acceptodds
Desk rejected at ICLR 2027

What Limits Latent Diffusion for Receptor-Conditioned Peptide Design? From Latent-Space Quality to Interface Guidance

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Receptor-conditioned peptide design requires sequences and all-atom conformations that fit a target protein interface. Existing methods, however, often rely on noisy latent sampling or uniform diffusion schedules that do not distinguish interface-critical residues from flexible regions. We introduce PepMALD, a manifold-anchored latent diffusion framework for interface-guided peptide design. Its Manifold-Anchored Hierarchical Autoencoder maps designable residues to deterministic latent states that capture local geometry, multi-level interactions, and complex-level context. Latent representation alignment and interface-guided adaptive diffusion then regulate denoising according to receptor proximity. This spatially adaptive process stabilizes interface residues earlier while preserving flexibility away from the binding surface. A dedicated Post-2026 set is designed to evaluate generalization to newly released receptor-peptide complexes. Across multiple benchmarks, PepMALD generates candidates with improved physical plausibility and interface compatibility. Mechanistic analyses associate these gains with a stable latent manifold and interface-first denoising, clarifying how receptor geometry guides residue-level generation. A colorectal cancer case study further indicates its potential for disease-relevant peptide design.

Primary area:
generative models
Keywords:
Peptide Design, Latent Diffusion, Geometric Representation Learning, Interface-Guided Generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_bKRDOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes PepMALD, a structure-based peptide design framework built upon latent diffusion. The authors identify limitations in both the latent space learned by conventional VAEs and the subsequent diffusion process, and introduce three methodological improvements. First, they propose a manifold-anchored hierarchical autoencoder that aggregates multi-level representations and global context, while replacing conventional KL regularization with L2 penalties on the E(3)-invariant sequence latent representations and the deviations of E(3)-equivariant structural latents from their corresponding Cα coordinates. Second, they introduce a latent representation alignment loss that encourages intermediate denoiser representations to recover the clean sequence latents. Third, they develop an interface-guided adaptive diffusion scheme that adjusts residue-specific noise schedules according to receptor proximity, allowing interface-proximal residues to stabilize earlier during generation. The authors evaluate PepMALD on sequence-structure co-design and binding conformation generation using two benchmarks, PepBench and PepBDB. PepMALD outperforms the evaluated baselines on several metrics, with particularly notable improvements in binding energy and structural accuracy on PepBench, although the gains on PepBDB are more modest.

Critical strengths

The most notable strength of this paper lies in its comprehensive experimental evaluation and analysis. Beyond conventional benchmarks covering both de novo peptide design and binding conformation generation across two datasets, the authors construct an additional post-cutoff test set to assess generalization to newly released complexes. The paper also presents extensive analyses, including independent validation using cofolding models, molecular dynamics simulations, latent space characterization, and investigations of intermediate denoiser representations and diffusion dynamics. Together, these experiments provide a broad assessment of the proposed method's effectiveness and offer insights into how its individual components contribute to the observed improvements.

Critical weaknesses

The most significant weakness is that some methodological choices lack convincing conceptual justification. First, it is unclear why conventional KL regularization is replaced with simple L2 penalties. KL regularization encourages a continuous latent distribution compatible with a Gaussian prior, whereas L2 regularization merely constrains latent magnitudes without explicitly promoting distributional continuity or smooth interpolation. This raises concerns about the suitability of the resulting latent space for diffusion-based generation. Second, the motivation for aligning intermediate denoiser representations with clean latents is also unclear. Since the standard diffusion noise-prediction objective already implicitly supervises the recovery of clean latents through the denoiser's output, the necessity and additional benefits of imposing such supervision on intermediate hidden representations remain insufficiently justified.

Questions for authors
  1. What is the definition of the RMSD metric in Table 1? Is it calculated with the native peptide structure, or with the boltz/chai-predicted structure? Is it calculated on CαC_\alpha, or on all atoms?
  2. How does IGAD address the training–sampling discrepancy caused by using clean latents during training but noisy latents to define the distance during inference?
  3. How does the proposed “manifold anchoring” differ from replacing KL with conventional L2 regularization?
  4. How sensitive is IGAD to the interface distance threshold and the choice of adaptive noise schedule?
  5. Does IGAD actually cause interface residues to stabilize earlier during generation? Could the authors compare the denoising trajectories of interface residues with and without IGAD?
4 comments

1. What is the definition of the RMSD metric in Table 1?

Thank you for asking. RMSD (Å) compares generated peptide CαC_\alpha coordinates with those in the native test-set receptor–peptide complex. It uses CαC_\alpha atoms only and the native reference, not Boltz/Chai predictions: RMSD=1N∑i=1N∣xigen−xinative∣22\mathrm{RMSD}=\sqrt{\frac{1}{N}\sum_{i=1}^{N}\\|x_i^{\mathrm{gen}}-x_i^{\mathrm{native}}\\|_2^2} where xix_i denotes corresponding CαC_\alpha coordinates and NN counts valid residue pairs.

