acceptodds
Under review as a conference paper at ICLR 2027

PepRAE: Adapting Protein Representations for Target-Conditioned Peptide Design

18 views

Abstract

Designing peptides for protein targets requires jointly generating amino-acid sequences and their bound three-dimensional conformations. Protein sequence and structure models provide protein-model representations, and the shared amino-acid alphabet of proteins and peptides makes these representations a natural starting point for peptide design. Yet these representations are not learned to serve as generative variables for target-conditioned peptide generation. This raises the question of how they can be adapted for this task. We introduce PepRAE, which performs this adaptation by learning feature latents from protein-model representations, while C-based spatial latents explicitly represent the peptide's bound geometry. The feature latents are learned through two reconstruction objectives. Representation reconstruction encourages the feature latents to support reconstruction of the protein-model representations. Peptide reconstruction provides downstream supervision by decoding peptide sequence and bound conformation from reconstructed features and spatial latents. After reconstruction training, target-conditioned latent diffusion jointly models peptide feature and spatial latents. Reverse diffusion followed by decoding generates peptide sequences and bound conformations. With Boltz-2 as the representation source, PepRAE achieves the highest amino-acid recovery, the lowest structural RMSDs, and the lowest Rosetta interface score () among the evaluated baselines on the LNR benchmark. To assess dependence on the representation source, we instantiate PepRAE with the pretrained sequence model ESM-2 and with LightEnc, our lightweight sequence encoder trained from scratch. Across the three separately trained variants, peptide reconstruction remains accurate and generation performance is broadly comparable. Together, these results show that the same formulation can learn feature latents from distinct protein-model representations for target-conditioned peptide design.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptide Design, Biomolecular Generative Modeling, Protein Representations, Generative Models

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

This paper proposes PepRAE for peptide generation given the protein target, binding site, and peptide length, which architecture-wide is a partially latent diffusion model. Its main contribution is developing a framework that can learn feature latents for protein targets and peptides from pretrained protein models via representation reconstruction and peptide reconstruction. PepRAE achieves competitive performance in amino acid recovery, reconstruction rmsd, and interface qualities.

Critical strengths
  1. The problem is well-defined and the partially latent architecture is well-motivated for joint modeling of sequence and structure (similar to proteina-complexa (https://research.nvidia.com/labs/genair/proteina-complexa/) ). The framework is carefully designed to solve the mismatch between complete-context representations available during training and target-only information available at inference.
  2. The benchmarks are carefully set to ensure fair comparison. PepRAE results are relatively strong compared with those baselines, showing the effectiveness of the proposed framework. And ablations on different loss terms and components also show the usefulness of different modules.
Critical weaknesses
  1. One of the main motivations for this work is claimed to be using the learned representation from pretrained protein model (e.g. Boltz-2) for this generation task. But the ablation result in Table 4 shows that even training PepRAE from LightEnc, a lightweight sequence encoder trained from scratch, can achieve fairly good performance, which wins in L-RMSD, M-RMSD, DockQ and ∆G, compared with other much stronger pretrained models. So this raises a question of how important it actually is to use those much stronger representations.
  2. The overall method description and framing lead me to view this work primarily as a peptide design model. In the benchmark, the authors show that PepRAE achieves strong native-recovery performance, including the best amino-acid recovery and lowest structural RMSDs relative to the reference peptides. However, the evaluation does not assess whether the generated peptide sequences are self-consistent with the generated structures, for example through independent refolding with an external structure predictor. Such a self-consistency check is a common and important validation in protein design, because native-recovery metrics alone do not establish that the designed sequences intrinsically support the generated conformations.
Questions for authors

Clarification:

  1. Can the author clarify whether the purpose of this work is for de novo design of protein-binding peptides, or is it just for native recovery? If it's the former, have the authors done any self-consistency evaluation to see whether the generated peptides are designable? An example reference is https://www.nature.com/articles/s42004-023-01029-7

Suggestion:

  1. It would be useful to provide a more detailed analysis of which specific information encoded in pretrained protein representations contributes to performance on this task, especially given the competitive results of the lightweight LightEnc baseline.
  2. It might be worth checking how similar the test peptide sequences are to peptides in the training set to exclude some possibilities for peptide-level leakage and memorization, which would also give a false-positive recovery.
10 comments
AuthorsOpenReview ↗

Response to Reviewer adDD — W1: Representation sources and adaptation objectives

Thank you for recognizing the task formulation, the treatment of the training–inference mismatch, and the supporting ablations. We first clarify the role of the representation sources, then report the additional structure-prediction and peptide-overlap analyses in the accompanying comments.

W1. Both pretrained sources and LightEnc support competitive generation.

Table 4 directly examines how generation performance depends on the representation source. It compares separately trained Boltz-2, ESM-2, and LightEnc variants with matched feature-latent widths, downstream architecture, objectives, and optimization settings (Section 4.5). Different sources lead on different metrics: ESM-2 has the highest reported AAR (40.41%), Boltz-2 the lowest C-RMSD (2.432 Å), and LightEnc the lowest L-RMSD and M-RMSD (1.187 and 1.408 Å), highest DockQ (0.746), and lowest Rosetta ΔG (−30.269 REU).

These results show that a large pretrained source is not required for competitive generation in the evaluated setting. The paper's claim is that the same adaptation formulation can learn feature latents for generation from the tested representation sources. LightEnc is itself trained: Appendix C specifies its masked-language-model training from random initialization on the training split, followed by freezing before PepRAE training.

The role of the adaptation objectives is tested separately in Table 2, with the Boltz-2 source and architecture held fixed. Removing representation reconstruction, No Lrep\mathcal L_{\rm rep}, changes C-RMSD from 2.432 to 2.612 Å, L-RMSD from 1.213 to 1.343 Å, and Rosetta ΔG from −30.161 to −27.946 REU. Residue reconstruction accuracy remains almost unchanged, at 99.89% and 99.81%, respectively (Table 13). Near-identical reconstruction accuracy therefore coexists with different reported generation results.

Restricting representation-decoder updates to representation reconstruction also yields less favorable generation means (Table 2). This control blocks peptide-loss updates to the representation-decoder parameters, while retaining those gradients through the decoder to the feature latents and encoder (Appendix G.1). Each configuration retrains its autoencoder and then trains three diffusion models with its selected autoencoder fixed.

