acceptodds
Under review as a conference paper at ICLR 2027

Site-Aware Receptor-Sequence-Conditioned Peptide Binder Generation

Abstract

Sequence-based peptide binder generation enables efficient design from readily  available receptor sequences. However, binding alone is insufficient for functional  peptides, whose biological activity often depends on engaging specific receptor  regions. Existing sequence-based methods either offer no direct control over which  receptor site is engaged or require the target region to be specified in advance,  limiting their use when the functionally relevant region is unknown. To address  this gap, we introduce **SitePepGen**, a **Site**-aware, receptor-sequence-conditioned  masked language model for **Pep**tide **Gen**eration that requires neither receptor  structures nor predefined functional binding sites. SitePepGen combines two  components: (1) a residue-level receptor binding-site prediction head, jointly  trained with peptide generation using shared ESM-2 representations to identify  putative peptide-binding regions; and (2) a site-aware candidate selection module  that ranks generated peptides by their physicochemical compatibility score with the  predicted receptor sites and the model’s generation confidence. In the evaluation  of receptor binding-site recovery using predicted receptor–peptide complexes,  SitePepGen improves macro-averaged Jaccard and F1 over the strongest baseline  by 3.62 and 3.56 percentage points, respectively, while maintaining comparable  sequence-plausibility diagnostics. For receptor binding-site prediction, SitePepGen  further improves AUROC and AUPRC over the strongest baseline by 0.95 and  1.69 percentage points, respectively. The code will be released at https\://  anonymous.4open.science/r/SitePepGen/.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide binder generation, Site-aware design, Receptor sequence conditioning

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

SitePepGen jointly learns peptide sequence generation and receptor binding-site prediction using a shared ESM-2 backbone. Candidates are ranked by sequence confidence and compatibility with predicted sites, without requiring receptor structures or predefined binding regions. For each generated peptide, LocalColabFold predicts a corresponding receptor-peptide complex. Results show improved receptor binding-site recovery over PepMLM.

Critical strengths
  1. Incorporating receptor binding-site information addresses a relevant limitation of sequence-conditioned peptide design. Generation and candidate selection require neither receptor structures nor predefined binding regions, making the approach applicable when such information is unavailable.

  2. Joint training connects residue-level site prediction with peptide generation through a shared ESM-2 backbone, with little additional parameter overhead. Site-aware candidate ranking also avoids complex-structure prediction within the selection loop.

  3. Experimental results show consistent improvements in receptor binding-site recovery over the evaluated baselines. Compared with PepMLM, SitePepGen improves macro Jaccard and F1 by 3.62 and 3.56 percentage points, respectively, with paired bootstrap confidence intervals supporting both gains. Separate site-prediction results and sequence-plausibility diagnostics provide complementary evidence for the approach.

Critical weaknesses
  1. Among 202 test queries, 74 have receptor sequences identical to those in training, while unseen receptors are defined solely by the absence of exact sequence matches. Evaluation using receptor sequence clustering is needed to assess generalization beyond closely related proteins.

  2. The compatibility term in Eq. (4) summarizes amino-acid composition, hydrophobicity, and charge, without representing residue ordering or geometric interactions. Peptides with identical compositions but different residue arrangements receive identical compatibility scores. Its effectiveness as a measure of site-specific binding compatibility requires stronger validation.

  3. Ablation experiments lack a comparison with PepMLM using the same imputation procedure, as well as a comparison that ranks the same candidate pool using only language-model scores. These controls would help isolate the contribution of receptor binding-site information.

  4. Structural metrics are averaged over each method’s structurally evaluable queries, but the corresponding sample counts and exclusion reasons are not clearly reported. If these subsets differ across methods, direct comparisons may be biased.

Questions for authors
  1. How similar are the test receptors to those in training beyond exact sequence matches? Do the gains persist under a split based on receptor sequence clustering?
  2. What supports the physicochemical compatibility score in Eq. (4), particularly the amino-acid composition term? Does using predicted sites outperform using random or shuffled sites?
  3. Could you include PepMLM with the same imputation procedure and a baseline that ranks the same candidate pool using only language-model scores? These comparisons would help isolate the contribution of receptor binding-site information.
  4. How many queries are structurally evaluable for each method, and how are unsuccessful predictions handled? Do the gains persist when all methods are evaluated on the same query subset?
7 comments
AuthorsOpenReview ↗

Receptor Similarity Stratification and Subset Evaluation

Q1:How similar are the test receptors to those in training beyond exact sequence matches? Do the gains persist under a split based on receptor sequence clustering?

