acceptodds
Under review as a conference paper at ICLR 2027

PepARD: Full-Atom Peptide Design via Geometric Latent Diffusion with Adaptive Reward Steering

Abstract

Target-aware peptide design requires simultaneously modeling molecular geometry, receptor interactions, and diverse biophysical constraints. However, existing methods often use coarse structural representations, separate peptide generation from pose prediction, or rely on fixed objectives during optimization. We propose PepARD, a unified framework for full-atom target-specific peptide design based on geometric latent diffusion. PepARD encodes peptide sequence and 3D structure into complementary invariant and -equivariant latent spaces, while its Adaptive Multi-Channel -Equivariant Graph Neural Network (AMEGNN) captures detailed receptor-peptide interactions with variable heavy-atom configurations. Receptor-conditioned diffusion enables both de novo sequence-structure generation and fixed-sequence binding pose prediction within a shared framework. To further improve physical plausibility, an Adaptive Reward Weight Network dynamically balances structural, steric, geometric, and sequence objectives during latent refinement. Experiments on PepBench and PepBDB demonstrate that PepARD produces diverse and physically plausible peptides with favorable binding properties and improves recovery of native-like binding conformations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide Design, Geometric Latent Diffusion, Equivariant Graph Neural Network

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 2 reviews, above 27% of the others.
Reviewer_83nPOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

PepARD introduces an all-atom peptide design framework capable of simultaneously generating the amino acid sequence and three-dimensional structure of a peptide, tailored to the binding pocket of a target protein. The method employs a dual-stream equivariant autoencoder to represent sequence semantics and molecular geometry, utilizing a receptor-conditioned diffusion model to generate candidate peptides. Additionally, adaptive reward steering is designed to balance structural plausibility, collision avoidance, and sequence quality.

Critical strengths

The authors introduce gradient guidance with a set of physical objectives to steer the generation towards valid regions. Instead of relying on a fixed set of reward weights, PepARD allows these weights to be adjusted throughout the optimization process. This approach may offer useful insights for multi-objective optimization.

Critical weaknesses
  1. Overall, the co-design performance of PepARD has several aspects severely inferior to existing methods. According to the co-design success rates reported in the paper, PepARD achieves 40.9% on PepBench, which is significantly lower than the 59.4% achieved by PepFGLD. The binding energy of PepARD on PepBench is -19.90, which is also significantly worse than the best baseline, achieving -34.35. Furthermore, across other metrics, the performance of the proposed method does not differ substantially from the baselines.

  2. Many of the proposed contents exist in previous work. The full-atom latent diffusion as well as the invariant and equivariant latent space exist in the baseline PepGLAD. The proposed AMEGNN and its RollerPooling are also adopted by the baseline PepGLAD. The physical guidance are widely adopted in baselines like PepTri and PepGLAD. It is weird to claim the methodologies in the previous baselines as core contributions. Also, the paper lacks a section of related work to discuss the difference between this work and previous ones.

Questions for authors

Could the authors provide qualitative or quantitative evidence illustrating how the adaptive reward weights evolve during the iterative optimization process, and how these shifts relate to the improvement of individual optimization objectives? For instance, do different target proteins or distinct optimization stages lead to systematically different weight-adjustment trajectories? Furthermore, compared to approaches that employ uniform or fixed weights, do these adaptive adjustments result in superior final structures?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 83nP

Thank you for the thoughtful review.

1. Co-design Performance (Weakness 1)

On PepBench, PepARD trails PepFGLD in reported success (40.9%40.9\% vs. 59.4%59.4\%), with similar ΔG\Delta G scores (−19.90-19.90 vs. −19.86-19.86). The diversity gains over PepTri are modest. UniMoMo reports a lower nominal ΔG\Delta G (−34.35 kcal/mol-34.35\,\mathrm{kcal/mol}) alongside more atomic clashes (>45%>45\% vs. <5%<5\% for PepARD) and lower contact consistency (0.4390.439 vs. 0.7910.791). On PepBDB, the reported success rates are close (50.2%50.2\% vs. 49.7%49.7\% for PepFGLD). We will clarify baseline sources, success criteria, and relaxation settings in Table 2.

