acceptodds
Under review as a conference paper at ICLR 2027

Feedback-Efficient Antibody Design within Discrete CDR Shape Space

Abstract

Antibody design is typically performed in a joint space of discrete residue identities and continuous atomic coordinates. Yet complementarity-determining regions (CDR) backbones recur within a few canonical conformations, suggesting that explicitly leveraging these structural patterns could make antibody design more efficient. We therefore formulate antibody design in a discrete CDR shape space constructed by clustering experimental CDR conformations by type, length and backbone similarity ( shapes). Over this discrete space we train a controllable generative model that, conditioned on a target epitope and a combination of six CDR shapes, produces antibodies that bind the specified epitope while realizing the prescribed conformations. We then perform feedback-guided Bayesian optimization over CDR-shape combinations at test time. Compared to direct generation with test-time scaling in continuous space, our method reduces the CDR structural deviation by under the same budget, while matching its performance with fewer evaluations. In short, a compact, biologically grounded search space preserves design capacity while making antibody optimization substantially more feedback-efficient.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Antibody Generation, Controllable Generation, CDR Shape Space, Structural Priors, Bayesian Optimization, Test-Time Search

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_ca8tOpenReview ↗
Rating
3(+1)
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents IgSeven for antibody design. It represents CDR backbone shapes using a finite vocabulary of structural motifs. Ig7G generates antibody sequences, structures, and docking poses under these shape conditions. Ig7O uses Bayesian optimization to search for CDR shape combinations based on property feedback.

Critical strengths
  1. Support for multiple tasks. The framework supports co-folding, inverse folding, antibody generation with specified CDR shapes, and optimization with one or multiple objectives. These tasks share the same conditioning interface. This range of applications is a strength of the work.
  2. Strong co-folding performance. The method achieves strong performance in epitope-constrained co-folding. These results demonstrate its ability to capture antibody structures and interactions with antigens. They also support the rationale behind the method design.
Critical weaknesses
  1. The central motivation needs stronger evidence. Figure 1 shows that cluster-assigned CDRs have higher amino acid recovery than unassigned CDRs. This correlation does not establish that conditioning generation on recurrent shapes improves design quality. Loop type, length, training-data frequency, and prediction difficulty may affect both cluster assignment and AAR. Higher sequence recovery also does not establish stronger antigen binding. The authors present this observation as motivation. However, further evidence is needed to show that the restricted structural interface improves design efficiency while preserving design capacity.
  2. The method and experimental protocols are difficult to follow. The paper describes the architecture and provides optimization pseudocode. However, the inputs, control channels, and sampling settings for each task remain unclear. The descriptions of co-folding and inverse folding mainly specify their inputs and outputs. They do not clearly state which sequence, structural, and latent controls are used. The paper also needs to explain how known sequences or structures are preserved during generation.
  3. CDR novelty and diversity are not established. The fixed frameworks, shape vocabulary, and reference-recovery objectives may favor designs close to existing antibodies. The paper does not quantify the novelty of generated CDRs or their similarity to training examples. These analyses are particularly important for CDR-H3. Reusing known motifs does not by itself imply memorization. Sequence and structural novelty analyses are needed to assess whether the restricted space supports diverse and novel designs.
Questions for authors
  1. Independent validation of designed complexes. Could the authors provide validation using methods that were not used during optimization? Replicated MD simulations in explicit solvent could provide additional evidence. Comparisons with matched baselines could assess CDR stability and the persistence of interface contacts. Such results would complement the current scores, although they would not directly establish binding affinity.
  2. Nanobody interface energy. For nanobodies, the best Rosetta energy from discrete search is −32.681. Complexa Best-of-N achieves −42.792. Could the authors explain why interface energy worsens while AAR and RMSD improve? Per-target results and energy curves across evaluation budgets would help identify the cause.
2 comments

The authors have addressed most of my concerns. However, I'm still concerned about the novelty and diversity of the generated sequences. Could the authors provide the novelty and diversity of other methods/baselines to compare with?

Thanks for the additional experiment. This addresses my last concern. I'm willing to raise my score to 3. Please add the results to the revised version.

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

The paper reformulates epitope-specific, fixed-framework antibody design as selection plus generative realization over a discrete structural vocabulary. CDR loops from SAbDab2 are grouped by type and length and clustered by pairwise backbone RMSD with DBSCAN, giving 772 shape tokens across six vocabularies of 130, 75, 370, 73, 10 and 114 tokens. An antibody is then specified by seven tokens, one framework plus six CDR shapes, inducing a space of roughly 2.5e8 combinations per paired target against about 2.8e62 possible sequences over 48 CDR residues.