Table 3 separately tests target-adapter supervision with the autoencoder fixed. PepRAE trains the adapter by latent alignment and stops denoising gradients at its output. No supervision removes alignment while retaining this gradient stop; Denoising-supervised removes alignment and allows denoising gradients to update the adapter (Section 4.4.2; Appendix G.2). C-RMSD is 2.432 Å for PepRAE, 2.767 Å for No supervision, and 2.780 Å for Denoising-supervised. The complete configuration has the best reported three-run mean on all seven generation metrics in each ablation table. These are means of run-level benchmark statistics; within each run, RMSDs are medians of per-target minima (Appendix D.1).

Tables 2–3 support the evaluated adaptation and conditioning objectives; Table 4 establishes applicability across the tested sources. LightEnc's competitive performance is consistent with the stated cross-source finding.

AuthorsOpenReview ↗

Response to Reviewer adDD — W2/Q1: Design task and consistency analysis

W2/Q1. Design task and sequence–structure consistency.

PepRAE is a de novo target-conditioned peptide sequence–structure generator. Given a target structure, binding site, and prescribed peptide length, it generates peptide feature and spatial states from Gaussian noise and decodes them into a sequence and bound conformation (Sections 3.1 and 3.4). The evaluated setting uses bound target coordinates, but does not supply the native peptide sequence or coordinates as peptide inputs. Following the peptide-design benchmark protocol in Section 4.1 and Appendix D, Tables 1 and 10 report native recovery alongside Rosetta interface scores, clashes, dihedral-distribution distances, and diversity.

Following your suggestion, we evaluated whether sequence-based complex prediction reproduces each candidate's own generated conformation, using Boltz-2. The protocol and results are below.

Candidate selection. For each of the 92 targets, we selected the candidate with the lowest Rosetta ΔG from each model's original 100-candidate pool. ΔG is the mean over three relaxations per candidate (Appendix D). Candidate identities were fixed before structure prediction, without using native-recovery scores or prediction results. PepRAE uses the same three diffusion runs as Table 1, evaluated separately; baselines use the models and candidate pools evaluated in the submission.

Complex prediction. We used Boltz-2 v2.2.1 with receptor MSAs and single-sequence peptide inputs, keeping the receptor construct identical across methods for each target. Boltz-2 is also the representation source for the main PepRAE configuration; in this experiment, it receives sequences and MSAs, without generated peptide coordinates, structural templates, or contact/pocket constraints. Each input received five predictions with seed 12, three recycling steps, and 200 sampling steps. We selected the prediction with the highest official confidence_score, rather than by agreement with the generated structure, and applied no relaxation to the predictions.

Comparison and metrics. Each selected prediction is compared with that candidate's original, unrelaxed generated structure. Native-sequence controls use the same prediction settings and are compared with their experimental structures. Both RMSDs measure peptide Cα coordinates: pocket fit aligns a fixed set of receptor-pocket Cα atoms without subsequently fitting the peptide; peptide fit aligns the peptide itself. The pocket residue set is fixed before prediction and used only for evaluation. Thus, pocket fit assesses agreement in the receptor frame, while peptide fit assesses the peptide's internal conformation.

Table R8(a). Boltz-2 geometric agreement.

MethodPeptide RMSD: pocket fit (Å) ↓Peptide RMSD: peptide fit (Å) ↓Pocket-fit RMSD ≤2 Å (%) ↑
PepGLAD9.1723.7951.09
PepFlow10.5042.4693.26
PepTri9.1132.7242.17
UniMoMo9.3802.3617.61
PRIME9.0312.9267.61
PepRAE7.328 ± 0.0642.613 ± 0.0608.70 ± 1.09
Native control0.8120.38479.35

Table R8(b). Confidence of the same selected predictions.

MethodPeptide pLDDT ↑Pocket–peptide PAE (Å) ↓
PepGLAD72.194.412
PepFlow71.484.559
PepTri70.324.259
UniMoMo72.474.593
PRIME74.214.239
PepRAE73.78 ± 1.183.918 ± 0.105
Native control88.802.660

Each run covers all 92 targets. RMSDs are cross-target medians. For each target, we compute peptide-averaged pLDDT (0–100) and mean bidirectional pocket–peptide PAE, then average these values across targets. The ≤2 Å column reports the percentage of targets meeting that geometric criterion. PepRAE reports mean ± sample SD of the three separately calculated run-level statistics; percentage SDs are in percentage points. Bold marks the best reported central value among the six design methods.

PepRAE has the lowest reported pocket-fit RMSD and pocket–peptide PAE among the evaluated design methods. Its pocket-fit RMSD is 7.328 Å, compared with 9.031 Å for PRIME, the next-lowest baseline. It also has the highest reported fraction within 2 Å: 8.70%, versus 7.61% for UniMoMo and PRIME; the native-sequence control reaches 79.35%. UniMoMo has the lowest peptide-fit RMSD, and PRIME the highest peptide pLDDT.

AuthorsOpenReview ↗

Response to Reviewer adDD — S1: Backbone information and decoding responses

PepRAE generates both peptide sequences and bound conformations, so we examined backbone geometry. The backbone angle ψ, defined by the atoms Ni,Ciα,Ci,Ni+1N_i,C_i^\alpha,C_i,N_{i+1}, describes rotation around Ciα−CiC_i^\alpha-C_i. We compared predictions using source features, residue identity alone, or a sequence window containing the central residue and two neighbors on each side. To account for amino-acid-specific angle preferences, we also predicted residuals around identity-conditioned label means. Finally, we tested decoding after changing the autoencoder's feature posterior means, with equally large random changes as controls. Table R12(a) reports both prediction analyses in separate columns; Table R12(b) reports the intervention.

Predicting backbone angles. With source models frozen, we fitted linear models on 9,113 training complexes and evaluated them on the original 1,064-complex validation split. Regularization was selected within training by receptor-cluster splits.

