acceptodds
Under review as a conference paper at ICLR 2027

PepDiT: Structure-Tokenized Diffusion Transformers for Target-Spot Driven Antimicrobial Peptide Generation

Abstract

The rapid emergence of multidrug-resistant pathogens represents an urgent global health challenge, necessitating the de novo design of selective antimicrobial peptides (AMPs) and target-specific mini-binders. However, computational peptide design remains bottlenecked by three fundamental limitations: (1) short polypeptides (– residues) exhibit pronounced conformational plasticity and lack stable hydrophobic globular cores, rendering conventional protein structural tokenizers ineffective; (2) existing generative models sample sequences blindly from global natural distributions without spatial-chemical guidance toward pathogen-specific surface interaction hotspots; and (3) unguided generation often yields membrane-disruptive peptides with severe mammalian cytotoxicity and red blood cell hemolysis. To overcome these challenges, we introduce PepDiT, a unified structure-tokenized diffusion transformer framework for target-spot-driven AMP and mini-binder design with multi-objective preference alignment. We first curate -1.2M, a dataset of 1.2 million diverse polypeptide conformations, and train PepTokenizer, an -equivariant discrete geometric vector quantizer with torsional-dihedral codebooks that captures peptide backbone dynamics with sub-Angstrom precision. Building upon these discrete structural tokens, we formulate , a multi-track Diffusion Transformer that co-diffuses continuous sequence embeddings and quantized 3D geometric tokens. PepDiT incorporates a geometric Target-Spot Cross-Attention Module that conditions generative trajectories directly on 3D electrostatic, hydrophobic, and steric surface hotspots of pathogenic targets (such as outer-membrane BamA complexes, LPS Lipid A clusters, and Penicillin-Binding Proteins). Furthermore, we establish a Multi-Objective Direct Preference Optimization (Pep-DPO) post-training protocol to navigate the Pareto frontier between antimicrobial potency (pMIC), low hemolysis (), serum stability, and target binding affinity (). To systematically evaluate computational polypeptide models, we establish PepBenchmark, a standardized 12-task benchmark spanning structural reconstruction, broad-spectrum bacterial killing (across 7 critical ESKAPE and fungal pathogens), safety indices, and target pocket engagement. Extensive experiments demonstrate that PepDiT establishes state-of-the-art performance across all 12 tasks, achieving an unprecedented Selectivity Index (), an improvement in target-spot binding hit rate, and picomolar-to-nanomolar binding affinities against challenging antimicrobial targets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Peptides Design

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_cbbHOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PepDiT, a diffusion model for antimicrobial peptide generation. PepDiT features a Target-Spot-Driven Hotspot Conditioning mechanism to enable the target-aware generation process. The paper also proposes a DPO-based post-training optimization to further improve the model performance. Critically, to train the PepDiT model, the paper curated PepDB-1.2M, a large scale peptide dataset containing MD conformational data, cyclic macrocycles, and AlphaFoldMultimer target complexes.

Critical strengths
  • The PepDB-1.2M is a large scale peptide dataset. This dataset merge several existing dataset with additional labels. This dataset could benefit the research community and further attract future researchers.
Critical weaknesses
  1. I did not found a significant technical novelty in this paper. The paper claim PepDiT is a "a multi-track Diffusion Transformer that co-models discrete 3D structural tokens and continuous sequence latents with target-spot cross-attention." (lines 122, 123). However, such multi-modal generation is widely studied in existing works [1,2,3]. Similar argument is applied on Pep-DPO. It is basically an application of well-established DPO algorithm.

  2. The paper claims the PepBench as a major contribution which includes 12 comprehensive tasks. However, 1) it is very unclear to me what are the 12 comprehensive tasks? Many tasks in section 4.1 are just evaluation metrics, such as RMSD, MIC, and MAE. These are not tasks. 2) It is not clear how the paper obtain certain values, for instance MIC and HC50 and Human Serum Stability half-life. Usually those values need to have wet-lab verification.

[1] Didi, Kieran, Zuobai Zhang, Guoqing Zhou, Danny Reidenbach, Zhonglin Cao, Sooyoung Cha, Tomas Geffner et al. "Scaling atomistic protein binder design with generative pretraining and test-time compute." In International Conference on Learning Representations, vol. 2026, pp. 61056-61109. 2026.

[2] Wang, Yongkang, Xuan Liu, Feng Huang, Zhankun Xiong, and Wen Zhang. "A multi-modal contrastive diffusion model for therapeutic peptide generation." In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 1, pp. 3-11. 2024.

[3] Campbell, Andrew, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. "Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design." arXiv preprint arXiv:2402.04997 (2024).

Questions for authors
  1. How is MIC and HC are obtained both in DPO stage (line 336 to 338) and evaluation stage for generated candidates.

  2. Can the author propose more ablation study on the tokenizer to verify the effectiveness. The tokenizer is one of the major contribution in this paper. But I did not see any ablation study. It would be helpful to understand this component by removing certain part and varying the codebook size and dimension.

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

The paper introduces PepDB-1.2M curated database and PepTokenizer trained on this corpus. Moreover the paper introduces peptide Diffusion Transformer architecture and Multi-Objective Preference Post-Training. The paper also formulates PepBenchmarch, a standardized 12-task benchmark.