Ig7G realizes a token specification. It builds on the Complexa flow-matching generator and autoencoder, adapts the denoiser to antibody-antigen complexes with a hotspot curriculum, then freezes it and trains a zero-initialized ControlNet branch. Structural controls are built by grafting each token's cluster medoid onto the framework template and embedding pairwise CA distances, with absolute-coordinate controls and antibody-antigen pairwise geometry excluded so the control fixes internal geometry without fixing the docking pose. The branch also accepts sequence prompts, which gives folding and inverse folding for free.

Ig7O performs test-time Bayesian optimization over token combinations with Ig7G frozen, using a random-forest plus upper-confidence-bound instantiation, a Gaussian process with a graph diffusion kernel and expected improvement, and expected hypervolume improvement for Pareto search. Each proposal is realized by Ig7G and scored by an external oracle, with 48 initialization evaluations followed by four per round.

Results cover epitope-constrained co-folding, shape realization under oracle token assignments, feedback-guided optimization against interface energy, CDR recovery and CDR RMSD objectives, inverse folding, and search-efficiency curves under matched oracle budgets with both a reference-based and a reference-free oracle.

Critical strengths
  • Epitope-constrained co-folding reaches 0.595 A CA RMSD, 0.643 DockQ and 0.719 success against a best baseline of 1.150, 0.361 and 0.469, with two of those baselines receiving explicit epitope constraints (L1000-1012).
  • Inverse folding runs the same control interface in reverse and stays competitive with specialized models using roughly 4,000 rather than 150,000 antibody training complexes (L936-943).
  • The optimization study separates continuous best-of-N, discrete random search and discrete Bayesian optimization at fixed budget, giving 2.637, 1.943 and 1.413 A (L957-969).
  • Both a reference-based and a reference-free oracle are reported, and the evaluation-budget framing fits wet-lab limits of 100 to 1000 candidates (L070-072).
Critical weaknesses
  • The 1.87-times and 13-times claims (L021-024) come from searching against CDR RMSD, which requires the native structure; the appendix calls this a diagnostic (L957-958) but the abstract does not.
  • On interface energy, the only reference-free objective, nanobodies regress by 23.63 percent, reported only as a signed gain-row entry (L461-464).
  • The reference-free confidence oracle gives 2.82 to 5.40 percent, an order of magnitude below the headline (L977-985).
  • Single-shot baseline rows (L426-429) sit in the same columns as per-metric bests over roughly 104 oracle-guided evaluations.
  • Nothing states that the clustering pool excluded the post-2025 test structures (L718-732, L848-863), while each test CDR is assigned to the cluster containing its own observed conformation (L381-382).
  • If test loops were in the pool, the medoid can sit a fraction of an angstrom from the answer, so the reported 1.369 A would measure copying rather than realization.
  • Coverage is unreported: only 43.2 to 59.7 percent of CDRs are cluster-assigned (L052-055), and 1,926 of roughly 2.5e8 combinations are realized anywhere in the database (L741-748).
Questions for authors
  • Were CDR loops from the 96 test complexes excluded when building the vocabularies and selecting medoids (L718-732)? If not, please report the RMSD between each test CDR and its assigned oracle medoid, and rerun the shape-realization table with vocabularies built from the training split only. This is the answer that most affects my assessment.
  • What are CDR RMSD, recovery, SeqNat and MolProbity for a single sample with tokens drawn without oracle information, for example by frequency-weighted sampling or the best member of one initialization batch (L380-387)? That is the number comparable to the baseline rows.
  • Why does the discrete space underperform continuous best-of-N on nanobody interface energy by 23.6 percent (L461-464), and would you state the reference-based nature of the CDR RMSD and recovery oracles in the abstract with the reference-free curves as the headline?
  • For the interface-energy-optimized designs, what do orthogonal properties look like, such as interface hydrophobicity, ESMFold2 interface pTM, SeqNat and MolProbity? Direct Rosetta optimization over about 104 evaluations is reward-hacking-prone (L1104-1112).
  • The motivating figure is correlational, since cluster membership is itself a native-likeness test and its association with recovery is close to tautological (L050-056). Is there evidence that prescribing a shape helps beyond what filtering generations for native-likeness would give?
  • How much of the optimization result comes from the 48-evaluation initialization versus the acquisition-guided steps, given that the initialization already carries most of the advantage (L977-985)?
  • Were the training cutoffs of the folding baselines checked against the post-2025 test structures (L861-863), and exactly what epitope information did each baseline receive?
  • All results come from a single generation run (L384-385). Can you report seed-to-seed spread for the headline rows, since with best-of-N reporting run-to-run variance determines whether the margins are real?
1 comment

We appreciate the reviewer for the effort and time in reviewing our work and for proposing detailed and valuable feedback. Our point-to-point responses are as follows.


RW1 / RW2 / RW3 / RQ3

The 1.87-times and 13-times claims come from searching against CDR RMSD, which requires the native structure.