For source model mm, we use HC(m)[i]\boldsymbol H_C^{(m)}[i] (Eq. (1)); sis_i is residue identity. Direct models predict yi=(sin⁡ψi,cos⁡ψi)\boldsymbol y_i=(\sin\psi_i,\cos\psi_i). For residual prediction, let Rfit\mathcal R_{\rm fit} index fitting residues and nsn_s count residues of identity ss. We compute each identity's mean label, subtract it, and fit a linear residual predictor gres(m)g_{\rm res}^{(m)}:

y‾s=1ns∑i∈Rfit:si=syi\overline{\boldsymbol y}_s=\frac1{n_s}\sum_{i\in\mathcal R_{\rm fit}:s_i=s}\boldsymbol y_i

ei=yi−y‾si\boldsymbol e_i=\boldsymbol y_i-\overline{\boldsymbol y}_{s_i}

y^i=y‾si+gres(m)(HC(m)[i]).\widehat{\boldsymbol y}_i=\overline{\boldsymbol y}_{s_i}+g_{\rm res}^{(m)}(\boldsymbol H_C^{(m)}[i]).

Angles are recovered by atan2. The identity-only reference uses a fitted identity model for direct prediction and y‾si\overline{\boldsymbol y}_{s_i} alone for residual prediction.

Table R12(a). ψ prediction error (degrees) ↓.

InputDirect predictionIdentity reference + residual
Residue identity only69.7768.36
Sequence window (±2 residues)61.1660.99
Boltz-2 source features39.4539.66
ESM-2 source features50.4150.51
LightEnc source features55.2355.26

All sources have lower errors than both references; Boltz-2 has the lowest reported error in both analyses.

Changing feature means before decoding. We perturb the autoencoder's eight-dimensional peptide feature posterior means μh,C[P]\boldsymbol\mu_{h,C}[P] entering Dξrep\mathcal D_\xi^{\rm rep}. Following Appendix I.2, aa is the amplitude and sh\boldsymbol s_h contains the training standard deviations of these means, one per channel. A linear classifier fitted on training data predicts helix, strand, or coil from standardized means (μh,i−mh)⊘sh(\boldsymbol\mu_{h,i}-\boldsymbol m_h)\oslash\boldsymbol s_h on active channels, with training mean mh\boldsymbol m_h. Here ⊘,⊙\oslash,\odot denote elementwise division/multiplication. With dactd_{\rm act} active channels out of dh=8d_h=8, its helix/coil weights wH,wC\boldsymbol w_H,\boldsymbol w_C define

dHC=dactwH−wC∣wH−wC∣2\boldsymbol d_{\rm HC}=\sqrt{d_{\rm act}}\frac{\boldsymbol w_H-\boldsymbol w_C}{\\|\boldsymbol w_H-\boldsymbol w_C\\|_2}

Δμh=a(sh⊙dHC)\Delta\boldsymbol\mu_h=a(\boldsymbol s_h\odot\boldsymbol d_{\rm HC})

μh,ipert=μh,i+Δμh(i∈P).\boldsymbol\mu_{h,i}^{\rm pert}=\boldsymbol\mu_{h,i}+\Delta\boldsymbol\mu_h\\ (i\in P).

We use a=1a=1. Four random perturbations vj\boldsymbol v_j satisfy ∣vj∣2=∣Δμh∣2\\|\boldsymbol v_j\\|_2=\\|\Delta\boldsymbol\mu_h\\|_2 in original latent coordinates. Target feature rows, spatial inputs, and autoencoder parameters stay fixed. Eq. (4) gives (GP′)pert=Dψmol(Dξrep(μh,Cpert),XCα;GT).(\mathcal G'_P)^{\rm pert}=\mathcal D_\psi^{\rm mol}\big(\mathcal D_\xi^{\rm rep}(\boldsymbol\mu_{h,C}^{\rm pert}),\boldsymbol X_C^\alpha;\mathcal G_T\big).

We reassign secondary structure from decoded coordinates. Let fH(v)f_H(v) be the helix fraction under perturbation vv, with fH(0)f_H(0) from unperturbed decoding. In percentage points, ΔfH(v)=100[fH(v)−fH(0)]\Delta f_H(v)=100[f_H(v)-f_H(0)]; the excess response is Δexcess=ΔfH(Δμh)−14∑j=14ΔfH(vj).\Delta_{\rm excess}=\Delta f_H(\Delta\boldsymbol\mu_h)-\frac14\sum_{j=1}^4\Delta f_H(\boldsymbol v_j).

Table R12(b). Boltz-2 source: helix-fraction change (percentage points).

PerturbationChange
Helix–coil direction+2.73
Mean of four norm-matched random directions+0.80
Excess response+1.93

We evaluate 63 complexes with complete Cα coordinates from a fixed 64-complex validation subset. After averaging random responses within complexes, we bootstrap the paired contrast over 56 PDB groups (1,000 draws). The pointwise 95% percentile-bootstrap interval is 0.41–3.74 percentage points. Decoded residue identities change at 71.1% of positions relative to unperturbed decoding, so this is a joint sequence–structure response.

The helix–coil response exceeds the average random response.

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

PepRAE tackles target-conditioned peptide design: given a target, binding site with bound coordinates, and prescribed length, jointly generate sequence and bound conformation. An autoencoder maps frozen protein-model features (Boltz-2, ESM-2, or a one-layer scratch encoder, LightEnc) to 8-dim per-residue feature latents, trained by representation reconstruction (L_rep) plus peptide reconstruction; spatial latents are Cα coordinates with only their scale learned. A target-conditioned equivariant latent diffusion model then generates both fields, with target conditioning from a linear adapter trained by an alignment loss.

On a 92-target filtered LNR benchmark with five retrained baselines and 100 candidates per target, the Boltz-2 variant reports the best AAR (39.73%), RMSDs (2.43/1.21/1.47 Å), ΔG (−30.16), IMP, TM-score, clash rates, and side-chain JSD, but is second on DockQ and F1_intra and has the second-lowest pose diversity. A matched three-source study trains Boltz-2, ESM-2, and LightEnc variants with three diffusion seeds each.

Critical strengths

The matched three-source study, as an artifact. Nine full training runs with matched latent width, objectives, architecture, and optimizer, reported with run-level tables, SDs, an honest single-run sensitivity check (Appendix E.2), and an explicit acknowledgment that PRIME shares the same two-stage design (Appendix K). This transparency is why the paper’s weaknesses are legible. It does not change my recommendation, since the experiment undercuts the paper’s thesis.

