acceptodds
Withdrawn from ICLR 2027

TTDesign: A Test-Time Design Paradigm for Protein Inverse Folding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Protein language models (PLMs) have advanced the development of protein inverse folding, which aims to design amino acid sequences that fold into a target backbone structure. However, their strong generalization capabilities do not always translate into effective designs for individual target structures. In this work, we propose TTDesign, a novel test-time design paradigm for protein inverse folding that tailors sequence designs to each target structure without requiring any model parameter updates. TTDesign establishes a self-optimization loop with sequence-structure consistency as its guiding objective, iteratively refining sequence designs through structural feedback. Furthermore, TTDesign incorporates a retrieval augmentation mechanism to enhance this process with external protein knowledge. Extensive experiments on CATH v4.2 demonstrate that TTDesign outperforms all baselines in sequence-structure consistency, achieving an average scTM-score of 0.835 compared with 0.802 for the strongest baseline. TTDesign also achieves the best performance on the TS50 benchmark, demonstrating the strong generalization ability of TTDesign.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein inverse folding, test time optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_yDZ2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper tackles protein inverse folding with a test-time approach: given a backbone, design a sequence that folds into it. The idea is to optimize each target separately at test time, with no weight updates.

TTDesign works in two steps. Foldseek retrieves the ten most similar chains from the CATH training set, and their aligned residues are blended with ESM-3's probabilities to initialize sequence logits. Those logits are then optimized for 200 steps: sample a sequence with Gumbel-Softmax, fold it with a frozen ESMFold, compute FAPE and distogram loss against the target, and backpropagate. The best candidate by TM-score is returned.

Contributions as claimed: a test-time design paradigm needing no fine-tuning, a self-optimization loop driven by structural feedback, a retrieval mechanism for external structural priors, and best-in-class results.

Validation: CATH v4.2, CATH v4.3 and TS50 against several baselines, using scTM, RMSD, pLDDT and a success rate. Designed sequences are folded with ESMFold2 rather than the in-loop ESMFold. There is also a breakdown by retrieval quality and an ablation separating the two components. Sequence recovery, runtime and variance are not reported.

Critical strengths
  1. The retrieval-quality analysis is useful and numerically consistent. Table 4 groups targets by retrieval quality, which is appropriate for this method. Its weighted averages match Table 1, and the reported percentages are correct. My concerns are about the evaluation design, not numerical reporting.

  2. The ablation shows that optimization helps without retrieval. Optimization alone improves ESM-3 from 0.799 to 0.814 scTM. This is a relatively clean comparison using the same base model and initialization source, suggesting that structural feedback provides some benefit on its own.

Critical weaknesses
  1. Both components lack novelty, follow much standard computational designing approach. E.g., these closely resemble prior work (that are also not cited): The optimization loop resembles AfDesign's fixed-backbone hallucination, including the same loss weights. The retrieval step is also similar to Refold, published seven months earlier. Neither is cited, and the paper does not clearly establish what is new or compare against these methods. It appears as careful engineering than actual methodological (or theoretical) novelty.

  2. The evaluation does not directly test the main claim. The method selects the candidate with the best TM-score and then reports scTM, which is nearly the same metric. Also no baseline uses the same 200-sample budget and ranking method. So the gains may come from "best-of-200" selection rather than the proposed loop. The paper does not rule this out.

Questions for authors
  1. How does TTDesign compare to ESM-3 with best-of-200 sampling, ranked by the same ESMFold TM-score?
  2. How does the method differ from AfDesign fixed-backbone hallucination, and from the retrieval approaches in PRISM and Refold (both are recent, 2025-'26)?
  3. What is the sequence recovery on CATH v4.2, CATH v4.3 and TS50? How do they compare with the closely related baselines?
  4. Was ESMFold2 run in single-sequence or MSA-conditioned mode for the reported numbers? This matters a lot for the foldability scores, especally in retrieve+generation based approaches like this one.
  5. What is the wall-clock cost per target, and where does performance saturate as T increases? How does it compare with the closely related baselines without any test-time guided refinements?
Reviewer_7a93OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TTDesign combines retrieval-based sequence initialization with test-time optimization for protein inverse folding. It updates sequence logits through a frozen ESMFold model using structural losses, while keeping model weights fixed. Experiments on CATH v4.2, CATH v4.3, and TS50 report improved structural consistency, with ESMFold2 used for final evaluation.

Critical strengths
  1. Improvements are consistent across the three benchmarks. On CATH v4.2, scTM-score increases from the strongest baseline’s 0.802 to 0.835, while RMSD decreases from 4.628 to 3.815 Å.

  2. The ablations show that retrieval and optimization work better together than either component alone. Using a separate folding model for final evaluation also provides useful evidence that the gains transfer beyond the optimization model.

Critical weaknesses
  1. The “new paradigm” claim needs better positioning. ESM-Design already optimizes sequences by backpropagating through ESMFold, while AFDesign supports FAPE/distogram-based fixed-backbone design. These closely related approaches are neither discussed nor compared, leaving the methodological advance unclear.

  2. The comparisons do not control test-time compute. TTDesign runs 200 folding forward/backward steps and selects the best candidate, but baseline sampling/search budgets, runtime, and memory are not reported. Although ProtInvTree is included, the results do not establish an advantage over simpler search strategies at comparable cost.

Questions for authors
  1. How does TTDesign compare with ESM-Design or AFDesign using comparable initialization and computation budgets?

  2. Do the gains remain against best-of-N ESM-3 sampling and ProtInvTree at equal wall-clock budgets? Please report runtime, peak memory, and variation across seeds.

  3. What is the exact rule for combining retrieval evidence with base probabilities, and what Gumbel-Softmax temperature or schedule is used? These details are missing from the current description.

Reviewer_HX1dOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper propsoes a test-time design paradigm, TTDesign, for protein inverse folding, which targets at specific target structures without model fine-tuning. For this purpose, a self-optimization loop and retrieval augmentation mechanism are adopted. Evaluations on related benchmarks show the effectiveness of the proposed method.

Critical strengths
  • This paper is clearly-motivated. Protain inverse folding for individual target structure is a practical but underexplored problem, which is worthy studying.
  • The proposed paradigm is generally reasonable, and combines the strengths of both traditional and deep learning methods.
  • The experimental evaluations are extensive and effective.
Critical weaknesses
  • The overall methodological novelty and significance are limited. Methodological novelty and significance. Gradient-based sequence optimization through backpropagation has been explored in prior work on protein design. The authors should more clearly distinguish the proposed method from existing approaches and substantiate its methodological novelty. Alternatively, they should better articulate the specific challenges addressed by the method and explain its significance for advancing protein inverse folding.

  • The fairness of computational comparisons is problematic. The method performs 200 rounds of folding and backpropagation for each target, which may incur substantial computational costs. However, the paper does not provide comparisons with competing methods under matched computational budgets. To establish the effectiveness and practical value of the proposed approach, the authors should include compute-matched baselines and report relevant computational costs, such as the number of model evaluations, GPU time, and total inference time.

Questions for authors

Please refer to the weakbesses points above.

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

Related papers

Discussion (0)

Sign in to comment.