We thank the reviewer for pointing out that the original abstract did not explicitly specify the scope of our headline results. We have revised the abstract to clarify that the 1.87× reduction in CDR structural deviation and 13× improvement in evaluation efficiency were obtained using reference-based structure-matching feedback. We have also added results under reference-free folding-confidence feedback, demonstrating a 7.79% improvement in optimization scores and 2.00× higher oracle-evaluation efficiency over continuous CDR Best-of-N. The revised abstract now explicitly distinguishes these two optimization settings and their corresponding results.

To further strengthen the reference-free evaluation, we have added complete ESMFold2 and Mol-1 optimization experiments covering all 71 paired-antibody and 25 nanobody targets (Section 5.3, Table 5, and Appendix D.3). At the same evaluation budget, IgSeven improves ESMFold2 and Mol-1 scores by 7.79% and 6.06% over continuous CDR Best-of-N, and by 3.41% and 1.99% over discrete random search, respectively. Moreover, COMBO achieves the same average scores with 2.00×/1.73× fewer oracle evaluations than continuous Best-of-N and 1.44×/1.24× fewer evaluations than discrete random search under ESMFold2/Mol-1. These results demonstrate the contributions of both discrete CDR-shape conditioning and feedback-guided optimization to improved effectiveness and evaluation efficiency under reference-free objectives.

Why does the discrete space underperform continuous best-of-N on nanobody interface energy by 23.6 percent?

We conducted an additional analysis of buried interface area (dSASA) on VHH targets. Specifically, IgSeven(COMBO) achieves a mean dSASA of 1,411 Ų, compared with 1,721 Ų for BoltzGen and 1,866 Ų for Complexa Best-of-N. However, COMBO maintains competitive area-normalized interface energy, achieving −2.279 REU per 100 Ų, compared with −2.112 for BoltzGen and −2.301 for Complexa.


RW4

Single-shot baseline rows sit in the same columns as per-metric bests over roughly 104 oracle-guided evaluations.

We thank the reviewer for pointing this out. We clarify that the single-shot baseline results are included only as references for general antibody design quality, rather than as directly comparable results to oracle-guided optimization. Our optimization comparisons are conducted exclusively under matched evaluation budgets or matched performance levels: we compare different search methods using the same oracle-evaluation budget, or measure the number of evaluations required to reach the same performance. We will further clarify this distinction in the revised manuscript to avoid confusion between single-shot generation and test-time optimization.

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

This paper formulates antibody design as selection and generative realization of discrete CDR-shape tokens derived from backbone conformations using DBSCAN clustering of experimental CDR loops from SAbDab2. A paired antibody is specified by one framework token and six CDR tokens, which significantly reduces the search space by at least two orders. Ig7G, built as a ControlNet branch on top of the Complexa flow-matching model, generates sequences and all-atom structures conditioned on these tokens. Ig7O performs Bayesian optimization over the token combinations at test time, using oracle feedback to guide the search without retraining the generator. The core idea is sound and well grounded in immunology. However, some improvements may be needed to address the current implementation where the central claim is about feedback efficiency in a design setting, yet the primary metrics are reference-based and the comparison against continuous-space search.

Critical strengths

The ControlNet-based control mechanism is well executed.The generator faithfully realizes specified shapes while producing reasonable sequence quality and the inverse-folding result in the appendix demonstrates control interface works bidirectionally with only 4k training complexes, while other models have had significantly larger training corpora.

The Bayesian optimization comparison within the discrete space is also justified in the results presented. The multi-objective Pareto extension using EHVI is a natural and non-trivial addition. The epitope-constrained co-folding results are also a nice addition to the results. The paper is well organized, the appendices are thorough, and the evaluation covers multiple tasks. Epitope-constrained co-folding, inverse folding, shape realization, and feedback-guided optimization are distinct tasks that each probe a different aspect of the system.

The CDR-level flexibility analysis is a genuine contribution to the structural vocabulary. The examples showing that loops assigned to the same cluster can form different antigen contacts through sequence variation directly address the concern that discretization over-constrains design.

Critical weaknesses

The primary comparison between discrete-space Ig7O and continuous-space Complexa Best-of-N confounds three factors, namely the search space, generator (Ig7G with trained ControlNet branch vs base model without it), and the search algorithm. However, the authors have not isolated (or even discussed) the effect of discretization from the effect of the stronger conditioning.

The primary optimization metrics are reference-based as the shape tokens come from clustering native conformations, the generator is trained on native complexes, and the oracle rewards closeness to the native structure. To me, the higher scores presented in the paper largely reflect training-distribution recovery rather than actual design quality, perhaps.

