acceptodds
Under review as a conference paper at ICLR 2027

PepRIG: Receptor-Guided Sequence–Structure Coupling for All-Atom Peptide Generation

Abstract

Peptides have diverse biological activities that support applications in targeted therapy, functional foods, and nutritional health. These activities depend on molecular interactions shaped by peptide sequence and structure, making their joint design important for developing peptides with desired functions. Generative models support joint peptide sequence–structure design, yet receptor compatibility requires residue identities and spatial arrangements to jointly establish complementary interactions at the binding interface. The central challenge is to coordinate these interdependent choices throughout generation, so that sequence and structure adapt together to the target receptor. To address this challenge, we introduce PepRIG, a receptor-guided all-atom diffusion framework that couples sequence and structure generation at the representation and potential levels. Specifically, Interface-Guided Co-Adaptation (IGCA) integrates multiscale receptor-interface information into shared representations, guiding sequence and structure updates with a common description of peptide–receptor interactions. Complementing this shared representation, Reciprocal Binding Potential (RBP) derives coupled gradients from a receptor-conditioned scalar potential, allowing each modality's correction to account for the other modality. Extensive experiments on LNR and PepBDB demonstrate that PepRIG outperforms strong baselines on key evaluation criteria, achieving state-of-the-art results across both benchmarks. Ablation studies demonstrate the complementary benefits of the two mechanisms, supporting shared receptor guidance for coupled sequence–structure generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Receptor-conditioned peptide design, Sequence–structure co-design, Latent diffusion models

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_YQXrOpenReview ↗
Rating
2
Confidence
4(+1)
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

PepRIG generates peptide sequences and all-atom conformations conditional on a fixed receptor pocket and peptide length. Shared multiscale interface features and gradients of a scalar latent compatibility potential couple semantic and geometric denoising. Evaluation measures reference recovery on LNR and PepBDB.

Critical strengths

Component ablations and independent-head replacements support the proposed coupling mechanisms (Tables 2-3). The full model improves sequence recovery and structural metrics over both replacements, providing evidence beyond simply supplying interface information. The evaluation extends beyond separate per-metric oracle maxima. Mean-candidate gains and same-candidate joint success support broader improvement on LNR. At Table 4's strictest thresholds, joint successes increase from 122 to 142 among 3,720 candidates per method. FastRelax preserves the LNR DockQ advantage.

Critical weaknesses
  1. Methodological Ambiguity & Lack of Innovation: The Reciprocal Binding Potential (RBP) trick is severely ill-defined. The Methods section fails to clarify whether RBP is a rule-based function, a traditional ML model, or a Deep Learning module, making the core algorithmic contribution opaque. Furthermore, the framework lacks novelty; target-binder co-adaptation and energy landscape guidance have already been achieved by prior work such as EpiCoCo (ICML 2026).

  2. Insufficient Empirical Validation: The experimental section lacks a clear, comprehensive main table, and the current baseline comparisons are insufficient to convincingly demonstrate the model's superiority.

Questions for authors
  1. Can you explicitly define the technical nature of the RBP mechanism? Please clarify whether it is a parameterized deep learning function, a rule-based formula, or another specific method.

  2. How exactly does your proposed framework theoretically and empirically differentiate its co-adaptation and energy landscape guidance from EpiCoCo (ICML 2026)?

  3. Can you provide a complete main results table with an expanded, rigorous set of baseline comparisons to justify your performance claims?

1 comment

We appreciate the time and care the reviewer put into evaluating our work. We respond point by point below.

1. Technical nature of RBP (W1, Q1)

RBP is a trainable neural module. A shared multilayer perceptron scores each peptide–receptor residue pair from the noisy peptide features, receptor features, distance encoding, and diffusion time. The pair scores form a single scalar potential (Eq. 7). Its negative gradients with respect to the semantic and geometric latents give the two denoising corrections (Eq. 8). With the autoencoder frozen, we train RBP jointly with the denoiser by supervising these corrections with the injected diffusion noise (Eq. 11). The scalar is a learned denoising potential, not a hand-designed energy or an affinity predictor.

2. Distinction from EpiCoCo (W1, Q2)

EpiCoCo adapts the MHC geometry alongside the epitope and steers sampling by contrasting predictions from a denoiser conditioned on high versus low affinity (Secs. 2.3–2.5). PepRIG holds the receptor fixed. IGCA recomputes a shared interface representation as the peptide changes, while RBP derives both peptide correction fields from one scalar learned without affinity labels. On a fixed interface graph, the cross-modal Jacobians of the gain-normalized corrections are transposes of one another (Eq. 9). The coupling is thus between peptide sequence and geometry, not between the peptide and a moving receptor.

