- 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
-
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.
-
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).
-
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
-
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.
-
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.
-
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).
-
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
- 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).
- Which baseline numbers in Tables 2 and 3 were rerun by the authors? Please specify the protocol, model sizes, and training corpora.
- 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?
- What are the wall-clock time and GPU cost per design under no guidance and under SMC (N=8, K=16)?
- 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
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.
| Method | All-atom | Native FR xyz | DockQ ↑ | DockQ_BB ↑ | SR ↑ | Cα RMSD ↓ | H3 RMSD ↓ | Complex BB ↓ | Complex AA ↓ | CDR-AAR ↑ |
|---|---|---|---|---|---|---|---|---|---|---|
| RFantibody | No | Yes | — | 0.4979 | — | 0.7834 | 0.9243 | 2.3978 | — | 0.4693 |
| - | - | - | - | - | - | - | - | - | - | - |
| Germinal | Yes | No | ||||||||
| DisAb-unbound | Yes | No | 0.2746 | 0.2702 | 0.6667 | 1.4050 | 2.1454 | 5.7614 | 6.0940 | 0.6085 |
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.
| Cohort | N | DisAb-VQ BB RMSD ↓ | GCP-VQ-VAE BB RMSD ↓ | DisAb-VQ AA RMSD ↓ |
|---|---|---|---|---|
| VH | 232 | 0.4073 | 0.4104 | 0.9909 |
| VL | 232 | 0.3704 | 0.3820 | 0.9246 |
| Antigen | 99 | 1.1538 | 0.6575 | 1.5365 |
| CAMEO chains | 400 | 0.5561 | 0.4128 | 1.1519 |
| CAMEO complexes | 183 | 0.6825 | — | 1.2354 |
| SAbDab complexes | 180 | 1.0860 | — | 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.
| Scope | DisAb-VQ active / total | Usage | GCP active / total | Usage |
|---|---|---|---|---|
| VH | 3290 / 8192 | 40.16% | 2634 / 4096 | 64.31% |
| VL | 2750 / 8192 | 33.57% | 2342 / 4096 | 57.18% |
| CAMEO chains | 8128 / 8192 | 99.22% | — | — |
| CAMEO complexes | 8167 / 8192 | 99.69% | — | — |
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.
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
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).
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.