A transparent path from sequence to source

Purpose and scope

The atlas supports scientific discovery by connecting peptide sequence strings with experimental, benchmark, and bibliographic context. Its data model keeps fields for sequence, name, chirality, origin, physicochemical category, terminal and chemical modifications, uptake description, model, cell line, cargo, patent, and PubMed identifier.

This is a positive-only discovery atlas. It excludes prediction-only candidates, generated negatives, and random decoys, so it is not suitable as a classifier training set without independently sourced negative examples.

Data provenance

The explorer uses CPPsite3 as its record-level core, then compares exact, case-preserved sequence strings against POSEIDON and the positive datasets published with CellPPD and CellPPD-Mod. Sources were retrieved on 1 November 2026.

• 7,271 searchable records in the merged atlas.
• 4,763 exact, case-preserved sequence strings.
• 6,788 core CPPsite3 experimental-context records.
• 483 sequence records contributed by the supplemental positive datasets after exact deduplication.
• 1,859 distinct valid PubMed identifiers; source placeholder 0 values are treated as missing.
• 6,881 records have a non-placeholder PMID (94.6% coverage).

Full machine-readable registry with source revisions: sources.json. Licenses and citations: ATTRIBUTION.txt. The imported CPPsite3 CSV is pinned by SHA-256; the other three source snapshots are pinned by Git commit.

Records contributed after exact deduplication

Evidence-class labels

Evidence class is an atlas-assigned descriptive label based on the source collection; it is not a quality score or a claim of independent validation.

• CPPsite3 curated record: a record curated by CPPsite3 from reported experimental literature.
• POSEIDON quantitative uptake: a POSEIDON row with a reported uptake value and experimental context.
• CellPPD positive benchmark: membership in an imported positive CellPPD file; negatives and generated decoys are excluded.
• CellPPD-Mod positive benchmark: membership in an imported positive modified-peptide training or validation file; negative sets are excluded.

How the interface works

1. Build-time normalization. Source CSV, semicolon-delimited, and positive RTF datasets are parsed without external packages; empty values are normalized and mapped to explicit property names.
2. Conservative inclusion. CPPsite3 records, POSEIDON uptake records, and only positive CellPPD/CellPPD-Mod examples are eligible. Predicted candidates, generated negatives, and randomly sampled decoys are excluded.
3. Exact deduplication. Sequence strings are trimmed and compared after removing formatting whitespace. Case and punctuation are retained because lowercase letters can encode D-amino acids or other modified-residue notation.
4. Provenance merge. When a sequence occurs in more than one source, all matching datasets are shown. Supplemental rows are added only when the exact string is absent from the core.
5. Browser-local indexing. Search text is created after the JSON loads. The application keeps search and filter state in the URL fragment rather than sending those values to the host; ordinary page and data requests still reach the hosting provider.
6. Combinable filters. Keyword, motif, case-preserved exact sequence, atlas precedence dataset, category, origin, chirality, architecture, and length filters are applied together. Broad motif matching ignores case for recall; exact matching retains case so D-residue notation is not conflated with uppercase forms.
7. Source-first details. Record values are displayed without scientific reinterpretation, and numeric PMIDs link directly to PubMed.
8. Portable export. A visitor can download the current filtered result set as CSV for local analysis.

Fact, explanation, and citation policy

Dataset facts are values or counts taken from, or calculated directly from, the attributed source files. Scientific claims in the primer are linked to peer-reviewed publications. Editorial explanation includes the page structure, plain-language framing, diagrams, and interpretive cautions written for this rebuild.

The interface does not silently convert a curator’s proposed mechanism into an established mechanism. Source field labels such as “uptake efficiency” and “proposed uptake mechanism” remain visibly qualified.

Security boundary

This implementation is static: it has no login, registration, database server, file upload, contact form, or server-side write surface. A restrictive meta Content Security Policy is declared, user-controlled data is inserted through text-safe DOM APIs, and runtime scripts and data requests are limited to this site.

Search and filter values are processed by the application in the browser and stored in the URL fragment. The hosting provider still receives ordinary page and data requests. The current host does not expose response-level CSP, HSTS, Referrer-Policy, or X-Content-Type-Options headers; the meta CSP cannot enforce frame-ancestors.

Privacy

This edition does not load third-party analytics or analytics cookies. Search and filter values are processed locally in the browser. The hosting infrastructure still receives the ordinary network request information needed to deliver pages and data files.

Known limitations

• Best effort, not universal completeness: “all” cannot be guaranteed because the literature changes, some datasets are unavailable in reusable form, and public resources apply different inclusion criteria.
• Source fidelity: source spelling, taxonomy, and curation choices are retained. This interface does not independently validate all 7,271 records.
• Evidence heterogeneity: an experimentally curated CPPsite3 record, a quantitative POSEIDON measurement, and a positive benchmark label carry different amounts of context.
• Shared upstream evidence: overlap between sources can reflect shared papers, inherited curation, or reused benchmark sets. A cross-source listing is not independent experimental replication.
• Sequence normalization: uniqueness is based on an exact, case-preserved string after formatting whitespace is removed. This avoids merging D-amino-acid notation with uppercase forms, but string identity still does not establish chemical identity.
• Literature horizon: the source authors report curation through April 2024; newer experiments may be absent.
• Positive-only scope: the atlas intentionally excludes negative and generated-decoy sets and should not be used alone to train or validate a CPP classifier.
• No similarity engine: this edition supports exact and substring sequence search, not BLAST, Smith–Waterman, structure search, or peptide mapping.
• No moderation workflow: community submissions, accounts, comments, events, and administrative editing are not part of this static edition.
• No clinical claims: inclusion in a curated CPP dataset does not establish safety, efficacy, or suitability for treatment.