acceptodds
Under review as a conference paper at ICLR 2027

Multi-Scenario Generative Protein Design via Preference Alignment

Abstract

Generative models can now design proteins that perform one function in one context: bind a given target, scaffold a given motif, catalyze a given reaction. However, many applications need a sequence that behaves differently depending on its context, like a binder that engages two unrelated targets, or a switch that adopts one fold on its own and another once a ligand is bound. Training a model for such multi-scenario tasks directly is impractical, since structural data of the same protein across contexts is scarce. To address this, we introduce Control-Proteina (Ctrl-P), which instead coordinates separately trained models at inference time, eliminating the need for training on multi-scenario data: each model expresses its preferences as a population of sequences and conditions its own generation on the preferences of the others, so that iterative refinement converges on designs that succeed in every context. In a multi-target protein–protein interaction benchmark that spans similar and dissimilar binding interfaces, Ctrl-P solves 18 of 20 target pairs, including 8 of the 10 challenging dissimilar pairs where the strongest baseline solves 2. We further demonstrate the generality of our approach by applying Ctrl-P to design allosteric motif switches that adopt specified conformations in response to a binding partner: it solves 8 of 12 tasks, compared to the strongest baseline’s 7, while finding more than twice as many successes per task on average. A case study on pan-arenavirus binder design shows that Ctrl-P generalizes beyond two scenarios, yielding candidates predicted to engage all three targets where no evaluated baseline yields any. Together, these results extend protein design toward programmable behavior across molecular contexts.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein design, generative models, multi-objective design

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 3 reviews, above 85% of the others.
Reviewer_QCxuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Ctrl-P, a preference alignment framework for multi-scenario protein design. Instead of training on scarce paired multi-context data, Ctrl-P coordinates multiple single-scenario generators by exchanging ProteinMPNN-derived sequence preferences during inference. Through an iterative negotiation process, it designs sequences that satisfy multiple objectives, achieving strong performance in multi-target binder design, motif switching, and multi-target viral binder design.

Critical strengths
  1. The paper addresses an important but challenging problem in protein design: generating proteins that satisfy multiple contextual objectives. The proposed preference alignment framework provides a practical alternative to directly training multi-scenario models, which is difficult due to the scarcity of paired multi-context structural data.
  2. The method demonstrates substantial improvements over existing baselines on challenging benchmarks, especially for dissimilar target pairs in multi-target binder design.
  3. The negotiation procedure, temperature-controlled preference strength, and preference population representation are well motivated and provide a reasonable mechanism for balancing exploration and exploitation during multi-scenario optimization.
Critical weaknesses
  1. The statement “we cannot always compare refolded structures to model-generated structures because of multiple contexts” is not convincing. In fact, for each designed sequence, one can obtain context-specific predicted complexes (e.g., design–target A and design–target B) and compare these structures against the corresponding model-generated structures. The current evaluation appears to rely heavily on confidence metrics such as pLDDT, which may not adequately capture whether the designed sequence adopts the intended conformation in each context. A direct structural comparison (e.g., backbone RMSD or interface-level metrics) between predicted and generated complexes would provide a more rigorous assessment.
  2. I question whether the preference-conditioned generator is genuinely performing generative design or whether it is functioning primarily as a folding model conditioned on sequence preferences. The ProteinMPNN-generated sequence populations provide an MSA-like representation that may already encode substantial structural information. Consequently, the role of the generative model during iterative sampling appears unclear: beyond providing an initial structural proposal, subsequent rounds seem to mainly optimize consistency between the structure and the supplied sequence preferences. It would therefore be important to clarify what additional design capability the generative model provides over a folding model with similar sequence-conditioning inputs.
  3. The benchmark lacks structural visualization and qualitative analysis. Since the benchmark is newly constructed and evaluation relies heavily on folding-model predictions, it is important to show representative target structures and successful designs. Folding models may have biases, particularly for challenging multi-scenario tasks, and confidence-based metrics alone may not fully reflect whether the generated structures are physically plausible. Structural inspection would provide additional evidence supporting the reported improvements.
Questions for authors

The definition of "scenario" can be more general. Can Ctrl-P handle multi-ligand scenarios where different ligands induce different states, and can it design protein–protein interaction-driven motif switches where different binding partners trigger distinct conformational states? Can it design (with protein target A, without protein target A)?

