acceptodds
Under review as a conference paper at ICLR 2027

MOSAIC: Conditional Structural Memory for Antibody and Nanobody Design

Abstract

Antibody and nanobody design requires jointly modeling local CDR sequence and structure and global antibody placement relative to a target antigen. Existing generators encode reusable structural patterns only implicitly and often struggle to coordinate these two objectives. Explicit retrieval offers reusable structural context, but similar antigens may support distinct antibody interfaces. We introduce MOSAIC, a conditional structural-memory method that augments a frozen flow-matching generator with diverse training examples selected using inference-time conditions. MOSAIC transfers retrieved sequence and layout features through separate, selectively gated pathways for CDR refinement and antibody pose correction. A null route allows the CDR adapter to abstain from memory use, while a calibrated support gate attenuates poorly supported pose corrections. Across VH–VL and VHH benchmarks, MOSAIC improves docking accuracy and success rates over its no-memory counterpart in both design and structure prediction, and achieves the best DockQ and success rates among the evaluated design methods. Full-test-set paired comparisons support the benefit of memory, while randomized-retrieval and component ablations on VH–VL support condition-dependent selection and complementary refinement pathways. These results demonstrate that explicit structural memory can improve antibody–antigen complex generation without updating the generative backbone.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody design, protein structure generation, structural memory, retrieval-augmented generation, flow matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_HmnqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper proposes a method MOSAIC, which retrieves "relevant" information (training example) to improve antibody CDR generative model's performance.

Critical strengths
  • The idea of retrieval (although this paper is not the first one to propose retrieval-based antibody generation) sounds relevant and novel.
Critical weaknesses
  • Biological rationale: It's not sufficiently biologically justified why that information can be used for memory. The paper describes that memory mechanism helps computationally, but why that information in the retrieval key should identify biologically relevant precedents for a new antibody-antigen design problem? Two structurally similar antigens can be recognized at different epitopes by antibodies with very different paratopes. It's also unclear why antibody framework similarity is the primary retrieval information for CDR sequence design.

  • Redundancy: If the generator was already trained on these antibody–antigen complexes, why retrieve the same training examples again at inference time? This redundancy only shows a marginal contribution (e.g., DockQ)

  • Evaluation: No comparison with conceptually similar baselines, such as RADab. Table 3 lacks metrics focused on CDRs, thus unable to evaluate whether the model actually improve CDR design.

  • Performance: In Table 3, the performance gain is indeed incremental (within error range)

Questions for authors

See Weakness.

1 comment

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns. Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would sincerely appreciate your consideration in updating your evaluations accordingly.

Reviewer_Yqi6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MOSAIC attaches a non-parametric memory to a frozen latent flow-matching antibody and nanobody generator and trains only the memory module, 7.6M parameters on a 166M backbone. A 938-dimensional query key, built from antigen CA geometry, a byte-level polynomial sketch of antigen and framework sequences, and chain and length context, retrieves up to four records from a bank of 12,363 training records by greedy maximum marginal relevance. Each record carries sequence and layout features only, with no Cartesian coordinates.

Two gated pathways consume the retrieved set. A CDR-restricted adapter at backbone layers 8 to 11 cross-attends over record values, mixes the resulting contexts through a learned router with an additional zero-context null route, and updates only canonical CDR positions. A pose head consumes pooled antibody tokens, the query key and a similarity-weighted summary of the retrieved keys, and emits a bounded rotation and translation about an antigen-derived anchor, multiplied by a support gate whose sigmoid midpoint is calibrated above the maximum support score observed under randomized retrieval. Memory training adds randomized-retrieval control terms and distillation from a privileged teacher that sees clean antibody coordinates and native identities.

Evaluation covers design on 93 VH-VL and 50 VHH targets, structure prediction on the same targets, Protenix-v2 refolding, paired memory-on and memory-off runs at seed 42 and at seeds 1 to 5, retrieval and component ablations on all 93 VH-VL targets, a width-matched no-memory control, and stratification by training-antigen identity. Headline design numbers are DockQ 0.187 and success rate 0.312 on VH-VL against 0.151 and 0.247 with memory off, and 0.156 and 0.240 on VHH against 0.114 and 0.120. The design protocol retains a native-derived epitope mask at inference, defined by antigen residues within 10 A of an antibody backbone atom in the native complex.

