Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript provides a comprehensive and well‑organized critical review on artificial‑intelligence‑assisted discovery and development of therapeutic peptides. It systematically covers different therapeutic peptide categories, molecular representations, benchmark quality, generative and structure‑aware modelling. This timely review is scientifically sound and supported by abundant recent references, with informative tables clarifying key concepts. Therefore, I would recommend the publication of this article in Biomolecules after minor revision. 1. Page 3, Table 1; Page 19, Table 3; Page 29, Table 5

On Page 3, Table 1 splits “immunomodulatory” into “immunomodulato / ry”. Multiple table headers and cells in Table 3 on Page 19 have excessive word breaks. It is suggested to appropriately adjust column widths, simplify text within cells, and set long descriptions to left alignment.

2. Page 3, Table 1; Page 31, Table 7

In Table 1, the row containing “Self-assembling and material-forming peptides” spans Pages 3–4. The first row of Table 7 is split across Pages 31–32. Although table headers are repeated on subsequent pages, some content is disconnected from its corresponding row labels. It is recommended to prevent page breaks within table rows and add the label “Table X (continued)” on continued pages.

3. Page 2, Line 46

Section 2 "Design Space" and Sections 5-8 have obvious repetitions in the discussion of various peptides, such as AMP, CPP, vaccine peptides, self-assembling peptides, etc., which appear repeatedly in multiple places, resulting in redundant descriptions. It is recommended to integrate the expressions or cross-reference them.

4. Page 6, Line 35; Page 25, Table 4; Page 25, Line 6

PDCNet, Pepti-Agent, Boltz-1, Chai-1, etc. are all preprints. It is recommended to reduce excessive reliance on preprints.

5. For greater intuitiveness and to help readers, please provide a table giving the detailed amino acid sequences of representative therapeutic Peptides.

Comments on the Quality of English Language

The English would benefit from professional language editing to improve clarity and readability.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

 (1) Summary :  
This is a well-intentioned and clearly argued critical review whose central theses are correct and worth stating:
(i) therapeutic peptides are chemically defined molecular constructs, not 20-letter strings; (ii) representation and evaluation are inseparable, with computational rigor (M0–M3) and translational depth (T0–T3) being independent axes. The peptide-specific reporting requirements are genuinely useful, and the domain statements on AMP shortcut learning and MHC “binding ≠ presentation ≠ immunogenicity” are accurate.

However, coverage of the most consequential 2022–2025 methodological advances is incomplete for a 2026 survey of this scope: RFdiffusion, RFpeptides, ProteinMPNN, ProtGPT2, ESM-2/ESM-3, AlphaFold3, ProGen/ProGen2, PepFlow and PepGLAD are absent from the abstract, Table 4 contents, and section summaries, while the representative-method list reads as a convenience sample rather than a systematic selection. With only 86 references across 8 classes, 5 representation families, workflows, datasets and a 16-domain checklist, the evidentiary base is too thin for the claimed breadth. The “Peptide-AI Evidence Grid” is a useful reframing but largely a repackaging of DOME/TRIPOD+AI/TRL-style maturity concepts, and its novelty should be calibrated accordingly. The core framework is sound and salvageable; the literature integration and novelty claims require more revision.

(2)Questions :
(i) What inclusion criteria governed the selection of methods in Tables 3–4 (HELM-GPT/CycloPepper included vs RFdiffusion/ProteinMPNN excluded)?
(ii) How do you position the grid against maturity/validation frameworks beyond DOME and TRIPOD+AI ?

(iii). Can you substantiate “the M3 column is almost empty” with an explicit tally?

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This manuscript is a review by a single author that organizes AI applications in peptide discovery by therapeutic class. I appreciate this apporach, and it would be an excellent resource for young researchers. The author argues that receptor agonists, PPI inhibitors, membrane-active peptides, delivery vehicles, antigens, conjugates, macrocycles, and self-assembling materials are distinct computational problems, each with unique input variables, label semantics, and validation experiments. The review contributes a Peptide-AI Evidence Grid assessing methodological rigor (M0–M3) separately from translational depth (T0–T3) and a peptide-specific reporting checklist based on DOME, TRIPOD+AI, and FAIR principles.

Manuscript is class based, linked with three relevant diagnostic questions, keepng the endpoitn in mind. Sections 4 and 5.3 includes arguments, addressing topics often overlooked in AI-design reviews, such as manufacturability and FDA guidance.

However, the manuscript fails to meet its own standards for evidence. How its evidence base was constructed? Suggetsed to be included: search strategy, inclusion criteria, or count of screened records. This affects the validity of claims made about thirty studies. With therapeutic relevance, provide details on what has reached the clinic. Including a table of AI-involved peptide candidates by development stage and outcomes would strengthen the argument.

A manuscript invoking FAIR should have recommended resources linked:

1. Table 2: URL, version and access date for every database (THPdb2, Peptipedia v2.0, PeptiVerse, CancerPPD2, PDCdb, CPPsite 2.0, IEDB, CyclicPepedia, CycPeptMPDB, CREMP, DRAMP 4.0, DBAASP v3, CAMPR4, PepBenchmark).
2. Tables 3 and 4: a code/repository column (PepMLM, EvoBind2, HydrAMP, RFpeptides, HELM-GPT, CycloPepper, CamSol-PTM, PeptideCLM, NetMHCpan). Mark absences as they support the argument.
3. DOME, TRIPOD+AI and FAIR: link the checklist pages, not only the citing articles.

Roughly 1/3 of the text can be omitted without loss of content. Table 1's failure-mode column is restated across 2.1–2.7, consider keeping one. The "X is not Y" antithesis is the idea = state it once, then instantiate per class not more. Merge 5.5 ->5.1; compress Section 11 into a numbered priority list. For the large Tables I suggest a clearer subsectioning, so it would not be a block, but something to be easily used.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The revisions have addressed the concerns.