acceptodds
Withdrawn from ICLR 2027

AMP-UGen: Promoting Antimicrobial Peptide Generation through Peptide Understanding

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing antimicrobial peptide (AMP) generation models primarily learn the probability distribution of known AMP sequences. Although this paradigm can produce diverse AMP-like candidates, generation is commonly driven by sequence likelihoods, property labels, or predictor scores, without explicitly incorporating the functional and physicochemical knowledge captured by peptide understanding. We introduce AMP-UGen, an understanding-driven generation framework that turns such knowledge into learnable signals through two complementary mechanisms, Understanding-to-Design Transfer (UDT) and Understanding-Feedback Policy Optimization (UFPO). UDT transfers design-semantic states extracted by the peptide understanding model to a masked diffusion model, allowing peptide understanding to guide sequence construction. UFPO reuses the same understanding component to evaluate generated candidates and converts its assessments into feedback for updating the generation policy. Among the evaluated generation models, AMP-UGen achieves the highest predicted-positive rates under the frozen HydrMIC and AMPlify predictors, reaching 56.72% and 82.34%, respectively. During property optimization, UFPO reduces charge MAE from 1.183 to 0.311 and hydrophobic-moment MAE from 0.0273 to 0.0163. These results show that AMP-UGen improves predicted AMP quality and property control by integrating peptide understanding into both sequence construction and policy optimization, demonstrating its potential as a general framework for peptide generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antimicrobial peptide generation, peptide understanding, masked diffusion models, controllable generation, reinforcement learning, computational biology

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3c8jOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper present AMP-UGen a model that combines AMP understanding and AMP generation. The paper first finetune an LLM to understand peptide embeddings from ESM2. Then the paper proposes Understanding-to-Design Transfer (UDT) to aggregate peptide design query features from the finetuned LLM. The aggregated features are then conditioned into a masked diffusion model to generate peptide. To further transfer the understanding knowlege for peptide generation, this paper propose Understanding-Feedback Policy Optimization (UFPO) to train the diffusion model align with the three customized reward goals. Experiments demonstrate the effectiveness of the proposed AMP-UGen model.

Critical strengths
  • This paper is well-written and easy to follow.
  • The paper is well motivated. The methodology well justified the motivation.
  • The idea os reusing the understanding model in UFPO is a brilliant idea.
Critical weaknesses

Although the paper propose some interesting ideas. I do have some major comments on the evaluation part.

  • The paper heavily rely on existing models for evaluation as mentioned from lines 416 to 419. I do not think they are wrong. But since every ML models has bias towards certain distribution, solely rely on them might not enough to estimate the performance of the model. There could be more other quantitative and qualitative studies. For instance, the pLDDT score in co-folding model to evaluate the folding of the sequence (alpha helix, beta hairpin, or random coil). MD simulation could also be applied on certain amount generated sequences for simulating the interaction between sequences and membrane.

  • It is unclear to me how are charge success and moment success defined in the paper. To be clear, normally AMPs carry a net positive charge between 3 to 9. It is not the higher the better. It is because the membranes are usully negative charged. Similarly, AMPs usually has 30% and 50% hydrophobic residues. Over 50% might cause the sequence insoluble in water.

  • Fig 3 does not make a lot of sense to me. Since UPFO directly optimize Net charges and Hydrophobic moment, the distribution naturally shift to the target. But this does not directly prove the model is genreating new AMPs that satisfy the target and predicted antimicrobial quality (stated in line 475).

  • AMP generation baselines could be more comprehensive. Some similar reward based model can be included such as,

Tang, Sophia, Yinuo Zhang, and Pranam Chatterjee. "Peptune: De novo generation of therapeutic peptides with multi-objective-guided discrete diffusion." ArXiv (2025): arXiv-2412.

Cao, Hanqun, Haosen Shi, Chenyu Wang, Sinno Pan, and Pheng-Ann Heng. "GLID 2^ 2 e: A gradient-free lightweight fine-tune approach for discrete biological sequence design." Advances in Neural Information Processing Systems 38 (2026): 139875-139898.

  • Some typo and grammar errors noted. Some symbol in method section is not well defined. The presentation can be further improved, such as the placement of Table 1.

Overall, this is an interesting paper. I put a weak rejection for now. But I am willing to increase my rating if the rebuttal can address most of my concern.

