acceptodds
Under review as a conference paper at ICLR 2027

Multimodal Latent Flow Matching for De Novo Nanobody Design

Abstract

Generative protein design must leverage heterogeneous sequence and structural data while supporting precise control over molecular interactions. We intro- duce Proteina Mixta, a multimodal latent flow-matching model for all-atom co- design and sequence-only generation. A shared autoencoder maps sequence-only, backbone–sequence, and all-atom inputs into continuous residue-level represen- tations, enabling one model to generate across modalities. By modeling dis- crete amino acid sequences through continuous latent dynamics, Proteina Mixta achieves strong performance in both all-atom monomer co-design and sequence- only generation. We extend Proteina Mixta to target-conditioned nanobody (a single-domain antibody) design by assigning synthetic annotations of antibody framework and complementarity-determining region (CDR) to general protein interfaces. Antibody-specific conditioning is then learned entirely from these synthetic examples: true CDR and framework labels are never seen during rep- resentation learning or generative training. The multimodal formulation further enables direct co-training of the structure-based generator with sequence-only antibody repertoires from the Observed Antibody Space (OAS) database. Training on liability-filtered OAS sequences increases effective CDR3 diversity by 26% and reduces sequence-liability rates by 13%. In a controlled comparison with common targets, VHH frameworks, CDR lengths, and sampling budgets, Proteina Mixta achieves up to 50-fold higher unique in silico success rates than prior open models, without sequence redesign. These results suggest that a shared latent space lets structure-free data supervise all-atom generation, opening a route to de novo structure design at the scale of sequence data.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
nanobody, design, generative model, flow matching, biology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_b5LZOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Proteina Mixta, a multimodal latent flow-matching model for protein and nanobody (VHH) design. The core idea is a shared autoencoder that maps sequence-only, backbone-sequence, and all-atom inputs into a single continuous residue-level latent space, paired with a single flow-matching network that generates across all three regimes. Building on La-Proteina, Mixta replaces its all-atom-only autoencoder with a multi-regime one trained with modality dropout, so structure-free sequences update the same generative field used for all-atom co-design.

The main contributions are: (1) a multimodal latent flow-matching model unifying sequence-only generation and all-atom co-design in one shared representation; (2) framework-conditioned nanobody design learned from pseudo-VHH supervision, where synthetic framework/CDR annotations on general protein interfaces yield framework-constrained VHH design without true CDR or framework annotations at any training stage; (3) co-training of the structure-based generator on sequence-only antibody repertoires (OAS), which improves CDR3 diversity and reduces sequence liabilities without refolding or inferred structural labels.

Critical strengths
  1. The shared autoencoder and product-space flow-matching design is conceptually clean and empirically supported. Fig. 3 shows that the Mixta autoencoder, unlike La-Proteina's, encodes sequence-only, backbone-sequence, and all-atom inputs into a structured joint latent space; under sequence-only encoding, Mixta preserves near-perfect sequence recovery and competitive all-atom RMSD (11.4 Å for La-Proteina vs. much lower for Mixta).
  2. The pseudo-VHH construction (Section 3.2, Fig. 2) transforms generic protein-protein interfaces into framework-conditioned design examples by assigning one chain as binder, placing three contiguous pseudo-CDRs in surface-exposed target-proximal regions, and assigning the most-contacting region as CDR3. This yields framework-constrained VHH design without any true CDR or framework annotations during training, which is a practically significant data-efficiency result. The benchmark (Section 4.2) holds targets, framework (hNbBCII10), CDR lengths (11/8/18), sampling budget (1,000 designs per target), and evaluation criteria fixed across methods, and separately reports oracle-free vs. oracle-based methods, sequence redesign, and inference cost.
Critical weaknesses

The primary success metric uses Proteinix-v2 as the evaluator (Eq. 6: iPAE ≤ 6 Å, pTM ≥ 0.83, ipTM ≥ 0.75, pDockQ2 ≥ 0.45), and Mixta's beam search directly optimizes negative mean interface PAE from the same Proteinix-v2 predictor (Eq. 46). The authors acknowledge this in Appendix F.2 ("This creates an advantage specific to this evaluation protocol: its reward directly optimizes one component of the success gate") and note that "an independent-oracle assessment remains necessary to test transfer." However, the main-text claims of 50-fold improvements and state-of-the-art performance rest on this coupled protocol. The oracle-validation table (Table 14) shows Proteinix-v2 rejects 70.28% of known PDB VHH complexes, so computational failure cannot establish absent binding, and passing does not establish binding. Without an independent structural oracle (e.g., AlphaFold3, RoseTTAFold2, or experimental validation) applied to the Mixta designs, the headline success rates are difficult to interpret as evidence of genuine design quality rather than predictor-specific optimization.