Response:

Thank you for this suggestion. We added receptor-sequence similarity stratification of the original test set using the original training set and trained models. This is a test-subset analysis; a separate receptor-cluster-disjoint split and retraining have not been performed.

We searched the 6,796 unique training receptor sequences using MMseqs2. A high-similarity training neighbor must satisfy local identity ≥40%, coverage of the shorter sequence ≥70%—with both coverages ≥70% for equal-length sequences—and E-value ≤0.001. Among the original 128 unseen queries, 114 had a detected training neighbor meeting these criteria and 14 did not. Grouping used receptor sequences only. Results are reported in Table 1.

Table 1. Results on 14 queries without detected high-similarity training neighbors.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
PepMLM30.4941.3940.2440.06
SitePepGen32.1344.7940.5641.62
Improvement+1.64+3.40+0.32+1.56

SitePepGen achieves 1.56 percentage points higher F1 on these 14 queries, providing supplementary evidence for site recovery on this subset.

AuthorsOpenReview ↗

Compatibility Score Motivation and Ranking Controls

Q2:What supports the physicochemical compatibility score in Eq. (4), particularly the amino-acid composition term? Does using predicted sites outperform using random or shuffled sites?

Response:

Thank you for this suggestion. Eq. (4) uses amino-acid compositional overlap, hydrophobicity, and charge to provide coarse site-level physicochemical information for auxiliary candidate ranking. Composition-based descriptors have been used for peptide–protein interaction prediction [1], and mutational studies have shown the roles of hydrophobicity, electrostatics, and residue identity in peptide recognition [2,3]. These studies motivate the descriptor choices.

We added two inference-time ranking controls. Within each run, all configurations use the same site-supervised SitePepGen checkpoint, the same candidate sequences, and their associated LM scores.

SitePepGen (LM-only+imputation) removes the entire site-compatibility term in Eq. (4) from ranking, selects candidates using only language-model (LM) scores, and subsequently performs X imputation. This control assesses whether adding the complete compatibility score improves candidate selection.

SitePepGen (random-site control) retains the original compatibility function and component weights but replaces predicted sites with an equal number of residues sampled uniformly without replacement from the same receptor. Composition-, hydrophobicity-, and charge-related compatibility scores are then recomputed using these random sites. This control assesses whether predicted sites provide more useful scoring information than random sites. Both site conditions use identical uniform normalization and empty-site handling. Each query receives one random-site draw per run.

SitePepGen uses the paper’s original configuration, combining LM scoring with compatibility computed from predicted sites. All three configurations retain identical X-imputation, structure-prediction, and contact-evaluation protocols. Results are averaged across paired runs on the same 202 queries.

Table 2. Ranking controls on the same SitePepGen candidate pool.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
SitePepGen (LM-only+imputation)49.8461.5663.7160.70
SitePepGen (random-site control)49.4461.3063.8060.68
SitePepGen50.2861.8364.5861.26

Adding predicted-site compatibility to LM-only ranking increases Jaccard from 49.84% to 50.28% (+0.44 percentage points) and F1 from 60.70% to 61.26% (+0.56 percentage points), supporting the empirical utility of Eq. (4) as a complete auxiliary ranking feature.

With the compatibility function retained, using predicted sites exceeds using random sites by 0.84 percentage points in Jaccard and 0.58 percentage points in F1, supporting the role of predicted-site information in candidate selection. These experiments assess the complete compatibility score, while the cited literature motivates the composition descriptor.

References

[1] Anindya AL et al. (2024). Deciphering peptide-protein interactions via composition-based prediction: a case study with survivin/BIRC5. Machine Learning: Science and Technology, 5, 025081. DOI

[2] Kruse T et al. (2024). Substrate recognition principles for the PP2A-B55 protein phosphatase. Science Advances, 10, eadp5491. DOI

[3] Benz C et al. (2025). Defining short linear motif binding determinants by phage display-based deep mutational scanning. Protein Science, 34, e70174. DOI

AuthorsOpenReview ↗

Matched PepMLM and LM-Only Ranking Controls

Q3:Could you include PepMLM with the same imputation procedure and a baseline that ranks the same candidate pool using only language-model scores? These comparisons would help isolate the contribution of receptor binding-site information.

Response:

Thank you for this suggestion. We added PepMLM with the same imputation procedure and an LM-only ranking control on the same SitePepGen candidate pool. Results are reported in Table 3.

PepMLM (LM-only+imputation) uses the same PepMLM model as the main table. It independently generates candidates with a candidate budget and sampling settings matched to SitePepGen, applies LM-only ranking, and performs one-pass X imputation.