2. Attribution and Contribution (Weakness 2)

We will revise the contribution statement to credit PepGLAD for full-atom latent diffusion with invariant and equivariant representations, and dyMEAN and PepGLAD for AMEGNN and RollerPooling. PepARD's main contribution is learning reward weights from the current candidate state and refinement step over ten post-diffusion updates. We will explain how this mechanism differs from the sampling guidance in PepGLAD, PepTri, and PepFGLD in a dedicated Related Work section.

3. Adaptive Reward Weights (Question)

Compared with uniform weights, adaptive weighting raises co-design success from 39.9%39.9\% to 40.9%40.9\% on PepBench and from 27.6%27.6\% to 50.2%50.2\% on PepBDB (Table 1). In FixSeq, DockQ increases from 0.5900.590 to 0.6430.643 on PepBench and from 0.3660.366 to 0.3860.386 on PepBDB, while PepBDB all-atom RMSD increases from 9.539.53 to 9.62 A˚9.62\,\mathring{\mathrm{A}} (Table 3).

Reviewer_83nPOpenReview ↗

Thanks

Thanks for the response, which partially addressed my concerns and thus I have raised the score. However, I still think the methodology novelty is quite limited, therefore I am still leaning towards rejection.

Reviewer_hk8HOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PepARD addresses receptor-conditioned peptide sequence–structure co-design and fixed-sequence pose generation. It combines a geometric autoencoder, sequence-only invariant latents, three-channel coordinate latents, and an AMEGNN-based latent diffusion model. An adaptive network predicts weights for latent-space rewards during ten post-diffusion refinement steps. Decoded structures undergo Rosetta FastRelax before evaluation. Experiments on PepBench and PepBDB report competitive results and improvements over variants without refinement or adaptive weighting.

Critical strengths
  • The sequence-only invariant encoder provides a clear way to condition FixSeq generation without accessing the reference peptide structure. Appendix A.2 explicitly documents the inference inputs and masking of the training-only RMSD reward.

  • The refinement ablations show meaningful improvements. On PepBench, FixSeq DockQ increases from 0.526 without refinement to 0.590 with uniform weights and 0.643 with adaptive weights (Table 3). PepBDB co-design success also improves from 0.320 without refinement to 0.502 for the full model (Table 1).

  • Appendix A.3 clearly distinguishes latent-space geometric proxies from decoded full-atom outputs, helping readers understand what the refinement directly optimizes.

Critical weaknesses
  • Table 2 needs protocol reconciliation. Its PepTri success and energy entries match PepTri’s published pre-relaxation results, while PepARD describes post-FastRelax evaluation (§2.4). Appendix C.2 also uses a different success definition. Whether these results share a common evaluation protocol is unclear.

  • The architectural contribution is insufficiently distinguished from PepGLAD/dyMEAN, which already provide the adaptive multi-channel backbone and related latent modeling machinery. The paper should identify its specific changes; removing equivariance does not isolate improvements over the existing architecture.

  • Equation 22 encourages stronger steering for above-baseline candidates while detaching the final reward. This heuristic does not directly establish better objective balancing. Comparisons only against uniform weights cannot distinguish state adaptation from better constant weights or different effective update magnitudes (Tables 1 and 3).

  • The clash reward considers peptide latent channels rather than receptor atoms, and the geometric rewards operate before decoding (§2.4). Decoded geometry and interface-clash measurements before and after FastRelax are needed to substantiate the claimed physical benefits.

Questions for authors
  1. Were all baselines evaluated on identical splits with the same metric definitions, candidate counts, and FastRelax settings? Please reconcile the PepTri entries with its original Table 1.

  2. Which AMEGNN and RollerPooling operations differ from the cited implementations? How does the proposed architecture compare with the original backbone under matched training?

  3. Does adaptive weighting outperform validation-tuned constant weights and a step-only weight schedule using the same refinement budget?