4 comments
AuthorsOpenReview ↗

Response to review

Thank you for your review. Please find below our response to your question and the concerns you raised.

  1. On the self-consistency RMSD criterion The table below shows the number of challenging tasks solved for Ctrl-P and Proteina-Complexa with beam search when including scRMSD as an additional criterion. Note that, in our negotiation algorithm, at any given round a design is produced by a generative model conditioned on a single target. Thus, for each design, there is only an associated structure for one of the two targets, and this is the structure we use for the self-consistency metric. The same also applies for the beam search baseline. The confidence metrics are evaluated against both targets, since they only depend on the sequence of the design. Overall, we see that the scRMSD does not tend to gate successes compared to the other confidence metrics. We have also added these results into Appendix D.12 of the paper. Given these results, we would opt to keep our current success criteria for the main results. Metrics that only depend on the design sequence are more general and allow us to evaluate against more baselines using the exact same criteria. For example, BindCraft does not have any generated structures to compute scRMSD against since it directly optimizes the design sequence. | | No gate | scRMSD < 3 Å | scRMSD < 2 Å | scRMSD < 1.5 Å | |---|---|---|---|---| | Ctrl-P | 8 | 8 | 8 | 8 | | Proteina-Complexa + beam search | 2 | 2 | 2 | 2 |

  2. On the advantage over a folding model Yes, we have evidence that the generative model adds value beyond the folding model. We measured successes on one of the hard multitarget tasks while allowing the structure to change more or less between rounds. As shown in the table below, successes peak when we allow the structure to change, as opposed to keeping the structure fixed and only redesigning the sequence (k controls the amount of redesign, where k=0 corresponds to fully redesigning the structure, and k=100 corresponds to no redesign and only inverse folding the sequence). In more detail, recall that we seed the design model's generation with the output of a folding model. This is because the design model itself is not good at folding. As described in more detail in Appendix D.3, we noise the folding model prediction, and use the generative model to denoise it. The parameter k here refers to the denoising step that we renoise to, so k=0 means fully renoising and k=100 means no renoising at all. We see that, if we only let the model do inverse folding, we find far fewer successes than if we allow it to change the structure. We have also added these results to Appendix D.3 of the paper. | k | 8 | 16 | 24 | 32 | 48 | 64 | 80 | 100 | |---|---|---|---|---|---|---|---|---| | Unique successes (Ctrl-P, DIS04) | 6 | 17 | 35 | 4 | 3 | 1 | 2 | 3 |

  3. On structural visualizations We have added several visualizations of target pairs and successful designs to Appendix D.13 of the paper. Overall we see that while the similar targets share binding interfaces, the dissimilar targets show limited alignment. The successful dual binders largely maintain their backbone geometry when binding to either target. Figure 16 shows the binding interfaces, where we can see that the binders use different residues to engage the dissimilar targets, while the similar targets tend to engage more of the same parts of the binder epitopes and residues.

Thank you again for your review, we appreciate the time and effort it takes. Please let us know if we have sufficiently addressed your concerns or if you have any follow-up questions.

Thank you for your responses. The result of "inverse folding only" is interesting and all my "critical weaknesses" are resolved. I’m still curious about the questions I raised in the review and would appreciate your thoughts on them.

Thank you, we are glad to have been able to address your concerns, and apologies for forgetting to respond to your question. Yes, the intention is for a “scenario” to be very general. As design models get better, we expect them to handle an increasing number of individual scenarios, such as protein binding, small molecule binding, nucleic acid binding, enzyme motif scaffolding, protein binding + motif scaffolding, etc. The strength of our approach is then that we can finetune any of these models to condition on preferences and thereby combine their scenarios. The models we coordinate could also be completely different, as long as they share the same preference conditioning interface. Here we restricted ourselves to bivalent binders and allosteric switches because of the scenarios supported by the models we fine-tuned. For example, the protein-protein interaction driven motif switches would require a model that can condition on a target and a motif at the same time, yet Complexa only supports target conditioning (we had to finetune a model for ligand binding + motif scaffolding for the allosteric switch experiments). Nevertheless, the fact that the same approach already works for both bivalent binders and allosteric switches gives us evidence that it can work for other combinations as well. Thus, as more generalist models become available, we aim to support these more diverse tasks as you suggested.

Thank you. All my concerns and questions are resolved. I will maintain my original positive score.

