acceptodds
Under review as a conference paper at ICLR 2027

Hybrid Diffusion–Schrödinger Dynamics with Unified Learning for Target-Conditioned Peptide Co-Design

Abstract

Target-conditioned peptide co-design is inherently challenging, as the generated peptides must exhibit biologically plausible sequences and structurally valid conformations. This difficulty is further amplified by the strong constraints imposed by the target protein, requiring physicochemical compatibility and geometric complementarity with the binding interface. Under these coupled constraints, existing methods often struggle to generate bindable peptides, resulting in limited binding success rates. In this paper, we propose PepHyD, a hybrid generative dynamics framework that formulates peptide co-design as a transition from broad sequence–structure exploration to target-directed transport. Specifically, PepHyD adapts the generative dynamics to the evolving state of generation, transitioning from diffusion-driven exploration of diverse peptide configurations to Schr\"odinger bridge transport that progressively concentrates intermediate representations toward target-compatible states. This design accommodates early-stage uncertainty while progressively enforcing target-specific constraints. Moreover, we introduce a unified target parameterization for both dynamics, allowing them to be learned within a shared denoising framework. Extensive experiments on LNR and BioDB demonstrate state-of-the-art performance, while evaluations on unseen targets and binding pockets further show strong generalization.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Target-conditioned peptide co-design, Sequence--structure co-design, Protein--peptide binding.

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_nH6vOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PepHyD, a hybrid generative framework that combines diffusion-based exploration with Schrödinger bridge transport for target-conditioned peptide co-design. The framework aims to progressively transition from diverse sequence–structure exploration to target-compatible peptide generation.

Critical strengths

Hybrid Generative Framework: Combining diffusion and Schrödinger bridge dynamics is an interesting approach to balancing structural exploration and target-specific refinement.

Critical weaknesses
  • Insufficient Motivation: The authors argue that diffusion enables broad exploration, while Schrödinger bridge dynamics guide generation toward target-compatible states. However, standard diffusion models already support coarse-to-fine generation, making the necessity of introducing bridge dynamics unclear. The improvements might simply result from changes in the noise schedule rather than the bridge formulation itself. The authors should provide controlled ablations to distinguish these effects. Furthermore, it remains unclear how the bridge dynamics explicitly enforce binding-interface constraints or offer advantages over conventional conditional diffusion.
  • Limited Methodological Novelty: If my understanding is correct, the most import contribution is the combination of diffusion and Schrödinger bridge. Both diffusion and Schrödinger bridge frameworks are well-established, and simply combining them without demonstrating clear advantages raises concerns about methodological novelty. Additionally, the authors present the importance of residue identity, physicochemical compatibility, and geometric complementarity as a contribution, although these are already well-established principles in molecular design. The authors should clarify their actual methodological contributions beyond combining existing approaches.
Questions for authors
  • What specific advantages do Schrödinger bridge dynamics provide over standard conditional diffusion? Could the improvements be explained by changes in the noise schedule?
  • How are binding-interface constraints explicitly incorporated into the bridge dynamics?
  • Figure 2 seems to have rendering problem. On chrome brower, if we zoom in to some extents, the figure will be corrupted completely.
Reviewer_dcqTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PepHyD, a framework for generating peptide sequences and all-atom structures conditioned on a target binding site. PepHyD first encodes peptides into a latent space using a geometric VAE, followed by a diffusion process that combines an initial diffusion stage and a Schrödinger bridge stage using the same denoising network. Results on BioDB and LNR show improvements in structural recovery and energy-based binding scores. The paper also includes ablation studies and a PTP1B case study.

Critical strengths
  1. Combining diffusion and Schrödinger bridge stages within a shared denoising network is an interesting design choice. The unified parameterization allows the two stages to be trained together.

  2. Experiments on BioDB and LNR show consistent improvements in RMSD and DockQ. Comparisons with pure diffusion, pure bridge, and different stage allocations also provide useful evidence for the benefit of the hybrid design.

Critical weaknesses
  1. The noise schedule is unclear. Section 3.2 and Algorithm 2 seem to suggest that increasing the diffusion ratio should retain more signal at the switching point under a fixed forward diffusion schedule, but Figure 2(a) shows the opposite trend. Could the authors clarify how the plotted signal relates to the boundary signal level �0��0� used in sampling?