We will add this definition to the evaluation-metrics description accompanying Table 1 in the main text, explicitly specifying the native reference and the use of CαC_{\alpha} atoms.

2. How does IGAD address the training–sampling discrepancy caused by using clean latents during training but noisy latents to define the distance during inference?

Thank you for highlighting this. Training defines forward noising from clean structural latents; the denoiser receives noisy latents and local time. Sampling recomputes scores from current noisy latents and the fixed receptor each reverse step. Gaussian distance scores χit\chi_i^t continuously set Kit=1−χit(1−Kmin⁡)K_i^t=1-\chi_i^t(1-K_{\min}), with Kmin⁡=0.5K_{\min}=0.5 and Kit∈[0.5,1]K_i^t\in[0.5,1]. The same current exponent determines local time, signal coefficients and noise variance, with scores reassessed each step.

We forward-corrupted clean latents for 93 complexes, repeated three times, and compared noisy versus clean score rankings by within-complex Spearman correlation, averaging repetitions before summarizing complexes.

Table 1

ttSpearman correlation
900.002 ± 0.257
700.119 ± 0.283
500.282 ± 0.254
300.475 ± 0.302
100.793 ± 0.181

Agreement increases as corruption decreases, supporting dynamic score/schedule updates without requiring clean-reference agreement at every step.

We will clarify in §3.3 how clean anchors determine forward noising during training and how dynamically recomputed anchors jointly determine local time conditioning and reverse-update coefficients during sampling.

3. Why replace KL regularization with L2, and how does “manifold anchoring” differ from conventional L2 regularization?

Thank you for this question. MHAE learns deterministic reconstruction codes; conditional diffusion learns their distribution. A Gaussian diffusion endpoint does not require Gaussian clean codes. For q=N(μ,diag⁡(σ2))q=\mathcal{N}(\mu,\operatorname{diag}(\sigma^2)) and p=N(m,I)p=\mathcal{N}(m,I), DKL(q∣p)=12∣μ−m∣22+12∑j(σj2−log⁡σj2−1)D_{\mathrm{KL}}(q\\|p)=\frac12\\|\mu-m\\|_2^2+\frac12\sum_j(\sigma_j^2-\log\sigma_j^2-1) With fixed covariance, the variance term is constant, leaving a quadratic mean penalty. This motivates deterministic encoding with magnitude/offset control, leaving distribution modeling to diffusion.

VAE/MHAE checkpoints were evaluated on 93 LNR complexes; H/Z denote perturbation strengths:

Table 2

ConditionMetricVAEMHAE
CleanCαC_\alpha RMSD (Å)0.5470.376
CleanBackbone RMSD (Å)1.1441.039
H=0.400Sequence retention (%)78.7392.92
H=0.400Sequence JS (nats)0.1450.046
H=0.400, Z=2.000 ÅSequence retention (%)78.5792.95

Retention uses unperturbed decoding as reference. MHAE improves clean reconstruction and sequence consistency.

Anchoring specifies geometric references: Lanch=12N∑i(λH∣ui∣22+λZ∣zi−ai∣22)\mathcal{L}_{\mathrm{anch}}=\frac{1}{2N}\sum_i(\lambda_H\\|u_i\\|_2^2+\lambda_Z\\|z_i-a_i\\|_2^2) ui,ziu_i,z_i are sequence/structural latents; aia_i is native CαC_\alpha. H uses zero-centered L2; Z uses a residue-specific reference, invariant to joint rotation/translation and allowing learned offsets.

We trained no-Z and backbone-center controls with the same MHAE architecture and H regularization. The center averages valid N/CαC_\alpha/C/O coordinates; original CαC_\alpha-anchored MHAE is the reference.

Table 3

Z anchoringCαC_\alpha RMSDBackbone RMSDBond MAEClashes
None0.6641.8090.52940.60
Backbone center0.5591.3250.2276.16
CαC_\alpha (PepMALD)0.3761.0390.1262.08

RMSD/MAE use Å; MAE uses ideal backbone bond lengths. Clashes are pairs <1.2 Å apart, excluding one/two-bond neighbors. Backbone-center versus no-Z improvements support residue-level geometric anchoring beyond H magnitude control.

We will clarify the motivation for deterministic regularization and residue-specific geometric anchoring in §3.1, and extend Appendix C.3 with the Gaussian KL decomposition and the reconstruction, perturbation, and anchoring-control results reported above.

4. How sensitive is IGAD to the interface distance threshold and the choice of adaptive noise schedule?

Thank you for asking about sensitivity. IGAD uses a continuous Gaussian score, χi=exp⁡[−di2/(2σa2)]\chi_i=\exp[-d_i^2/(2\sigma_a^2)], with distance did_i and width σa\sigma_a in normalized pocket coordinates. We varied the default width 1.00 by ±25% and replaced linear Ki=1−0.5χiK_i=1-0.5\chi_i with quadratic Ki=1−0.5χi2K_i=1-0.5\chi_i^2, both bounded in [0.5,1].