Critical strengths
  • The paired memory-on and memory-off runs share checkpoint, seed and budget on the complete test sets, and improve DockQ on 81 of 93 VH-VL and 40 of 50 VHH targets (L384-390).
  • Gains hold at every one of five generation seeds, and a width-matched no-memory control reaches only 0.152 and 0.120 DockQ against 0.187 and 0.156 (L423-431).
  • Disabling both pathways provably recovers the backbone computation under identical noise and solver settings (L817-824), so the comparison isolates the module rather than a training difference.
  • The information boundary is documented honestly: the native-derived interface location is disclosed (L702-709), temporal separation is not claimed as homology separation (L727-735), and the retrieval key's inputs are enumerated (L798-806).
Critical weaknesses
  • Only MOSAIC and MOSAIC (off) receive the native-derived epitope mask and epitope-centred cropping, while baselines follow their own protocols (L284-291), so the abstract's best-among-methods claim (L020-022) is confounded.
  • The memory-off model alone already beats all five baselines by three to five times on buried-surface-area error, 177.3 and 248.9 against 672.1 to 1135.0, which points at the conditioning rather than the method.
  • The structure-prediction table carries the same asymmetry, yet the text claims the lowest iRMS and LRMS without restating it (L370-372). Single-record retrieval gives 0.182 DockQ and 0.290 success against 0.186 and 0.312, and dropping the support gate gives 0.181 and 0.290, each a two-success difference on 93 targets.
  • No paired tests or confidence intervals are reported for any ablation row, as the paper itself concedes (L949-956). Fixed-key controls that shuffle the retrieved content still reach 0.164 to 0.170 against 0.150 for memory-off, so much of the gain survives when the content is wrong.
  • The single-pathway ablations are strongly super-additive, 0.155 and 0.153 alone against 0.186 together and 0.150 for neither, which is unexplained and may reflect train-inference mismatch.
Questions for authors
  • Which Table 1 baselines receive any epitope or hotspot specification, given that BoltzGen and Germinal are epitope-targeted methods (L284-291)? Can you report MOSAIC and MOSAIC (off) on all 143 targets with the native epitope replaced by a predicted or non-native one?
  • Why do Table 1 and Table 5 differ for the same configuration when the appendix states the same checkpoint, seed and integration budget (L897-901), and which should a reader treat as the result?
  • On which split was the maximum randomized-retrieval support score of 0.404011190 measured (L228-231)? If on the evaluation targets, the gate midpoint is a test-dependent hyperparameter; how do results change at 0.35 and 0.50?
  • Can you release per-target DockQ for every ablation row and report paired bootstrap intervals or sign tests for MOSAIC versus single-record retrieval and versus no support gate (L949-956)?
  • Does retraining the memory module with one pathway permanently disabled reproduce the inference-time ablation numbers, or are those rows measuring an out-of-distribution condition for the pathway that remains?
  • The pose head reads the query key and a retrieved-key summary rather than record values (L801-809), and record values contain no coordinates (L172-178). What is the DockQ if the retrieved-key summary is replaced by a learned constant vector, keeping the gate and the bounds?
  • For the 30 targets at or below 30 percent antigen identity (L1013-1016), what is the identity distribution of the records actually retrieved, and how often does the retrieved antibody share the test antibody's CDR-length signature?
2 comments

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns. Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would sincerely appreciate your consideration in updating your evaluations accordingly.

Dear Reviewer,

We would appreciate your feedback on our additional controls and clarifications, particularly the independently retrained single-pathway models, gate-sensitivity analysis, paired ablation statistics, and learned constant-summary control. We have explicitly scoped our claims to native-epitope-conditioned design.

Your comments helped us clarify the sources and limitations of the improvements. Please let us know if any concerns remain that we could clarify. Thank you for your time and consideration.

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

The paper introduces MOSAIC, a conditional-memory method that augments a frozen latent flow-matching generator for antibody and nanobody design. It retrieves up to four diverse training examples. One gated pathway uses their sequence and layout features to refine CDR representations, while another adjusts the antibody's position and orientation relative to the antigen. The retrieved values contain no Cartesian coordinates. The main contribution is this two-pathway memory module rather than a new flow-matching objective.

The authors evaluate design and structure prediction on 93 VH–VL antibodies and 50 VHH nanobodies. Structure prediction uses the complete antibody sequence, whereas design uses a native-derived epitope and may involve epitope-dependent antigen cropping. The design results therefore apply to native-epitope-conditioned generation, not blind antigen-only design.