Worth preserving in revision: the pool-level metrics (Clash_in 0.46%, Clash_out 1.51%, JSD_sc 0.302) average over all 100 candidates and so cannot be inflated by one lucky sample. Margins over PepGLAD, PepFlow, and PepTri are large and seed-stable. These are the submission’s strongest numbers and sit in Appendix E.3.

Critical weaknesses

1. The methodological contribution is incremental relative to prior work, and the component that distinguishes it is not supported by the submission’s own cross-source results.

The two-stage latent-diffusion formulation for target-conditioned peptide co-design is established by PepGLAD and UniMoMo. The practice of deriving generative latents from a frozen pretrained encoder followed by learned projections, in the specific setting of peptide binder design, is established by PRIME, which Appendix K acknowledges shares the same feature-plus-spatial latent structure. The broader recipe of autoencoding pretrained representations into a diffusion latent space, including the convention of comparing several pretrained backbones, is established by RAE, FAE, and RepTok. The peptide decoder, equivariant denoiser, residue-block formulation, target-centred coordinate frame, lower-bounded reverse-step coefficient, and evaluation code are adopted from UniMoMo. The novel contribution therefore reduces to the representation-reconstruction objective and the residue-aligned extraction interface.

The evidence provided does not establish that this residual contribution is responsible for the reported performance. In Table 4, the LightEnc variant, a single-layer transformer trained from random initialisation on the training split of this benchmark, attains the best L-RMSD, M-RMSD, DockQ, and ΔG among the three sources, while ESM-2 attains the best AAR and Boltz-2 leads only on C-RMSD. If a scratch-trained single-layer encoder is interchangeable with a frozen state-of-the-art structure-prediction model, the premise that protein-model representations constitute the valuable ingredient is not supported, and the framing of the title, abstract, and introduction requires revision.

2. The reconstruction evidence is largely determined by architectural choices and therefore carries limited diagnostic weight, and an internal comparison indicates the magnitude of the contribution made by candidate-pool aggregation.

Equation 3 fixes the spatial posterior mean to the observed Cα coordinate, so the spatial pathway reproduces its own input rather than reconstructing it. The feature latent is required to convey little beyond residue identity, approximately 4.3 bits, across eight continuous dimensions, which the reported linear-probe accuracy of 99.66% confirms it does. The finding stated in the abstract and Section 4.5, that peptide reconstruction remains accurate across all three sources at 99.89 to 100.00% identity accuracy and 0.471 to 0.492 Å Cα RMSD, is therefore close to guaranteed by the construction and provides little evidence regarding the success of representation adaptation. The single reconstruction diagnostic that does separate the three sources, source-feature cosine similarity of 0.779, 0.868, and 0.951, is explicitly declared non-comparable across sources in Appendix H.1. The cross-source analysis is thus left without a diagnostic that is simultaneously informative and comparable.

3. The comparative claims are not supported at the stated margins given the submission’s own reported variance, and the evaluation protocol favours the behaviour the model exhibits.

AAR is reported as 39.73 ± 0.82 against a single-run PepTri value of 39.23, and IMP as 29.35 ± 3.92 against 28.26 for UniMoMo; neither constitutes a separation. C-RMSD and L-RMSD differ from UniMoMo by approximately 1.6 and 1.4 standard deviations respectively, while DockQ and F1_intra are lower than UniMoMo. The sensitivity analysis in Appendix E.2 shows the AAR ranking falling to second and DockQ to third when averaging across diffusion runs is removed.

4. The principal claim depends on baseline reproductions that are unreplicated, unvalidated, and in at least one case internally inconsistent.

Each of the five baselines is retrained once by the authors using self-selected settings, without replication and without any demonstration that the reproduction recovers published performance on any benchmark. The reported PepFlow row contains an inconsistency that cannot be reconciled with the submission’s own metric definitions.

Equation 43 defines co-design diversity as Dcodesign=DseqDposeD_\text{codesign} = \sqrt{D_\text{seq} D_\text{pose}}, and Appendix D.4 states that this quantity is computed per target before averaging across targets. Under that ordering, the Cauchy–Schwarz inequality requires

Dcodesign‾=1n∑jxjyj≤Dseq‾⋅Dpose‾.\overline{D_\text{codesign}} = \tfrac{1}{n}\sum\nolimits_j \sqrt{x_j y_j} \le \sqrt{\overline{D_\text{seq}} \cdot \overline{D_\text{pose}}}. Evaluating this bound for every method in Tables 1 and 10, PepFlow exceeds its own upper bound by 0.028. The violation is robust to rounding: at the extremes of the rounding intervals the bound is 0.5367 while the reported value is at least 0.5635. Reconciliation would require Dseq≥0.651D_\text{seq} \ge 0.651 or Dpose≥0.541D_\text{pose} \ge 0.541. Every other method, and all nine PepRAE, ESM-2, and LightEnc run-level rows in Tables 9c and 12b, satisfies the bound with a Jensen gap in the range 0.009 to 0.028; PepFlow is the sole exception. At least one of its three reported diversity values is therefore incorrect.

Questions for authors
  1. Please report the learned-embedding control: replace the frozen source features HCH_C with learned per-residue embeddings of the same width, holding architecture, data, optimization, candidate budget, and evaluation fixed, and report AAR, C-RMSD, L-RMSD, M-RMSD, DockQ, and ΔG over three diffusion seeds.

  2. What is the empirical per-channel standard deviation of the peptide feature posterior means mu_h,C[P] on the training split, and what is the variance of the clean feature state U^(0)_h,P relative to that of the clean spatial state U^(0)_x,P?

  3. Table 10 reports PepFlow with Dseq=0.588D_\mathrm{seq}=0.588, Dpose=0.489D_\mathrm{pose}=0.489, and Dcodesign=0.564D_\mathrm{codesign}=0.564, but Equation 43 under the per-target-then-average ordering of Appendix D.4 bounds the third by 0.588×0.489=0.536\sqrt{0.588 \times 0.489} = 0.536 via Cauchy–Schwarz. Which of the three values is incorrect?