Using the same PepMALD weights, initial states, random perturbations and 100 sampling steps, we generated one sample per setting for each of 93 complexes.

Table 4

σa\sigma_aMappingCαC_\alpha RMSD (Å) ↓DockQ ↑
1.00Linear (default)8.555 ± 6.1270.413 ± 0.230
0.75Linear8.256 ± 6.2460.431 ± 0.235
1.25Linear8.206 ± 6.1180.432 ± 0.232
1.00Quadratic8.687 ± 6.3810.414 ± 0.236

Reducing/increasing width lowered RMSD by 3.5%/4.1% and increased DockQ by 0.018/0.019. Quadratic scheduling increased RMSD by 1.5% with little DockQ change. These results show that IGAD maintains similar generation performance across the tested distance parameters and schedule mappings, indicating robustness to these parameter choices.

5. Why is LRA needed in addition to noise prediction?

Thank you for requesting this distinction. Noise prediction supervises the output used to recover clean latents from noisy inputs and diffusion coefficients. LRA directly supervises hidden states to standardized sequence latents from frozen MHAE: LLRA=1∣Vp∣∑i∈Vp∣Wr(Rθ,it,ℓr)−u~i∣22\mathcal{L}_{\mathrm{LRA}}=\frac{1}{|V_p|}\sum_{i\in V_p}\\|W_r(R_{\theta,i}^{t,\ell_r})-\tilde{u}_i\\|_2^2 where VpV_p contains peptide residues and RR is the last intermediate representation. The clean reference is fixed across timesteps, encouraging consistently accessible clean-latent information.

The w/o LRA ablation (Fig. 13) retains noise prediction and evaluates this added supervision. Table 5 summarizes linear-readout results from frozen denoiser representations (Fig. 4A and Appendix H). The coefficient of determination, R2R^2, measures clean-latent recovery relative to mean prediction: 1 indicates perfect recovery, 0 matches the mean baseline, and negative values indicate worse performance.

Table 5

ttPepMALD R2R^2w/o LRA R2R^2
900.1190.140
700.2970.281
500.5580.551
300.8510.818
100.9780.951

PepMALD has higher recoverability at most evaluated timestep. Together with the w/o LRA ablation, this supports complementary representation supervision.

6. Does IGAD cause interface residues to stabilize earlier during generation?

Thank you for suggesting this comparison. We compared PepMALD with the w/o IGAD model on 93 test complexes using three runs per model, with matched receptor conditions, initial states, random perturbations and 100 sampling steps.

Interface residues were defined by a native minimum peptide–receptor CαC_{\alpha} distance below 8 Å, with grouping fixed throughout. The 92 complexes containing interface residues were included. Table 6 reports the distance to each trajectory's own final state and displacement over five steps, measured using structural latent coordinates in Å. Runs were averaged within each complex before computing the mean and standard deviation across complexes.

Table 6

ProgressDistance, w/o IGADDistance, PepMALDDisplacement, w/o IGADDisplacement, PepMALD
20%10.022 ± 2.3438.726 ± 2.1347.231 ± 1.8105.485 ± 1.195
40%7.326 ± 1.6176.499 ± 1.4424.594 ± 1.1903.624 ± 0.853
60%5.235 ± 1.3104.653 ± 0.9783.311 ± 0.7672.647 ± 0.590
80%2.834 ± 0.6592.482 ± 0.4962.329 ± 0.5901.966 ± 0.420
90%1.542 ± 0.3811.328 ± 0.3041.691 ± 0.3981.436 ± 0.292

At 20% progress, PepMALD reduced the distance to its own final state from 10.022 to 8.726 Å and five-step displacement from 7.231 to 5.485 Å relative to w/o IGAD. The same pattern held at every reported stage: interface residues were closer to their respective final states and underwent smaller positional changes, supporting earlier stabilization in these trajectories.

We will add this trajectory comparison to §4.4 after the IGAD coefficient analysis.

Reviewer_bKRDOpenReview ↗

Thanks

Thanks for the detailed responses, which should help the readers better understand the paper. I decided to maintain the current score of weak acceptance.

We sincerely thank you for your time, thoughtful feedback, and valuable guidance throughout the review process.

Reviewer_nvDaOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces PepMALD for receptor-conditioned peptide sequence–structure co-design. It combines a deterministic hierarchical autoencoder with latent representation alignment and an interface-dependent diffusion schedule, aiming to better capture peptide representations and receptor geometry during generation. The method is evaluated on standard benchmarks and a temporally filtered test set, with additional ablation studies, independent structure prediction, and physical refinement experiments.

Critical strengths
  1. PepMALD combines sequence–structure representation learning with receptor-aware denoising in a sensible way. The ablation results support the contributions of its three core components, with improvements across different aspects of generation quality.

  2. PepMALD shows clear improvements over the baselines on PepBench, particularly in success rate and RMSD. The independent structure prediction and Rosetta relaxation experiments provide useful additional checks beyond the main benchmark evaluation. Overall, I think the experiments are quite thorough.