EpiCoCo tests its guidance on pMHC binding energy, where adding contrastive guidance to co-modeling and affinity prompting changes the reported score from −42.67 to −45.20 REU (its Table 2). Our Table 4 instead tests whether the two corrections benefit from sharing a potential. Relative to two separate gradient-producing potentials, a shared potential raises DockQ from 0.6500 to 0.6531 on LNR and from 0.5011 to 0.5044 on PepBDB, and reduces Local RMSD on both. These ablations test different mechanisms on different benchmarks. They do not rank the two systems head-to-head.

3. Expanded main comparison (W2, Q3)

We added UniMoMo-single (ICML 2025) on LNR and PepBDB, extending the comparison of joint all-atom peptide sequence and structure design methods from three to four baselines per benchmark. The results below combine the unrelaxed scores in Table 1 (40 candidates per target) with the published UniMoMo-single values. Appendix A.15 identifies the source and aggregation for the added results. Bold marks the best reported value per metric within each benchmark.

LNR (93 targets)

MethodGDT-TS ↑Contact-F1 ↑DockQ ↑Local RMSD (Å) ↓Seq. diversity ↑
PepGLAD0.71480.79360.60731.21470.9138
PepFlow0.73680.79470.53641.06480.8265
PepTri0.74010.82830.62821.18110.9062
UniMoMo-single0.620.610.571.980.92
PepRIG0.75440.85570.65311.05400.9046

PepBDB (142 targets)

MethodGDT-TS ↑Contact-F1 ↑DockQ ↑Local RMSD (Å) ↓Seq. diversity ↑
PepGLAD0.51670.51480.43721.91360.9143
PepFlow0.63680.70460.41671.26480.6659
PepTri0.63730.74930.51171.44350.9231
UniMoMo-single0.480.480.442.780.90
PepRIG0.65780.77730.50441.35390.9153

PepRIG has the highest reported GDT-TS and Contact-F1 on both benchmarks, together with the highest DockQ and lowest Local RMSD on LNR. On PepBDB, PepTri has the highest DockQ and PepFlow the lowest Local RMSD.

Reviewer_Zae4OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PepRIG, a receptor-guided latent diffusion framework for joint peptide sequence and all-atom structure generation. The method introduces two components: Interface-Guided Co-Adaptation (IGCA) and Reciprocal Binding Potential (RBP). IGCA aggregates multiscale receptor-interface information and incorporates a shared residual into semantic and geometric denoising branches. RBP derives correction fields for both modalities from a shared scalar potential to encourage coupled sequence–structure refinement. Experiments on LNR and PepBDB demonstrate improvements in peptide structural quality over existing baselines. The authors also provide component ablations, independent-head comparisons, and additional analyses of generation quality.

Critical strengths
  • The paper addresses an important limitation of existing peptide generation methods, where sequence and structure may be generated without sufficiently considering their joint compatibility with the target receptor. To address this, the authors propose IGCA and RBP, which incorporate receptor-interface context and reciprocal coupling into the denoising process. These components are intuitively motivated and naturally integrated into the existing latent diffusion framework.

  • The paper demonstrates improvements in peptide structural quality on representative benchmarks. The ablation studies clearly examine the individual and combined contributions of IGCA and RBP, while additional analyses further support the experimental findings. Overall, the experimental results and ablation analyses are thoroughly presented and clearly explained.
Critical weaknesses

W1. Unclear binding relevance of the Reciprocal Binding Potential

Although RBP is designed to improve receptor–peptide binding compatibility, its potential is trained using Gaussian noise prediction rather than binding affinity or interaction energy. It remains unclear whether the learned potential captures meaningful binding preferences beyond serving as an auxiliary denoising objective.


W2. Insufficient evaluation of sequence generation and receptor-binding compatibility

Although the paper focuses on joint sequence–structure generation, Sliding-AAR is omitted from the main comparison, and the reported sequence recovery improvements are marginal. Furthermore, unlike PepTri, the paper does not report binding energy (ΔG) or binding Success Rate. Given the mixed DockQ results across datasets, it remains unclear whether the improvements in structural quality translate into better sequence generation and receptor-binding compatibility.

Questions for authors

Q1. Quantitative evidence for the motivating limitation

Could the authors quantitatively demonstrate the limitations of existing methods in coordinating peptide sequence and structure with the receptor interface? How does PepRIG specifically address these limitations beyond improvements in conventional structural metrics?


Q2. Sequence generation and binding evaluation

Could the authors include Sliding-AAR in the main baseline comparison and report binding energy (ΔG) and binding Success Rate using the same protocol as PepTri?


Q3. Computational efficiency

Could the authors report training time, inference time, GPU memory usage, and parameter counts compared with other baseline models?

1 comment

Thank you for your thoughtful feedback. We address your concerns below.