Questions for authors
  1. Can the authors provide success rates for the reported Mixta designs under an independent structure predictor not used during search (e.g., AlphaFold3, RoseTTAFold2, or OpenDDE with a different protocol)? Given that the search reward directly optimizes one component of the evaluation gate, an independent-oracle assessment would materially change my confidence in the headline 50-fold claim. If experimental validation is unavailable, a cross-predictor comparison on a subset of targets would be a meaningful substitute.

  2. Table 2 shows interface success drops from 15.77% to 10.84% while CDR3 diversity rises 26% and liability rate falls 13%. How should practitioners weigh these competing objectives? Is there a mixing ratio or filtering strategy that preserves interface success while retaining the diversity/developability gains?

  3. Appendix H shows RL post-training raises PD-L1 success from 0.1% to 83.0% but yields zero successes on four held-out targets. Does this mean sequence-only RL must be run per target, and if so, what is the compute cost per target relative to the structure-based pipeline?

  4. Fig. 3c and Appendix B.5 show that latent warm-up improves Mixta's co-designability and diversity but degrades every metric for La-Proteina. The authors suggest this is because the warm-up stage is effectively sequence-only latent generation, which La-Proteina was not trained on. Can this be made more precise—e.g., by measuring the alignment between the sequence-latent trajectory and the all-atom latent trajectory at the handoff point?

3 comments
AuthorsOpenReview ↗

Addressing Weaknesses

Weakness: Predictor dependence and experimental validation

We thank the reviewer for highlighting this limitation. We have added the independent-predictor assessment identified in Appendix F.2 and narrowed the main-text claims.

Direct generation and beam search require different interpretations. Direct generation uses no Protenix-v2 feedback during sampling, whereas beam search explicitly optimizes a component of the Protenix-v2 success gate. Both are evaluated using that gate, so their reported success rates remain evaluator-dependent. We now state this distinction prominently and qualify the headline improvement and state-of-the-art claims as computational results under the specified evaluation protocol.

Evaluation with a predictor not optimized by search is now included in Table 1. We rescored every method, including Mixta beam search, using ESMFold2, which is not used in the main VHH generation or selection pipeline. The method ranking is generally preserved, including for the search arm. This supports robustness beyond the optimized predictor, while remaining computational evidence rather than validation of binding.

Oracle validation does not resolve the coupling concern. Protenix-v2 performed best among the six predictors evaluated on 508 experimentally determined complexes (Appendix F.2), motivating its use as our primary evaluator. Nevertheless, its rejection of 70.28% of known complexes demonstrates limited sensitivity. Neither this benchmark nor cross-predictor agreement establishes that passing designs bind experimentally.

Experimental validation remains outstanding. Experiments are in progress but the emphasis on the work presented is our method development and in silico efforts, and we make no experimentally validated binding claim. The revised manuscript distinguishes predictor-defined success, robustness across predictors, and experimental binding.

AuthorsOpenReview ↗

Answering Questions pt 1

Q1: Cross-predictor evaluation

We thank the reviewer. We have added cross-predictor evaluation to Table 1, including both direct generation and beam search.

  1. Evaluation protocol. We rescore every method using ESMFold2, which is not used in Mixta’s main VHH generation or selection pipeline. This evaluation also uses MSA conditioning rather than the template-conditioned protocol of our primary evaluator, testing robustness across both predictor and conditioning protocol.

  2. The direct-generation advantage persists. Across seven targets, Mixta produces 396 TNP-clean unique successes, compared with 83 for Proteina-Nanocuerpo, 36 for RFantibody, and 33 for BoltzGen with Boltz-IF—approximately 4.8-fold more than the strongest baseline in this block. [Insert corresponding success rates and evaluation budgets.] These results support an advantage under a second predictor; they do not independently establish the original 50-fold magnitude. We now scope that headline explicitly to its original evaluation protocol.

  3. The search improvement transfers across predictors. Under ESMFold2 evaluation, Mixta search produces 546 TNP-clean unique successes versus 396 for direct generation, a 37.9% increase [insert corresponding success rates and budgets]. Thus, the observed search benefit extends beyond Protenix-v2, although its magnitude remains evaluator-dependent. Mixta search also achieves a similar success count to the ESMFold2 gradient-design baseline (546 versus 559), which directly optimizes this evaluator.

  4. Scope of the evidence. Protenix-v2 remains our primary evaluator because it performed best among six predictors on 508 experimentally determined complexes under the template-conditioned, MSA-free comparison in Appendix F.2. The new results provide evidence of cross-predictor robustness, not experimental validation of binding.

Q2: Weighing OAS benefits and computational success

We thank the reviewer. OAS co-training offers sequence-level benefits with a computational success trade-off. We distinguish three practical considerations.

  1. The benefits persist among designs that pass. Beyond the aggregate diversity and liability improvements in Table 2, our new analysis (Appendix F.4, Table 19) shows that among gate-passing designs, CDR3 families per 100 increase by 13.4%, the largest-family fraction decreases by 24.2%, and sequence liabilities decrease by 15.5%. These benefits therefore extend to the candidates retained by computational screening. Nevertheless, the pass-rate decrease from 15.77% to 10.84% means approximately 31% fewer passing candidates at a fixed generation budget.