Critical strengths
  1. A reasonable architecture for combining local refinement with global placement. Retrieved sequence and layout patterns may help refine CDRs without providing reliable guidance for global antibody placement, motivating separate pathways for the two updates. Separate pathways, residue-level abstention, and bounded pose corrections are sensible choices. The memory module adds only about 4.6% trainable parameters to the frozen backbone, offering a lightweight way to reuse training examples.
  2. Several meaningful controls support a docking benefit. Memory-on/off comparisons cover both complete test sets. Retrieval and pathway controls are evaluated on all 93 VH–VL targets, and a parameter-matched baseline is tested on both antibody types. Across five generation seeds, mean DockQ improves from 0.140 to 0.183 for VH–VL and from 0.112 to 0.151 for VHH. Independent Protenix-v2 refolding also preserves improvements.
Critical weaknesses
  1. The headline VHH results are inconsistent. Table 2 reports DockQ 0.114 to 0.156, whereas Section 4.5 reports a gain of 0.035 and Table 10's weighted means give approximately 0.1215 to 0.1560. More importantly, Table 2 implies 6/50 to 12/50 successes, but Table 10 implies 11/50 to 12/50, despite the same stated seed and integration budget. Rounding cannot explain this difference. Multi-seed results support the improvement direction, but consistent per-target outcomes are needed to establish its size.

  2. Practical gains and retrieval-specific benefits remain limited or uncertain. Design DockQ improves from 0.151 to 0.187 for VH–VL and from 0.114 to 0.156 for VHH, but success rates of 31.2% and 24.0% mean most interfaces remain below the acceptable threshold. Medium/high-quality yields would better establish utility. VHH prediction is also mixed: DockQ improves over the backbone, but TM-score/lDDT stay unchanged and H3 RMSD, iRMS, and LRMS worsen. Protenix-v2 has higher DockQ and success rate; AlphaFold 3 has higher DockQ and the same success rate.

    Fixed-key value shuffling still outperforms memory off, but introduces mismatched inputs. A matched key-only adapter would better isolate the value of retrieved content; matching of adaptation and distillation is also unclear. The gate midpoint/temperature (0.42/0.03) and component-wise pose bounds (0.35 radians/0.50 nm) lack clear calibration or selection details and sensitivity tests. Gate and pathway removal establish component benefits, not robustness of these settings. Finally, low antigen identity does not establish antibody/CDR/interface separation from training.

Questions for authors
  1. Can Table 2, Figure 4, and Table 10 be reconciled using the same per-target results and configurations, with directly counted successes and paired confidence intervals?
  2. What fraction of designs reaches medium or high DockQ quality, and which targets actually cross the acceptable threshold with memory enabled?
  3. What architecture and initialization checkpoint does the flow backbone use, and how is it related to the La-Proteina autoencoder?
  4. How were the gate midpoint/temperature (0.42/0.03) and the rotation/translation bounds in Eq. (7) (0.35/0.50) selected? Were they chosen on independent validation data, and do nearby settings preserve the gains?
  5. Can the authors provide test-to-training antibody/CDR similarity and overlap summaries across all training stages and the memory bank, complementing the antigen-identity analysis?
3 comments

Dear Reviewer,

Thank you again for your valuable comments. As the discussion period is nearing its end, we would like to kindly ask whether our responses have adequately addressed your concerns. Please let us know if any further clarification is needed. If our responses have resolved your concerns, we would sincerely appreciate your consideration in updating your evaluations accordingly.

The authors’ response has addressed most of my concerns, and I have therefore increased my overall score.

Thank you for your careful review, constructive feedback, and updated assessment. Your comments helped us improve the consistency of our reported results, strengthen the experimental controls, and clarify the scope and limitations of our claims. We appreciate your time and thoughtful engagement throughout the discussion.

Comments on the paper

1 comment

Dear Reviewer,

We would appreciate your feedback on whether our rebuttal addresses your concerns. We have clarified the retrieval rationale, added a RADAb-style local refinement comparison and independently retrained pathway controls, and provided paired confidence intervals and CDR-specific metrics. We have also narrowed our claims to improvements in interface geometry and docking accuracy.

If any concerns remain, we would be happy to clarify them. Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.