Critical weaknesses

The main text states that complete experimental settings are provided in Appendix F, but several method-specific hyperparameters, including the IGAD settings and LRA loss weight, are not specified in the manuscript. Providing these details would improve reproducibility.

Questions for authors
  1. At inference time, are the binding site and peptide length provided? How is the receptor region selected in the experiments, and does this require information from the native peptide?

  2. Is diversity calculated over all generated candidates or only successful ones? Given the lower diversity of PepMALD on PepBench, does this difference still hold when considering only successful designs under the same sampling budget?

3 comments

1. What are the key hyperparameters for PepMALD?

We thank the reviewer for highlighting the need for these details. The key training and model hyperparameters, including the loss weights and the settings for MHAE, LRA, and IGAD, are summarized below.

Table 1. Key hyperparameters of PepMALD.

ComponentHyperparameterValue
OptimizationOptimizerAdamW
OptimizationLearning rate1×10−41\times10^{-4}
OptimizationWeight decay0.01
MHAEEmbedding / hidden dimension128 / 128
MHAESequence latent dimension dHd_H8
MHAESequence masking ratio ρ\rho0.25
MHAEReconstruction balance η\eta0.5
MHAESequence latent regularization coefficient βH\beta_H1×10−41\times10^{-4}
MHAEStructural latent regularization coefficient βZ\beta_Z1×10−31\times10^{-3}
MHAEFull-atom coordinate reconstruction weight λX\lambda_X1.0
MHAECαC_\alpha coordinate reconstruction weight λCα\lambda_{C_\alpha}1.0
MHAEBackbone bond-length reconstruction weight λbb\lambda_{\mathrm{bb}}1.0
MHAESide-chain bond-length reconstruction weight λsc\lambda_{\mathrm{sc}}1.0
MHAESide-chain torsion reconstruction weight λQ\lambda_{\mathcal{Q}}0.5
DiffusionDenoiser hidden dimension128
DiffusionDiffusion and reverse sampling steps TT100
DiffusionSequence noise prediction loss weight ωH\omega_H0.375
LRAAlignment loss weight λR\lambda_R0.1
LRALayer used for alignment ℓr\ell_rFinal denoiser hidden layer
LRAProjection dimension dHd_H8
IGADMinimum schedule exponent Kmin⁡\mathcal{K}_{\min}0.5
IGADGaussian distance scale σa\sigma_a1.0 in normalized coordinates

2. Are the binding site and peptide length provided at inference time, and does receptor region selection use native peptide information?

Yes. Both the receptor binding region and peptide length are provided at inference time in our experiments.

The receptor region is defined from the native complex by selecting receptor residues whose CβC_\beta atoms are within 10 Å of any native peptide CβC_\beta atom. For residues without a CβC_\beta atom, a virtual CβC_\beta is constructed from the backbone geometry. The peptide length is set to the number of residues in the native peptide. Thus, pocket selection and length specification use native peptide information in the benchmark.

During generation, peptide sequence identities are masked, and the sequence and structure latent variables are initialized from Gaussian noise. Coordinate normalization uses receptor geometry, and IGAD updates its interface-dependent schedule using the current generated structural latent state.

For practical design, users can specify the receptor pocket and a desired peptide length or length range. Our evaluation therefore concerns sequence–structure co-design conditioned on a specified binding region and peptide length.

3. Is diversity calculated over all candidates, and does PepMALD’s lower diversity persist among successful designs under the same sampling budget?

The diversity reported in Table 1 of the manuscript is calculated over all generated candidates.

To address the reviewer’s question, we selected successful candidates using ΔG<0\Delta G<0 and recalculated diversity using the original metrics under the same generation budget.

Table 2. Diversity among successful candidates under the same generation budget.

ModelSuccessful candidatesSuccess rateSequence diversityStructural diversityJoint diversity
PepMALD2,27361.10%0.96770.40870.5780
PepGLAD1,75347.12%0.96580.56450.6969
PepMimic1,94552.28%0.96880.50400.6550
PepTri1,79548.25%0.96530.58660.7010

Diversity is calculated among successful candidates for each target and then averaged across targets with at least two successful candidates. The corresponding numbers of targets are 91, 90, 88, and 87, respectively. Successful-candidate counts and success rates include all generated results.

PepMALD’s mean joint diversity remains lower than that of the three baselines when considering only successful candidates. Sequence diversity is similar across the four models, while the main difference lies in structural diversity. PepMALD achieves a higher success rate, with a more concentrated structural distribution among its successful candidates.

We will add the hyperparameter table to Appendix F and clarify in the experimental-setup paragraph of §4.1 that benchmark inference is conditioned on a receptor binding region and peptide length derived from the native complex.

Thank you for the detailed clarifications and additional experiments. These responses have addressed most of my concerns. I will maintain my original score of 4.

Thank you for taking the time to consider our responses and for your thoughtful feedback throughout the review process.