  2. OAS introduces a distribution shift, and computational confidence has limits. OAS supplies heavy-chain sequences without paired light chains or antigen context, including conventional heavy chains rather than exclusively VHHs. Appendix F.4 characterizes their distribution shift from our benchmark. This shift may challenge the structure-prediction oracle, whose gates also reject 70.28% of known PDB VHH–antigen complexes (Appendix F.2). The lower pass rate may therefore reflect both design limitations and predictor sensitivity; we cannot attribute it solely to either. Sequence-only flow matching nevertheless enables selected repertoire properties to transfer to all-atom generation without first predicting structures for the added corpus.

  3. Our mixing sweep suggests configuration-dependent benefits, not a universal optimum. On PD-L1, the checkpoint-sweep results were:

OAS fractionTraining-corpus filteringTop-3 / median interface success
0%—16.5% / 3.00%
1%Filtered15.6% / 3.85%
10%Filtered20.2% / 8.45%
25%Unfiltered10.9% / 1.40%

Filtered 10% OAS thus improves computational success in this setting, showing that a reduction is not inevitable. However, this single-target sweep does not establish a generally optimal ratio or simultaneous preservation of all objectives. The 25% setting also changes filtering, so its decline cannot be attributed to mixing ratio alone. These filters apply to the training corpus, not generated candidates.

Practical recommendation. For maximizing gate-passing yield under the Table 2 protocol, the non-OAS model remains the supported default. For prioritizing sequence diversity and fewer liabilities, OAS co-training offers measurable benefits even within the passing pool. Applying interface and liability filters followed by diversity selection is a reasonable triage strategy, but we have not demonstrated that it restores the original yield. We therefore present OAS as a controllable distributional intervention with measured trade-offs, rather than a universal improvement in binding or developability.

AuthorsOpenReview ↗

Answering Questions pt 2

Q3: Multi-target RL and compute

This is an open question. So far sequence-only RL has been per target: neither checkpoint in Appendix H transfers to the held-out targets (Table 22). We also tried multi-target sequence-only RL, but the improvement did not carry across targets. A likely reason is that a per-target run can concentrate on a narrow successful region for one antigen, whereas a shared model has to balance several targets under the same offline RL objective; we have not isolated the cause, and we note this as future work in Appendix H.

On compute, RL is a one-time cost, after which the updated model can be resampled at the cost of ordinary sampling, whereas test-time beam search does not update the weights and pays its cost again for every design. Because sequence-only generation has a low base success rate, search needs many expensive designs per success, so RL is cheaper per successful design, and this is why we see promise in making multi-target RL work.

Q4: Latent warm-up trajectories and sampling ablations

We thank the reviewer for suggesting this analysis. We have examined sequence-only versus joint latent trajectories, expanded Appendix B.5 with more sampling ablations and intermediate latent readouts, to provide a deeper understanding of latent warm-up generation.

For the latent trajectory comparison between sequence-only and all-atom generation, we generate 100 paired samples per mode at length [50, 150, 250], using identical initial priors and shared per-step latent noise within each pair. At handoff twarm=0.3t_{warm} = 0.3, we compare the cosine similarity and relative displacement (relative distance normalized by the magnitude of all-atom latent) between latent states as below. However, this diagnosis overlooks the fact that the sequence latent in La-Proteina is poorly defined. We therefore report it here as an exploratory analysis rather than establishing causal evidence for latent warm-up improvement.

LengthMixta cosine ↑La-Proteina cosine ↑Mixta relative displacement ↓La-Proteina relative displacement ↓
500.9020.7640.4400.723
1500.9020.7450.4350.756
2500.8810.7570.4710.733

In fact, the manuscript ablations show that the current improvement is a useful sweet spot, whereas the widely observed benefit of latent warm-up is improved diversity. In Figure 9a, at warm-up noise ηzwarm=0.05\eta_z^{warm} = 0.05, increasing twarmt_{warm} from 0.2 to 0.3 raises co-designability from 75.4% to 84.2% and structural diversity from 65.0% to 71.0%. At 0.4, diversity increases further to 73.8%, while co-designability falls to 74.8%. At 0.5, both deteriorate.

We further provide a more comprehensive ablation in Figure 9b by sweeping the warm-up latent temperature ηzwarm\eta_z^{warm}, handoff time twarmt_{warm}, backbone temperature ηCα\eta_{C_\alpha}, joint latent temperature ηz\eta_z, and number of warm-up steps. Across the 117 evaluated configurations, increased diversity is the broader benefit, while co-designability remains similar or decreases. Nevertheless, a favorable range of settings are still shown to improve both metrics, indicating that latent warm-up offers a useful sweet spot rather than a uniform improvement across sampling configurations.

