acceptodds
Under review as a conference paper at ICLR 2027

Retrieved Templates as Revisable Priors for Protein Inverse Folding

Abstract

Structural retrieval can provide strong target-specific evidence for protein inverse folding, yet retrieved sequences are neither uniformly reliable nor necessarily compatible with the target backbone. We investigate two questions: which retrieved residues should be trusted, and how should this uncertain evidence enter a generative model? We introduce a discrete generative method that treats aligned templates as a revisable sequence prior. The method combines global structural similarity, local geometric consistency, and inter-template agreement to estimate residue-level reliability. These estimates define a partially masked source distribution, which a time conditioned model iteratively refines while remaining conditioned on the target structure and aligned templates. We further train on model-generated intermediate states to match the states encountered during refinement. Our method achieves sequence recovery of , , and on CATH 4.2, TS50, and TS500, respectively. It also improves over direct template transfer by percentage points on CATH 4.2. ESMFold evaluation of the top five among 20 generated candidates yields mean scTM scores of , , and on the three benchmarks. These results suggest that retrieval is most useful when treated as informative but editable evidence.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein inverse folding, retrieval-augmented generation, discrete flow matching

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_D2qcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the inverse folding problem. The proposed method uses retrieved templates as priors and refines the prior sequence gradually. The proposed method achieves state-of-the-art performance in the several benchmarks.

Critical strengths
  • The proposed methods achieve state-of-the-art performance in several important benchmarks.
  • Ablation studies are implemented to prove the effectiveness of the proposed designs.
Critical weaknesses
  • The proposed method uses a paradigm of prior-then-refine. This paradigm has been investigated in Bridge-IF [1]. The difference between the proposed method and Bridge should be discussed.

  • In Section 4.5, the structural consistency should be evaluated against other baselines.

  • In Table 2, the effectiveness of iterative refinement (T=5) is limited. The reason should be discussed.

  • The overall experimental results are metrics. More insightful cases and structural results should be discussed.

[1] Bridge-if: Learning inverse protein folding with markov bridges.

Questions for authors

See Weaknesses

Reviewer_DAzGOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

Summary:

The paper works on protein inverse folding: design a sequence that fits a given backbone. It retrieves structurally similar proteins and aligns their residues to the target. The core argument is that retrieved sequences are useful but unevenly reliable, so they should be treated as editable evidence rather than fixed answers.

Contributions:

Two claims. Retrieval should set the starting state of generation, not just condition the predictor, so the initial sequence is sampled from the template prior mixed with mask and every residue stays revisable. And trust should be decided per residue using local geometric agreement, since a single global similarity score is constant across an alignment and cannot tell a conserved region from a mismatched one.

Validation:

Three standard benchmarks measured by recovery and perplexity, plus ESMFold refolding. The main experiment is an ablation separating retrieval-as-initialization from retrieval-as-conditioning. The model is also compared against direct template copying and against an oracle that uses native labels to pick the better template. A separate analysis shows the local reliability score ranks correct template residues better than the global score does.

Critical strengths

1. The ablation separating initialization from conditioning is clean and informative. Table 2 reports 37.80, 51.34, 59.61, then 66.26, using the same backbone and the same template pairs throughout. It shows that retrieval-as-source and retrieval-as-conditioning both contribute and are complementary.

2. The label-assisted oracle addresses the copying concern directly. The authors build a non-deployable upper bound that uses native labels to pick the better of two templates. It reaches 57.50%, and the model reaches 66.65%. Copying is the natural objection to any retrieval method, and this answers it with evidence rather than argument.

Critical weaknesses

1. Leakage controls may be insufficient to support the reported results.

The filtering removes same-PDB entries and structures with TM-score > 0.99, but does not include sequence identity filtering or specify the retrieval database. The top-1 template alone achieves 53.54% recovery, raising concerns that the reported 66.65% may partly reflect database coverage rather than generalization. This is important because it potentially affects all reported results.

