acceptodds
Under review as a conference paper at ICLR 2027

ADEL: Truth-Anchored Distillation with Engh-Huber-informed Strain and Linkage Chemistry for Macrocycle Design

Abstract

Designing macrocyclic peptide binders requires more than generating a plausible peptide structure: the ring must close between chemically compatible atoms, and for some linkage types, the identities of these atoms depend on the sequence being designed. In this work, we propose **ADEL**, an all-atom sequence–structure co-design diffusion model that learns to satisfy this coupled geometric and chemical constraint. Our approach combines truth-anchored distillation with Engh–Huber-informed closure geometry and explicit linkage-chemistry supervision. Instead of directly distilling the projection of a sampling-time closure corrector, which we show provides a training signal redundant with the existing closure loss, we distill the model's denoising response after projection and renoising. Because this teacher is produced by the model itself, we introduce a per-atom truth gate that retains a teacher target only when it is closer to the reference structure than the student prediction, preventing the self-distillation objective from drifting during training. We further supervise the local geometry around the closing bond and the residue identities required by sequence-dependent linkages. We retrain six published methods on the same data split and evaluate all generated designs using a common type-aware protocol. On leakage-controlled CPSea targets, our model achieves a macrocycle closure rate of 47.2%, compared with 35.4% for the strongest retrained baseline, while maintaining 99.8% structural validity. Since not all baselines can explicitly represent sequence-dependent linkages, we interpret the aggregate comparison as multi-chemistry coverage rather than a fully matched constraint-solving benchmark. Across head-to-tail, isopeptide, and disulfide linkages, ADEL maintains comparable closure performance within a single co-design framework.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cyclic peptide design, constrained diffusion models, sequence–structure co-design, self-distillation, generative protein design

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_r5VFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ADEL, an all-atom diffusion-based framework for joint sequence and structure design of macrocyclic peptide binders. The method addresses the challenge of simultaneously satisfying geometric constraints for ring closure and sequence-dependent chemical requirements associated with different cyclization linkages. ADEL integrates three main components into a unified training framework: (1) truth-anchored distillation, which uses corrected and re-denoised predictions as teacher targets, retaining only atom-level predictions that are closer to the native structure than the student's predictions; (2) Engh–Huber-informed geometric constraints, which regulate closure distances and local bond geometry; and (3) linkage-chemistry supervision, which enforces the residue identities required for different cyclization mechanisms. The authors evaluate ADEL on 226 held-out targets from the CPSea dataset against seven retrained baseline methods. ADEL achieves a macrocycle closure rate of 47.2%, compared with 35.4% for the strongest baseline, while maintaining 99.8% structural validity. The framework supports three linkage types—head-to-tail, isopeptide, and disulfide—and also demonstrates improvements in computationally predicted binding energies. These results suggest that integrating geometric and chemical constraints into sequence–structure co-design can improve the computational design of macrocyclic peptide binders.

Critical strengths

1. Practical relevance of integrating geometric and chemical constraints

The paper addresses a relevant challenge in macrocyclic peptide design by jointly considering geometric constraints for ring closure and sequence-dependent chemical requirements. Integrating these constraints into a unified sequence–structure co-design framework provides a practically meaningful approach to generating macrocycles with different linkage chemistries, including head-to-tail, isopeptide, and disulfide linkages.

2. Encouraging computational performance in joint closure and structural validity

Although the fairness of baseline comparisons may be subject to debate due to differences in supported linkage chemistries and evaluation conditions, the proposed model demonstrates encouraging computational performance in its own right. In particular, ADEL achieves a joint closure-and-validity rate of 47.1%, with 99.8% of generated structures satisfying the defined structural validity criteria. These results demonstrate that ADEL can generate macrocyclic peptide structures satisfying both ring-closure and structural validity criteria, while highlighting the remaining challenges in achieving consistent closure across all generated designs.

3. Comprehensive experimental evaluation and analysis

The paper provides extensive experimental analyses, including comparisons with seven retrained baselines, component-wise ablation studies, multiple random seeds, and detailed failure analyses. These evaluations help characterize the contributions of individual components and provide useful insights into the behavior and limitations of the proposed framework.

Critical weaknesses

1. Limited Methodological Novelty

The methodological novelty of the proposed framework appears limited. Most components rely on relatively straightforward extensions of existing techniques, including geometric constraint losses, residue-level supervision, and correction-based self-distillation. Furthermore, several theoretical results, particularly Lemma 2 and Proposition J.4, follow rather directly from the definitions of the proposed selection mechanisms. While the integration of these components is practically meaningful, the paper does not clearly establish a substantial new methodological or theoretical contribution beyond their combination.