Our interpretation is that latent warm-up provides a noisy but sequence-informed initial condition for subsequent joint denoising. Why particular settings improve both metrics remains unresolved. One possible reason for this ambiguity is that the sampling schedule introduces a transition not explicitly covered during training: the model is never learned to perform joint denoising given partially denoised sequence latent. The revised manuscript therefore presents the sweet spot empirically and motivates future training-time schedules tailored to transitions between regimes in multi-regime diffusion.

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

This paper introduces Proteina Mixta, a generative model designed for target-conditioned de novo nanobody (VHH) design, where the goal is to generate a 3D docked nanobody structure and sequence that binds to a specific target antigen. The model extends La-Proteina by training its autoencoder in mixed modes to accommodate sequence-only, backbone-sequence, and all-atom inputs within a shared latent space. For the flow-matching model, it supports both coupled (Cα,zr)(\text{C}_\alpha, z_r) generation on 3D structures as well as, structure-free zseqz_{\text{seq}} flow matching on sequence databases. To evaluate the model's capabilities, the authors assess its performance on unconditional monomer generation across varying lengths, as well as on a benchmark of seven target antigens for de novo nanobody design

Critical strengths
  • The proposed method for supporting multiple modalities in the model is definitely very interesting. I found the Pseudo-VHH construction to also be very well-motivated, allowing the generative model to learn docking and loop-inpainting without requiring direct, scarce antibody-antigen complex training data.
  • The baseline alation for the monomer task is very clear and rigorous and provide a clear analysis of the performance benefits of a multi-regime latent space on general protein structure generation.
  • Overall, the paper is well-written, and easy to follow.
Critical weaknesses

I am willing to increase my score if the authors address my questions/concerns below:

  • The effetc of OAS sequence co-training (Section 4.3, Tables 2 and 9) is not clear to me. I thought one of the central motivations of the paper is that sequence-only representations allow co-training on databases like OAS. However, Table 2 reveals that adding 10% OAS sequence-only data actually degrades 3D structural binding success. Additionally, the main Table 1 model doesn’t include OAS co-training (Table 9). Can the authors clarify the paper's primary narrative on the utility of sequence-only flow matching for the VHH design task? It would also be helpful if the authors could clarify the motivation of sequence only design given the results of Table 2.

  • The effect of the latent warm-up is also not clear in my opinion. On page 6 the paper claims that “Mixta also maintains a mixed secondary-structure distribution”. However, Table 4 shows a severe bias towards alpha helices when using the latent-warmup schedule. The β\beta-sheet seems to collapse while alpha-helices increase significantly (Table 4). Can this collapse be the reason for the better performance of this model on co-designability metrics? The authors should qualify this main-text claim. It would also be helpful to include the ground-truth secondary structure composition of natural proteins in Table 4 as a reference.

  • I found the ablation of Proteina Mixta vs.Nanocuerpo to be confounded. The paper attributes the performance gap between Proteina Mixta and the Proteina-Nanocuerpo baseline in Table 1 to the multimodal latent space. However, the models differ in several other significant ways:

    1. Mixta reintroduces 7 triangular multiplicative pair-update blocks in the flow trunk.
    2. Its VAE is fine-tuned on SAbDab2.
    3. It uses a different optimizer (Muon vs. Adam) and training curriculum.
    4. Finally, Mixta uses a specific multi-branch checkpoint selection, as described in Appendix E.3. I believe hese differences have not been ablated independently. Therefore, it is had for me to see if the VHH gains stem from the multimodal autoencoder or simply from the heavier trunk, fine-tuning, and the specific checkpoint weight selection arithmetic. It would be helpful if the authors could include these separate ablations.
  • The weight arithmetic (Appendix E.3 and Table 10) also helps the results in favor of Mixta compared to the baselines I believe. For 4 of the 7 targets in Table 1 the authors tuned a post-hoc checkpoint merging protocol. This target-specific weight arithmetic boosted PD-L1 success from 8.7% (unmixed) to 35.9%, for example ( Table 10). However, none of the baselines benefited from this kind of checkpoint tuning which I believe makes the comparison unfair. Also I find the fact that Mixta uses the evaluation oracle Protenix-v2 for the beam-search loop as well to be concerining. Can the authors use an independent structure predictor as oracle to validate the significance of the improvements they find?

Questions for authors
  1. For the monomer baseline in Table 4, for how many training steps was the retrained`La-Proteina model run?
  2. Why were the monomer evaluations in Section 4.1 restricted to short proteins when the original La-Proteina paper demonstrated structural generation on proteins up to length 500–800?
3 comments
AuthorsOpenReview ↗

Addressing Weaknesses pt 1

W1: Sequence-only training and OAS co-training

We thank the reviewer for identifying this ambiguity. We have revised the narrative to distinguish sequence-only training from external OAS co-training.

The Table 1 model already includes sequence-only training. 10% of its training examples use the sequence-only regime, but it includes no external OAS data. Table 2 therefore tests adding an external sequence corpus, rather than enabling sequence-only flow matching.

