acceptodds
Under review as a conference paper at ICLR 2027

Structure-Contrastive Reveal: Guiding Generation Order in Protein Inverse Folding

Abstract

Protein inverse folding aims to design amino-acid sequences compatible with a given protein backbone structure. Many recent models generate sequences iteratively, predicting amino acids at unresolved positions and filling only a subset of these positions at each decoding step. Decoded residues become context for later predictions. Thus, the order in which positions are selected can affect the final generation trajectory. Existing position-selection policies usually prioritize positions where the model is most confident. However, high confidence does not necessarily mean that a prediction is supported by the target structure. Since inverse folding is conditioned on a target structure, identifying predictions that actually rely on structural information is pivotal. Intuitively, if the predicted amino acid at a position is supported by the structure, weakening the structural information should reduce the probability assigned to that prediction. Motivated by this intuition, we explore comparing the probability of the same predicted amino acid under the original and weakened structural conditions, using the probability decrease as a proxy for structural dependence. In this work, we introduce Structure-Contrastive Reveal (SCR), which combines this probability decrease with predictive confidence to decide which residue positions to decode first, while using the predictions under the original structural condition for the selected positions. Extensive experiments show that SCR consistently improves structural quality across different inverse-folding models and further enhances the state-of-the-art baseline under matched search configurations. An extension to trajectory-conditioned motion generation further suggests that condition-dependent reveal selection can generalize beyond protein inverse folding.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein inverse folding, structure-conditioned generation, protein design, iterative decoding

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

The paper proposes Structure-Contrastive Reveal (SCR), an inference-time position-selection policy for iterative protein inverse folding. At each step, SCR samples amino-acid proposals under intact structural conditioning and measures their log-probability drops under weakened conditioning within a high-confidence candidate pool. Confidence and contrast rankings are combined with equal weight to determine which positions to reveal first. Revealed amino acids always come from the intact-condition predictions. Experiments cover NanoIF, frozen ESM3, integration with ProtInvTree, intervention timing, and trajectory-conditioned motion generation.

Critical strengths
  1. The method is simple and requires no additional training. SCR preserves intact-condition amino-acid proposals and changes only reveal-position selection. Global SCR shares one weakened-condition evaluation per decoding step, giving it practical potential in compatible decoders.
  2. Both instantiations show reported improvements, with attention to computational cost. In Table 1, Mean TM improves from 0.7570 to 0.8104 on NanoIF and from 0.8253 to 0.8539 on frozen ESM3; gains also occur with ProtInvTree. Figure 2 compares quality against measured generation time. Several experiments include paired confidence intervals and acknowledge nonsignificant differences. These results support empirical value under the evaluated settings.
Critical weaknesses
  1. The central motivation is insufficiently established: perturbation sensitivity does not imply reveal priority. The probability contrast directly measures sensitivity to a specified input modification; it does not establish proposal correctness or the benefit of committing that position earlier. Sensitive positions might instead benefit from additional sequence context before commitment. Figure 3 shows perturbation responses, and Figure 4 shows persistent effects of early intervention, but neither directly tests whether high-sensitivity positions should precede low-sensitivity positions at comparable confidence. NanoIF’s contraction of neighboring N/C/O coordinates to 25% of their original displacement from Cα creates nonphysical geometry, further mixing structural dependence with sensitivity to unusual inputs. The successful edge-attenuation variant provides partial support, but does not establish the direction of prioritization. Since SCR’s methodological increment is to use established condition-contrastive reasoning for position selection, demonstrating the necessity of this rule is central. The Random/Entropy baselines in Table 4 do not replace matched high-grounding-first, low-grounding-first, and shuffled-grounding controls within the same candidate pool.

  2. Limited model coverage leaves practical competitiveness and applicability insufficiently established. Protein experiments use a self-built NanoIF and one public ESM3 checkpoint. Table 1 reports AAR gains of 34.50% to 38.86% and 39.16% to 40.92%, but lacks broader established-system comparisons under a common protocol to contextualize the final performance. ProtInvTree tests complementarity with search but reuses the same two backbones, and its ESMFold reward shares the folding model used for structural evaluation. Motion offers preliminary transfer evidence rather than replacing core inverse-folding coverage: Targeted Grounding has the best trajectory error, Shuffled Grounding has the best FID, and the reported trajectory-error confidence interval is against Confidence rather than the grounding controls. These extensions broaden the settings without resolving the central ordering assumption or reliability across backbones.