SitePepGen (LM-only+imputation) uses candidates generated by SitePepGen, ranks them solely by LM scores, and applies the same X-imputation procedure as in the paper. SitePepGen uses the paper’s main configuration, combining LM scoring with predicted-site compatibility for candidate ranking and following the original X-imputation and structural-evaluation procedures.

The two SitePepGen ranking configurations use the same checkpoint and candidate pool within each run, with identical X-imputation, structure-prediction, and contact-evaluation protocols. Only the inclusion of site compatibility in ranking changes, and results are averaged across paired runs. All configurations are evaluated on the same 202 queries.

Table 3. Matched imputation and LM-only ranking controls.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
PepMLM (LM-only+imputation)47.1558.7861.2158.10
SitePepGen (LM-only+imputation)49.8461.5663.7160.70
SitePepGen50.2861.8364.5861.26

Under the common LM-only ranking and imputation pipeline, SitePepGen exceeds PepMLM by 2.69 percentage points in Jaccard, with F1 increasing from 58.10% to 60.70% (+2.60 percentage points). This reflects the model-level difference after matching inference.

Adding site compatibility on the same SitePepGen candidate pool increases Jaccard from 49.84% to 50.28% (+0.44 percentage points) and F1 from 60.70% to 61.26% (+0.56 percentage points), reflecting the additional effect of predicted-site information in candidate ranking.

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

Author proposed a method to generate target site specific peptides through pLM and post hoc site checks.

Critical strengths
  1. Joint site prediction and peptide generation are combined with lightweight candidate reranking, avoiding structure prediction during generation.
Critical weaknesses

see questions

Questions for authors
  1. How is this site-specific generation if the predicted binding site does not directly condition peptide sampling?

  2. The compatibility score only considers AA composition, hydrophobicity, and charge. How can it distinguish peptides with different binding interfaces but identical compositions?

  3. In Table 4, X imputation contributes more than site-aware selection. How much improvement remains if PepMLM uses the same sampling and post-processing pipeline?

  4. Why is moPPIt-v3 not included in the benchmark? It provides explicit site-specific generation, unlike PepMLM.

  5. How reliable is the site-recovery evaluation with only one AF2-Multimer prediction per peptide? Did you test multiple independent predictions or validate the pipeline using reference peptides?

  6. Your unseen-receptor split only excludes exact sequence matches. What happens under a homology-aware split?

14 comments
AuthorsOpenReview ↗

Site-Aware Generation without Direct Site-Conditioned Sampling

Q1:How is this site-specific generation if the predicted binding site does not directly condition peptide sampling?

Response:

Thank you for this suggestion. Regarding site-specific generation, we would like to clarify how site information is used in our method. SitePepGen incorporates receptor-site information through joint representation learning and candidate selection. During training, site supervision shapes peptide-generation representations through the shared ESM-2 backbone. During inference, predicted sites inform candidate ranking through compatibility scoring. The predicted site is not an additional input to the peptide sampler.

This site-aware generation design allows the model to learn and use binding-region information from receptor sequences without requiring user-provided target sites. Model comparisons under a common inference pipeline and site-aware ranking controls on the same candidate pool are reported together in Table 1 in our response to Q3.

AuthorsOpenReview ↗

Compatibility Scoring for Peptides with Identical Compositions

Q2:The compatibility score only considers AA composition, hydrophobicity, and charge. How can it distinguish peptides with different binding interfaces but identical compositions?

Response:

Thank you for pointing this out. Peptides with identical amino-acid counts but different residue orders have identical composition, hydrophobicity, and charge descriptors in Eq. (4), and therefore receive the same compatibility score. This term provides coarse site-level physicochemical information without explicitly modeling order-dependent residue pairing or interface geometry.

The complete ranking also includes the position-dependent LM score in Eq. (5), so peptides with identical compositions but different residue orders may still receive different overall scores.

To assess the role of predicted-site information in ranking, we constructed SitePepGen (random-site control). This control retains the scoring function and component weights in Eq. (4), replacing only the predicted sites with an equal number of residues sampled uniformly without replacement from the same receptor. Composition-, hydrophobicity-, and charge-related compatibility scores are then recomputed using these random sites. SitePepGen uses the paper’s original configuration, computing compatibility from predicted sites. Both configurations combine compatibility with LM scores for candidate ranking.

Random and predicted sites use identical uniform normalization and empty-site handling. Each query receives one random-site draw per run. The two configurations use the same checkpoint, candidate pool, and LM scores within each run, with identical X-imputation, structure-prediction, and contact-evaluation protocols. Results are averaged across paired runs on the same 202 queries.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
SitePepGen (random-site control)49.4461.3063.8060.68
SitePepGen50.2861.8364.5861.26

With the compatibility function retained, using predicted sites exceeds using random sites by 0.84 percentage points in Jaccard and 0.58 percentage points in F1, supporting the empirical utility of predicted-site information for candidate ranking.