  4. Please report AAR against a composition-matched random-sequence baseline under the identical best-of-100 protocol, stratified by the length bins of Figure 4.

  5. Please report recovery metrics as a function of candidate budget N∈{1,5,10,25,100}N \in \{1, 5, 10, 25, 100\}, together with per-candidate means rather than minima for RMSD and ΔG, and paired per-target Wilcoxon signed-rank tests against UniMoMo and PepTri across all 92 targets.

18 comments
AuthorsOpenReview ↗

Response to Reviewer qWsu — W1: Contribution and cross-source evidence

Thank you for the detailed review. We first clarify what the cross-source experiment tests and how it relates to the training ablations.

W1. LightEnc's results are consistent with the claim made in the paper.

The statement in Critical Strengths that this experiment "undercuts the paper's thesis" does not follow from Table 4. The Abstract and Section 4.5 explicitly report broadly comparable generation across separately trained Boltz-2, ESM-2, and LightEnc variants. Their stated conclusion is that "the same formulation can learn feature latents from distinct protein-model representations for target-conditioned peptide design." LightEnc's competitive performance is part of this finding. The paper does not claim that pretraining is necessary or that pretrained sources must outperform LightEnc. Appendix C describes LightEnc's masked-language-model training on the training split and its subsequent freezing before PepRAE training.

Table 4 changes the representation source while retaining the adaptation pathway. Tables 2–3 instead test the training choices within that pathway. A lack of a consistent pretrained-source advantage does not negate either the cross-source applicability result or the fixed-source ablation results.

Sections 2–3 and Appendix K identify the inherited generation framework and geometric components. The method adapts residue-aligned source features through a representation decoder whose reconstructed features feed the peptide decoder. Representation and peptide reconstruction jointly supervise this pathway. For target-only conditioning, an adapter is trained to match complete-context target feature posterior means; alignment updates the adapter, while denoising gradients stop at its output.

These choices are tested in the submitted experiments. In Table 2, restricting representation-decoder updates to representation reconstruction changes C-RMSD from 2.432 to 2.682 Å and Rosetta ΔG from −30.161 to −28.692 REU. This control still propagates peptide-reconstruction gradients to the feature latents and encoder; it prevents that loss from updating only the representation-decoder parameters (Appendix G.1).

Table 3 holds the autoencoder fixed and changes adapter supervision. C-RMSD is 2.432 Å with alignment, compared with 2.767 Å without adapter supervision and 2.780 Å with denoising supervision. In each table, the full configuration has the best reported three-run mean on all seven generation metrics. These are means of run-level benchmark statistics; each run's RMSD statistic is the median of per-target minima (Appendix D.1).

The ablations therefore provide evidence for the evaluated reconstruction and conditioning choices. Table 4 addresses a different question: whether the formulation also works with other representation sources. LightEnc's performance does not contradict that conclusion.

AuthorsOpenReview ↗

Response to Reviewer qWsu — W2: Reconstruction diagnostics and generation evidence

W2. Near-identical residue reconstruction accuracy does not imply the same generation performance.

Table 2 tests representation reconstruction (Lrep\mathcal L_{\rm rep}) with the Boltz-2 source, architecture, latent width, spatial-pathway design, training data, and evaluation protocol held the same. The full model and the No Lrep\mathcal L_{\rm rep} control both recover residue identities almost exactly, but their generation results differ:

ConfigurationResidue accuracy (%) ↑C-RMSD (Å) ↓L-RMSD (Å) ↓Rosetta ΔG (REU) ↓
PepRAE99.892.432 ± 0.1141.213 ± 0.063−30.161 ± 0.249
No Lrep\mathcal L_{\rm rep}99.812.612 ± 0.1881.343 ± 0.053−27.946 ± 0.615

Residue accuracy is the reconstruction result in Table 13; the other columns are generation results from Tables 2 and 14. Each configuration retrains its autoencoder and then trains three diffusion models with its selected autoencoder fixed. Generation values are the mean ± sample SD of the three run-level statistics; RMSDs use the per-run aggregation defined in Appendix D.1.

The full configuration has better reported means on all seven generation metrics in Table 2. This supports representation reconstruction as a training choice in the evaluated setting, even though residue-identity reconstruction is already near-exact without it. Identity reconstruction accuracy alone therefore does not account for the observed generation differences.

Section 4.3 characterizes conditional peptide reconstruction from encoded source features and Cα-anchored spatial inputs. Observed Cα coordinates provide the spatial posterior means in Equation 3; they do not directly specify the decoded residue identities or all atomic coordinates. Reconstruction is a diagnostic of this decoding interface. Evidence for generation comes separately from the generation ablations and benchmark, where diffusion generates peptide feature and spatial states from noise (Section 3.4).

The 99.66% linear-probe accuracy establishes that residue identity is linearly readable. The review's approximately 4.3-bit figure is the maximum entropy of a 20-class identity label, not a measurement of information retained in the continuous feature latents. The probe result cannot establish that those latents contain only identity information.

Finally, the cross-source study does have common output-level measurements. Source-feature cosine scores are within-source reconstruction diagnostics because their reference features differ. Table 4 compares generated peptides using the same AAR, RMSD, DockQ, and Rosetta ΔG definitions across all three sources. The cross-source conclusions therefore do not depend on treating source-specific cosine scores as comparable.

AuthorsOpenReview ↗

Response to Reviewer qWsu — W3: Benchmark results and run variation

W3. Benchmark results and run variation.

Table 1 reports numerical rankings under the common evaluation protocol, with PepRAE summarized by the mean of three separately computed diffusion-run statistics. Table 9 provides the individual diffusion runs and sample SDs. Each PepRAE run has lower C-RMSD, L-RMSD, and ΔG than the best baseline values reported in Table 1, all obtained by UniMoMo:

Model/runC-RMSD (Å) ↓L-RMSD (Å) ↓Rosetta ΔG (REU) ↓
PepRAE Run 12.3951.155−30.431
PepRAE Run 22.5601.205−29.942
PepRAE Run 32.3401.280−30.110
UniMoMo (Table 1)2.611.30−28.70

These are the three diffusion runs sharing the selected Boltz-2 autoencoder, compared with the baseline models reported in Table 1. Within each run, RMSDs are medians of per-target minima and ΔG is the mean of per-target minima (Appendix D.1). The rankings on these three metrics hold both individually and after averaging across runs.