I'd have liked to see more analyses of the vocabulary itself, e.g., what fraction of experimental CDRs fall outside all DBSCAN clusters, and the distribution shift between native training backbones and grafted medoid backbones at inference. This would be valuable to inform the generalization of the approach rather than the data itself.

The paper does not use a binding-related oracle while the one free energy metric included shows a significant degradation for nanobodies relative to the continuous baseline. Furthermore, no variance estimates are reported for any result despite stochastically generating the backbones and no experimental validation is attempted despite the paper’s explicit motivation around wet-lab evaluation budgets.

Questions for authors

I would recommend running a Best-of-N through Ig7G with randomly sampled shape tokens and compare the results against Best-of-N from the base Complexa generator without conditioning? This would isolate the contribution of the discrete search space from the contribution of the ControlNet.

What fraction of CDR loops in your SAbDab2 dataset are assigned to noise by DBSCAN and excluded from the vocabulary, and what happens to design quality when the target requires a CDR conformation not in the vocabulary? Can you report generation variance by running Ig7G multiple times on the same token combination and reporting the distribution of CDR RMSD?

The paper motivates the discrete space by its value for wet-lab campaigns, but the optimization never targets a binding-relevant objective, and the one energy metric shows the method not optimizing it for nanobodies. Why is no binding-related oracle used in the optimization loop?

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

This paper introduces IgSeven, a controllable antibody design framework based on a discrete CDR backbone-shape space. The authors cluster experimentally observed CDRs in SAbDab2 by CDR type, length, and backbone RMSD, obtaining 772 shape tokens. A paired antibody is specified by a fixed framework and six CDR tokens, while a nanobody uses three heavy-chain CDR tokens. Ig7G adapts a Complexa flow-matching generator and adds a ControlNet branch to generate CDR sequences, full-atom structures, and docking poses conditioned on the antigen, epitope, framework, and structural/sequence controls.

The paper reports lower CDR RMSD when the true CDR shapes are supplied as oracle conditions and better best-so-far values for several objectives using discrete search and Bayesian optimization. In particular, the paper reports a reduction from 2.637 Å for continuous-space Best-of-N to 1.413 Å for COMBO under one matched-budget setting, and claims that comparable performance can be reached with 13 times fewer evaluations in part of the search-efficiency study.

Critical strengths
  1. A biologically motivated compositional representation. Clustering CDR backbones by type and length gives the discrete tokens a meaningful structural prior. Combining six tokens is more interpretable and controllable than searching directly over unconstrained sequences and continuous coordinates, while still allowing sequence and docking flexibility.

  2. A modular generation-and-search framework. Ig7G realizes structural conditions and Ig7O uses external feedback to select them. The paper compares random search, tree ensembles, and COMBO within the same discrete space, and supplements native-reference CDR RMSD with a reference-free ESMFold2 reward.

Critical weaknesses

The central efficiency comparison confounds the discrete representation, the control branch, and the generator. Continuous Best-of-N uses the antibody-adapted Complexa generator (mathcal G_{ heta_{mathrm{Ab}}}), whereas discrete random search and BO use Ig7G with an additional trained ControlNet branch. Although the base denoiser is shared, these are not the same generator. Therefore, the improvement from 2.637 Å to 1.413 Å cannot be attributed solely to the CDR-shape space or Bayesian optimization. Discrete random search only shows that BO improves over random sampling within Ig7G; it does not rule out the control branch as the main source of gain. A matched continuous-latent baseline with comparable control capacity, or separate ablations of the control branch, representation, and search strategy, is needed, together with generation/evaluation cost.

Questions for authors
  1. Can continuous Best-of-N be compared against Ig7G under matched ControlNet/structural-control capacity? Please add ablations that disable the Ig7G control branch, add comparable control capacity to continuous latent search, and separate representation from acquisition effects, while reporting generation time, structure-evaluation time, and total compute.

  2. Was the CDR shape vocabulary and its medoids constructed using only training structures or all SAbDab2 structures? If test structures were included, please rerun with a strict train-only vocabulary and report token coverage, OOV rates, shape realization, and search results.

  3. Please separate native-reference CDR RMSD/AAR from test-time-available objectives and add at least one independent affinity/binding proxy or experimental validation. Clarify whether Rosetta (Delta G) or ESMFold2 rewards were used in training, hyperparameter selection, or search-protocol selection, to rule out oracle overfitting.

  4. Can the authors evaluate additional frameworks or cross-scaffold generalization? At minimum, report per-target results and uncertainty separately for paired antibodies and nanobodies rather than only aggregate best values.

Comments on the paper

1 comment
Area_Chair_12xGOpenReview ↗

Please respond

Hi reviewers,

Thanks for your work so far on this paper. Please respond to the authors before the end of the discussion period if you have not done so already, even if to acknowledge their response.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.