  2. Equation 59 defines BNR as the fraction of candidates beating the native peptide, but Table 4 shows Rosetta BNR increasing from 27.97% to 46.15% as the sampling budget grows from 10 to 100. A larger candidate pool alone does not explain this increase in a candidate-level rate. This needs clarification.

Questions for authors
  1. Is BNR calculated over individual candidates, or as the fraction of targets with at least one candidate that beats the native peptide? Please clarify the numerator and denominator used for the reported results.

  2. Could the authors clarify how the boundary signal retention shown in Figure 2(a) relates to αˉK\bar{\alpha}_K in Section 3.2, and whether the diffusion noise schedule is rescaled when changing the diffusion ratio?

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

The paper addresses target-conditioned peptide co-design: given a protein target and a defined binding site, jointly generate a peptide sequence and its 3D structure that is physicochemically compatible and geometrically complementary to the interface. The authors argue that existing co-design models produce structurally plausible peptides that frequently fail to bind, and propose PepHyD, which splits the latent generative trajectory into two regimes. The early portion (Z_T → Z_K) uses diffusion for broad exploration; the later portion (Z_K → Z_0) uses a Schrödinger-bridge process that the authors describe as transporting samples toward target-compatible states. Both regimes are written in a common conditional Gaussian form (Eq. 20) and trained with a single shared denoising network via an EDM-style preconditioned target parameterization (Eq. 23–24, Appendix B.3); the default configuration is T = 300, ρ_D = 0.3 (90 diffusion steps / 210 bridge steps), p_B = 0.7. Evaluation covers two benchmarks (LNR, BioDB) against six co-design baselines, reporting best binding quality with success rates of 85.05% and 87.01% and the best DockQ, RMSD_Cα and BNR values (Table 1). The authors further report an average success rate of 86.7% across three computationally predicted PTP1B pockets not seen in training, exceeding UniMoMo by 28.4 points (Table 3), and 50 ns MD simulations indicating lower pose and internal RMSD and better native-contact retention than a UniMoMo complex (§4.2, Fig. 4). Ablations vary ρ_D, p_B and sample budget N.

Critical strengths
  • Matched internal controls for the central design choice. Pure Diffusion and Pure Bridge (Table 2) hold VAE, backbone, training data and candidate count fixed and vary only the dynamics. This is a stronger internal control than most papers in this subfield provide, and it is precisely what makes the additional control I request in W2 cheap to run.
  • Honest test of sampling-budget dependence (Table 4). The most common and most legitimate suspicion about generative binder-design results — that success rates are an artifact of generating many candidates and post-screening — is directly addressed: SR is already 87.83% at N = 10 and stays in the 86–88% band through N = 100. This is a reporting practice the subfield should adopt.
  • Three independent scoring functions reported for BNR (Rosetta, MM-GBSA, EvoEF2; Table 1). The interpretation is inadequate (W5), but because all three are reported, the inconsistency is discoverable by a reader. Reporting the unfavourable columns is a real virtue.
  • Leakage control inside BioDB. Receptor clustering at MMseqs2 40% identity with cluster-level separation of train/val/test (§4 Dataset, Table 6) is above the average care level for this task — with the LNR caveat in W3.
  • Large, non-marginal improvement in steric validity. Clash_in 0.83 / Clash_out 0.68 on BioDB versus PepBFN 43.98 / 15.05 and PepHAR 10.48 / 13.30 is an order-of-magnitude gap that is not explainable by run-to-run variation (it is, however, on par with UniMoMo at 1.46 / 1.15).
  • The unified-parameterization derivation is correct and transparently presented. Eq. 37–49 are, as far as I can verify, algebraically correct, and the preconditioning coefficients follow. The clarity of this derivation is what made W1 checkable at all.
Critical weaknesses