Reviewer_dtwpOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper asks what limits the generative quality of latent diffusion for receptor-conditioned peptide design, and identifies three bottlenecks (§1): (i) latent space quality — VAE posterior sampling yields collapsed representations with low effective dimensionality, and residue-level latents do not retain receptor-interface information in decodable form; (ii) encoder–denoiser misalignment — a denoiser trained purely on noise prediction follows reverse trajectories that drift off the encoder's latent manifold; (iii) a uniform diffusion schedule — applying the same noise level to every residue ignores interface heterogeneity and prevents early stabilization of key interfacial residues.

The authors propose PepMALD, with one module per bottleneck. MHAE (§3.1) replaces posterior sampling with a deterministic projection of each residue onto a fixed point of the latent manifold, aggregating multi-layer hidden states with complex-level global context under an anchoring regularizer (Eq. 3). LRA (§3.2) explicitly aligns intermediate denoiser representations to a normalized sequence-latent reference (Eq. 5). IGAD (§3.3) maps an interface anchoring score to a residue-level schedule exponent, giving ᾱ^IGAD_{t,i} = (ᾱ_t)^{K_i}. They additionally construct a Post-2026 temporal holdout (410 complexes from 189 PDB entries, receptor identity < 0.40).

Reported evidence covers PepBench and PepBDB (Table 1), the Post-2026 set (Fig. 2B), latent-space diagnostics (§4.2, App. G), reverse-dynamics analyses (§4.3, App. H–I), a seven-way ablation (Fig. 13), external validation with Chai-1, Boltz, Rosetta relax and restrained MD (App. O–P), and an APC–β-catenin case study (§4.4).

Critical strengths
  • Latent space is measured, not asserted (§4.2, App. G). The paper probes the latent space directly rather than inferring its quality from downstream generation: frozen-latent low-capacity linear probes with coordinate and receptor-structure access explicitly blocked (App. G.4) give interface-contact AUROC 0.64 → 0.78 and receptor-distance R² 0.02 → 0.22; active dimension fraction 0.5 → 1.0, effective rank 3.49 → 7.19, MMD-to-prior 0.042 → 0.022. Crucially, Fig. 3B and Figs. 7A/7C/7D measure these quantities on the actual baselines' latents, not only against an in-house VAE (participation ratio 0.856 vs. the VAE-based baselines). I raised a strawman-baseline concern on first reading and withdrew it after Appendix G — the cross-model measurement is already there.
  • External validation is designed to be independent of the model's own scoring (App. O–P). Two independent structure predictors (Chai-1, Boltz), Rosetta relax, and OpenMM restrained MD are all run, and App. O frames the question as reproduction of the design mode rather than of the native pose — the methodologically correct question for a design model, and one that avoids the self-scoring circularity common in this subfield. Reporting Spearman 0.714/0.720 between the two predictors is a proper evaluator-bias control.
  • Leakage control is concrete and adequate (App. F.1–F.2). MMseqs2 40% receptor clustering, cluster-level splitting, removal of LNR receptors falling in the same cluster, and peptide/receptor redundancy filtering. The Post-2026 set applies the same receptor-identity < 0.40 filter. Baselines are reproduced under a unified data-processing pipeline.
Critical weaknesses

The title's diagnostic claim is never tested. Location: title; §1 contribution list; §4.2–4.3; Fig. 13. The paper declares three limiting factors and then shows that a model which removes them performs better. But "latent space quality limits generation quality" is a statement about a mediating relationship: it requires varying latent quality and showing generation quality tracks it. The present evidence cannot distinguish that claim from the far weaker "MHAE is simply a better autoencoder". §4.2–4.3 are design verification — do my modules behave as designed — not causal diagnosis; these are different experiments. The ablation (Fig. 13) shows only that each module contributes, and even that rests on an arbitrary composite score (see W6). Status: demonstrated gap — the experiment is absent, not merely weak. Minimal fix, either: (a) build 3–4 latent spaces of deliberately different quality (vary βH / βZ / anchoring strength or latent width), train the same denoiser on each, and plot (effective rank, interface AUROC) against (ΔG, DockQ, Success) as a dose–response — every measurement tool is already in the paper, so the cost is 3–4 retrainings; or (b) reframe the title and abstract as a method paper (e.g. "PepMALD: Manifold-Anchored Latent Diffusion with Interface-Guided Schedules") and explicitly rename §4.2–4.3 as design verification. Either fully resolves it, which is why this is not unconditionally Critical.

