acceptodds
Withdrawn from ICLR 2027

C2S: Condition-to-Simplex Protein Design with Response-Aware Sequence Selection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Protein design seeks sequences that realize desired behaviors, yet structure-conditioned methods largely optimize compatibility with prescribed conformations rather than behavior under a specified context. We introduce Condition-to-Simplex (C2S), which recasts condition-specific design as control over a simplex of structural states. C2S translates a functional specification into continuous state preferences and introduces -FiLM to act on state contrasts, steering sequence probabilities according to the desired regime rather than a fixed conformation. A response model then resolves the remaining choice in functional space, selecting candidates by their predicted behavior under the specified context. Across structural-control and measured-response tasks, C2S improves both request following and functional target matching. This formulation provides a path from designing for structural states to designing for context-dependent protein behavior.Our code is available at https://anonymous.4open.science/r/papercode.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein design, Multistate inverse folding, Conditional generation, Protein function prediction, Response-aware sequence selection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_aCTAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes using one functional request to drive the whole design. The user says what the protein should do and under which conditions, and the same request both steers which sequences get generated and selects the final one. The state-contrast part is supported by theory, but the gains are small compared to the baselines.

Critical strengths
  • The control is clean and the mapping from concentration to simplex weight is a nice idea.
  • The technical evaluation is careful and negative results are also reported.
Critical weaknesses
  • The main baseline compares a model trained on the assay against inverse-folding likelihood which was never meant to predict a concentration-dependent response. It mostly shows that a supervised predictor helps.
  • The pool comparison is against random, which is a floor and not a real baseline, and the fixed-concentration test only works in one of the three target groups. The main novelty provides the weakest numbers.
  • ProteinMPNN outerforms their own pool on their own task, and under the matched protocols it wins everywhere. Two of the other conformational sources also have better point estimates.
  • On the motif transfer the target likelihood drops while the direction margin grows. That can mean the model is just pushing the two states apart without actually fitting the requested one better, and nothing in the paper separates these two cases.
  • Boltz-2 numbers/plot are never reported.
  • Diversity and novelty are never looked at as the request changes.
Questions for authors
  1. How does the error change as the budget goes from 1 to 19? Somewhere in between the request effect should disappear, and it would help to see where.
  2. Could you show the preference against the request weight directly, instead of only the slope gains? I would like to see if the response is smooth or driven by the endpoints.
Reviewer_5pXLOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes C2S, a function-conditioned protein sequence design framework that maps functional requests to weights over bound and unbound structures. A π-FiLM module modifies DynamicMPNN representations. The authors evaluate structural-state controllability on protein conformational ensembles and functional selection using PYL1–ABI1, PDZ, and Tite-Seq datasets. While the proposed state-contrast modulation is technically interesting, I have substantial concerns about the underlying motivation and experimental validation. In particular, the relationship between functional objectives and structural-state weights is not sufficiently justified, and the reported results do not convincingly establish an advantage of C2S over simpler inverse-folding methods.

Critical strengths
  • The manuscript is easy to follow, and the plots are useful for understanding the method.
  • π-FiLM introduces only 11K trainable parameters and improves structural preference controllability over scalar-FiLM at unseen weights. This demonstrates a potentially useful mechanism for multistate inverse folding.
Critical weaknesses
  1. Insufficient justification for the function-to-structure mapping. The paper directly maps binding occupancy or normalized functional response to structural-state weights. However, binding does not necessarily involve conformational changes, and there is no general relationship between functional response and conformational populations. The physical meaning of these weights remains unclear. Moreover, the PDZ experiment always uses π = [0, 1], making the need for multistate conditioning questionable. Meanwhile, Related work on directly optimizing inverse-folding models toward desired properties, such as ProteinDPO and EnhancedMPNN, should also be discussed.
  2. Insufficient evidence that the proposed method improves protein design. The functional evaluations rely heavily on separately trained supervised response predictors and existing mutation libraries. These predictors are trained on limited protein diversity: PYL1 contains only 3,211 single mutants of one protein, PDZ covers five domains and seven interactions, and the Tite-Seq source library contains approximately 1,600 designed antibody variants. Such settings raise concerns about generalization to unseen proteins and de novo designs.
Questions for authors
  1. How is the mapping between functional response and structural-state weights justified? Is there evidence that these weights correspond to actual conformational preferences?
  2. What is the independent contribution of π-FiLM to design performance? Could the authors report the exact number of training sequences for each response predictor and evaluate generalization to unseen proteins? A direct comparison against bound-only ProteinMPNN/DynamicMPNN under the same response selector, together with independent structural (such as AlphaFold) or functional evaluation (Boltz-2), would materially strengthen the claims.
  3. The independent Boltz-2 structural evaluations show no clear improvement. More importantly, Appendix F reports that a simpler ProteinMPNN endpoint baseline outperforms C2S on PYL1 (Top-8 error: 0.228 vs. 0.241).
  4. Minor: There is a rendering problem in line 230 for DtϕD_t^\phi
Reviewer_kng5OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes C2S, which maps a functional request (ligand concentration or binding partner) to a weight π over the unbound and bound structures of a protein. π is used to weight the features of a frozen DynamicMPNN, together with a small learned module (π-FiLM). A separately trained response predictor then picks the final variant closest to the target response. The authors report that the generator’s outputs follow the requested weight more closely than matched controls, and that response-aware selection on PYL1 and PDZ gives lower target error than structural ranking.

Critical strengths
  1. The problem direction is reasonable: designing for condition-dependent function rather than a single fixed structure.

  2. The evaluation protocol is fairly careful (parameter-matched scalar-FiLM control, held-out positions/domains, paired bootstrap).

Critical weaknesses

W1: Motivation. The introduction defines the gap as “the request does not control both generation and selection” and proposes controlling both as the solution, without showing where the usual generate-score-select pipeline actually fails. It also does not explain why function should be expressed as a weight between two structures.

W2: Preference for a structural state is measured by the difference in inverse-folding log-likelihoods under the two structures. This reflects compatibility with a backbone rather than conformational population. The only other check, Boltz-2, is a single-structure predictor and shows no clear gain. It is therefore unclear whether the designed sequences actually shift the conformational equilibrium.

W3: The “unified formulation” that “closes the loop” between generation and selection looks largely disconnected. The generator and the response predictor are trained separately and interact only through the candidate pool and the shared request value. There is no joint objective or feedback.

W4: The reported functional gains (e.g., PYL1 19.1%) do not seem attributable to the proposed mechanism. The evaluation task is much narrower than the “design” framing suggests. In PYL1 and PDZ, each decision is to pick one of at most 19 already-measured substitutions at a single position. Moreover, the improvements compare a supervised selector against structural ranking on the same candidate pool, so they mainly reflect the selector. When the generator is made request-free (π = 0.5), the error is on average about the same (0.242 vs 0.241, Table 11), so the benefit of conditioning on the request is not clearly shown.

W5: Writing. The introduction is hard to follow, and some terms are not defined clearly in the main text.

Questions for authors

Q1: Can the claimed shift in state preference be validated with a conformational sampling model (e.g., AF2 with MSA subsampling or BioEmu), by estimating the open/closed population of designed sequences at different π?

Q2: At Top-8, the request-free (π = 0.5) pool and the request-conditioned pool give essentially the same (0.242 vs 0.241, Table 11). What does conditioning the generator on the request add beyond a generic generator ranking followed by the selector?

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

Related papers

Discussion (0)

Sign in to comment.