OAS provides a sequence-level benefit, with a lower computational pass rate. Adding OAS reduces measured 3D success from 15.77% to 10.84%. However, the new passing-subset analysis (Appendix F.4, Table 19) shows greater CDR3 diversity and fewer sequence liabilities even among designs that pass. This supports distributional steering, not improved binding success. Predictor limitations and distribution shift may contribute to the lower pass rate (Appendices F.2–F.4), but we cannot separate these effects from design limitations.

The primary contribution is a unified model supporting sequence-only, backbone–sequence, and all-atom generation. Sequence-only flow matching enables training on repertoire data without first predicting structures for that corpus. The OAS experiment demonstrates this capability and its trade-offs; it is not the basis of our main VHH success claim. Likewise, sequence-only VHH generation demonstrates a further available generation mode, rather than superiority over structure-conditioned design. We have revised Section 4.3 to make these claims and limitations explicit.

W2: Helical enrichment and latent warmup

We thank the reviewer for raising this question. First, we would like to clarify that our claim refers to the model itself, more specifically performance under default sampling that is more closely aligned with training schedules. Mixta retains a mixed secondary structure composition ratio of 55/12/32% (helix/sheet/coil), whose helix fraction is slightly lower than the matched La-Proteina baseline of 59/11/30%. The additional helical enrichment occurs only with the latent warmup, which is an interesting out-of-distribution schedule we found that works on Mixta but not La-Proteina. La-Proteina also tends to generate more helices despite its poor generative quality.

Second, even though that latent warmup raises the helix ratio to 72%, it is still within reasonable range compared to other co-design methods, which include 72-73% for the original paper La-Proteina, 73% for APM [1], 78% for ProteinGenerator [2] and near 90% for FoldFlow [3]. While the natural PDB data has roughly 39% and our training AFDB has 32% of helical structures, the helical bias is a commonly observed limitation of existing co-design methods [4].

We then provide a further analysis in why latent warmup tends to encourage helical bias. Latent warmup partially denoises the sequence latent before joint denoising, providing a noisy sequence-informed prior. We examined 100 length-150 trajectories following the exact warm-up configuration. By feeding the partially denoised latent along the trajectories into the decoder, we found increasingly preference towards residues with strong helix propensity: Alanine (A), Glutamate (E), and Arginine (R), with decoder amino acid type probabilities increase from 5.37/5.02/6.1% (at t=0) to 11.12/21.68/19.39% (at t=0.3).

We agree that more helical structures may contribute to Mixta’s observed co-designability increase from default to latent warmup. The DPLM-2 paper raises a similar concern towards the correlation between helix-rich composition and favorable self-consistency RMSD. However, Mixta’s diversity also increases, which means that it is not a global structural collapse to a small group of helix-enriched structures.

[1] Chen, R., et al. (2025). An all-atom generative model for designing protein complexes. In Proceedings of the 42nd International Conference on Machine Learning (PMLR, Vol. 267, pp. 9516–9545).

[2] Lisanza, S. L., et al. (2025). Multistate and functional protein design using RoseTTAFold sequence space diffusion. Nature Biotechnology, 43(8), 1288–1298.

[3] Bose, J., et al. (2024). SE(3)-stochastic flow matching for protein backbone generation. In The Twelfth International Conference on Learning Representations (ICLR 2024).

[4] Li, A. J., & Kortemme, T. (2024). ProteinZen: Combining latent and SE(3) flow matching for all-atom protein generation. Machine Learning in Structural Biology Workshop, NeurIPS 2024.

AuthorsOpenReview ↗

Addressing Weaknesses pt 2

W3: Mixta versus Nanocuerpo and component ablations

We thank the reviewer and agree that Nanocuerpo versus Mixta is a system-level comparison, not a controlled ablation of the multimodal latent space. We have revised the manuscript to make this distinction explicit.

The models represent different approaches. Nanocuerpo fine-tunes Proteina-Complexa, retaining its pretrained architecture and latent representation. Mixta is a standalone model that revisits the autoencoder, latent representation, flow architecture, and training procedure to support multimodal generation. Table 1 compares these complete systems; it does not isolate which changes cause the performance gap.

We have added a controlled ablation of triangular pair updates. Two from-scratch runs use the same autoencoder, training mixture, optimizer, curriculum, and matched initialization of shared parameters, differing only in the pair-update blocks. Evaluation of 1.02M designs across four targets with identical sampling and confidence gates gives:

TargetWithout pair updatesWith pair updates
BHRF10.42%9.65%
PD-L10.73%6.67%
TrkA0.27%3.05%
Insulin2.53%11.57%
Pooled1.14%7.79%

Triangular updates substantially improve computational success when training Mixta from scratch on the combined monomer, binder, and VHH mixture. However, they are not required for strong VHH performance: Nanocuerpo achieves 27.1% success on PD-L1 without triangular updates by fine-tuning a pretrained model. These results establish the value of triangular updates in our from-scratch setting, while showing that strong performance also depends on pretraining and training history.