IGAD's stated mechanism is at chance level on held-out cases, yet is asserted as established. Location: App. I, lines 1493–1505. On the independent test set, positive high–low effective-α gaps occur in 52.7% of cases and positive anchor–α correlations in 53.8% (t = 70); the training-distribution figures are 54.4%/53.5%. The trajectory-error criterion is better but also drops out of sample (78.9%/76.3% → 65.6%/66.7%). Even the showcased case 3WBN gives mean effective α of 0.379 vs. 0.348 for high- vs. low-anchor residues — a gap of 0.031. Yet the abstract and §4.3 assert that IGAD "modulates local transition kernels according to interface anchoring" and produces "interface-first denoising". Three problems compound: (i) the monotone trend in Fig. 4B may be an artifact of aggregating across cases when roughly half of individual cases point the other way; (ii) no null rate is given — interface residues are intrinsically more constrained by the receptor condition, so any receptor-conditioned model may exceed 50% on trajectory separation; Fig. 9D compares trajectories against baselines, but the percentage analysis of Fig. 10 is PepMALD-only; (iii) there is an internal tension — §4.3 states the spatial contrast is strongest at early sampling stages, but at t → T the latent z̃^t entering the anchor score (Eq. 6, s = t) is nearly pure N(0, I) noise, i.e. the mechanism is claimed to be strongest where its input signal is least trustworthy. The authors partly acknowledge this by limiting Prop. 3 to a conditional per-score-update guarantee. Status: demonstrated inconsistency between the authors' own reported numbers and their conclusion statements — not a speculative concern. This matters because C4-b carries one of the three contributions; if the mechanism is at chance, IGAD's measured contribution (largest Success drop in the w/o-IGAD ablation) is likely operating through a different route, e.g. plain regularization from reduced effective noise on interface residues. Minimal fix: (i) compute the four percentages of Fig. 10 for PepGLAD/PepMimic/PepTri under the identical protocol to establish the null rate; (ii) binomial-test 52.7%/53.8% against 50%; (iii) revise §4.3 and the abstract to "observed at the aggregate level, with limited case-level consistency"; (iv) report the rank correlation between χ^t and χ^0 as a function of t.

Questions for authors

Can you provide a dose–response between latent quality and generation quality — e.g. 3–4 latent spaces produced by varying β_H / β_Z / anchoring strength or latent width, each with the same denoiser trained on top, plotting effective rank and interface AUROC against ΔG / DockQ / Success? If this is infeasible in the rebuttal window, would you instead commit to the reframing in W1(b) — a method-paper title and abstract, with §4.2–4.3 relabeled as design verification? Either answer would move my score up; (a) more than (b).

What are the four percentages of Fig. 10 for PepGLAD, PepMimic and PepTri under the identical protocol? Without this null rate, I cannot tell whether 65.6% reflects IGAD or reflects the fact that interface residues are more constrained in any receptor-conditioned model. Relatedly, are 52.7% and 53.8% distinguishable from 50% under a binomial test at your sample size?

4 comments

1. Dose–response between latent quality and generation quality

Thank you for the suggestion. Following option (a), we conducted an additional experiment relating latent-space quality to generation quality. Using the same data, architecture, and training protocol, we trained three autoencoders with βH=10−5,10−3,10−2\beta_H = 10^{-5}, 10^{-3}, 10^{-2} and separately trained identically configured denoisers on their latent spaces. We measured the effective rank of standardized HH, interface AUROC, and reconstruction RMSD on 93 LNR test targets. For each newly trained model, we generated 40 candidates per target on LNR.

Table 1

Model / settingEffective rank of standardized HH ↑Interface AUROC ↑Reconstruction CαC_\alpha RMSD (Å) ↓ΔG\Delta G ↓DockQ ↑Success ↑
βH=10−5\beta_H = 10^{-5}6.4800.7320.426-16.050.66354.78%
βH=10−4\beta_H = 10^{-4} (PepMALD)7.4770.7730.376-22.720.68961.10%
βH=10−3\beta_H = 10^{-3}7.3940.7410.428-19.170.67856.05%
βH=10−2\beta_H = 10^{-2}7.2100.7230.436-17.590.66955.38%

As shown in Table 1, among the three newly trained models, effective rank increases in the order 10−510^{-5}, 10−210^{-2}, 10−310^{-3}, with corresponding improvements in ΔG\Delta G, DockQ, and Success. For example, changing βH\beta_H from 10−510^{-5} to 10−310^{-3} increases effective rank from 6.480 to 7.394, while ΔG\Delta G improves from -16.050 to -19.174, DockQ from 0.663 to 0.678, and Success from 54.78% to 56.05%. Interface AUROC shows a weaker relationship with generation performance; effective rank provides the more consistent correspondence.

The βH=10−5\beta_H=10^{-5} and βH=10−3\beta_H=10^{-3} models have reconstruction CαC_\alpha RMSDs of 0.426 Å and 0.428 Å, respectively. The latter improves effective rank and all three generation metrics without lowering reconstruction RMSD.

We will add a new subsection immediately after §4.2 to present this dose–response experiment and discuss the observed relationship between latent-space quality and generation performance across the tested settings.

2. Compute the four percentages of Fig. 10 for PepGLAD/PepMimic/PepTri under the identical protocol to establish the null rate

Thank you for highlighting the missing baseline comparison. We computed the four Fig. 10 percentages under the same protocol: cases with positive low-minus-high error AUC (PAP_A), positive low-minus-high error at t=70t=70 (PEP_E), positive high-minus-low effective-α\alpha (PαP_\alpha), and positive reference-anchor–effective-α\alpha Pearson correlation (PrP_r) at t=70t=70.