Critical strengths

The paper clearly introduces a problem of generating antimicrobial peptides.

Critical weaknesses
  1. The Authors cite HydraAMP (Szymczak et al., 2023) and HydraAMP-v2 (Szymczak et al., 2025) and FoldFlow-2 (ICLR 2025) (Bose et al., 2025), with the following bibliography entries:

Michal Szymczak, Matthew Nelson, et al. Hydraamp: High-throughput de novo generation of antimicrobial peptides with controlled toxicity and broad-spectrum potency. Nature Machine Intelligence, 5:845–857, 2023.

Michal Szymczak, Matthew Nelson, et al. HydraAMP-v2: Disentangled multi-property diffusion transformers for non-hemolytic antimicrobial design. Nature Machine Intelligence, 7:112–126, 2025.

Joey Bose, Avishek Bose, Hugo Darracq, Leo Chen, et al. FoldFlow-2: Riemannian flow matching on SE(3) manifolds for generative macromolecular assembly. International Conference on Learning Representations (ICLR), 2025.

Neither of these papers I can find in the internet. Nevertheless, the Authors include HydraAMP-v2 and FoldFlow-2 as a baseline in Table 2. This raises serious concerns about the validity of the entire evaluation process.

  1. In Multi-Objective Polypeptide Reward Function The Authors introduce a reward RpMIC(s)\mathcal{R}_{pMIC}(s) which uses predicted MIC. There is not written how the MIC is predicted. Is it predicted with some surrogate model. Which one exactly? Additionally in experiments The Authors also claim to measure MIC of predicted candidates. There is not written how they measure the MIC.
Questions for authors

See weaknesses.

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

The paper introduces PepDiT, a framework for generating antimicrobial peptides conditioned on a pathogen's 3D binding site, assembled from four claimed contributions. PepDB-1.2M is a structural corpus of 1.2×1061.2 \times 10^6 polypeptide conformations; PepTokenizer is an SE(3)\mathrm{SE}(3)-equivariant vector quantizer discretizing those structures into a K=1024K = 1024 codebook; a dual-track Diffusion Transformer then flow-matches sequence simplex embeddings ut\mathbf{u}_t alongside structural token embeddings zt\mathbf{z}_t, with cross-attention to target surface features; and Pep-DPO applies preference alignment under a fixed-weight scalar reward. The authors additionally propose PepBenchmark, described as 12 tasks across four domains. Evidence is confined to two tables: reconstruction fidelity of 0.42 A˚0.42\,\text{Å} RMSD, and Rosetta-scored binding on three targets giving a 58.6%58.6\% hit rate against BamA, reported as an 8.4×8.4\times improvement. All quantities are computationally predicted; the submission includes no wet-lab validation, no ablations, no released code, and no appendix.

Critical strengths

1. The motivation and neediness are sound. Tokenizers trained on globular structures from AFDB and the PDB do degrade on short peptides that never settle into a stable fold, and existing AMP generators do condition on scalar activity labels rather than on binding-pocket geometry. The BamA lateral gate, PBP2a, and Lipid A are reasonable clinical targets to pursue.

Critical weaknesses

1. The target structure has no public coordinates. PDB 7T16, named in §4.1 as the A. baumannii BamA target, is held by RCSB under status WDRN: withdrawn and never released.

2. The lone comparative figure cannot be obtained from the table. ESM-3 at 9.6% gives 6.1×, PepFlow at 14.8% gives 4.0×, and neither amounts to the 8.4× claimed throughout.

3. The equations are substantively, not cosmetically, wrong. For example: Eq. 3 sets aside the invariants constructed in Eqs. 1-2 while separately accepting s_i, so the tokens end up neither invariant nor purely structural; Eq. 7 passes no conditioning from c_tgt to the cross-attention, the only sublayer that sees the target; please check other Equations carefully too.

4. The method cannot deliver the Pareto behavior it advertises. Eq. 12 is fixed-weight scalarization, which returns one point, whereas Figure 1(d) promises a "4-dimensional Pareto frontier."

5. The corpus description is inaccurate. Its four components total exactly 1,200,000, and the largest is credited to complexes from AFDB, which holds predicted monomers. With 450,000 MD ensemble members and a split filtered on sequence identity alone, conformers of one peptide can fall on both sides.

6. Many citations fail to resolve, and several misreport real work. Six of them supply baseline figures in Tables 1 and 2; the full list is going to the AC. The effect compounds, as Section 2 characterizes prior work in ways its own citations do not bear out.

7. Nothing in the paper can therefore be verified. The method as specified cannot be run, and the experiments reported do not test the claims made for it.

Questions for authors

Unprofessional behavior. In iclr 2027 guideline, https://iclr.cc/Conferences/2027/AIPolicyForAuthors

Ultimately the paper’s authors are responsible for the contents of their submissions. Consequently, a substantial falsehood, instance of plagiarism, or misrepresentation produced by an LLM would be considered a Code of Ethics violation on the part of the paper’s authors, and might lead to desk rejection of the paper.

There are too many hallucinations in references then I always feel uncomfortable and am confused in reading this paper. Of course, many related works are mismatched to corresponded references then this paper cannot propose its methodological background properly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.