On the other factors. VAE fine-tuning on SAbDab2 yielded no measurable gain in our initial experiments; we retained it because we observed no degradation. Appendix E.3 separately reports the effect of checkpoint construction relative to its parent checkpoints, holding the architecture and latent space fixed. A full factorial study over trunk, autoencoder, optimizer, curriculum, and checkpoint selection exceeds our compute budget. We therefore report the controlled comparisons above and acknowledge that they do not fully disentangle the final performance gap.

What we claim for the multimodal latent space. The contribution is additional capability alongside strong structural generation: Mixta supports sequence-only, backbone–sequence, and all-atom generation, and can use structure-free sequence data to steer the sequence properties of generated structures. Nanocuerpo, with its inherited architecture and latent representation, does not support these additional modes; increasing trunk capacity alone would not introduce them. These demonstrated capabilities motivate Mixta’s redesign as a standalone model. We report its VHH performance as a result of the complete system, rather than evidence that the multimodal latent space alone causes the improvement over Nanocuerpo.

W4: Checkpoint selection and independent evaluation

We thank the reviewer. We agree that checkpoint selection and reuse of the evaluation predictor require explicit accounting.

Checkpoint merging is an additional model-selection step. The four targets were used to select a single merged model, which is then applied unchanged to all seven targets; we do not construct target-specific weights. Nevertheless, these four targets inform model selection, and merging provides substantial gains—for example, from 8.7% to 35.9% on PD-L1. We have clarified this distinction and separate results on selection targets from those on the remaining three targets.

The baselines did not receive equivalent checkpoint selection. We have access to their released final weights, but not the compatible intermediate checkpoints needed to apply the same procedure. This limits the fairness of attributing the final performance gap to the model alone. We therefore report unmerged Mixta alongside the merged model, with individual parent-checkpoint results in Appendix E.3 and Table 10. This makes the contribution of merging visible and allows comparison without that additional step.

We have added evaluation with a predictor not optimized by search. Table 1 now reports ESMFold2 rescoring for every method, including Mixta search. [Insert the principal quantitative result.] Protenix-v2 remains our primary evaluator because it performed best on our experimental-structure benchmark (Appendix F.2); ESMFold2 provides a computationally tractable test of whether improvements extend beyond the optimized predictor. We explicitly acknowledge that Protenix-guided search evaluated with Protenix measures predictor-specific optimization, and that neither predictor establishes experimental binding.

AuthorsOpenReview ↗

Answering Questions

Q1: Training duration and checkpoint selection

Both models are trained for 160K steps, with the best checkpoint selected based on best insilico generation performance (at 80K steps for La-Proteina and 110K steps for Mixta).

Q2: Sequence-length coverage and experimental scope

The original La-Proteina model was trained up to length 512 at its second training stage and 800 at its third training stage. That’s why it is able to do structural generation up to length 500-800.

Instead of claiming state-of-the-art generation across all lengths, the monomer experiment mainly aims to answer the question “whether a multi-regime latent space retains all-atom co-design quality while additionally supporting sequence training and generation”.

Since Mixta is trained up to length 256, we retrain La-Proteina on the same data to ensure strict experimental control. This length range is sufficient for our antibody application purpose. Antibody typically has length up to 130, which is well covered by the training length distribution.

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

Proteina Mixta uses flow matching in a shared representation learned from protein sequences and structures. Synthetic loop/framework annotations on general protein interfaces provide design supervision. Antibody sequences supply additional structure-free training. The authors report improved computational design success on several targets. Sequence co-training increases loop diversity and reduces sequence liabilities, while lowering predicted interface success.

Critical strengths
  1. Learning from structure-free sequences addresses an important data bottleneck without requiring predicted repertoire structures as labels.
  2. Monomer ablations support the representation, co-designability improves.
  3. Direct nanobody generation shows substantial computational gains without sequence redesign.
Critical weaknesses
  1. The final model cannot be recreated from the description. The nanobody results use a model assembled from several checkpoints. Recreating it requires the exact parameter groups and selection mask Q, but these are missing despite being promised in E.3 and the reproducibility statement.
  2. Sequence co-training may not improve the pool of successful designs. Adding OAS sequences improves diversity and reduces undesirable sequence motifs across generated outputs, while computational success falls from 15.77% to 10.84% (Table 2). The table does not show whether designs that pass are more diverse or have fewer liabilities. The practical benefit therefore remains uncertain.
  3. Search is evaluated using the predictor it optimizes. Search improves a Protenix score that also helps determine whether designs pass. This limits evidence for improvements beyond that predictor. H.1's claim of independent evaluation needs to be restricted to the separate RL experiment. The main results also use target-specific sampling settings which means efficiency claims need clear accounting of tuning costs.
  4. The RL explanation is confounded and some summaries disagree with tables. Table 24 changes the redundancy penalty, composition reward, KL weight and training duration together, so H.4 cannot attribute the diversity change to the redundancy penalty alone. Also, 4.4/H.7 contradict Table 22's small baseline-success counts, and G.2 says TNP-clean measurements are unavailable although Table 20 reports them. These require reconciliation.
Questions for authors

How do synthetic CDR labels transfer to real nanobodies? Training labels loops by contact count rather than sequence order. What evidence shows that distinguishing CDR1/2/3 helps?

