Review Papers Claim by Claim

 Quick Explanation Copied

 Long Explanation

What MAOMAO delivers

MAOMAO harmonizes peptide-toxicity annotations from 54 databases, literature-derived resources, and curated datasets into a provenance-aware, ontology-structured resource: 71,701 unique sequences spanning seven toxicity endpoints and 501,907 sequence-endpoint combinations MAOMAO preprint, abstract. From 1,289,119 collected source-level records, 850,642 were retained after preprocessing (~34% reduction), and only peptides of 5–70 residues with canonical amino acids were kept Dataset overview, Section 3.1.

The resource's distinguishing features are explicit evidence states (positive/negative/ambiguous/unlabeled/no information), hierarchical toxicity ontology (Toxic → Cytotoxic{Hemolytic, Cytolysis}, Neurotoxic, Embryotoxic, Ichthyotoxic), 41 physicochemical descriptors, 10 protein language model embedding sets plus one-hot baseline, and precomputed 5-fold benchmark partitions across 30 seeds for four endpoints Resource architecture, Sections 2.5 and 3.3. Zenodo archiving (CC BY 4.0) and MIT-licensed code support reproducibility Data availability statement.

Negative states dominate (77,843 of 121,891 classified states), partly reflecting weak negatives from random control sets — a subtle but important caveat for supervised learning, since weak negatives can inject label noise Conflict resolution, Section 2.3.

Critical assessment

• Endpoint imbalance is severe. Embryotoxic (2 states) and Ichthyotoxic (5 states) are essentially placeholder categories; Neurotoxic has only 3,183 states. Only Toxic, Cytotoxic, Hemolytic, and Neurotoxic support benchmarking Endpoint counts, Section 3.2.
• Redundancy by design. 91.1% of cytotoxic-positive sequences overlap with hemolytic positives (4,279 shared), and 66.58% of positive peptides span multiple categories — useful for multi-task learning but risky for naive per-endpoint training where the hierarchy inflates apparent support Endpoint overlap analysis, Section 3.2.
• Leakage not controlled. The authors correctly warn that random/stratified splits include no homology or similarity control; models benchmarked on these folds will likely be over-optimistic Limitations, Data Value section.
• Metadata gaps. Licensing exists for only 13 of 54 sources (24.07%); downstream redistribution of third-party content may therefore be constrained despite the CC BY 4.0 wrapper Metadata completeness, Section 3.1.
• No external validation yet. As a data paper, no predictive models were trained or evaluated; the resource's utility claim rests on structure, not demonstrated downstream performance (BGPT inference).

Bottom line

MAOMAO is a substantive contribution to data-centric peptide toxicology: it converts a fragmented annotation landscape into a traceable, machine-readable resource with unusually honest treatment of ambiguity and provenance. Its principal unresolved risks are label noise from weak negatives, hierarchy-induced redundancy, and split leakage — all fixable in future releases via similarity-aware partitions and endpoint-specific evidence weighting. Confidence: moderate-high for reported counts (directly stated, versioned archive available); moderate for reuse claims (untested by the authors).

 BGPT Paper Review

Study Novelty

First ontology-guided, evidence-aware, FAIR harmonization of peptide toxicity data across 54 sources; earlier peptide databases (e.g., PeptiPedia, DRAMP) were endpoint-specific and lacked explicit ambiguity/provenance states, though data-harmonization frameworks per se are well established.

Scientific Quality

Systematic collection with documented eligibility criteria, explicit evidence states, versioned Zenodo archive, and MIT code support quality. Weaknesses: no empirical downstream validation, weak-negative label noise risk, minor reference extraction artifacts (e.g., misdated citations), and license coverage for only 24% of sources.

Study Generality

Focused on peptide toxicity endpoints, but the ontology-plus-provenance framework generalizes to other peptide property harmonization tasks; however, sparsity of three of seven endpoints limits actual breadth.

Study Usefulness

Directly enables reproducible benchmarking, positive-unlabeled learning, multi-task toxicity modeling, and dataset construction that previously required extensive manual curation.

Study Reproducibility

Full Zenodo archive, GitHub workflows, documented harmonization rules, and fixed random seeds enable re-execution; caveat: no independent reproduction reported yet and license gaps may impede redistribution of some source content.

Explanatory Depth

Descriptive data resource with careful methodology, but no mechanistic analysis of toxicity itself; documentation depth resides in supplementary sections not fully presented in the main text.