PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction

Name: Verified Rigor Record — PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction
Item: PeptiVerse: A Unified Platform for Therapeutic Peptide Property Prediction
Rating: 100
Author: Αlpha¹

Zhang, Y.; Tang, S.; Chen, T.; Mahood, E.; Vincoff, S.; Chatterjee, P.

Paper source

Read PDF

Αlpha¹ Rigor Review

Rigor Review Verified

Critical

High-priority

Action items to resolve · Critical / High

Confidence: moderateEngine: 7.6.0Source: full text

This Alpha1 Rigor Review uses AI Rigor Reviewers trained on a curated corpus of high-fidelity and retracted papers, with expert supervision and curation. It can still make mistakes; verify each finding against the source before relying on it.

PeptiVerse presents a well-motivated computational platform for therapeutic peptide property prediction, with strong scientific premise, adequate resource identification, and transparent reporting. However, the study is weakened by incomplete statistical reporting (threshold p-values, no confidence intervals, no assumption verification) and missing outlier handling documentation, which limit the reader's ability to fully assess the reliability of the reported performance.

All three independent reviewers broadly agreed on most dimensions; the main divergence was on statistical analysis (1 fail, 2 pass), resolved to warn based on a checklist-level calculation showing 3 of 7 sub-criteria adequate. The statistics verification component checked only 2 tests (both consistent) and does not validate the bulk of reported results. Citation verification found no retracted or unlocatable references among 76 checked.

12 major claims checked against the paper's own evidence: all adequately supported.