Questions for authors

Why high-grounding positions should be revealed first: Could you further explain why positions that are more sensitive to structural conditioning should be revealed earlier rather than deferred until more sequence context is available? Comparing high-grounding-first, low-grounding-first, and shuffled-grounding rankings while retaining the confidence filter and decoding settings would directly test this rationale.

Transferability of the grounding insight: Could you examine this phenomenon in two additional compatible inverse-folding models? Beyond final performance, it would be informative to test whether grounding shows similar relationships with prediction correctness and the benefit of earlier revelation. This would help determine whether it reflects a transferable design insight or behavior specific to particular models and perturbations.

Reliability of NanoIF for mechanism analysis: NanoIF’s sequence recovery is relatively limited. Could you clarify whether it is adequately trained, contextualize its performance, and assess whether the grounding analysis depends on backbone quality? Repeating key analyses across training checkpoints or with a stronger compatible backbone would help determine whether the conclusions reflect useful structural information rather than unstable perturbation responses in a weaker model.

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

This paper proposes Structure-Contrastive Reveal (SCR), a training-free strategy for improving iterative protein inverse folding through structure-aware residue selection. The method estimates the dependence of amino acid predictions on structural conditioning by measuring changes in predictive log-probabilities after weakening the structural input. This contrastive signal is combined with prediction confidence to determine which residue positions should be revealed first. The approach is evaluated on a geometric inverse-folding model, NanoIF, and the pretrained ESM3 model. Experiments demonstrate improvements over conventional confidence-based decoding and additional gains when integrated with ProtInvTree. The authors further examine computational trade-offs, generation-order effects, and transfer to conditional human motion generation.

Critical strengths
  1. A simple and practical training-free decoding strategy.

SCR introduces a straightforward mechanism that improves generation without modifying the parameters of the underlying models. The separation between residue-position selection and amino acid prediction is particularly appealing, as it allows the method to be applied to different pretrained generative architectures.

  1. Consistent improvements supported by diverse experiments.

The authors evaluate SCR across two substantially different inverse-folding models and demonstrate further benefits when combined with reward-guided tree search. The experiments examining computational efficiency and early versus late application provide useful evidence that reveal-order decisions can have persistent effects on generation quality.

Critical weaknesses
  1. The structural grounding interpretation requires stronger validation.

A decrease in amino acid probability following structural perturbation does not necessarily indicate biologically meaningful structural dependence. Such changes may also reflect sensitivity to perturbed or out-of-distribution structural inputs. Although the paper includes perturbation controls, the extent to which the signal measures genuine structural support remains unclear.

  1. The generality of the approach is not fully established.

The proposed framework requires different structural weakening procedures for different model architectures, and negative contrastive scores are handled differently for NanoIF and ESM3. These design choices are not fully isolated. Furthermore, while the motion-generation extension demonstrates some improvements, targeted structural grounding does not consistently outperform all alternative strategies across every reported metric.

Questions for authors

How sensitive is SCR to the magnitude and type of structural perturbation? Can the authors provide a more direct analysis separating meaningful structural dependence from generic perturbation sensitivity?

How much of the improvement remains when targeted grounding is compared against randomized or shuffled grounding under matched computational budgets? Can paired uncertainty estimates be provided for these comparisons?

How sensitive are the conclusions to the use of ESMFold as the structural evaluator? Would the observed improvements remain consistent under an independent structure prediction model?

Reviewer_sfE9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Structure-Contrastive Reveal (SCR), a training-free decoding strategy for protein inverse folding. Instead of revealing residues solely based on prediction confidence, SCR measures how the probability of a sampled amino acid changes when the structural conditioning is weakened. This contrastive signal is combined with confidence to determine which positions should be revealed first, while the actual amino acids are still sampled under the original structural condition. The method is evaluated on CATH 4.2 using a custom-trained NanoIF model and a frozen ESM3 model, with additional experiments incorporating SCR into ProtInvTree. The authors also investigate how early reveal decisions affect subsequent generation and extend the approach to trajectory-conditioned human motion generation.