What produces the monomer improvements? Can existing comparisons separate the benefit of the autoencoder objective from mixed sequence/structure training of the flow model?

How novel are the generated interfaces relative to training examples?

3 comments
AuthorsOpenReview ↗

Addressing Weaknesses

W1: Reconstructing the final model

We thank the reviewer for identifying this omission. The parameter-group and selection-mask definitions were inadvertently commented out in the submitted version.

Revised Appendix E.3 (changes in purple) specifies the exact tensor-name suffixes defining both residual groups, the selection mask QQ, and the scope of the final update. The code release includes a loadable mask, checkpoint-construction scripts, and SHA-256 hashes for every parent and merged checkpoint. Together, these specify how to reconstruct the final model. We also updated the reproducibility statement.

W2: OAS benefits among successful designs

We agree that the original table did not establish whether sequence-level benefits extend to successful designs. We added this analysis in Appendix F.4, Table 19.

The benefits persist among gate-passing designs. CDR3 families per 100 passing designs increase from 66.76 to 75.73 (+13.4%, p=0.0097p=0.0097), while the largest-family fraction decreases from 0.1476 to 0.1119 (−24.2%, p=0.0218p=0.0218). Sequence liabilities decrease by 15.5% among passing designs and 14.4% among failing designs. OAS thus improves diversity and sequence quality within the passing subset, although the lower pass rate (15.77% to 10.84%) remains a trade-off.

This ablation tests additional OAS data, not sequence-only training itself. Both arms include sequence-only, backbone–sequence, and all-atom training, with otherwise matched settings (Sec. 4; Appendix, Mixta Training).

The pass-rate decrease does not by itself establish reduced binding. Our confidence gates reject 357/508 known PDB VHH–antigen complexes (Appendix F.2). The distribution shift characterized in Appendix F.4 may also challenge the structure-prediction oracle. We distinguish the demonstrated sequence-level benefits from in silico interface performance.

W3: Predictor dependence and tuning costs

We agree that evaluating search with the predictor it optimizes limits the conclusions, and that target-specific tuning must be included in efficiency accounting. We made three changes.

  1. Clarified predictor dependence. Following Proteina-Complexa, the main search optimizes Protenix interface PAE, and Protenix also contributes to the success gates. We now state this explicitly in the main text and restrict H.1’s independence claim to the separate RL experiment, which uses ESMFold2-Fast for reward and Protenix-v2 for evaluation. Validation against experimental structures (Appendix F.2) assesses oracle reliability but does not remove this dependence.

  2. Added evaluation with a predictor not optimized by search. Table 1 reports ESMFold2 rescoring for every method, including Mixta search. This supports improvements beyond the optimized predictor, while remaining a computational assessment rather than experimental validation of binding.

  3. Clarified tuning scope and shared-setting performance. Tuning is a grid over two scalar noise scales with shared weights and no retraining: at most four temperature pairs per target, one always the default. The shared-versus-tuned comparison (Appendix Table 26, extending Table 24) shows that shared search settings reproduce the tuned results on all four development targets, with tuning changing only the three remaining targets. For direct generation, one universal pair across all seven targets recovers most of the tuned performance.

W4: Corrections and consistency

We thank the reviewer for these catches. All were correct and are fixed in the revised manuscript (changes in purple).

Confounded comparison (Appendix H.4, Table 24). The checkpoints differ in the redundancy penalty, composition-reward weight, KL weight, and selected checkpoint step. H.4 now lists these differences and describes the comparison as uncontrolled. The difference in CDR3 family counts is no longer attributed to any single setting. We removed the statement about evaluation hardware, which Table 24 did not report.

Text versus Table 22 (Sec. 4.4, Appendix H.7). The text stated that baselines had no successes on the other targets, but Table 22 reports:

  • Insulin: AbLang2, 1/1000; Mixta, 2/1000.
  • PD-L1: one success each for Mixta and both ADAPT variants; none for AbLang2.

Sec. 4.4 and H.7 now state these counts. The RL checkpoints still yield no passing held-out designs. We also corrected the second checkpoint’s rate to 42.2% (422/1000) in the Appendix H introduction and described the starting checkpoint as on par with AbLang2 and ADAPT rather than indistinguishable.

TNP-clean rates in Appendix G.2. The statement that TNP-clean rates were unavailable for shared-temperature populations was incorrect. G.2 now reports them from Table 20: the shared (0.05, 0.02) pair retains 90.1% of tuned TNP-clean successes (1,143 versus 1,268).