1. Binding relevance of RBP (W1)

RBP learns a receptor-conditioned signal that helps recover reference bound poses. Gaussian noise perturbs peptide latents from bound complexes, so denoising recovers the peptide's placement relative to the receptor as well as its internal structure. The learned gradients enter the reverse sampling updates directly (Eq. 10). Compared with two separate potentials that also derive corrections from gradients, the shared potential raised DockQ from 0.6500 to 0.6531 on LNR and from 0.5011 to 0.5044 on PepBDB, with lower Local RMSD on both (Table 4). This supports binding-relevant structural guidance in the learned corrections. The scalar itself has not been evaluated for ranking poses or affinities.

2. Quantifying joint recovery (Q1)

A peptide must place residues suited to the intended contacts at the right positions on the receptor. Separate best-of-40 scores can select different candidates, so high scores for sequence, structure and pose do not show that any one peptide meets all three requirements. On LNR, PepTri's median within-target Spearman correlations of DockQ with Sliding-AAR and negative Local RMSD were 0.00 and 0.07 (Fig. 8). At Sliding-AAR ≥0.30\geq 0.30, Local RMSD ≤1.5\leq 1.5 Å and DockQ ≥0.49\geq 0.49, only 122 of 3,720 PepTri candidates met all three criteria, covering 21 of 93 targets.

PepRIG addresses this gap during denoising. IGCA recomputes a shared interface representation from the current peptide pose for both updates. RBP makes the sequence correction respond to placement and the coordinate correction respond to the residue state. Under the same thresholds, 142 of 3,720 PepRIG candidates met all three criteria, covering 26 of 93 targets. Joint recovery rose from 3.28% to 3.82%, a gain of 0.54 percentage points (pointwise paired-target 95% CI [0.11,1.05][0.11,1.05]). At the loosest thresholds, both methods covered 51 targets and the interval for the rate difference included zero. Table 5 reports the full comparison.

3. Sliding-AAR (W2, Q2)

We added a matched Sliding-AAR comparison with PepTri in Table 10. Averaging each target's best score among 40 candidates, PepRIG improved from 0.3609 to 0.3659 on LNR and decreased from 0.3431 to 0.3406 on PepBDB. Sliding-AAR counts exact residue matches after offset alignment, so it does not credit geometry improvements that leave decoded identities unchanged.

4. Rosetta interface energy and success rate (W2, Q2)

Under PepTri's code-default FastRelax and Rosetta scoring protocol, PepRIG had lower mean target-minimum ΔG\Delta G and higher success rates on PepBDB, while PepTri led on LNR. The paired 95% intervals for the mean energy difference included zero on both datasets. We use separated-partner repacking and average the lowest ΔG\Delta G among 40 candidates per target. Success is the fraction of all candidates below PepTri's cutoff of 0 Rosetta energy units (REU), with −5-5 REU as a stricter cutoff.

DatasetModelMean min. ΔG\Delta G (REU)ΔG<−5\Delta G<-5 (%)ΔG<0\Delta G<0 (%)
LNRPepTri−18.769237.1553.95
LNRPepRIG−18.543335.7049.14
PepBDBPepTri−15.782524.8847.17
PepBDBPepRIG−15.942727.0647.59

5. Computational cost (Q3)

We profiled PepTri and PepRIG on one NVIDIA A800 GPU.

DatasetModelTotal (M)Trainable (M)Train (s/step)Train peak alloc. (GiB)Gen. (s/40)Gen. peak process (GiB)
LNRPepTri5.64851.90530.1035.48716.5015.613
LNRPepRIG6.81912.23450.2958.52216.5002.303
PepBDBPepTri5.64861.90530.1312.62618.7776.551
PepBDBPepRIG6.81792.23450.3049.43019.5032.334

Compared with PepTri, PepRIG used 59% and 64% less peak GPU process memory during generation on LNR and PepBDB, respectively. Generation time was nearly identical on LNR and slightly higher on PepBDB. PepRIG took longer per training step and used more peak allocated memory. See Table 13 and Appendix A.14 for the profiling setup.

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

This work introduces PepRIG, a receptor-guided all-atom diffusion framework that couples sequence and structure generation at both the representation and potential levels. Specifically, it contains two modules. IGCA extracts multiscale shared representations from the current peptide–receptor interface and uses them to guide both sequence and structure updates. RBP constructs a receptor-conditioned shared scalar potential and derives gradients with respect to semantic and geometric variables, providing coupled correction directions for the two modalities.

