acceptodds
Under review as a conference paper at ICLR 2027

Distogram-Based Structural Loss: A Plug-In Module for Protein Inverse Folding

Abstract

Protein inverse folding aims to design amino acid sequences that fold into given backbone structures. However, most inverse folding models are trained primarily to recover native sequences, an objective that does not directly assess structural consistency between generated sequences and the target backbone. Here, we introduce DISCO loss, a model-agnostic auxiliary objective that incorporates pairwise distance information from the target backbone into inverse folding training. A jointly trained auxiliary module predicts residue-pair distance distributions using representations derived from inverse folding outputs, supervised by discretized distances computed from the target backbone. Combined with the sequence recovery loss, DISCO loss provides structural supervision without full 3D reconstruction or separate surrogate pretraining. On de novo protein backbones, DISCO loss improves structural self-consistency across inverse folding models with distinct generation mechanisms, with improvements also observed in long-range contact recovery. Additional evaluations show that sequence recovery performance and diversity are largely maintained. These results demonstrate the effectiveness of auxiliary supervision based on target-backbone distances in improving the structural consistency of generated sequences.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Inverse folding, Self-consistency, Distogram

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_4PFvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes DISCO loss, a jointly trained auxiliary distogram-prediction module for protein inverse folding. Logit-derived single/pair representations and frozen ESM-2 features predict the target-backbone distogram, providing structural supervision without folding-model feedback during training. The auxiliary module is removed at inference. Experiments on ProteinMPNN, PiFold, and Grade-IF report improved structural consistency on RFdiffusion-generated backbones under AF2 and ESMFold, together with contact, secondary-structure, recovery, diversity, amino-acid-composition, and Rosetta-energy analyses.

Critical strengths

The auxiliary target fits the task and adds no inference-stage module. The distogram is available from the conditioning backbone, and joint training avoids a separately pretrained structural surrogate or folding-model feedback. Several complementary evaluations support empirical benefits. The paper uses AF2 and ESMFold, paired confidence intervals, and comparisons with AFDistill and a FAPE-based variant. Appendices further analyze long-range contacts, secondary-structure recovery, composition, diversity, and Rosetta energies, and report model-dependent recovery losses rather than presenting uniformly positive outcomes.

Critical weaknesses

The methodological contribution is incremental, and the motivation does not directly establish the failure mode being addressed. The auxiliary module adapts established distogram-prediction components to inverse-folding supervision. Its main contribution is the jointly trained, removable structural head, so evidence connecting it to actual inverse-folding errors is important. Figure 1 instead uses native sequences with 10% random substitutions. Such substitutions need not reflect correlated, context-dependent model errors, and the description does not adequately specify backbone selection, sample size, or substitution details. A matched-recovery analysis of actual model outputs would more directly establish the motivating problem. The recovery–structure trade-off and broader applicability remain insufficiently characterized. Table 3 reports a 3.88-point CATH recovery decrease for ProteinMPNN and recovery decreases for Grade-IF on the three full benchmarks, while PiFold improves. Appendix C.1 already sweeps temperature and loss weight, but evaluates structural gains on de novo backbones, while recovery is measured separately on natural-protein benchmarks. Appendices D.1–D.2 and G identify improvements in contacts, secondary structure, composition, and Rosetta energies. The remaining gap is a joint recovery–structure analysis across objective weights on held-out native backbones, where both metrics are defined. Lower native recovery is not itself evidence of worse design; a joint analysis would clarify the available trade-off. Evaluation is restricted to single-chain designs, and applicability to additional strong inverse-folding backbones is not established. These limitations constrain the significance of an otherwise useful incremental training objective.

Questions for authors
  1. Figure 1 protocol and applicability of the motivation: Could you detail the procedure behind Figure 1, including the sources of native sequences and backbones, sample size, position selection, substitution rules, and structural evaluation settings? At matched recovery, are the patterns of structural quality consistent between controlled mutations of native sequences and actual model-generated sequences? This comparison would help establish whether the motivation extends to practical inverse-folding outputs.

  2. Causes and implications of reduced recovery: Could you analyze why adding DISCO reduces recovery for some models? Are these changes accompanied by improved structural consistency on the same targets, or are there cases where both recovery and structural quality decline? Paired results and representative examples, building on the existing composition and structural analyses, would help distinguish a beneficial design trade-off from a loss of sequence-prediction capability.