The AAR and IMP statements refer to the reported point estimates, with PepRAE summarized by its three-run mean; they do not claim statistically significant separation. Appendix E.2 also reports the single-run comparison, including the changes in AAR and DockQ rankings.

The candidate-pool evaluation follows the UniMoMo/PepTri convention described in Section 4.1 and Appendix D.1. All methods use the same 92 targets, prescribed peptide lengths, 100-candidate budget, and metric-specific aggregation. The benchmark also evaluates the full candidate pools: PepRAE's intrapeptide and target–peptide clash rates are 0.46% and 1.51%, computed over all 100 candidates within each target and then averaged across targets. Its side-chain dihedral-distribution distance is 0.302 (Appendix E.3). These are the lowest reported values among the evaluated methods and provide complementary evidence on clash frequency and side-chain distributions across the generated pools.

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

PepRAE adapts protein-model representations for target-conditioned peptide sequence–structure generation. It learns feature latents through representation and peptide reconstruction, then jointly generates feature and Cα-based spatial latents using conditional diffusion. A target adapter provides target-only conditioning. Experiments on LNR show competitive candidate-pool performance, while separately trained Boltz-2, ESM-2, and LightEnc variants demonstrate applicability across representation sources.

Critical strengths
  1. The framework coherently combines representation adaptation, explicit spatial modeling, and target-only conditioning.
  2. Ablations support the main training choices, and cross-source experiments demonstrate the formulation’s flexibility.
  3. Baselines share a common training dataset and evaluation protocol, with aggregation rules and scope limitations clearly disclosed.
Critical weaknesses
  1. The main evaluation aggregates AAR, DockQ, RMSDs, and Rosetta scores independently over each target’s 100-candidate pool. Consequently, favorable values across columns may correspond to different peptides, leaving unclear whether individual candidates simultaneously exhibit strong sequence recovery, structural recovery, and interface scores. This protocol is explicitly disclosed and shared by all baselines, so it supports comparisons of metric-specific pool performance. However, it provides limited evidence about the joint quality of individual designs. Candidate-level analyses, such as paired metric distributions or the fraction of candidates meeting predefined structural and interface criteria simultaneously, would help establish whether the reported strengths coexist within the same peptides.

  2. Beyond whether high-quality candidates exist, practical use requires identifying them without access to the native peptide sequence or binding pose. Although Rosetta provides a possible ranking criterion, the paper does not report benchmark-wide DockQ or RMSD results for candidates selected by Rosetta score. Figure 6 reports paired scores for two examples but does not explain how those candidates were selected. Applying a common native-independent ranking rule to the existing pools would clarify whether PepRAE’s pool-level improvements translate into better selected candidates relative to the baselines. The authors acknowledge native-independent selection as future work, but its effectiveness remains an important uncertainty when assessing practical utility.

Questions for authors
  1. Could you provide candidate-level analyses showing how sequence recovery, structural recovery, and interface scores relate within the same peptides? In particular, what fraction of candidates simultaneously satisfy predefined structural and interface quality criteria, compared with the baselines?

  2. Using Rosetta ΔG as a common ranking criterion, could you report DockQ and RMSD results for the top-ranked candidate and the mean results across the five highest-ranked candidates for each target, compared with the baselines? Please also clarify the candidate selection rule used in Figure 6 and whether structural metrics are evaluated before or after relaxation.

11 comments
AuthorsOpenReview ↗

Response to Reviewer synp (1/2): Joint candidate quality

Thank you for your positive assessment of the framework, ablations, and common evaluation protocol. We first address your concern about the quality of individual candidates. The accompanying response addresses Rosetta-ranked candidate selection and clarifies how the examples in Figure 6 were selected and scored.

W1/Q1. Joint structural and interface quality.

For each candidate, we check DockQ ≥ 0.23 and Rosetta ΔG < 0, and report the fractions meeting each criterion and both together. The DockQ cutoff follows the acceptable-or-better category in the official DockQ documentation. The ΔG < 0 criterion follows PepGLAD, one of the baselines evaluated in our paper. We apply the same criteria to every method.

For each target, we compute these fractions over all 100 candidates, then average them equally across the 92 targets (Section 4.1), without Top-k filtering. As specified in Appendix D, DockQ is measured before relaxation; each candidate's ΔG is averaged over three independent Rosetta relaxations.

Table R1. Candidate fractions satisfying individual and joint criteria (%).

MethodDockQ ≥ 0.23ΔG < 0Both criteria
PepGLAD71.0232.7622.55
PepFlow73.5257.5343.03
PepTri72.7152.1036.49
UniMoMo73.1573.2751.33
PRIME62.6858.3539.35
PepRAE76.40 ± 1.1575.51 ± 0.7856.82 ± 1.53

These results reuse the same three PepRAE diffusion runs and candidate pools as the main benchmark in Table 1. We report the mean ± sample SD of the three separately computed run-level fractions (Appendices D.1 and E.2). Baselines use the models and candidate pools evaluated in the submission. SDs are in percentage points; bold marks the highest reported central value.

PepRAE has the highest joint fraction among the evaluated methods: 56.82%, compared with 51.33% for UniMoMo, the strongest baseline on this measure, a difference of 5.49 percentage points. Whereas Table 1 aggregates metrics independently (Appendix D.1), Table R1 reports the fraction of candidates with both acceptable-or-better docking quality and a negative Rosetta score.

AuthorsOpenReview ↗

Response to Reviewer synp (2/2): Rosetta-ranked selection and clarification of Figure 6

Below we report the Rosetta-ranked comparison you requested and then address your questions about the examples in Figure 6.

W2/Q2. Structural quality after Rosetta-based selection.

Following your suggested rule, we rank each target's 100 candidates by increasing Rosetta ΔG. Top-1 evaluates the first candidate; Top-5 takes the arithmetic mean of the first five candidates' metrics within each target. Tables R2(a–b) then average these values equally across the 92 targets. Selection uses ΔG alone, without the native peptide sequence or pose; all metrics refer to the same selected candidate IDs.