partialSection 2.4 and Table 1Reviewer 1
PeptiVerse achieves state-of-the-art performance across diverse property prediction tasks.
PeptiVerse outperforms prior methods on several tasks (e.g., permeability, solubility) but is not the best on SMILES binding affinity; the paper explains this is due to evaluation protocol differences. The claim is partially supported but caveated.
Evidence: Table 1 shows PeptiVerse achieves higher F1/AUC on many tasks, but on SMILES binding affinity, PepLand reports higher Spearman correlation (0.768 vs 0.582).
“PeptiVerse achieves competitive or superior performance under a unified evaluation protocol.”
partialAbstract, Introduction, Results 2.4Reviewer 2
PeptiVerse is the first unified framework that predicts multiple peptide properties for both amino acid sequences and SMILES.
The paper presents PeptiVerse as the first unified multimodal platform, but acknowledges prior systems like PepLand and PeptideDashboard; novelty claim is partially supported by a head-to-head comparison but the 'first' claim may be overstated.
Evidence: Table 1 shows PeptiVerse outperforms or matches prior tools on several benchmarks, but the paper itself notes PepLand covers a 'limited subset of properties' and is modality-specific.
“PeptiVerse is, to our knowledge, the first unified framework capable of predicting multiple physicochemical and developability properties for both amino acid sequence inputs and SMILES-encoded peptide inputs.”
partialAbstractReviewers 2, 3
PeptiVerse delivers state-of-the-art performance across diverse property prediction tasks.
PeptiVerse achieves competitive or superior F1 and Spearman ρ on benchmarks, but the comparison is against prior tools that used different split strategies (random vs. similarity-aware), making direct superiority claims slightly overstated.
Evidence: Table 1 shows PeptiVerse outperforms PeptideBERT and PepLand on several tasks under similarity-based splits.
“delivers state-of-the-art performance across diverse property prediction tasks”
supportedAbstract and Section 2.4Reviewer 1
PeptiVerse is the first unified framework for predicting multiple properties for both amino acid and SMILES peptide inputs.
The paper cites prior tools that are limited to one modality or property, and PeptiVerse demonstrably handles both. The claim is supported.
Evidence: The paper provides a comparison table (Table 1) and discusses prior tools in Sections 1 and 2.4.
“PeptiVerse is, to our knowledge, the first unified framework capable of predicting multiple physicochemical and developability properties for both amino acid sequence inputs and SMILES-encoded peptide inputs.”
supportedSection 2.3, paragraph 2Reviewer 1
Structure prediction confidence (ipTM) does not correlate with peptide binding affinity.
The paper presents analysis showing negligible correlation (|ρ| ≈ 0.05) for both amino acid and SMILES inputs, with sufficient sample sizes. The claim is supported.
Evidence: Section 2.3 reports Spearman ρ ≈ 0.05 for 1,436 and 1,597 pairs.
“ipTM showed negligible association with experimental binding affinity across either peptide representation (|ρ| ≈ 0.05).”
supportedSection 2.3, paragraph 3Reviewer 1
PeptiVerse binding affinity predictions correlate with experimental measurements (ρ=0.58, 0.56, p<10^-3).
The reported correlations are statistically significant and based on a reasonable sample size. The claim is supported, though the exact p-value is not given.
Evidence: Section 2.3 reports Spearman ρ = 0.58 (AA) and 0.56 (SMILES) with p < 10^-3.
“achieving Spearman ρ = 0.58 for amino acid sequence inputs and ρ = 0.56 for SMILES-based inputs (all p < 10^-3).”
supportedSection 2.2, paragraph 3Reviewer 1
Simple models (e.g., XGBoost) with pretrained embeddings are sufficient for peptide property prediction, reducing computational cost while improving generalizability.
The paper shows that multiple architectures perform similarly, and XGBoost is consistently competitive. The claim is supported by the results in Section 2.2.
Evidence: Section 2.2 states that tree-based boosting models achieve competitive performance and offer advantages in stability and efficiency.
“tree-based boosting models such as XGBoost consistently achieve competitive performance across tasks while offering advantages in training stability, computational efficiency, and ease of deployment.”
supportedResults 2.3Reviewer 2
Binding affinity predictions from PeptiVerse show statistically significant agreement with experimental measurements.
Spearman ρ = 0.58 (AA) and 0.56 (SMILES), both with p < 0.001, clearly support this claim.
Evidence: Results 2.3 reports ρ = 0.58 and 0.56, both p < 10^-3.
“PeptiVerse binding affinity predictions show statistically significant agreement with experimental measurements across both representation modalities, achieving Spearman ρ = 0.58 for amino acid sequence inputs and ρ = 0.56 for SMILES-based inputs (all p < 10^-3).”
supportedResults 2.3Reviewer 2
Structure-based confidence (ipTM) is an insufficient proxy for peptide-protein binding strength.
The paper presents a direct negative result: ipTM vs. experimental affinity shows negligible correlation (|ρ| ≈ 0.05) across 1436/1597 pairs, supporting the claim.
Evidence: Results 2.3: 'ipTM showed negligible association with experimental binding affinity across either peptide representation (; |ρ| ≈ 0.05)'
“ipTM showed negligible association with experimental binding affinity across either peptide representation (; |ρ| ≈ 0.05), indicating that structure confidence metrics are insufficient proxies for peptide–protein binding strength.”
supportedResults 2.2Reviewer 2
Simple models (e.g., XGBoost) generalize comparably to deep architectures when using high-quality embeddings.
Results 2.2 and Figure 3 show that across tasks, performance differences between architectures are modest, with XGBoost often competitive; this is supported by empirical results.
Evidence: Results 2.2: 'overall performance differences between architectures were modest when trained on fixed, information-rich embeddings derived from ESM-2 and PeptideCLM' and 'tree-based boosting models such as XGBoost consistently achieve competitive performance across tasks.'
“overall performance differences between architectures were modest when trained on fixed, information-rich embeddings derived from ESM-2 and PeptideCLM”
supportedAbstractReviewer 3
PeptiVerse is a unified platform for therapeutic peptide property prediction supporting both amino acid sequences and SMILES representations.
The paper describes and implements the platform, providing a web interface and open-source code that accept both input types.
Evidence: Methods section 4.1 and Results section 2.5 describe the dual-input architecture and web interface.
“PeptiVerse accepts either amino acid sequences or chemically modified peptide SMILES, delivers state-of-the-art performance across diverse property prediction tasks, and provides both a web interface and open-source implementation”
supportedDiscussionReviewer 3
Simple models (e.g., XGBoost) on frozen embeddings are sufficient for practical peptide property prediction.
The paper shows that across tasks, lightweight classifiers achieve competitive performance with deep architectures when using strong pretrained embeddings, supporting this claim.
Evidence: Results section 2.2: 'tree-based boosting models such as XGBoost consistently achieve competitive performance across tasks' and Figure 3 shows best models are often XGBoost or SVM.
“PeptiVerse shows that foundational embeddings paired with simple, well-regularized classifiers are sufficient (and often superior) for practical peptide property prediction.”

Recomputed 2 tests: 2 consistent, 0 inconsistent, 2 via agent-written checks.

1 integrity concern flagged (0 high).

