 targets.

MethodAll-atomNative FR xyzDockQ ↑DockQ_BB ↑SR ↑Cα RMSD ↓H3 RMSD ↓Complex AA ↓CDR-AAR ↑Complex BB ↓
RFantibodyNoYes—0.4979—0.78340.9243—0.46932.3978
-----------
GerminalYesNo0.12570.10280.11112.05164.36828.12030.24477.7951
DisAb-unboundYesNo0.27460.27020.66671.40502.14546.09400.60855.7614

At the shared 20-candidate budget, DisAb-unbound covers acceptable poses for 24/36 targets, compared with 4/36 for Germinal. Its DockQ is 0.2746 versus 0.1257, and complex backbone RMSD is 5.7614 versus 7.7951 Å. Paper update: Table 3 and the folding appendix now include Germinal; the SMC appendix reports the matched-budget gains. Would these additions address your concerns?

Comments on the paper

2 comments

General Response

We thank the reviewers for recognizing our novel tokenizer—the first, to our knowledge, to jointly encode multi-chain atom14 structures into residue-level discrete tokens while preserving inter-chain geometry (AFNg, waj4, kEqC, yRqo), the practical isolated-antigen design setting without native framework coordinates (waj4, kEqC, yRqo), and unified sequence–structure modeling (waj4). We have strengthened the evidence through supplementary comparisons using the existing DisAb models and made focused revisions to the manuscript.

  1. Antigen-token provenance and encoding independence. (AFNg W1/Q1; kEqC W1/Q1)

We clarified that antigen conditioning tokens are computed by encoding the antigen alone, separately from the antibody. We audited all 232 production inputs and performed four encoder-level tests: all inputs pass the conditioning audit, and antibody removal or perturbation changes zero antigen tokens. Section 3 now states the separate encoding scopes explicitly.

  1. Stronger comparisons and multi-candidate evaluation. (waj4 W1/W2; kEqC W4; yRqo W1/Q1)

We added AbX and evaluated all compared methods at a shared best-of-20 candidate budget, alongside the single-candidate results. Multi-candidate generation and screening are used in UniMoMo, ICML 2025; AbDPO, NeurIPS 2024; AnewOmni, bioRxiv 2026. In practical discovery, several antibodies are generated and tested to identify promising designs. Best-of-20 therefore measures the quality accessible from a candidate pool, complementing single-candidate performance; each metric selects its optimum from the same 20 draws. With unified Chothia retraining of dyMEAN and AbFlow, DisAb-bound reaches Aligned DockQ 0.7682, compared with 0.7192/0.7019, and H3 RMSD 2.2719 Å, compared with 3.0956/3.2201 Å. The revised tables identify all-atom output, framework-coordinate conditions, and candidate budgets.

  1. SMC scoring budget and statistical significance. (AFNg W2/Q3; waj4 W5; kEqC W3; yRqo W4/Q3)

We compared SMC with independent sampling on 232 bound targets, allocating 16 Boltz-2 candidate evaluations per target to each method. SMC improves Aligned DockQ 0.7020→0.7113, Cα RMSD 1.4647→1.3876 Å, and CDR-AAR 0.5088→0.5665. Their paired PDB-cluster 95% improvement intervals exclude zero: [0.0035,0.0155], [0.0413,0.1178] Å, and [0.0474,0.0687]. Sampling/scoring time is 312 versus 307 s/target. The revised SMC analysis reports this significant matched-budget result.

  1. Docking, folding-based evidence, and unbound candidate coverage. (AFNg Q2/Q3; waj4 W2; kEqC W2/Q2; yRqo W2)

We added randomized-input HDOCK redocking, Boltz-2 scRMSD/ipTM/iPAE evaluation, and best-of-20 unbound comparisons. On 197 common targets, DisAb has the highest oracle HDOCK DockQ among the evaluated methods (0.3500). DisAb-unbound generates an acceptable pose for 24/36 targets within 20 candidates.

The revised paper keeps the original method and organization, updates the main comparison tables, and adds concise folding, docking and matched-budget results. Encoder/oracle diagnostics and implementation checks are detailed in the responses where relevant. We hope these additions address the reviewers’ concerns and welcome their assessment.

AuthorsOpenReview ↗

Rebuttal Acknowledgments

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair for guiding the discussion and the reviewers for the time and care they devoted to our paper. Their questions and suggestions helped us strengthen both the experiments and the manuscript.

DisAb represents multi-chain all-atom geometry with one structural token per residue while preserving the relative placement of protein chains. Reviewers described this tokenizer as "a useful technical contribution" (Reviewer AFNg) and observed that it "fills a real gap" (Reviewer yRqo). With this representation, DisAb can jointly generate antibody sequences and structures, including a complete bound complex from an isolated antigen without native antibody framework coordinates. Reviewers recognized this task as "a challenging unbound antibody design setting" (Reviewer waj4) that is "well motivated and rarely attempted" (Reviewer yRqo). In the additional best-of-20 evaluation, DisAb-bound achieved the highest Aligned DockQ, backbone DockQ, VH+VL TM-score, and CDR-AAR, as well as the lowest Cα and H3 backbone RMSDs, among methods using no native framework coordinates. In the unbound setting, DisAb generated at least one acceptable docking pose among 20 candidates for 24 of 36 targets.

During the discussion, Reviewer waj4's suggestions led us to complete a controlled comparison with continuous-latent flow matching and an independent AlphaFold3 evaluation. We were encouraged that Reviewer waj4 found these experiments addressed most of their major concerns and raised their rating. Reviewer kEqC also recognized that the additional experiments strengthened the paper. Our follow-up exchange with Reviewer kEqC further helped us clarify how SMC selects a design without access to the native structure. We are also grateful to Reviewers AFNg and yRqo for their thoughtful initial reviews. Although we did not have a further exchange with them, we are confident that the additional evidence addresses their principal concerns.

In particular, we verified antigen-only conditioning across all 232 bound inputs, added best-of-20 comparisons with AbX and the Chothia-retrained dyMEAN and AbFlow baselines, and conducted further evaluations involving native-structure conditioning, HDOCK redocking, and SMC against independent sampling at a matched scoring budget. Together, these additional analyses provide further evidence for the robustness of our conditioning strategy and the effectiveness of our generation and selection framework.

Once again, we thank the Area Chair for coordinating the review and all four reviewers for their constructive feedback. Their suggestions gave us concrete ways to test the modeling choices behind DisAb and present the evidence for its contributions more clearly.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.