Reviewer_49XrOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Ctrl-P, a preference-alignment framework for multi-scenario protein design. It coordinates generative models through sequence-population-based preference conditioning and iterative negotiation. Experiments demonstrate promising results in multi-target binder design, allosteric motif switching, and three-target protein design.

Critical strengths
  • The proposed preference-alignment mechanism enables coordination across multiple scenarios without requiring paired multi-scenario training data.

  • Ctrl-P substantially outperforms existing baselines on multi-target binder design, particularly for dissimilar target pairs (8/10 versus 2/10), and demonstrates applicability across different design tasks.

Critical weaknesses
  • The evaluation primarily relies on computational structure-prediction metrics, without experimental confirmation. The relatively small benchmarks also limit the evidence for generalization.

  • The framework combines preference conditioning, folding-model seeding, and iterative search. Stronger controlled ablations are needed to clarify the independent contribution of preference alignment.

Questions for authors
  1. Can the authors provide additional validation across more diverse targets or independent prediction models?

  2. Can the authors compare Ctrl-P with an otherwise identical search framework without preference conditioning?

  3. How does Ctrl-P scale to more than three scenarios?

1 comment
AuthorsOpenReview ↗

Response to review

Thank you for your review. We'd like to respond to your questions and a few of the points you raised below.

  1. On additional validation In order to provide further validation, we re-evaluated all the designs produced by the two best methods (Ctrl-P and Complexa with beam search) using AlphaFold2 with the same AlphaProteo success criteria used in previous work such as Complexa (ipAE < 7, pLDDT > 0.9). We also curated an additional 20 dissimilar target pairs for a total of 30. The results are presented in the table below. We see that the Boltz-2 successes are largely validated by the AF2 successes, and that on the larger set of challenging targets, Ctrl-P maintains its advantage over baselines with 19/30 tasks solved against 9/30 (or 20/30 vs. 2/30 under the AF2 success criteria). We have added these results to Appendix D.11. See also our reply to a similar point from reviewer yB7Q. | Method | Boltz2 (tasks solved / 30) | AF2 (tasks solved / 30) | |---|---|---| | Proteina-Complexa + beam search | 9 | 2 | | Ctrl-P (cif recycling) | 19 | 20 |
  2. On method ablations In appendix D.4, we compare Ctrl-P using our varying temperature schedule against two fixed temperature schedules. The fixed infinite temperature setting corresponds exactly to the ablation you are asking about, where everything else is identical except that the model does not condition on any populations (the model is trained to ignore the population conditioning when given an infinite temperature input). With the population conditioning, the method only solves 2/10 dissimilar multi-target binding tasks, the same performance as the Complexa with beam search baseline. Thus, population conditioning matters, and so does how it is scheduled: the varying temperature schedule outperforms every fixed temperature.
  3. On more than three scenarios In this work we mainly focused on two scenarios for two reasons. First, even this setting was not adequately addressed by existing methods, and second, it already proved to be a powerful paradigm, allowing us to address both bivalent binders and allosteric fold-switching proteins. While we do not train for it directly, we did find that our method can handle more than two scenarios, as demonstrated by the viral case study. We did not evaluate on more than three scenarios so we cannot say how the method would perform, but the three scenario case gives us reason to believe it can extrapolate. We leave this as a limitation to be addressed in future work, and we have also made this limitation clearer in the manuscript, highlighted in blue in the rebuttal draft. See also our reply to reviewer yB7Q.

Thank you again for your time and effort in reviewing our work. Please let us know if this addresses your concerns or if you have any follow-up questions.

Reviewer_yB7QOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This work introduces Ctrl‑P (Control‑Proteina), an inference‑time framework for multi‑scenario generative protein design, where a single protein sequence must exhibit distinct desired behaviours under different molecular contexts (e.g. binding multiple unrelated targets, ligand‑dependent allosteric motif switching). Directly training models for multi‑scenario tasks is infeasible due to scarcity of experimental structures of one protein across multiple contexts. Instead of modifying diffusion/SDE sampling trajectories (common prior compositional approaches), Ctrl‑P fine‑tunes a base protein generative model (Proteina‑Complexa) to accept sequence‑population preference conditioning: sequence populations sampled via inverse folding represent what other scenario‑specialized models “prefer”. A stochastic negotiation search iteratively exchanges these sequence‑population preferences between per‑scenario generators to find a shared sequence satisfying all scenarios, optimizing for the minimum objective value across all conditions.