Rigor

Eight Dimensions6 pass

12 major claims3 partial9 supported

10 copyedit issues flagged (1 major): mostly consistency, clarity, punctuation.

Checked 76 references: 41 verified — 35 not checked.

Reproducibility

3 data/code links checked; 3 live.

dataHugging FaceLIVEHTTP 200
https://huggingface.co/ChatterjeeLab/PeptiVerse
Resolves to Hugging Face (data repository).
dataHugging FaceLIVEHTTP 200
https://huggingface.co/spaces/ChatterjeeLab/PeptiVerse
Resolves to Hugging Face (data repository).
codeGitHubLIVEHTTP 200
https://github.com/ynuozhang/Peptiverse_R.git
Resolves to GitHub (code repository).

Alpha¹ Rigor Action ItemsFinal pass before journal submission

As a post-publication audit, the published work is robust in its overall design and transparency, but an informed reader should weigh the incomplete statistical reporting (threshold p-values vs. exact values, missing confidence intervals, unverified assumptions) and the internal contradiction between text and Table 1 for binding affinity Spearman ρ values. These gaps are not severe enough to warrant retraction but weaken the precision of the reported performance claims and could benefit from a correction or erratum.

1.
HIGHstatistics
Resolve the internal contradiction between the text in Section 2.3 (Spearman ρ = 0.58 for AA, 0.56 for SMILES) and Table 1 (ρ = 0.557 for AA, 0.582 for SMILES) by using the exact, unrounded values in both places.
This inconsistency undermines the reader's trust in the reported results and could be perceived as a reporting error.
2.
HIGHstatistics
Report exact p-values (e.g., p = 2.3e-4) instead of threshold statements (p < 10^-3) for all Spearman correlation results in Section 2.3.
Threshold p-values convey less information and prevent readers from assessing the strength of evidence.
3.
HIGHstatistics
Add 95% confidence intervals for all reported effect sizes (Spearman ρ, Pearson r, F1, AUC) throughout Section 2.3 and Table 1.
Confidence intervals are essential for readers to gauge the precision and stability of performance estimates.
4.
HIGHrigor
Add an explicit statement on outlier handling for each property dataset in Section 4.1 (e.g., whether any extreme values were excluded, winsorized, or retained and why).
The absence of outlier handling documentation is a recognized gap across all three reviewers and raises concerns about data preprocessing decisions.
5.
MEDIUMstatistics
Add a brief statement in Section 4.2 (or a supplement) verifying that regression model assumptions were checked (e.g., normality of residuals, homoscedasticity) or justify the use of non-parametric metrics if assumptions were not met.
Assumption verification is a standard expectation for inferential statistics and is currently unreported.
6.
MEDIUMstatistics
Add individual data points (e.g., scatter plots or dot plots) to Figures 3-5, or explicitly define what the error bars represent (SD, SEM, or CI) in the figure legends.
Bar charts without individual data points or defined error bars obscure the distribution of the data and are a commonly criticized reporting practice.
7.
MEDIUMdata code
Obtain a permanent DOI for the code repository (e.g., via Zenodo) and update the Data and Code Availability section accordingly.
A DOI ensures persistent and citable access to the code, which a GitHub URL alone does not guarantee.
8.
MEDIUMreporting
Correct the author list on the title page to include spaces or commas between names (e.g., 'Zhang Yinuo, Tang Sophia, ...').
The current concatenated format appears to be a formatting artifact that must be fixed before formal publication.
9.
MEDIUMreporting
Remove duplicate URLs throughout the Data and Code Availability section and Reference list.
Duplicate URLs clutter the text and suggest a formatting or copy-paste error.
10.
MEDIUMreporting
Remove the extra space before the comma in the Abstract (change 'PeptiVerse , a universal' to 'PeptiVerse, a universal').
This is a minor punctuation error that should be corrected for a clean presentation.
11.
MEDIUMreporting
Correct the URL typo 'githubz' to 'github.com' in the OpenFold3 reference.
A broken reference URL undermines reproducibility and reflects poorly on the manuscript's preparation.
12.
LOWreporting
Add a brief note about the reproducibility of the computational environment (e.g., a conda environment file or Docker container) to the Data and Code Availability section.
Specifying the exact computational environment further supports independent re-use and reproducibility.

Full text (59,621 chars)Analyzed: Jul 7, 2026

Αlpha¹ assesses methodological rigor, not the importance of the findings. See how we calibrate →