In addition, SitePepGen (LM-only+imputation) in Table 1 in our response to Q3 retains the same model and candidate pool, removing only the entire site-compatibility term from ranking. Adding this term increases average F1 from 60.70% to 61.26% (+0.56 percentage points). This control supports the overall utility of compatibility as an auxiliary ranking feature.

AuthorsOpenReview ↗

Gains over PepMLM after Matching Sampling and Post-Processing

Q3:In Table 4, X imputation contributes more than site-aware selection. How much improvement remains if PepMLM uses the same sampling and post-processing pipeline?

Response:

Thank you for raising this question. After matching candidate sampling, LM-only ranking, and X imputation, SitePepGen (LM-only+imputation) still exceeds PepMLM (LM-only+imputation) by 2.60 percentage points in F1 and 2.69 percentage points in Jaccard, as shown in Table 1.

PepMLM (LM-only+imputation) uses the same PepMLM model as the main table. It independently generates candidates with a candidate budget and sampling settings matched to SitePepGen, ranks candidates solely by LM scores, and performs one-pass X imputation.

SitePepGen (LM-only+imputation) retains SitePepGen’s model parameters and generated candidates, removes the entire site-compatibility term from ranking, and uses only LM scores with the same X-imputation procedure. These two configurations therefore use matched inference pipelines, each operating on candidates generated by its own model.

The table also includes the paper’s main SitePepGen configuration for reference, which combines LM scoring with predicted-site compatibility. The two SitePepGen ranking configurations use the same checkpoint and candidate pool within each run, with identical imputation and evaluation protocols, and results are averaged across paired runs. All configurations are evaluated on the same 202 queries.

Table 1. Matched inference and site-aware ranking controls.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
PepMLM (LM-only+imputation)47.1558.7861.2158.10
SitePepGen (LM-only+imputation)49.8461.5663.7160.70
SitePepGen50.2861.8364.5861.26

Under this common pipeline, Jaccard increases from PepMLM’s 47.15% to 49.84%, and F1 increases from 58.10% to 60.70%. SitePepGen therefore retains higher site-recovery performance after matching sampling, ranking, and imputation.

Reviewer_g7XoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SitePepGen augments a receptor-conditioned masked peptide model with a residue-level binding-site head on a shared ESM-2 backbone. A fixed physicochemical compatibility heuristic and LM score rank sampled peptides, followed by imputation of unresolved residues. Generation uses receptor sequence and peptide length, without requiring an input structure or specified target site. The paper evaluates predicted receptor-contact overlap, sequence diagnostics, and residue-level site prediction.

Critical strengths
  1. The objective is well motivated. Predicting where a peptide binds is a useful addition to sequence plausibility. Incorporating a site objective into a sequence-only generator is a straightforward, interpretable extension of the PepMLM.
  2. The overall improvement is supported on the reported test set. Results are also separated by exact receptor-sequence overlap with training.
Critical weaknesses
  1. Generalization beyond closely related receptors remains unclear. The seen/unseen analysis is useful, but defines unseen receptors solely by the absence of an exact full-sequence match in training. The 128 unseen queries may therefore still contain close homologs or conserved binding domains represented in training.
  2. The ablations only partially isolate the contribution of site awareness. Table 4 supports benefits from site supervision and the complete inference pipeline. However, the selection and imputation variants all use the site-supervised model, and selection combines site compatibility with LM scoring over multiple candidates. Without matched MLM-only and LM-only-ranking controls, it remains unclear how much of the overall 3.56-point F1 gain specifically requires site information.
Questions for authors
  1. For the 128 unseen queries, could you report each receptor’s maximum sequence identity to a training receptor, together with alignment coverage, and stratify the improvement over PepMLM by similarity? If feasible, could you also assess overlap at the binding-domain/interface level or provide a receptor-family-disjoint evaluation?
  2. Could you compare the MLM-only and site-supervised models using the same candidate budget, LM-only ranking, and identical X imputation?
4 comments
AuthorsOpenReview ↗

Receptor Similarity Analysis and Matched Inference Controls

Q1:For the 128 unseen queries, could you report each receptor’s maximum sequence identity to a training receptor, together with alignment coverage, and stratify the improvement over PepMLM by similarity? If feasible, could you also assess overlap at the binding-domain/interface level or provide a receptor-family-disjoint evaluation?