AuthorsOpenReview ↗

Answering Questions

Q1: Pseudo-CDR labeling

We thank the reviewer. Pseudo-CDR labels define slots for framework-conditioned design, not true antibody CDR boundaries. The labeling rule determines whether those slots have distinct interaction roles.

  • Why distinguish CDR identities? Framework-conditioned nanobody design specifies CDR1/2/3 as the regions to generate. These loops have different roles and properties, motivating distinct labels here and in prior antibody modeling. All models use CDR-specific label dropout to avoid overfitting to these identities and also learn generic CDR representations.

  • What contact ordering achieves. On TED dimers with identical loop placement, sequence-position labels make CDR1/2/3 interchangeable: CDR3 is the strongest-contact loop in approximately one-third of cases, i.e. chance. Contact-based labels raise this to 76%, qualitatively resembling SAbDab nanobodies. However, pseudo-CDR2 has too few contacts and pseudo-CDR3 is shorter than natural CDR3, so we do not claim natural-loop equivalence.

  • Downstream comparison. We compare contact-ordered labels, CDR-identity conditioning, and CDR-weighted hotspots against position-ordered labels without CDR-identity features, holding initialisation, autoencoder, and compute fixed. Proteina-Nanocuerpo generates 1,000 designs per target, evaluated with the Protenix-v2 gate:

Unique success rateWithoutWith
PD-L126.7%27.0%
BHRF11.5%3.0%
Insulin2.0%3.3%
TrkA0%3.6%

Success is target-dependent, with no significant overall difference across targets. This evaluates the combined labeling-and-conditioning recipe, not contact ordering alone.

Q2: Autoencoder and generative objectives

We agree that our comparisons establish the benefit of the combined framework without fully separating the autoencoder and generative objectives. These components are coupled: the multi-regime autoencoder provides the representations needed for multi-regime flow training, rather than serving as an interchangeable replacement for the original La-Proteina AE.

What the existing ablation establishes. The autoencoder can affect generation through its latent distribution, including in all-atom mode. Under the same multi-regime diffusion setup, the simple-masked and packing AEs perform slightly worse than Mixta AE (Appendix Table 4). Alongside the distinct latent layout in Appendix Figure 6, this suggests that latent space matters, but does not identify which properties drive performance.

Two factors complicate attribution. First, changing the autoencoder also changes the decoder, so decoded all-atom quality cannot isolate latent geometry. Second, flow matching jointly denoises Cα coordinates and latents, so latent-only analysis cannot capture joint denoising difficulty. Extending analyses of diffusability [1] to this coupled setting requires a larger controlled study, which we leave to future work.

A possible mechanism, not a causal conclusion. One plausible explanation for the monomer gains is regularization from complete structural-information dropout in sequence-only training. In the default regime, Cα coordinates are denoised ahead of latents, allowing sequence prediction to rely heavily on structural cues. Sequence-only training instead requires the shared model to learn evolutionary sequence patterns, potentially reducing shortcuts based on local sequence–structure correlations, such as repeatedly favoring a small set of helix-associated residues. This is consistent with Figure 3(c), where sequence diversity improves more than all-atom structural diversity.

[1] Skorokhodov, I., et al. (2025). Improving the diffusability of autoencoders. ICML, PMLR 267, pp. 55876–55905.

Q3: Interface novelty

CDR3 forms the bulk of the generated interface, so we assess its sequence and backbone similarity to training interfaces. The pseudo-VHH pathway constructs CDR-like segments from generic TED dimers, not antibody paratopes. We search the underlying training chains, not only segments selected during stochastic pseudo-CDR placement.

Sequence novelty. Across 15,240 unique generated CDR3s, mean maximum sequence identity against all 153M 18-residue windows from 587,597 training dimers is 56.4%, close to the composition-preserving shuffle baseline (56.1%), with no exact matches and none above 80% identity. Searching all windows avoids underestimating similarity by restricting the reference to selected pseudo-CDRs.

Backbone novelty. Against 2.97M interface-proximal training loops, generated CDR3s have a mean nearest-neighbor backbone RMSD of 1.45 Å, versus 0.91 Å for held-out loops. Only 18% have a sub-angstrom match, versus 65% of held-out loops. Gate-passing designs are no closer to training examples than other designs.

Together, these results support sequence and backbone novelty in the dominant interface region relative to the pseudo-VHH training corpus.

Dear Authors,

Thank you for documenting the checkpoint mask in the appendix and showing that OAS improves diversity among passing designs. I stand by my original score because the main comparison still leaves unclear how much multimodal training contributes to the higher fraction of designs that pass. Also, the rebuttal’s reference to a code release is misleading (" The code release includes a loadable mask, checkpoint-construction scripts, and SHA-256 hashes for every parent and merged checkpoint."). I could find any release code files.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.