The defining mechanism, "Schrödinger bridge transport toward target-compatible states", is not instantiated anywhere in the method, and the paper's own derivation shows this (Appendix B.3, Eq. 37–42; §3.2 Eq. 14–20; Algorithm 2 lines 1–6). During training the bridge's paired endpoint Z_K is not external information: Eq. 37 defines it as a diffusion-noised version of Z_0, Z_K = √ᾱ_K Z_0 + √(1−ᾱ_K) ε_K. Substituting into the bridge posterior (Eq. 38) yields Eq. 40–41, i.e. Z_t^B = a_t^B Z_0 + √(V_t^B) ε — a single Gaussian forward process conditioned on Z_0 alone. The paper itself states in Eq. 20 that both regimes share the same conditional Gaussian form. What actually differs between the two regimes is therefore only (a) the coefficient schedules (a_t, V_t), (b) the reverse transition kernel (Eq. 22 vs Eq. 21), and (c) the training-time sampling weight p_B. Target information enters both regimes identically, through the input U^b of the same F_θ. This is the decisive break from the I²SB analogy invoked in §2.2: in I²SB, X_1 is a degraded observation supplied externally at inference and carries information; here Algorithm 2 lines 1–5 show Z_K is generated by the model itself from Z_T ~ N(0, I), so zero new information is injected at the bridge boundary. This threatens C1 directly and, through it, the Abstract, Introduction, Conclusion and contribution 2, all of which rest on this narrative. Minimal fix (two options, the first recommended and cheap): (a) narrow the claim — drop "target-directed Schrödinger transport" and re-describe the method as a two-segment noise schedule with regime-specific posteriors plus EDM-style unified preconditioning; or (b) instantiate the mechanism — inject target information at the bridge endpoint that the diffusion segment cannot access (e.g. a Z_K prior derived from the target interface) and show that this injection causes the gain. (b) is effectively a different paper, so (a) is the proportionate fix. Note this is a demonstrated flaw, not a missing experiment: it follows from the submitted equations.

The margin over the only competitive baseline is single-run and ~2 points, and a simpler alternative explanation is not excluded (Table 1; Table 2). Two parts. (2a) The evidence is currently unassessable. The only number supporting C1 is Pure Diffusion 84.83 → PepHyD 87.01 SR, i.e. +2.18 points. With 147 targets × 40 candidates, within-target candidates are strongly correlated, so effective sample size is closer to the number of targets. The ρ_D dependence is non-monotonic and peaks sharply at the adopted default (0.1 → 85.84, 0.3 → 87.01, 0.5 → 85.75, 0.7 → 80.49), which is compatible with ~1-point run-to-run variation. No seeds, variances or significance tests appear anywhere in the paper, so this is not a claim that the effect is noise — it is that the claim cannot currently be confirmed or refuted. (2b) A simpler explanation is open. In Table 1 the SR ranking is almost perfectly anti-correlated with Clash_out: 17.85 / 20.08 / 18.07 (PepFlow / PepHAR / PepBFN) → SR 31.42 / 27.20 / 34.32; 1.46 / 1.15 (UniMoMo / PepHyD) → SR 82.77 / 85.05. Since SR is defined as the fraction of candidates with Rosetta ΔG < 0 (Eq. 58, §C.4), SR plausibly measures in substantial part "does the pose avoid clashing at the interface". If so, "the bridge reverse kernel samples the late trajectory at lower effective noise, yielding smoother latents, fewer clashes, and hence higher SR/DockQ" explains the result with no appeal to target compatibility. Minimal fix: (i) rerun PepHyD, UniMoMo, Pure Diffusion and all of Table 2 with ≥3 seeds and report target-level paired tests (paired bootstrap or Wilcoxon signed-rank) on SR and DockQ; (ii) add a temperature-matched control: Pure Diffusion with the posterior variance of its last 210 steps lowered to match the bridge kernel. If that control is statistically indistinguishable from PepHyD, C1 is refuted; if it is distinguishable, C1 acquires its first real evidence. This requires no retraining — only a change to sampling — so the request is proportionate.

Questions for authors

Given Eq. 37, is there any point at which the bridge segment uses target information that the diffusion segment does not? As I read Eq. 37 → 40–41, Z_K is a diffusion-noised Z_0 at training time and a model-generated latent at inference (Algorithm 2 lines 1–5), so the bridge marginal reduces to a Z_0-conditional Gaussian and the target enters only via U^b in F_θ, identically in both regimes. If I have misread the derivation, please point to the specific equation; otherwise, would you agree to re-state the mechanism claim as a two-segment noise schedule with regime-specific posteriors? Doing the latter, together with Q2 and Q3, would move me toward a positive trajectory.

Can you report the temperature-matched control — Pure Diffusion whose final 210 sampling steps use the bridge kernel's posterior variance — and PepHyD, UniMoMo and Pure Diffusion over ≥3 seeds with target-level paired tests on SR and DockQ? This needs no retraining. If the paired test shows a significant PepHyD advantage over the temperature-matched control, that would be the first direct evidence for the two-regime design and I would raise my score. If it does not, the result should be reported as a sampling-temperature effect unrelated to Schrödinger bridges.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.