acceptodds
Under review as a conference paper at ICLR 2027

STENCIL: Alignment Fidelity Outweighs Template Selection in Retrieval-Augmented Protein Inverse Folding

Abstract

Retrieval-augmented inverse folding conditions design on similar templates, and recent work invests in retrieving more and better ones. Within the template-retrieval setting studied here, candidate discovery already provides informative structural evidence. We therefore focus on the next post-retrieval bottleneck: conditional on an admissible template, can its residues be registered precisely at the target positions? STENCIL addresses this question by exhaustively re-aligning retained candidates by exact structural superposition, using a single template, and decoding in one deterministic pass with 3.2M parameters. On split-safe CATH 4.2 it reaches 69.342% residue-pooled recovery, the highest among methods with runnable deterministic re-evaluations under our protocol. The decisive test leaves the template, its coverage, its TM-score and its composition untouched, and permutes only where its residues land; that alone costs 26.95 recovery points. A complete-case paired 2x2 analysis finds that the alignment contrast is statistically larger than the selection contrast. The sequence-identity analysis bounds the regime in which retrieval helps.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein inverse folding, protein design, retrieval-augmented generation, structural template

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_1pxaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the role of residue-level alignment fidelity in retrieval-augmented protein inverse folding. Unlike previous approaches that mainly focus on improving template retrieval, selection, or fusion, the authors argue that accurately aligning retrieved template residues to target backbone positions is a more important bottleneck. To address this, the paper proposes STENCIL, which first retrieves structural templates using Foldseek and then performs structural re-alignment using TM-align. The results are used to construct template sequence evidence, which is incorporated into a graph-based decoder. The main experiments are conducted on CATH dataset, they achieved notable results on it. The authors further conduct several analyses to investigate the importance of alignment fidelity.

Overall, I think the paper addresses an interesting and practically relevant problem, and the controlled experiments provide useful insights into retrieval-augmented protein inverse folding. However, I have several concerns regarding the generality of the central claim, the dependence on highly similar retrieved templates, the methodological novelty, and whether the experimental results sufficiently establish the advantages of the proposed method.

Critical strengths
  1. The paper investigates a meaningful aspect of retrieval-augmented protein inverse folding. While existing approaches often focus on finding better templates or integrating more retrieved information, this paper emphasizes the importance of accurately aligning template residues to target positions. I think this is a reasonable research direction, as inaccurate residue correspondences could directly affect how retrieved information contributes to sequence prediction.

  2. The paper includes several useful controlled experiments to investigate the role of alignment. In particular, the correspondence corruption experiments and the paired 2×2 comparison of template selection and alignment sources provide evidence that residue-level correspondence is important for downstream recovery.

Critical weaknesses
  1. I am not fully convinced that the experiments sufficiently support the central claim that alignment fidelity outweighs template selection. STENCIL's performance strongly depends on the availability of highly similar templates, and it even underperforms the template-free baseline in low-identity settings. Moreover, the ablation uses a checkpoint trained with exact-alignment evidence, which may favor this alignment strategy due to training–inference distribution shift. The correspondence corruption experiments also do not directly establish that TM-align produces more accurate alignments than Foldseek. I think stronger evaluations across different template identity ranges, independently trained alignment strategies, and direct correspondence accuracy measurements would be helpful.

  2. The experimental comparisons and overall presentation could be improved. Some closely related methods, such as PRISM and Refold, are not directly re-evaluated, making the performance comparisons less conclusive. In addition, the considerable gap between micro and median recovery, together with the computational cost of exact re-alignment, raises questions about the practical advantage of the proposed method. I also have some concerns about the presentation and formatting. The related work section could be more comprehensive, and the section heading fonts appear slightly different from the standard ICLR template. I think the authors should clarify whether the official formatting requirements have been followed.

Questions for authors
  1. Could the authors provide a more detailed evaluation across different template sequence identity ranges? In particular, does STENCIL still outperform retrieval-based and template-free baselines when highly similar templates are excluded?

  2. Does the conclusion that alignment fidelity outweighs template selection still hold when models are independently trained with different alignment strategies? I am also curious whether the 517 complete-case queries are representative of the full validation set.

  3. Could the authors provide an independent evaluation of correspondence accuracy between TM-align and Foldseek? I think it would be helpful to investigate whether improvements in alignment accuracy correlate with downstream recovery gains.

  4. Could the authors further investigate the accuracy–computation trade-off? In particular, how does recovery change when fewer candidates undergo exact re-alignment, and can similar performance be achieved with a smaller computational budget?

Reviewer_e6SxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates the impact of template selection and residue mapping on sequence composition in protein inverse folding method. The paper focuses on the question: If conditional on an admissible template, can its residues be registered precisely at the target positions? This problem was solved by thoroughly realigning the retained candidates through precise structural overlay, using a single template, and decoding in a single deterministic pass using 3.2M parameters. Natural sequence recovery rate of 69.342% was achieved on the CATH4.2 dataset.

Critical strengths

The article uses a single template, approximately 3.2 million parameters, and a single deterministic decoding. Table 3 shows that the validation set recovery rate of the complete model is 73.07%, higher than the 66.64% for directly copying template residues at the covered locations. This indicates that the model can correct template evidence rather than simply transferring it verbatim. This result supports the practical value of its simplified architecture.