Questions for authors
  • How is the charge and moment success defined.
Reviewer_3RB6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose a peptide generation model called AMP-UGen. It couples a peptide understanding model, with a masked diffusion generator. The language model used is a prot2text-v2, finetuned on AMP domain instructions. The training data is derived from ESCAPE dataset. Virtual tokens from the language model based on the design query are used for conditioning the diffusion generator. The sampled sequences are evaluated post generation using a composite reward that balances physicochemical properties along with log-likelihood from the text model. Since it uses a masked diffusion model, it uses a pseudo log-likelihood ratio. When compared against baselines, it produces the highest antimicrobial activity, and high uniqueness, novelty and hemolytic rate. They also show that UFPO allows finegrained control of physicochemical properties.

Critical strengths
  • Paper is well written and easy to follow
  • The method allows precise property control which is very important for AMP design
  • The surrogate is somewhat novel for diffusion models in the AMP domain
Critical weaknesses
  • The novelty is limited as it is a straightforward application of a known technique to the AMP domain
  • No wet lab experiments are done
  • The evaluation is confined to the sequence level with no folding, stability or membrane penetration experiments done even insilico
  • There is a drop in activity after UFPO that makes UFPO not so attractive. The tradeoff between reward and activity loss is not very clear.
Questions for authors
  • I understand that wet lab validation is beyond the scope of this paper. But is it possible to do insilico simulations to check whether the sequences fold to stable structures and do MD simulations?
  • Can you evaluate the tradeoff between UFPO and activity loss? How should this be balanced?
Reviewer_qmfmOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces AMP-UGen, a framework that uses peptide understanding to guide antimicrobial peptide (AMP) generation. The authors first adapt a protein understanding model to the AMP domain. Understanding-to-Design Transfer (UDT) then passes its design representations to a masked diffusion generator, while Understanding-Feedback Policy Optimization (UFPO) combines feedback from the same understanding model with calculated physicochemical properties to refine generation. The experiments assess peptide understanding, candidate quality, property control, and the contribution of the main components. The results show improvements in several predicted activity metrics over the evaluated baselines, as well as better adherence to requested properties through UFPO.

Critical strengths
  1. Reusing the understanding model for both generation conditioning and candidate evaluation is an appealing aspect of the work. UDT makes its representations available during sequence construction, while UFPO uses its assessments to guide subsequent optimization. This integration is methodologically interesting and provides a coherent way to connect functional knowledge with the physicochemical requirements of AMP design.

  2. Some ablation experiments are useful. For example, removing condition inputs, freezing the Peptide Design Queries (PDQs), or omitting understanding feedback reduces performance on the corresponding measures, which demonstrate the practical value of these components. Multiple frozen evaluators and repeated generation runs also give the reported improvements a broader basis than a single scoring model would provide.

Critical weaknesses
  1. The current ablations show that conditioning and interface adaptation matter, but leave open what AMP understanding knowledge contributes beyond general conditional encoding. The added value of UFPO's understanding-based reward over conventional predictor feedback also needs clearer justification. This makes it difficult to assess the framework's distinctive methodological contribution.

  2. The trade-off between property control and predicted activity deserves more discussion. In Table 2(b), UFPO improves constraint satisfaction but lowers HydrMIC and AMPlify positive rates from 56.72% to 37.35% and from 82.34% to 71.45%, respectively, compared with the base model. It would help to explain which design objectives justify this trade-off and how the overall benefit should be assessed.

  3. The evidence for semantic controllability remains limited to a few charge and hydrophobic-moment targets, with candidate-level success reported mainly for two configurations. These results support control within the evaluated settings, but do not yet establish how well the model handles unseen property combinations, other functional requirements, or richer design instructions.

  4. The study is lack of comprehensive baseline comparsions and the evaluations of the generated AMPs. For example, the related-work section discusses relevant models including TG-CDDPM, AMP-Designer, AMPGen, ProteinDT, DPLM, and CtrlProt, etc. However, the generation experiments mainly compare against AMP-Diffusion, Diff-AMP, HydrAMP, and an autoregressive AMP-SFT generator.

Questions for authors
  1. What do the authors mean by "peptide understanding knowledge" in this framework? Could a concrete example show what UDT contributes beyond activity labels, sequence length, and numerical property targets, and how this affects generation?
  2. What evidence supports using the same understanding model for condition interpretation and candidate evaluation, beyond the benefit of adding an AMP-related reward?
  3. How do the authors explain the improvement in property satisfaction alongside the decline in predicted activity in Table 2(b)? How should this trade-off be assessed when deciding whether UFPO improves on the base model?
  4. How do the evaluation targets relate to the UFPO training distribution, and what evidence supports generalization to unseen combinations? Are the design targets and settings comparable across generation baselines, or could differences influence the conclusions?

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

Related papers

Discussion (0)

Sign in to comment.