Table 2

DatasetModelPAP_APEP_EPαP_\alphaPrP_r
TrainingPepMALD75.54%71.40%53.12%52.52%
PepGLAD75.51%71.37%0.00%—
PepMimic68.37%68.46%0.00%—
PepTri72.91%69.57%0.00%—
Training-distributionPepMALD82.46%78.95%53.51%48.25%
PepGLAD83.33%78.07%0.00%—
PepMimic85.09%76.32%0.00%—
PepTri80.70%76.32%0.00%—
IndependentPepMALD65.59%65.59%52.69%54.84%
PepGLAD66.67%64.52%0.00%—
PepMimic68.82%61.29%0.00%—
PepTri68.82%64.52%0.00%—

The training-distribution set comprises validation complexes. Each percentage uses the metric’s valid cases. Uniform schedules yield zero α\alpha gaps and undefined correlations (—).

Table 2 shows that the baselines also exhibit trajectory separation: their independent-set PAP_A values range from 66.67% to 68.82%, compared with 65.59% for PepMALD. We therefore agree that the original PepMALD-only percentages cannot attribute this separation to IGAD. These percentages also capture only the direction of the difference, not how closely high-anchor residues approach the reference latent state.

We compared PepMALD, w/o IGAD, and the three baselines on 93 test targets using the original high-/low-anchor grouping and reference-latent error definitions. Table 3 reports error AUC over sampling, error at t=70t=70, and endpoint error.

Table 3

ModelHigh-anchor AUCLow-anchor AUCHigh-anchor error, t=70t=70High-anchor endpoint error
PepMALD8.821 ± 0.22310.480 ± 0.1868.918 ± 0.2827.808 ± 0.178
w/o IGAD9.510 ± 0.18010.806 ± 0.0269.641 ± 0.3308.019 ± 0.266
PepGLAD10.166 ± 0.16912.120 ± 0.11010.973 ± 0.0988.010 ± 0.270
PepMimic10.751 ± 0.14012.675 ± 0.19711.143 ± 0.0449.341 ± 0.444
PepTri10.003 ± 0.13211.410 ± 0.08610.681 ± 0.1648.174 ± 0.038

In Table 3, PepMALD’s high-anchor error AUC is lower than its low-anchor AUC (8.821 versus 10.480), indicating greater average proximity to the reference latent state during sampling. Compared with w/o IGAD, PepMALD has lower high-anchor AUC (8.821 versus 9.510), error at t=70t=70 (8.918 versus 9.641), and endpoint error (7.808 versus 8.019). All three high-anchor metrics are also lower than those of the three baselines. The reported trajectories support the aggregate interface-first denoising pattern defined by reference-latent trajectory errors in the manuscript.

We will update Fig. 10A in Appendix I with the comparisons in Table 2 and revise the accompanying text to clarify that these percentages alone cannot attribute trajectory separation to IGAD. We will add Table 3 and its analysis to Appendix I to support the observed aggregate interface-first denoising pattern.

3. Binomial tests and aggregate-level interpretation of IGAD

Thank you for this suggestion. We performed two-sided exact binomial tests against 50% using the original 93 independent cases at t=70t=70 (Table 4). Neither proportion differed significantly from 50%. These results do not establish consistent correspondence between reference anchors and effective-α\alpha across individual cases.

Table 4

MetricPositive casesPercentagep-value
Positive high-minus-low effective-α\alpha49/9352.688%0.679
Positive reference-anchor–effective-α\alpha correlation51/9354.839%0.407

Table 3 supports the pattern at the level of average trajectories: PepMALD has lower high-anchor than low-anchor AUC and lower high-anchor AUC and error at t=70t=70 than w/o IGAD. These observations support the aggregate interface-first denoising pattern defined by reference-latent errors.

We therefore accept the requested qualification: interface-prioritized error reduction is observed at the aggregate level, with limited case-level consistency. The abstract will qualify the statements about earlier interface stabilization and interface-first denoising accordingly. Section 4.4 will distinguish the conditional modulation of local diffusion coefficients by current anchor scores from the observed average trajectory behavior.

4. Rank correlation between χt\chi^t and χ0\chi^0 as a function of tt

Thank you for this suggestion. On 93 test targets, we computed within-complex Spearman correlations between dynamic anchors χt\chi^t, calculated from the current generated state, and clean reference anchors χ0\chi^0, calculated from the reference complex and held fixed throughout sampling.

Table 5

Diffusion time ttSpearman correlation
1000.003 ± 0.035
900.013 ± 0.048
700.037 ± 0.028
500.023 ± 0.040
300.069 ± 0.060
100.084 ± 0.025
00.100 ± 0.037

Sampling proceeds from t=100t=100 to t=0t=0. The correlations are near zero at early sampling stages and show a modest increase at later stages, although they remain weak and the increase is not monotonic.