  4. Are the coordinate statistics in Equation 24 pocket-centered? Where are gradients stopped in Algorithm 2 to avoid differentiating through the unrolled refinement trajectory?

2 comments
AuthorsOpenReview ↗

Response to Reviewer hk8H

Thank you for the careful review and for highlighting the target-blind FixSeq protocol and refinement gains.

1. Evaluation Protocol and PepTri (Weakness 1 & Question 1)

We obtained the PepTri results in Table 2 through our own reproduction on PepBench and PepBDB. At the reported precision, the Success and ΔG\Delta G values match the pre-relaxation results in PepTri's Table 1.

Evaluation differences. PepTri uses OpenMM with Amber14 for post-relaxation and defines Success as ΔG<−5 REU\Delta G < -5\,\mathrm{REU}, whereas PepARD uses Rosetta FastRelax and requires both ΔG≤ΔGnative\Delta G \leq \Delta G_{\mathrm{native}} and a clash ratio below 0.150.15 (Appendix C.2).

Clarification and revision. We will explicitly label the PepTri results as our reproduction and clarify the evaluation settings, including data splits, candidate counts, relaxation stages, and scoring rules. The reported 61.9% relative improvement on PepBDB (Success: 0.502 vs. 0.310) will be retained only if confirmed under a unified evaluation protocol.

2. Backbone and Architectural Changes (Weakness 2 & Question 2)

PepARD builds on dyMEAN's adaptive multi-channel encoder and PepGLAD's geometric latent diffusion. dyMEAN already incorporates weighted channel distances and channel-dependent pooling.

Architectural distinctions. In Section 2.2, PepARD uses RBF features of pairwise channel distances alongside atom and edge-type information. Equation (4) aggregates messages over valid channel pairs, using upper-triangular pooling for self-edges, while coordinate masking keeps the receptor fixed.

Contribution and clarification. We will clarify these distinctions and acknowledge the inherited components. Our equivariance ablation compares two PepARD variants; isolating the backbone modifications would require a matched backbone-swap experiment. Our primary contribution lies in combining sequence-only FixSeq conditioning with task-masked adaptive latent refinement.

3. Adaptive-Weight Ablation (Weakness 3 & Question 3)

Equation (22) provides a surrogate that strengthens local steering when the detached terminal reward exceeds its moving baseline.

Experimental evidence.

TaskDatasetMetricNo RefinementUniform WeightsAdaptive Weights
Co-designPepBenchSuccess Rate ↑0.3900.3990.409
Co-designPepBDBSuccess Rate ↑0.3200.2760.502
FixSeqPepBenchDockQ ↑0.5260.5900.643
FixSeqPepBDBDockQ ↑0.3520.3660.386

Interpretation and limitations. These results demonstrate improvements over the tested uniform-weight baseline but do not fully isolate the benefit of state dependence.

A more controlled comparison would evaluate tuned constant weights, step-only schedules, and state-dependent weights using identical denoised candidates, rewards, refinement steps, and effective update norms.

We will restrict our claims to what the current experiments establish.

4. Coordinate Statistics and Stop-Gradients (Question 4)

Coordinate statistics. Diffusion operates on pocket-centered coordinates. However, Algorithm 2 restores physical coordinates at line 16 without explicitly specifying the coordinate frame used by Equation (24). Directly using restored coordinates could make coordinate-norm statistics dependent on the global origin. We will verify and clarify the coordinate frame.

Gradient boundaries. Line 26 detaches the terminal reward and moving baseline. The surrogate in Section 2.5 also treats the input state of each refinement step as fixed during backpropagation, while preserving gradients through the current reward weights to ϕ\phi.

We will verify these gradient boundaries against the implementation and clarify them in Algorithm 2.

5. Decoded Geometry and Interface Clashes (Weakness 4)

Latent versus atomic geometry. The clash and bond proxies operate on peptide latent channels, penalizing close channel pairs and deviations in consecutive CαC_{\alpha} distances. However, receptor–peptide clashes must be assessed on decoded atomic structures.