Critical strengths
  1. The distinction between prediction confidence and dependence on the target structure provides an interesting perspective on iterative protein inverse folding. Using structural perturbations to guide reveal order is a simple idea that does not require retraining the underlying model. The improvements are also meaningful. Local SCR increases NanoIF's Mean TM-score from 0.7570 to 0.8104, while Global SCR improves ESM3 from 0.8253 to 0.8539. Its compatibility with ProtInvTree further shows that the method can be incorporated into an existing test-time search framework.

  2. The paper includes useful analyses of how reveal-order decisions influence generation. In particular, the experiments in Figure 4 show that changing the reveal strategy early in decoding can affect the final sequence even after the model returns to confidence-based selection. The evaluation also covers different structural representations, alternative perturbation strategies, and compute–quality trade-offs. These experiments provide a useful starting point for understanding the role of generation order in iterative conditional models.

Critical weaknesses
  1. I am not convinced that the main improvements can be attributed to the proposed structural grounding signal rather than simply perturbing the confidence-based reveal order. In Appendix Table 3, Random Structure Probe increases NanoIF's Mean TM-score from 0.7570 to 0.7868, accounting for roughly 90% of the improvement achieved by Global SCR at 0.7900. A similar issue appears in the HumanML3D experiments, where Shuffled Grounding and Random-Temporal Grounding also improve over Confidence, and Shuffled Grounding even achieves a better FID than Targeted Grounding. Although Local SCR performs better than the random probe and targeted grounding achieves the lowest motion trajectory error, the current experiments do not adequately separate the contribution of structure-specific information from the effect of rearranging the decoding order. Most importantly, the protein experiments lack a direct shuffled-grounding control within the same high-confidence candidate pool. Without this comparison, the central claim that structural dependence provides a better reveal criterion remains insufficiently established.

  2. The definition and implementation of structural grounding also leave several important design choices unexplained. Equation (2) evaluates the probability change of a single stochastically sampled amino acid, whereas confidence is computed from the maximum probability over the entire distribution. This makes the grounding score sensitive to sampling noise, particularly when a low-probability amino acid is selected. NanoIF additionally clips negative contrasts to zero, while ESM3 retains them, without an ablation of this difference. The structural perturbations are also substantially different across the two models. NanoIF preserves all Cα coordinates even under global weakening, whereas ESM3 removes the entire residue-level structure-token track. The strong results of Edge-Attenuated SCR, which reaches 0.8058 Mean TM and 0.4010 recovery on NanoIF, further suggest that the choice of perturbation is important. Yet the paper does not systematically examine these alternatives or establish why the proposed implementations should be preferred. Given the additional inference cost, stronger comparisons with simpler perturbation strategies and compute-matched decoding baselines are needed to justify the proposed approach.

Questions for authors
  1. Could the authors evaluate Shuffled Grounding on the main CATH 4.2 experiments for both NanoIF and ESM3? In particular, I would like to see the grounding ranks randomly permuted within the same top-4q confidence candidate pool, while keeping the reveal schedule and sampling settings unchanged. Appendix Table 3 already shows that Random Structure Probe approaches Global SCR on NanoIF. How much improvement remains when the relationship between structural perturbation and the selected residue is removed entirely?

  2. Why is the grounding score evaluated using one stochastically sampled amino acid rather than the full predicted distribution? How would the results change if the contrast were computed using KL divergence between the intact and weakened distributions, or the probability change of the highest-confidence amino acid? Could the authors also report the effect of retaining versus clipping negative contrast values for both NanoIF and ESM3?

  3. NanoIF's global structural weakening preserves all Cα coordinates, whereas ESM3's global weakening masks all structural tokens. How do the authors justify these different perturbation strengths? Appendix Table 3 also shows that Edge-Attenuated SCR substantially outperforms Global SCR on NanoIF. Could the authors compare these perturbation strategies under matched forward-pass budgets and clarify why coordinate contraction was selected as the main implementation?

  4. How does SCR compare with confidence-based decoding under the same total inference budget? Figure 2 examines different decoding step counts, but it does not establish whether SCR is preferable to generating several independent candidates with the cheaper baseline and selecting among them using a model-based score. A comparison under matched wall-clock time or forward-pass budgets would help establish the practical value of the additional structural evaluations.

  5. The protein evaluation is limited to CATH 4.2 and two model implementations, one of which is custom-trained. Does the method generalize to an independent backbone benchmark or another established pretrained inverse-folding model? It would also be useful to evaluate structural quality with a second folding predictor, since the main results rely on ESMFold-based assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.