The original early spatial contrast statement was based on the fixed-reference-anchor coefficient analysis in Fig. 4B: high–low coefficient gaps were 0.132, 0.151, and 0.029 at t=90t=90, 70, and 10, respectively. These measure conditional coefficient contrasts, whereas Table 5 measures dynamic-to-reference rank agreement. Larger early-to-intermediate conditional contrasts can therefore coexist with weak rank correspondence.

We also analyzed PepMALD’s complete trajectories using the manuscript’s high-/low-anchor groups and reference-latent errors. Table 6 reports the time-averaged low-minus-high error within three equal sampling-progress intervals.

Table 6

Stage (sampling progress)Mean low-minus-high reference error
Early (0–33.3%)2.009 ± 0.108
Middle (33.3–66.7%)1.719 ± 0.046
Late (66.7–100%)1.197 ± 0.084

The error separation is positive throughout these stages and largest in the early-stage average, supporting a more pronounced average high-/low-anchor trajectory separation early in sampling. We will clarify in §4.4 that the early spatial contrast refers to conditional coefficient differences and observed average trajectory separation, while explicitly acknowledging the limited correspondence between dynamic and clean reference anchor rankings.

Dear Reviewer dtwp,

We hope this message finds you well. Thank you again for your suggestions. We followed option (a), which you indicated would be most informative, and trained identically configured denoisers on three additional latent spaces. Across these settings, higher effective rank was associated with lower ΔG\Delta G, higher DockQ, and higher Success, including for two settings with nearly identical reconstruction errors (§4.3, Table 2).

For IGAD, we added the four Fig. 10 statistics for all three baselines, performed the requested binomial tests, and reported how the rank correlation between dynamic and reference anchors changes during sampling. We also clarified the aggregate interpretation of interface-first denoising in the revised manuscript.

When you have a chance, we would appreciate hearing whether these additions address your two concerns. Thank you for your time and thoughtful review.

Sincerely,

The Authors

Comments on the paper

1 comment

We thank all three reviewers for their careful reading and constructive feedback, and the chairs and committee members for their time and effort throughout the review process.

We present the first study of how latent-space quality affects the diffusion process in receptor-conditioned peptide design. PepMALD improves latent-space quality with a manifold-anchored autoencoder and adapts diffusion to receptor-interface geometry. In response to the reviewers’ questions, we added experiments to examine how latent quality relates to generation quality and how interface residues behave during denoising. We also clarified the rationale for our design choices and the evaluation conditions. We summarize the main findings and revisions below.

1. Latent quality and generation quality. We varied βH\beta_H and trained denoisers with identical configurations on three additional latent spaces. Across these settings, higher effective rank was associated with lower ΔG\Delta G, higher DockQ, and higher Success on 93 LNR targets. The same trend was observed in two settings with nearly identical reconstruction RMSDs (0.426 and 0.428 Å). Section 4.3 now reports these comparisons, including interface AUROC (Table 2).

2. Interface-first denoising. Figure 10 now includes all four statistics for PepGLAD, PepMimic, and PepTri, providing a baseline for interpreting trajectory separation. We tested the two coefficient-based proportions against 50% and reported the rank correlation between dynamic and reference anchors across sampling. We characterize the aggregate interface-first pattern through the lower average reference-latent errors of high-anchor residues (§4.4 and Appendix I).

3. Interface-residue stabilization and schedule sensitivity. We compared PepMALD with w/o IGAD on the 92 test complexes containing interface residues. At each reported stage, PepMALD’s interface residues remained closer to their final states and changed less over the five sampling steps (§4.4, Table 3). The revised manuscript also explains how anchor scores are updated during sampling and reports sensitivity analyses for the distance scale and schedule mapping (response to Reviewer bKRD).

4. Geometric anchoring and latent representation alignment. We clarified the rationale for deterministic regularization and compared three structural anchoring choices. Using each residue’s native CαC_\alpha as the anchor produced the lowest reconstruction errors and fewest severe clashes (Appendix C.3). Compared with w/o LRA, PepMALD also showed higher clean-latent recoverability at four of the five evaluated timesteps (response to Reviewer bKRD, Table 5).

5. Benchmark conditions and reproducibility. We clarified that the native complex is used to define the binding region and peptide length at inference, specified that RMSD is calculated from generated and native peptide CαC_\alpha coordinates, and added the key model and training hyperparameters (§4.1 and Appendix F.3).

6. Diversity among successful designs. We recalculated diversity for candidates with ΔG<0\Delta G<0 using the same sampling budget. Sequence diversity was similar across the four methods, while PepMALD’s lower joint diversity among successful designs primarily reflected lower structural diversity (response to Reviewer nvDa, Table 2).

Changes are highlighted in blue in the revised manuscript. Our individual responses provide the full experimental results and address each reviewer’s questions in detail.

We again thank everyone involved in the review process for their time and consideration.

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

Related papers

Discussion (0)

Sign in to comment.