Response:

Thank you for this suggestion. We further stratified the 128 test queries without exact receptor-sequence matches in training according to their sequence similarity to training receptors. Using MMseqs2, we searched the 6,796 unique training receptor sequences. A high-similarity training neighbor was defined as a hit satisfying local identity ≥40%, coverage of the shorter sequence ≥70%—with both coverages ≥70% for equal-length sequences—and E-value ≤0.001.

Of these queries, 114 had a detected training neighbor meeting the criteria and 14 did not. Grouping used receptor sequences only, without peptide sequences, site labels, or model performance. We extracted these 14 queries from the main experimental results while retaining the original inference settings. Results are reported in Table 1.

Table 1. Results on 14 queries without detected high-similarity training neighbors, using the main experimental settings.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
PepMLM30.4941.3940.2440.06
SitePepGen M432.1344.7940.5641.62
Improvement+1.64+3.40+0.32+1.56

SitePepGen improves Jaccard and F1 by 1.64 and 1.56 percentage points, respectively, providing supplementary evidence for site recovery on queries without detected high-similarity training neighbors. This analysis stratifies the original test set using the original training set and trained models.

Q2:Could you compare the MLM-only and site-supervised models using the same candidate budget, LM-only ranking, and identical X imputation?

Response:

Thank you for this suggestion. We added two controls: PepMLM with the same sampling and post-processing pipeline, and LM-only ranking on the same SitePepGen candidate pool.

The new control uses the same PepMLM model as the main table, with screening and X imputation added. Its F1 increases from 57.70% to 58.10%. These results therefore differ in their inference pipelines rather than the underlying model.

PepMLM (LM-only+imputation) and SitePepGen (LM-only+imputation) each generate their own candidate pool with matched candidate budgets and sampling settings. Both rank candidates solely by LM scores and perform one-pass X imputation. This comparison assesses model performance under a common inference pipeline.

SitePepGen (LM-only+imputation) retains SitePepGen’s model parameters and generated candidates, removing only the entire site-compatibility term from ranking. SitePepGen uses the paper’s main configuration, combining LM scoring with predicted-site compatibility. The two configurations use the same checkpoint and candidate pool within each run, with identical X-imputation, structure-prediction, and contact-evaluation protocols. Only the inclusion of site compatibility changes. Results are averaged across paired runs, with all configurations evaluated on the same 202 queries.

Table 2. Matched inference and site-aware ranking controls.

MethodJaccard (%)Precision (%)Recall (%)F1 (%)
PepMLM (LM-only+imputation)47.1558.7861.2158.10
SitePepGen (LM-only+imputation)49.8461.5663.7160.70
SitePepGen50.2861.8364.5861.26

Under the common LM-only ranking and imputation pipeline, SitePepGen achieves an F1 score 2.60 percentage points higher than PepMLM. Adding site compatibility on the same SitePepGen candidate pool yields an observed further 0.56-point improvement. SitePepGen retains gains of 3.13 percentage points in Jaccard and 3.16 percentage points in F1 over PepMLM with the enhanced pipeline. These comparisons separately show the model-level difference under a common inference pipeline and the additional utility of site-aware ranking within the same model.

Dear Reviewer,

Thank you for your careful review and valuable feedback on our paper.

We have provided detailed responses to your comments and conducted additional experiments to address your main concerns. As the public discussion period draws to a close, we would greatly appreciate your feedback on our responses and the additional results. If you have any remaining concerns or would like further clarification, we will do our best to address your questions.

Thank you again for your time and consideration.

Best regards,

The Authors

AuthorsOpenReview ↗

Revised Manuscript and Summary of Updates

Dear Reviewer,

Thank you for your constructive feedback. We have uploaded a revised manuscript incorporating the additional analyses and clarifications discussed in our responses. The revisions are marked in blue.

Section 4.5 and Table 4 now include matched inference controls. Under matched candidate budgets, sampling settings, LM-only ranking, and X imputation, SitePepGen exceeds PepMLM by 2.60 percentage points in F1. We describe this as a model-level difference under common inference settings. Adding predicted-site compatibility to the same SitePepGen candidate pool provides a further 0.56-percentage-point F1 improvement. Appendix A.11 provides additional settings and ranking controls.

Appendix A.9 reports the receptor-sequence similarity screening. Among the 128 exact-sequence-unseen queries, 114 have a detected training neighbor under the specified criteria and 14 do not. Results on the latter subset are reported as a descriptive analysis, without claiming homology-disjoint generalization. We distinguish exact-sequence novelty from homology-disjoint evaluation.

We appreciate your suggestions, which helped clarify the experimental comparisons and the scope of our conclusions.

Best regards, The Authors

Thank you for your reply. All my concerns have been addressed, and I will maintain my positive rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.