The authors may also consider RadDiff (arXiv:2512.00126), which uses identity and substring filtering. RadDiff follows a closely related approach, using residue-level structural alignment and position-specific amino acid profiles as priors for discrete denoising, and reports 67.14% recovery and 2.46 perplexity on CATH 4.2. Relative to this work, the main distinction appears to be reliability-weighted mixing with MASK and a local-geometry reliability score. A direct comparison would help clarify the novelty and contribution.

2. The reported improvement is difficult to attribute to the proposed method.

The improvement over DualMPNN is only 1.14 percentage points. However, the additional surface module is not ablated, results lack multiple seeds or significance tests, and the iterative refinement highlighted in the title contributes only 0.39 percentage points. These limitations make it difficult to determine whether the gains arise from the proposed method, additional components, or run-to-run variation. Stronger ablations and statistical evidence would help substantiate the central claim.

Questions for authors
  1. Which structure database do you retrieve from, and at what date?
  2. What is the template-target sequence identity distribution, and does your margin over DualMPNN hold at 30% identity or below?
  3. What is recovery on CATH 4.2 with the MSMS surface module removed, everything else fixed?
  4. Over three seeds, what are the mean and standard deviation on CATH 4.2, and what does a paired bootstrap over the 1,120 test proteins give for your method versus DualMPNN?
  5. Table 3 reports 60.23 / 64.98 / 66.31 / 66.65, but the text at p. 8 reports 60.23 / 60.51 / 61.07. Which is correct?
  6. How does your source-state formulation differ from RadDiff (arXiv:2512.00126), which builds a residue-aligned amino acid profile as a prior for discrete denoising?
Reviewer_SxLeOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a retrieval-guided protein inverse-folding model that combines partially masked template-based sequence initialization with structure- and template-conditioned prediction. Retrieved information is weighted using global structural similarity, local geometry, and inter-template agreement. The main methodological contributions appear to be explicit local reliability estimates, a revisable discrete sequence prior, and training on model-generated trajectories. The method is evaluated through baseline comparisons and component ablations, using perplexity and sequence recovery.

Critical strengths
  • Combining global structural similarity with local geometry and inter-template agreement addresses the possibility that a globally similar template may provide unreliable guidance at individual positions. This gives the method a clear rationale for selectively incorporating retrieved information.

  • A revisable discrete sequence prior provides a concrete mechanism for incorporating template sequences while allowing subsequent predictions to depart from them. Training on model-generated trajectories is a complementary design choice intended to align training with the refinement process used at inference.

Critical weaknesses
  • The proposed method differs from DualMPNN in several potentially consequential ways, including local reliability weighting and the use of two retrieved templates rather than the closest template alone. Consequently, the reported gains cannot be attributed specifically to the revisable discrete prior. A matched comparison controlling template inputs and reliability information is needed to establish this contribution. The novelty discussion should also give a more precise account of DualMPNN’s existing mechanisms and distinguish them from the proposed advances.

  • Section 4.4 reports sequence recovery of 60.51% and 61.07% for global TM-score weighting and global plus local geometry, respectively, whereas Table 3 reports 64.98% and 66.31%. These differences alter both absolute performance and the apparent contribution of local geometry, making the component-level conclusions difficult to assess. I also could not find a code repository link in the submission, which limits independent reproduction and verification of the results.

Questions for authors
  • Can you provide a controlled comparison of DualMPNN, DualMPNN augmented with the same template inputs and local reliability weighting, and the proposed method, with retrieval resources and evaluation settings matched? Please also quantify the effect of using one versus two templates. To isolate the revisable discrete prior specifically, can it be ablated while keeping template conditioning and reliability weighting fixed?

  • Do the conflicting recovery values in Section 4.4 and Table 3 correspond to different experimental settings or runs? Please verify them against the original records, identify the correct values or label the differing settings, and revise the component-level interpretation accordingly.

  • What explains the differing trends in perplexity and sequence recovery in Table 1? These metrics need not rank methods identically, but please clarify whether they use the same evaluation examples and conditioning information, and specify the decoding procedure used to compute recovery.

  • Is an anonymized implementation available for review? If so, please provide the link, including the retrieval pipeline and training/evaluation configurations needed to reproduce the main comparisons and ablations.

1 comment

Thank you for the detailed response and additional experiments, they have addressed my concerns. I will keep my positive evaluation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.