The presentation is generally understandable, but the extensive theoretical exposition sometimes obscures the main methodological insights. Several results that follow relatively directly from the definitions are presented with substantial formal development, making the paper more difficult to follow than necessary.

2. Insufficient Validation of Epitope-Specific Binding Poses

Although ADEL is designed to generate macrocyclic peptide binders conditioned on a specified epitope, the evaluation does not sufficiently establish whether the generated peptides form plausible binding interfaces at the intended sites. The reported CαC_\alpha-RMSD measures peptide conformational similarity after optimal superposition and therefore does not assess receptor-relative placement. While Rosetta interface energies and interaction-related metrics provide useful evidence of interface quality, they do not directly establish epitope-specific binding. Additional analyses of peptide–epitope contacts, interface localization, and receptor-relative binding configurations would strengthen the evaluation of ADEL's target-conditioned binder design capability.

3. Limited Analysis of the Training Stabilization Mechanism

While the authors provide empirical evidence that truth anchoring improves training stability, the mechanism connecting the coordinate-level teacher-selection rule to stable parameter optimization remains insufficiently characterized. Appendix J discusses the role of gradient magnitude, Huber loss, and gradient clipping, while Lemma 2 establishes a coordinate-level descent property. However, the paper does not directly quantify how truth anchoring affects parameter-gradient norms or teacher-target drift during training. Additional diagnostics would help clarify the distinct contribution of truth anchoring to the observed training stability.

Questions for authors

Q1. Validation of Epitope-Specific Binding Poses

Can the authors provide additional quantitative evidence that ADEL-generated peptides form plausible binding interfaces at the specified epitopes?

In particular, I would appreciate an evaluation using DockQ, including its underlying components (Fnat, interface RMSD, and ligand RMSD), to assess the accuracy of the generated binding poses relative to experimentally determined complexes. Additionally, epitope contact precision and coverage would help establish whether the generated peptides are appropriately localized to the intended binding sites. Since the generated peptide sequences may differ from their native counterparts, the authors should clarify the residue-mapping procedure used for DockQ evaluation and interpret reference-based pose metrics alongside epitope-localization metrics.

Since ADEL is designed for de novo binder generation, I recognize that reproducing the native binding pose is not necessarily required, as alternative binding configurations may also be valid. Nevertheless, these complementary metrics would provide stronger evidence of the model's ability to generate meaningful epitope-conditioned binding configurations.

A comparison with baseline methods under the same evaluation protocol would be particularly valuable.

Q2. Mechanism of Training Stabilization by Truth Anchoring

The ablation study provides empirical evidence that removing the truth anchor leads to unstable training. However, the underlying mechanism remains insufficiently characterized. Lemma 2 establishes a coordinate-level descent property but does not directly guarantee stable neural network optimization.

Could the authors clarify the following points?

  1. Gradient behavior: How does truth anchoring affect parameter-gradient norms and variance during training? Given that gradient clipping is mentioned in Appendix J, were identical clipping settings applied to the anchored and unanchored models?

  2. Teacher-target drift: Is there direct evidence that truth anchoring prevents teacher predictions from drifting away from the ground-truth structures during training?

  3. Mechanism of stabilization: Can the authors distinguish the contribution of truth anchoring from that of existing optimization safeguards, such as Huber loss and gradient clipping?

These analyses would help clarify the specific mechanism underlying the observed training stabilization.

Q3. Clarification of the Methodological Contribution

While the integration of geometric constraints, linkage-chemistry supervision, and truth-anchored distillation is practically meaningful, the methodological novelty of the proposed framework remains somewhat unclear.

Could the authors clarify the key methodological insight that distinguishes ADEL from existing constraint-based generative models and self-distillation approaches?

In particular, Lemma 2 and Proposition J.4 appear to follow rather directly from the definitions of the proposed selection mechanisms. Beyond these theoretical properties, what is the principal methodological contribution of truth-anchored distillation?

A clearer explanation of the distinctive algorithmic insight, supported by the existing ablation results, would help establish the contribution beyond the effective integration of known techniques.