We use the same three PepRAE diffusion runs and candidate pools as the main benchmark in Table 1 (Section 4.1). Scoring follows Appendix D: structural metrics before relaxation, and ΔG averaged over three independent relaxations per candidate. PepRAE reports mean ± sample SD of three separately evaluated run-level statistics (Appendices D.1 and E.2), without pooling runs. Baselines retain their original models and pools. RMSDs are in Å; ΔG is in REU. Bold marks the best reported central value.

Table R2(a). ΔG-ranked Top-1: means across targets.

MethodDockQ ↑C-RMSD ↓L-RMSD ↓M-RMSD ↓ΔG ↓
PepGLAD0.37168.5984.4144.407-22.028
PepFlow0.48307.5492.7652.921-21.054
PepTri0.38019.0652.9513.058-15.692
UniMoMo0.50796.9922.5992.661-28.703
PRIME0.45417.9023.1363.329-27.835
PepRAE0.5253 ± 0.01316.831 ± 0.2412.488 ± 0.0512.686 ± 0.040-30.161 ± 0.249

Table R2(b). ΔG-ranked Top-5: means across targets.

MethodDockQ ↑C-RMSD ↓L-RMSD ↓M-RMSD ↓ΔG ↓
PepGLAD0.35358.8614.6014.559-16.965
PepFlow0.48177.4822.7392.922-14.122
PepTri0.39698.4102.9653.071-11.983
UniMoMo0.46738.3653.1123.285-25.225
PRIME0.44618.2713.0013.213-23.951
PepRAE0.5118 ± 0.00936.991 ± 0.2952.544 ± 0.0232.752 ± 0.028-26.347 ± 0.093

For Top-1, PepRAE has the highest mean DockQ and lowest mean C- and L-RMSD; M-RMSD is close to UniMoMo (2.686 versus 2.661 Å). For Top-5, it leads on the reported means of all four structural metrics. Its mean ΔG is lowest under both selections. The favorable structural results therefore extend to candidates selected by ΔG.

Comparison with Table 1. Section 4.1 and Appendix D.1 follow the UniMoMo/PepTri evaluation convention: Table 1 averages per-target DockQ maxima and takes the median of per-target RMSD minima. ΔG-selected candidates need not attain those extrema. The new RMSD means also change the cross-target summary, so differences from Table 1 reflect both selection and aggregation.

Tables R2(c–d) report RMSD medians across targets for these same selections. For Top-5, we first average the five candidates within each target, then take the median across targets. PepRAE has the lowest reported values in all six comparisons. Top-1 ΔG uses the same selection and aggregation as Table 1.

Table R2(c). ΔG-ranked Top-1: RMSD medians across targets (Å).

MethodC-RMSD ↓L-RMSD ↓M-RMSD ↓
PepGLAD7.5253.6853.671
PepFlow4.8002.5002.636
PepTri7.3102.7502.734
UniMoMo4.9552.3252.368
PRIME6.0252.7552.856
PepRAE4.420 ± 0.5272.270 ± 0.1802.344 ± 0.095

Table R2(d). ΔG-ranked Top-5: RMSD medians across targets (Å).

MethodC-RMSD ↓L-RMSD ↓M-RMSD ↓
PepGLAD8.2704.2584.224
PepFlow5.8422.4972.575
PepTri7.7542.7262.823
UniMoMo6.9082.4962.647
PRIME7.3862.7802.921
PepRAE5.505 ± 0.4882.367 ± 0.0612.487 ± 0.039

Regarding your questions about Figure 6. We manually chose a 5-residue peptide for 3T6J and a 16-residue peptide for 6FQ4 to illustrate different lengths. Candidates were not selected by metric rankings; neither has the highest DockQ or lowest ΔG in its pool. Their DockQ ordering is qualitatively consistent with Figure 4's length-stratified trend.

The displayed structures and DockQ values are before relaxation; ΔG is the mean over three independent relaxations of the same candidate, following Appendix D. Thus, the paired scores refer to the same candidate at the specified evaluation stages. Table R2 provides the benchmark-wide selection analysis, complementing these qualitative examples.

AuthorsOpenReview ↗

Looking forward to your response

Dear Reviewer synp,

We have posted additional analyses for W1/Q1 and W2/Q2 and would appreciate your assessment before the discussion closes.

Whether structure and Rosetta score are favorable in the same candidate. Table R1 evaluates both conditions on each individual peptide, using all 100 candidates per target. The fraction meeting both DockQ ≥ 0.23 and ΔG < 0 is 56.82% for PepRAE versus 51.33% for UniMoMo, the strongest baseline on this measure.

Whether good structures can be selected without the native answer. Table R2 applies your proposed ΔG ranking to the existing pools and reports both Top-1 and Top-5 results. PepRAE has the highest mean DockQ and lowest mean C- and L-RMSD for Top-1; for Top-5, it has the best reported means for all four structural metrics. Selection uses ΔG alone, and every metric is evaluated on the same selected candidate IDs.

We also clarified that Figure 6 shows manually chosen examples of different lengths, without selection by metric ranking. The displayed structures and DockQ values are from before relaxation; ΔG is averaged over three relaxations of the same candidate.

Could you let us know how these results affect your assessment of individual-candidate quality and score-based selection, and whether any concerns remain?

Comments on the paper

3 comments
AuthorsOpenReview ↗

Manuscript update: corrections to PepFlow co-design diversity and ablation IMP results

We have updated this comment to include an additional correction to the IMP results for the No Lrep\mathcal{L}_{\mathrm{rep}} ablation, identified after our previous manuscript update. We have uploaded a revised PDF incorporating both corrections described below.

PepFlow co-design diversity — Table 10 (Appendix E.3). As described in our previous update, the PepFlow co-design-diversity entry has been corrected from 0.564 to 0.521, following the issue identified by Reviewer qWsu and discussed in our response to W4/Q3. This was a transcription error in preparing the table. PepFlow's sequence and pose diversities remain 0.588 and 0.489, respectively. The correction changes the reported value and its numerical differences from other methods, but does not change the ranking in that column.