Critical strengths
  1. The two proposed modules have clear roles, and the ablation experiments demonstrate their effectiveness.
  2. The experimental analysis is relatively comprehensive. The authors also report negative results rather than omitting them, such as the decrease in DockQ on PepBDB and the increase in interface crowding.
Critical weaknesses
  1. The central motivation of the paper is to make sequence and structure jointly adapt better to the receptor. However, on PepBDB, the DockQ of PepRIG decreases from 0.5117 for PepTri to 0.5044.
  2. Based on the reported results, the statement in the abstract that PepRIG achieves “state-of-the-art results across both benchmarks” appears somewhat overstated.
Questions for authors

The main goal of PepRIG is to improve receptor-guided sequence–structure compatibility. Why is its DockQ lower than PepTri on PepBDB? Since both IGCA and RBP explicitly use receptor-interface information, why do the external clash rate and heavy-atom overlap both become worse? Is sequence–structure coupling the central contribution of the paper? If so, why is the improvement in sequence recovery relatively limited? Does the benefit of RBP mainly come from the shared scalar potential itself, or from introducing an additional gradient-based correction branch?

1 comment

We thank the reviewer for the careful reading and constructive feedback. We respond to each comment below.

1. PepBDB DockQ and the abstract

The PepBDB decrease reflects a difference between recovering the peptide’s internal contacts and recovering its exact placement on the receptor. IGCA pools receptor features around the evolving peptide at multiple scales, and RBP couples sequence and coordinate updates through a shared potential built from residue-pair features and distances. This design guides the peptide to fit its receptor neighborhood without prescribing a fixed native contact map. Nearby placements can fit similar neighborhoods while recovering different native contacts. PepBDB reference interfaces have fewer native contacts per peptide residue than LNR (median 2.76 versus 4.00), whereas the denser LNR contact patterns constrain placement through more simultaneous residue-pair relationships.

We measured both contact metrics on each method’s highest-DockQ unrelaxed candidate for each target. On PepBDB, mean intra-peptide contact F1 increased from 0.4611 to 0.4991, but peptide–receptor contact F1 decreased from 0.3874 to 0.3738. On LNR, both contact metrics improved, with interface contact F1 rising from 0.4430 to 0.4767. This difference in native interface recovery is consistent with best-of-40 DockQ rising from 0.6282 to 0.6531 on LNR and falling from 0.5117 to 0.5044 on PepBDB (paired difference −0.0073, 95% CI [−0.0164, 0.0021]).

PepRIG improved internal structural recovery on both benchmarks, including lower median Local RMSD (Table 1). Mean DockQ across all 40 PepBDB candidates also rose from 0.2843 to 0.2875. The revised abstract highlights the structural gains on both benchmarks and the additional improvement in receptor-relative placement on LNR.

2. External clashes and interface overlap

IGCA and RBP coordinate sequence and structure denoising using residue-level interface geometry. The objective guides peptide placement but does not explicitly penalize peptide–receptor atom pairs that come too close. An update that improves the overall pose can therefore still place part of the peptide backbone too close to the receptor or leave side-chain atoms overlapping at the interface.

3. Coupling and sequence recovery

Sequence–structure coupling is the central contribution. PepRIG retains the backbone's sequence supervision and adds coupled corrections in latent space. In RBP, the coordinate correction depends on the continuous residue representation, so a change in that representation can refine the geometry without changing the decoded amino acid. A receptor neighborhood can also accommodate different residue identities. For example, substituting one nonpolar residue for another can preserve a hydrophobic contact after local geometric adjustment. Sliding-AAR counts only exact residue matches after offset alignment and gives no credit for either type of refinement.

Best-of-40 Sliding-AAR rose from 0.3609 to 0.3659 on LNR and fell from 0.3431 to 0.3406 on PepBDB (Table 10). The joint test asks whether the same candidate meets the thresholds for Sliding-AAR, Local RMSD and DockQ. At the strictest setting in Table 5, the number of LNR candidates meeting all three thresholds increased from 122 to 142 out of 3,720 (0.54 percentage points, pointwise paired-target 95% CI [0.11,1.05][0.11,1.05]). Coverage increased from 21 to 26 out of 93 targets. At the loosest setting, coverage was unchanged and the interval included zero.

4. Shared potential versus an extra correction branch

RBP benefits from both gradient-based correction and the coupling induced by a shared scalar potential.

Independent heads predict the sequence and coordinate corrections directly. The two-potential control derives each correction from a separate scalar, whereas RBP takes both gradients from the same scalar. On a fixed interface graph, the shared potential gives the corrections the reciprocity relation in Eq. 9, linking how the sequence correction responds to geometry with how the coordinate correction responds to sequence.

In Table 4, switching from independent heads to two separate potentials increased DockQ by 0.0031 on both benchmarks. Sharing the potential yielded a further gain of 0.0031 on LNR and 0.0033 on PepBDB, while reducing Local RMSD by 0.0357 Å and 0.0194 Å, respectively.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.