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

The motivation for this work is that designing macrocyclic peptides requires a synergy between the target design and inverse folding part of conventional design pipelines. Such peptides require a ring which is obtained by covalently binding two specific atoms and this constraint is hard to satisfy unless the sequence is known. The authors propose distilling the second denoising call of a corrector based on predict-project-renoise, while ensuring it remains close to the structure. The pipeline applies two geometrics constraints: closing bond length and C-alpha pseudo-angles flanking it, and ensures correct chemistry by supervising using cross entropy on residue linkage positions. All of these are added as a training objective of an EDM style denoiser, introduced in a staged schedule.

Critical strengths

Appendix J6 motivates the problem quite well (‘right shape, wrong element’).

The baselines are expansive and authors retrain seven published methods on CPSea split under type aware closure and a validity gate.

Critical weaknesses

Structural accuracy needs to be clearly reported so we can understand what the method costs in terms of quality. Only one metrics Ca-RMSD is reported in a few tables and ADEL performs lower than the boltzgen backbone it’s trained on (2.56 vs 2.45) on it. Fairness: ADEL initialises from pretrained BoltzGen weights while the baselines do not and in fact another method, RFPeptides, is excluded from comparison because of this exact reason. Moreover,  as per section 2.3 ADEL is trained on the very criteria it is evaluated on whereas no baseline is. And ADEL does not score higher on quantities which are outside its objective e.g. Ca-RMSD or Table 12’s w deviation, or Table 15’s head-to-tail closure). Multiple values are reported as ‘our full model’ in Table 1,6,23 and need conciliation or explanation. Moreover, it seems that the final numbers are best-of-three seeds whereas the baselines were only run a single time.  The authors consistently misattribute predict-project-renoise (PPR) by Rochman-sharabi and Louppe, 2026 to Huang et. al (2024).

Questions for authors

Please report Ca-RMSD for stage-1,2,3 checkpoints and Table 6. Please also report LDDT, T M-score, DockQ, interface RMSD/LDDT-PLI. Please report time and memory during inference against baselines so that  we can justify the cost for distillation.  Section L.1 defines the headline metrics( as being derived from ‘closed and valid’ so residue identity at linkage (third objective) is never checked. Can you confirm this is not the case? Counting a closed ring without checking whether correct chemistry is enforced goes against the central motivation of the paper.

From what checkpoint does Stage 1 initialise, and was that checkpoint trained on the PDB? Were the retrained baselines — BoltzGen in particular — initialised from released weights or from random initialisation? Given Table 24's "The base model inherits the BoltzGen backbone", please reconcile this with §K.8's framing and with the exclusion of RFpeptides from the matched comparison for using released weights.

How is the closing distance resolved for a design that does not carry the residue the native linkage requires at position i or j? Is a fallback atom pair used, and if so which? Specifically, how does PepFlow reach 0.261 closure on the 142 isopeptide targets and 0.365 on the 17 disulfide targets with no interface for a linkage specification? Could you report a strict closure column (required residues present and distance satisfied) alongside the current one?

1 comment

We sincerely thank the reviewer for their thoughtful comments and constructive feedback. We would greatly appreciate it if the reviewer could take a moment to consider our rebuttal and let us know whether it addresses their concerns.

We would be more than happy to provide further clarifications, additional analyses, or evidence if any questions or concerns remain. We look forward to your feedback and appreciate your time and consideration.

Reviewer_8SPTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ADEL, an all-atom sequence–structure codesign diffusion model that learns to satisfy this coupled geometric and chemical constraint. It provides a training signal redundant with the existing closure loss.

Critical strengths
  1. The task is interesting and the results seem to be promising.
  2. The idea of truth-anchored distillation with Engh–Huber-informed closure geometry and explicit linkage-chemistry supervision is reasonable.
Critical weaknesses
  1. Too many words are unclear .e.g., "Cα-referenced", "which we show provides", make this paper hard to read.
  2. The overall figure is also unclear. Too much details make it hard to get the key point.
Questions for authors
  1. Too many words are unclear .e.g., "Cα-referenced", "which we show provides", make this paper hard to read.
  2. The overall figure is also unclear. Too much details make it hard to get the key point.
1 comment

We sincerely thank the reviewer for their thoughtful comments and constructive feedback. We would greatly appreciate it if the reviewer could take a moment to consider our rebuttal and let us know whether it addresses their concerns.

We would be more than happy to provide further clarifications, additional analyses, or evidence if any questions or concerns remain. We look forward to your feedback and appreciate your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.