No Lrep\mathcal{L}_{\mathrm{rep}} IMP — Table 2 and Table 14 (Appendix G.1). The reported mean IMP for this ablation has been corrected from 25.00% to 26.45% in both tables. The corresponding sample standard deviation in Table 14 has been corrected from 1.26 to 0.63 percentage points. The corrected result is therefore a mean IMP of 26.45%, with a sample standard deviation of 0.63 percentage points across the three diffusion runs.

The IMP correction changes the magnitude of the differences between configurations, but preserves their ordering in Tables 2 and 14. In particular, the full PepRAE model retains a higher mean IMP (29.35%) than the No Lrep\mathcal{L}_{\mathrm{rep}} ablation (26.45%). The interpretation of this ablation remains unchanged.

AuthorsOpenReview ↗

Summary of rebuttal results and manuscript revisions

Manuscript update. We have uploaded the revised paper and added the manuscript locations for the experiments discussed below. Changes are marked in blue. The R numbers refer to tables in the discussion.

Candidate quality and selection (Section 4.6). Across the full pools, 56.82% of PepRAE candidates meet both criteria (DockQ ≥ 0.23 and Rosetta ΔG < 0), compared with 51.33% for UniMoMo. Ranking each target’s candidates by ΔG gives PepRAE a mean DockQ of 0.5253 for the first candidate and 0.5118 averaged over the first five. PepRAE values are averaged across the same three diffusion runs, with selection performed within each run. The follow-up analysis uses targets as the denominator. The first candidate meets both criteria for 78.62% of targets with PepRAE and 78.26% with UniMoMo. At least one of the first five meets both criteria for 90.22% of targets with either method.

Agreement with sequence-based structure predictions (Section 4.6). For each target, we fixed the design selected by ΔG before prediction and compared the predicted complex with that design’s own unrelaxed structure. PepRAE’s pocket-aligned peptide Cα RMSD was 7.328 Å with Boltz-2 and 7.568 Å with AF2-Multimer. The lowest baseline values were 9.031 Å (PRIME) and 8.842 Å (PepTri), respectively. For PepRAE, we took the median across targets within each run and averaged the three medians. AF2-Multimer was not used to extract PepRAE’s input features.

Representation adaptation (Section 4.5). PepRAE uses a common residue-level interface and trains a separate model for each feature source (Table 4). Table 2 tests the reconstruction objective while holding the Boltz-2 source fixed. Table 22 replaces the Boltz-2 features with a learned amino-acid lookup. This changes the encoder input and its reconstruction target while holding the downstream architecture fixed. Compared with the learned amino-acid lookup, the Boltz-2 configuration gives lower values for all three RMSDs, higher DockQ, and lower Rosetta ΔG (−30.161 versus −25.995 REU). AAR is slightly higher with the lookup (40.00% versus 39.73%).

Candidate budget and paired comparisons. Figure 3 plots the results for 1, 5, 10, 25, and 100 candidates per target. At 100 candidates, the paired tests favor PepRAE over UniMoMo and PepTri on the three RMSDs and Rosetta ΔG under both per-target summaries. For these comparisons, the largest Holm-adjusted p-value is 0.041 (Tables 35–36).

AnalysisRebuttalRevised paper
Learned amino-acid embeddingR11App. L, Table 22, p. 42
Joint fraction across candidatesR1App. M, Table 23, p. 42
Target coverage after Rosetta selectionR13App. M, Table 24, p. 43
Quality of Rosetta-ranked candidatesR2App. M, Tables 25–28, pp. 43–44
Candidate budgetsR5Fig. 3, p. 10 and App. N, Tables 29–33, pp. 44–45
All-candidate meansR6App. N, Table 34, p. 45
Paired tests across 92 targetsR7App. O, Tables 35–36, p. 46
Boltz-2 structure predictionsR8App. P, Tables 37–38, pp. 46–47
AF2-Multimer structure predictionsR10App. P, Tables 39–40, p. 47
Composition-matched sequence control by lengthR4App. Q, Table 41, p. 48 (length bins in Fig. 5, p. 31)
Training-peptide overlap and AARR9App. Q, Table 42 and preceding text, p. 48
Latent scalesR3App. R, Tables 43–44, pp. 48–49
Backbone-angle probes and decoder perturbationsR12App. R, Tables 45–46, p. 49

Figure references from the discussion. The peptide-length analysis cited as Figure 4 in qWsu’s Q4 is now Figure 5 (Appendix E.4, p. 31). The additional budget curves requested in Q5 appear in Figure 3 (p. 10). The two examples discussed with synp as Figure 6 are now Figure 7 (Appendix I.3, p. 38).

The revised paper incorporates the previously announced corrections to PepFlow co-design diversity (0.521, Table 10) and to IMP for the ablation without representation reconstruction (26.45%, Tables 2 and 14, with a sample SD of 0.63 percentage points in Table 14).

AuthorsOpenReview ↗

Closing note on the revised manuscript

Dear Area Chair and Reviewers,

Thank you for the discussion. The revised manuscript incorporates the additional experiments and analyses, with changes marked in blue.

  • Representation adaptation. Section 4.5 clarifies how PepRAE adapts features from different sources through a common interface, with a separate model trained for each source. Tables 2 and 22 test the reconstruction objective and the choice of input features, respectively.
  • Candidate selection and structural agreement. Section 4.6 summarizes the results after Rosetta ranking. Table 24 reports the fraction of targets with at least one selected candidate meeting both criteria. Tables 37–40 compare sequence-based structure predictions with the selected designs.
  • Candidate budgets and paired comparisons. Figure 3 shows the results across five candidate budgets. Tables 34–36 report all-candidate means and paired tests across 92 targets.

In the discussion, synp said the additional analyses largely addressed the concerns in the review and expressed willingness to raise the rating. adDD found that the new results strengthened the case that pretrained embeddings help with generation quality and was willing to raise the rating.

qWsu confirmed that the requested experiments were included and maintained the original recommendation, citing reservations about the methodological contribution relative to ICLR’s expectations. PepRAE addresses a representation-learning problem within our selected ICLR area, applications to physical sciences (physics, chemistry, biology, etc.). Its contribution is a learned interface for adapting protein-model representations to target-conditioned peptide generation.

We appreciate your consideration of the revised manuscript and the discussion in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.