acceptodds
Withdrawn from ICLR 2027

Structure-aware language modeling enhances peptide conditional generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent studies have demonstrated that protein language models can effectively drive receptor-conditioned peptide generation. However, existing language-model-based approaches rely almost exclusively on sequence context to learn compatibility between peptide binders and their target proteins. It remains unclear whether incorporating target protein structural information can yield additional benefits, particularly with respect to the rationality of spatial conformations. Furthermore, most language-model-based methods do not explicitly account for the receptor site where peptides are expected to bind, yet such binding sites are critical for functional peptides. To bridge this gap, we investigate how structure-aware protein language models contribute to receptor-conditioned peptide design and introduce binding-site constraints within the language-model paradigm to enable site-specific generation. Experimental results show that combining receptor sequences with structural conditions effectively improves the spatial conformational rationality of peptide–protein complexes and the complementarity of binding interfaces. At the local scale, introducing binding-site constraints shifts generated candidate peptides toward the target region, indicating that binding-site prompts exert a soft directional bias. Code and data will be released at https://github.com/anonymous/VersaPep.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein language models, conditional generation, structure-aware modeling, binding-site-guided generation, peptide design, biomolecular generation

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

This paper proposes VersaPep, a structure-aware protein language model for receptor-conditioned peptide generation. Built upon SaProt, the method introduces Gated Cross-Chain Attention (CCA) to integrate receptor information and Binding-Site FiLM (BS-FiLM) to guide peptide generation toward specific binding regions. Experiment on the test set shows improved predicted complex quality compared to the baselines. Ablation studies demonstrate the complementary benefits of receptor sequence and structural information, while binding-site conditioning provides a directional bias toward target regions without substantially affecting overall complex confidence.

Critical strengths
  1. Well-motivated problem. The paper addresses an important limitation of existing protein language model-based peptide generation methods by incorporating structural information and explicit binding-site guidance. The distinction between global structural conditioning and local binding-site guidance is meaningful and well motivated.

  2. Reasonable and flexible framework design. The proposed framework integrates receptor sequence, structural tokens, and binding-site information into a unified language modeling framework. The use of gated cross-chain attention and BS-FiLM provides a flexible way to incorporate both global and local conditioning signals.

  3. Comprehensive ablation studies. The paper includes extensive ablations to investigate the contributions of receptor sequence, receptor structure, peptide-side structural information, and binding-site guidance. In particular, the experiments with true, predicted, and incorrect binding sites provide useful insights into how local conditioning affects peptide localization.

Critical weaknesses
  1. Limited improvements without peptide structural priors.

The performance gains appear to depend on the availability of peptide-side 3Di tokens. When these tokens are not provided, the improvement in mean ipTM over PepMLM decreases from 0.023 to only 0.003. Moreover, the variants without peptide 3Di do not outperform PepMLM in the independent DockQ evaluation. Since peptide structural information is generally unavailable in practical de novo design, these results raise concerns about the effectiveness of the proposed method in realistic generation settings.

  1. Uncontrolled training-data differences and potential benchmark bias.

Although all methods are evaluated on the same test set, the paper does not clarify whether the baselines were retrained on the proposed training split or evaluated using their original checkpoints. The original PepMLM and DiffPepBuilder models were trained on datasets constructed using different sources and filtering criteria. Since the test set is derived from the same data construction pipeline as VersaPep's training set, potential differences in training-data distributions may confound the reported performance comparisons.

  1. Limited effectiveness of binding-site guidance.

The effectiveness of binding-site guidance remains questionable. Table 4 shows only a marginal reduction (0.20 Å) in peptide-to-pocket distance with true binding-site conditioning, while peptides guided by incorrect sites still remain closer to the true pocket. Furthermore, Fig. 4c shows that predicted binding-site guidance underperforms no guidance at several thresholds. The unusually short inter-chain heavy-atom distances (1.30 Å and 1.97 Å) in Fig. 5 also raise concerns about potential steric clashes. More rigorous localization and structural validity evaluations are needed.

Questions for authors
  1. Are the peptide 3Di tokens used during inference derived from native bound structures? How can the performance gains be interpreted when such structural information is unavailable in practical de novo peptide design?

  2. Are PepMLM and DiffPepBuilder retrained on the same training set as VersaPep? If not, are potential distribution differences examined?

  3. How do the authors explain the marginal improvement in binding-site localization (0.20 Å in Table 4) ? Do the unusually short inter-chain distances in Fig. 5 indicate steric clashes?

I am open to reconsidering my rating if the authors can adequately clarify or address the concerns raised above during the rebuttal.

Reviewer_28JYOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents VersaPep, a SaProt-based model for generating peptide sequences conditioned on a target protein's sequence and structural tokens, with optional peptide 3Di tokens and binding-site information. The model incorporates gated cross-chain attention (CCA) and feature-wise linear modulation (FiLM) to enhance peptide representations. VersaPep achieves higher scores than PepMLM and DiffPepBuilder on the reported AlphaFold-Multimer confidence metrics.