Critical strengths
  1. Novel, sampler‑agnostic composition paradigm for multi‑scenario protein design. Rather than intervene on the generative SDE/diffusion process (which breaks hand‑tuned protein‑model sampling heuristics), coordination happens via a conditioning interface on top of black‑box generative models.

  2. Strong empirical performance especially on difficult, dissimilar‑interface multi‑target tasks. Most prior methods struggle for pairs with dissimilar binding interfaces; Ctrl‑P obtains large gains on this hard subset, and also demonstrates capability extending beyond two scenarios in the viral case study.

Critical weaknesses
  1. Success metrics are entirely derived from AlphaFold‑family predicted structures and confidence scores. It remains unknown whether the promising in‑silico multi‑scenario behaviours translate to real wet‑lab protein function. Protein design in‑silico metrics frequently overestimate real‑world performance, so the practical real‑world utility of Ctrl‑P cannot be confirmed from the paper’s evidence alone. This substantially limits confidence in the real‑world impact of the presented results.

  2. While the three‑target case study works empirically, the preference‑conditioning training only exposes the model to single populations, not concatenated pooled populations from multiple other scenarios.

Questions for authors

Your evaluation fully depends on folding‑model predictions as ground truth for multi‑scenario behaviour. To what extent can existing prediction tools distinguish true multi‑context behaviour from sequences that only appear to satisfy multiple scenarios due to AlphaFold‑family hallucination?

2 comments
AuthorsOpenReview ↗

Response to review

Thank you for your review. We'd like to respond to your question and a few of the points you raised below.

  1. On folding model confidence metrics In this work we present a new computational approach for designing proteins with different behaviours in different contexts, such as binding multiple targets. Experimental validation will always remain the gold standard, and we acknowledge this as a limitation in our manuscript. In silico success is a necessary first step, since designs are only tested in the wetlab once they look promising computationally. Based on such evaluations, our method already shows a significant advantage over all baselines we compared to. Moreover, while the folding model confidence scores that we use, such as the ipSAE and pLDDT, are not perfect predictors of binding, many works have shown that selecting designs based on these metrics reliably leads to experimental successes. For example, BindCraft [1] directly optimizes these metrics and is one of the most experimentally successful design methods. Closer to our multi-scenario setting, [2] were able to design species-specific TrkA binders using an approach similar to our Proteina Complexa + beam search baseline, which we outperform in silico. They selected designs with high confidence in human and mouse TrkA, but low confidence against TrkB and TrkC. They found that 18 of 94 tested designs exhibited the desired specificity experimentally. To provide more confidence in these metrics, we also re-evaluate the two best methods (Ctrl-P and Complexa with beam search) using a second independent folding model, AlphaFold2. We also evaluated on an expanded set of 30 difficult tasks. We see that Ctrl-P maintains its advantage under both Boltz2 and AF2 success criteria, and if anything does even better under AF2. We have added these results to Appendix D.11. See also our reply to a similar concern from reviewer 49Xr. | Method | Boltz2 (tasks solved / 30) | AF2 (tasks solved / 30) | |---|---|---| | Proteina-Complexa + beam search | 9 | 2 | | Ctrl-P (cif recycling) | 19 | 20 |
  2. On more than two scenarios In this work we mainly focused on two scenarios for two reasons. First, even this setting was not adequately addressed by existing methods, and second, it already proved to be a powerful paradigm, allowing us to address both bivalent binders and allosteric fold-switching proteins. While we do not train for it directly, we did find that our method can handle more than two scenarios, as demonstrated by the viral case study. We agree that directly training for more than two scenarios, or evaluating on more than three scenarios, remains a limitation which we hope to address in future work. We have also made this clearer in the manuscript. See also our reply to a similar question from reviewer 49Xr.

Thanks again for your time and effort in reviewing our work. Please let us know if this addresses your concerns and if you have any follow-up questions.

References

[1] Pacesa, Martin, et al. "BindCraft: one-shot design of functional protein binders." BioRxiv (2024): 2024-09.

[2] Corley, N. "Biomolecular design through latent reasoning with RFD4-Proteina." preprint (2026).

Thanks for the rebuttal, this has solved my problems, I raised my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.