Controlled evaluation. To isolate refinement effects, a controlled comparison should decode identical denoised candidates with and without latent refinement, then evaluate both before and after the same FastRelax procedure.

Relevant metrics include interfacial and intrapeptide heavy-atom clash rates and backbone bond outliers. We will clarify the distinction between latent-space penalties and post-relaxation geometric quality.

Thank you for the clarifications. After considering the responses, I am revising my rating to 2. My concern about the comparability of the baseline results remains unresolved.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Reviewer 1yQQ

Thank you for the detailed review.

1. Evaluation Protocol and PepTri (Weaknesses 5–7 & Question 1)

We reran PepTri for the ΔG\Delta G and Success entries in Table 2. The HSRN and dyMEAN FixSeq entries match PepGLAD Table 3. PepARD uses the 93-complex PepBench LNR split and a 30% sequence-identity cutoff for PepBDB. Table 2 averages three PepARD runs evaluated after Rosetta FastRelax.

Appendix C.2 defines success as ΔG≤ΔGnative\Delta G \le \Delta G_{\mathrm{native}} and clash ratio below 0.15. PepGLAD uses ΔG<0\Delta G < 0 and different diversity and consistency metrics. Common scoring across all rows remains unverified. We will document each row’s provenance, split, checkpoint, sampling budget, aggregation, and relaxation settings. The 0.005 PepBDB gap to PepFGLD is descriptive. Isolating the effect of refinement requires paired results with and without refinement, both before and after matched FastRelax.

2. Adaptive Refinement and Steering (Weaknesses 1, 3, 4 & Question 2)

No refinement, uniform weights, and adaptive weights yield PepBDB CoDesign Success of 0.320, 0.276, and 0.502, and PepBench FixSeq DockQ of 0.526, 0.590, and 0.643, respectively. The reported weight control is uniform.

Eq. (22) gives outcome-conditioned credit to steering magnitude through a detached advantage. Normalizing the terminal score leaves the update sensitive to individual reward scales. We would compare validation-tuned fixed weights, learned constant weights, and a step-only schedule on identical candidates with K=10K=10 updates. Matching update norms would test directional benefit, and rescaling rewards would test sensitivity.

3. Architecture and RollerPooling (Weakness 2 & Question 3)

PepARD builds on PepGLAD’s latent diffusion and dyMEAN’s multi-channel geometry. Its main contribution is state-dependent refinement. RollerPooling aggregates valid channel-pair messages with an upper-triangular self-edge mask. With uniform normalization, it is a masked mean over those pairs. dyMEAN instead pools coordinate scaling coefficients.

The current ablation changes the equivariant updates jointly. A direct sequence embedding, a single CαC_\alpha channel, and a mean over all valid self-pairs would separately test the sequence encoder, three-channel geometry, and triangular mask.

4. Coordinate Frame and Rigid Transformations (Question 4)

Algorithm 2 restores the pocket offset, and Eq. (24) leaves centering implicit. Absolute norms depend on translation. Centering at the pocket centroid crc_r gives ∣Qz+t−(Qcr+t)∣2=∣z−cr∣2\\|Qz+t-(Qc_r+t)\\|_2=\\|z-c_r\\|_2. This supports equivariant updates given invariant rewards, matched sampled pairs, and norm-based clipping. We will verify the implementation before making a claim about the full refinement procedure.

5. FixSeq Training and Inference (Question 5)

FixSeq clamps Ht=Hk=H0H_t=H_k=H_0 during sampling and refinement and masks the reference reward at inference. Algorithm 2 lists joint noising and loss during training. The ZZ-only formulation would keep Ht=H0H_t=H_0 and minimize E[∣ϵ^Z−ϵZ∣F2]\mathbb{E}[\\|\widehat{\epsilon}_Z-\epsilon_Z\\|_F^2]. We will clarify which training objective was used for the reported checkpoint.

Thank you for the clarifications. After considering the responses, I am maintaining my rating of 1. The comparability of the baseline results and the specific benefit of state-dependent refinement remain insufficiently established.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.