Critical strengths
  • The paper takes a systematic approach to examining the contributions of different components to peptide design.
Critical weaknesses
  • The problem formulation needs clarification, as it affects both the choice of baselines and the interpretation of the results. If the goal is peptide sequence design conditioned on both target and peptide structural tokens, comparisons with inverse-folding methods would be appropriate, while accounting for the difference between 3Di tokens and full backbone inputs. If the goal is de novo peptide design, additional task-matched baselines are needed to establish the method's effectiveness.

  • The architectural contribution is not sufficiently established. The gains in Table 2 appear marginal: adding CCA and BS-FiLM increases the mean maximum ipTM from 0.563 to 0.570, while CCA alone yields 0.560. Although some other metrics improve, the practical significance and robustness of these gains remain unclear. Given that cross-chain attention and FiLM are established techniques, the architectural contribution should be supported by either a well-substantiated design rationale or convincing empirical benefits; the current evidence does not sufficiently establish either.

Questions for authors

Q1: Could you clarify the intended problem setting? In what practical scenarios would peptide structural tokens be available when designing new peptide sequences? What additional value does the proposed architecture offer over simply fine-tuning SaProt for conditional generation?

Q2: Could you broaden the comparison to include additional task-matched peptide design methods, or explicitly explain why relevant methods are not comparable?

Q3: Could you clarify the experimental settings to establish whether the comparisons are fair? In particular, does the full VersaPep model use the same five-round iterative procedure described for the variant without supplied peptide 3Di tokens? Are the numbers of generated candidates matched across methods?

Q4: Are 3Di tokens computed for the receptor and peptide separately or jointly?

Q5: Is the method limited to short peptide binders, or can it also be applied to designing longer protein binders?

Q6: (Optional) Do the learned representations remain useful for downstream tasks? Could you provide an analysis of their utility?

Q7: (Optional) The figures could be presented more clearly.

Reviewer_Ccv1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PAPER SUMMARY

The paper introduces VersaPep, a SaProt-based model for receptor-conditioned peptide generation. Gated cross-chain attention transfers receptor information to peptide representations, while binding-site FiLM introduces optional local guidance. The experiments separate receptor sequence, receptor structural tokens, peptide structural tokens, and binding-site information. Evaluation on 300 targets uses AlphaFold-Multimer confidence, with additional docking and interface analyses in the appendix.

Critical strengths
  1. The conditioning analysis addresses a useful scientific question. Distinguishing receptor sequence and structure from peptide-side structural information is more informative than reporting only an aggregate improvement. The true-site, predicted-site, no-site, and wrong-site comparisons also provide a meaningful test of whether local guidance changes predicted placement.

  2. The paper includes checks beyond AlphaFold confidence and explicitly acknowledges that predicted structures do not establish experimental binding. The receptor-only setting is particularly relevant to prospective generation, and its inclusion is a useful starting point for evaluating the practical contribution.

Critical weaknesses
  1. The strongest results do not yet establish a convincing advantage in the practically important setting without supplied peptide structure. Table 1 reports mean ipTM of 0.469 for VersaPep, 0.449 without peptide 3Di, and 0.446 for PepMLM. The independent structural results are mixed: Table 7 reports mean DockQ of 0.254 for full VersaPep versus 0.246 for PepMLM, but 0.243 and 0.239 for the two no-peptide-3Di variants. Best acceptable rates are 0.593 for full VersaPep and 0.611 for PepMLM. Appendix C further reports that none of the DiffPepDock pairwise differences survives BH-FDR correction. These results do not show that the method is ineffective; they limit the strength of the claim that its structural conditioning improves generated complex quality beyond the confidence predictor. The conditioning inputs and candidate budgets need to be matched before the headline comparison can support a broader generation claim.

  2. The site-guidance result is small in the aggregate and needs a clearer estimate of uncertainty and practical magnitude. Table 4 gives mean minimum distance to the true pocket of 3.920 Å with true-site guidance and 4.120 Å without guidance, a 0.200 Å difference. The two visual examples are much more dramatic than this average. Confidence-thresholded comparisons may also contain different candidate subsets across settings. A target-paired analysis and an interface/contact recovery measure would better establish whether the effect reflects useful localization rather than a small shift in predicted distance. This concern does not require wet-lab validation; it concerns the existing computational claim.

Questions for authors
  1. What is the exact source of peptide 3Di at test time in the full model, and which corresponding information is provided to each baseline? Please distinguish the direct no-peptide-3Di and iterative variants consistently across Tables 1, 3, and 7, including the number of generated and evaluated candidates.

  2. Under matched receptor-only inputs and candidate budgets, what are the paired target-level effects and confidence intervals against PepMLM for DockQ and pocket/contact recovery? Which corrected comparisons in Appendix C specifically support the claimed receptor-structure benefit?

  3. Can the binding-site effect be reported on a common target/candidate protocol, with paired uncertainty and sample counts at each confidence threshold? Also clarify the clustering unit in the 80%-identity split: receptor, peptide, concatenated complex, or a grouping that constrains both chains.

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

Related papers

Discussion (0)

Sign in to comment.