Critical weaknesses
  1. The paper uses precise residue mapping as its core design motivation, but the recent baseline DualMPNN has already used TM-align to generate residue mappings and applied them to cross-branch information interaction. Therefore, this paper needs to clearly distinguish between existing alignment mechanisms and its own newly added processing steps. I believe the paper does not clearly describe this point.

  2. The paper is extremely difficult to read. The writing is riddled with repetitive constraints and pre-existing justifications, while key methods and experimental conditions remain unclear, significantly reducing readability and verifiability. The paper frequently emphasizes the protocol to which the results belong, which numbers cannot be compared, which experiments are only diagnostic, and which claims the authors did not make. These explanations are scattered throughout the main text, table notes, discussion, and multiple appendices, failing to form a concise and consistent set of experimental definitions. Most importantly, The paper is filled with a defensive writing style characteristic of AI, and I sincerely hope the author will carefully read and revise this paper before submission.

Questions for authors
  1. Compared to DualMPNN, where does the performance improvement primarily come from?

  2. DualMPNN already uses TM-align to establish residue correspondences. Please clarify the key differences in template retrieval, rearrangement, and decoder presented in this paper, and explain whether both use the same template database, candidate pool, and admission rules. Can you provide a controlled comparison to distinguish between full candidate realignment and rearrangement, and the respective benefits brought by changes in the decoder?

Reviewer_1oYqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how residue correspondence affects retrieval-augmented protein inverse folding. STENCIL retrieves structural candidates, re-aligns every retained candidate with TM-align, selects one template by the resulting TM-score, and places its amino-acid tokens at the corresponding query positions. A four-block residue graph network integrates this evidence through cross-attention and predicts the sequence in one deterministic pass. On CATH 4.2, it reports 69.342% micro recovery, compared with 65.700% for the authors’ retrained DualMPNN. Token-corruption experiments and a paired 2 × 2 analysis support the importance of accurate correspondence. The paper also shows that retrieval benefits depend strongly on template sequence identity.

Critical strengths
  1. Substantial recovery improvement under the reported protocol. STENCIL improves micro recovery over the retrained DualMPNN by 3.642 percentage points, from 65.700% to 69.342%, and improves macro recovery from 63.117% to 65.748%. It achieves this with approximately 3.2 million parameters and one deterministic forward pass. The paper also reports the unfavorable median comparison with KW-Design, which makes the performance discussion more balanced.
  2. Useful and well-controlled mechanism analysis. The full-test token-permutation experiment reports a 26.95-point recovery drop while retaining the template and other evidence fields. The complete-case paired analysis separates selection and mapping, with the reported mapping contrast of 2.577 points exceeding the selection contrast of 0.563 points. Direct template copying is also 6.43 points below the full model on the validation intervention, supporting learned correction of template evidence. These experiments provide more insight than a main benchmark table alone.
Critical weaknesses
  1. The methodological novelty and attribution remain limited. Structural retrieval, TM-align, message passing, and attention-based template conditioning are established ingredients. The main distinction appears to be exhaustive candidate re-alignment and a simpler representation of template evidence. The paper does not establish how much of the gain over DualMPNN comes from these choices under a matched implementation and retrieval budget. The internal 2 × 2 experiment supports a mechanism within STENCIL, but does not isolate the source of the entire 3.642-point improvement over DualMPNN. Ablations of cross-attention and the template representation would help establish the architectural contribution.
  2. The presentation needs more precise technical detail. The discussion repeatedly emphasizes the alignment-versus-selection message, while the network description does not specify the full message-passing block, the organization of template attention, or the mapping from quality features to attention bias. This makes the model harder to reconstruct and assess than the high-level pipeline suggests. The validation perturbation curve, synchronized interventions, and full-test permutation result also use different cohorts; their distinctions should be easier to follow.
  3. Dependence on informative templates limits the evidence for broader design utility. High recovery is closely associated with high template sequence identity. On the paired subset with template identity below 20%, STENCIL reaches 42.75%, below the separately trained geometry-only model at 45.88%. Exact-match and near-copy filtering does not remove all near-homologous templates. TS50/TS500 are secondary evidence because of training overlap and unmatched baseline protocols. Although fold-back results provide useful additional evidence, they use one prediction oracle on validation data. The current evidence therefore leaves performance on orphan proteins, template-poor targets, and genuinely novel structures unresolved.
Questions for authors
  1. Could you compare STENCIL with DualMPNN using matched candidate pools, admissibility rules, re-alignment budgets, and evaluation masks? A controlled comparison that changes the mapping, template representation, and decoder separately would clarify which changes explain the improvement. Please report end-to-end cost alongside recovery, since exact re-alignment accounts for 71.6% of warm-cache runtime.
  2. Could you evaluate several orphan proteins or template-poor targets against DualMPNN and a geometry-only model? Please define the selection criteria and report template availability, sequence recovery, and fold-back structural consistency to assess performance without informative templates.
  3. Please provide the equations or pseudocode for the message-passing block and template cross-attention, including Q/K/V organization, masking, and any quality-to-bias mapping. What happens at uncovered positions or when no template is available? Since retraining without quality features changes recovery by only −0.048 points with a confidence interval spanning zero at K = 1, which components are necessary for the reported accuracy?
  4. Please clarify the relationship between the common validity-mask statement in the main text and the method-specific coordinate-validity cohorts mentioned in the appendix. Can the main comparison be reported on a strict intersection cohort? For the 517-query paired analysis, please explain how excluded cases differ from included cases and report template-availability failures separately, so readers can judge the scope of the alignment-versus-selection conclusion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.