1 comment

Dear Reviewer,

Thank you for highlighting both the strengths of our work and the areas where it could be improved. We appreciated hearing that you found the paper clear, easy to follow, and informative. Your feedback also helped us see where our reasoning needed stronger support. In our rebuttal, we focused on the concerns about soundness that you identified as minor or fixable and provided additional evidence and clarification in response. We were encouraged that another reviewer raised their score after considering our responses.

When you have a chance, we would be grateful to hear whether our additional evidence and clarifications address the concerns you raised. We would be happy to discuss any points that you feel would benefit from further clarification.

Best regards,
The Authors

Reviewer_x67AOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the problem of inverse protein folding. The authors propose a plug-in module (i.e., a structural alignment loss )for any existing inverse folding model. The experiments demonstrate improvements in structural self-consistency and long-range contact recovery across three representative models, including ProteinMPNN, PiFold, and Grade-IF.

Critical strengths
  • This paper is easy to follow.
  • This paper studies an important problem in protein design.
  • The proposed DISCO loss is easy to implement.
Critical weaknesses
  • The novelty is limited. The only modification is the auxiliary loss. However, the motivation to improve structural alignment has been analyzed in different methods. For example, [1] used DPO to fine-tune an inverse folding model using feedback from a protein folding model.

  • Several more recent baselines are ignored. To name a few [2,3,4]

  • The performance improvement is not significant. In Table 1, the main improvement is on PiFold and Grade-IF. However, the previous metrics of these two models are weak.

[1] Protein Inverse Folding From Structure Feedback

[2] Uniif: Unified molecule inverse folding

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

[4] Structure-informed language models are protein designers

Questions for authors

See Weakness

Reviewer_pdbhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

DISCO adds distance-based supervision to protein inverse folding. An auxiliary network combines logit-derived features with frozen ESM-2 representations to predict target-backbone distance bins. It is trained jointly with the inverse folding model and removed at inference. Experiments on ProteinMPNN, PiFold, and Grade-IF show improved self-consistency on RFdiffusion backbones, especially for PiFold. The paper also examines contact recovery, sequence recovery, cost, diversity, and Rosetta energy.

Critical strengths
  1. DISCO uses existing backbone labels without separate surrogate pretraining or full structure prediction during training, and adds no inference cost.

  2. Both AlphaFold2 and ESMFold support improvements in the final designed sequences. For PiFold at length 200, AF2-based scTM rises from 0.564 to 0.795, alongside better long-range contact recovery.

Critical weaknesses
  1. Structural self-consistency is evaluated only on RFdiffusion backbones. Natural-data sequence recovery and frozen-logit probes do not establish whether the final designs recover natural structures. Evaluation on held-out natural backbones, with training overlap excluded, is needed to assess generalization.

  2. The benefit of distogram classification over simpler geometric objectives remains unclear. A distance-regression baseline using the same features, comparable network capacity, and a tuned MSE weight would help isolate this contribution, particularly given PiFold's reliance on ESM features.

Questions for authors
  1. How does DISCO perform on held-out natural backbones in terms of scTM, scRMSD, and contact recovery?

  2. What motivates distance binning and cross-entropy rather than direct distance regression?

  3. Why was Grade-IF excluded from the comparison with other structural supervision methods?

1 comment

Dear Reviewer,

Thank you for highlighting both the strengths of our work and the areas where it could be improved. We appreciated hearing that you found the paper clear, easy to follow, and informative. Your feedback also helped us see where our reasoning needed stronger support. In our rebuttal, we focused on the concerns about soundness that you identified as minor or fixable and provided additional evidence and clarification in response. We were encouraged that another reviewer raised their score after considering our responses.

When you have a chance, we would be grateful to hear whether our additional evidence and clarifications address the concerns you raised. We would be happy to discuss any points that you feel would benefit from further clarification.

Best regards,
The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Kindly Requesting Your Feedback on Our Rebuttal

Dear Reviewers,

Thank you so much for the time and care you’ve put into reviewing our paper. Your comments have been really helpful, and we’ve done our best to address each of them in our responses.

When you have a moment, we’d love to hear whether our responses resolve your concerns, or whether anything still seems unclear. We’re happy to provide more clarification or additional results if that would help, and we’d welcome any further discussion.

Thank you again for your thoughtful feedback and support.

Warm regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.