Methodology

Data sources

Only sources this project actually uses. There is no paper-DOI ingest and no live FDA/EMA API.

• PubChem — curated CIDs, compound-page links, and name-based SMILES lookup when a structure is missing
• ChEMBL — curated ChEMBL IDs and links. Bioactivity tables are not synced.
• Literature — public-summary seed notes and family descriptions. A citation is shown only when a field exists; DOIs are not invented.
• ClinicalTrials.gov — curated search queries that open trial listings. Trial records are not imported.
• FDA / EMA — curated agency strings on selected clinical-overlay rows. Label APIs are not called.
• Manual seed data — families, compounds, SMILES, tags, and relations. Default dataProvenance is seed.

Entity types

Nodes are classified from catalog role plus identifiers and tags. Types are not forced when they do not fit.

• Observed molecules — Verified drug, Known compound, Literature compound; payload/linker/ADC with external IDs
• Abstractions — Abstract scaffold; Abstract pharmacophore only when tags/name include pharmacophore or warhead
• Exploratory structures — Rule-generated SAR analog, Exploratory analog

A parent node is not automatically a literature MCS or shared chemical core. Family centers are labeled as exploration scaffolds.

Structure standardization

• SMILES are assigned from seed data, the known-SMILES map, PubChem lookup, or parent inheritance.
• Inherited SMILES are not independently assigned structures.
• InChI Key, PubChem CID, and ChEMBL ID are shown only when present.

Fingerprinting

• Type: Morgan fingerprint (RDKit circular fingerprint)
• Radius: 2
• Bits: 2048

Similarity

• Metric: Tanimoto (shared bits / union bits on packed uint8 fingerprints)
• Weekly mesh threshold: 0.45

Seed similarity edges, the weekly mesh, add-compound estimates, and family-compare matrices use this configuration. Edges labeled Manual are the exception.

Structural similarity does not predict biological equivalence.

Clustering

Used in CNS chemometric analysis only. There is no catalog-wide Butina clustering.

• Algorithm: Butina-style sphere exclusion
• Cutoff: 0.55 (Tanimoto)
• Medoid: the member with the highest average similarity to other cluster members
• Approved-neighbor band: 0.45–0.75

SAR relationships

Family-map derived_from edges are manually curated parent→child paths. They are not claimed literature reaction schemes.

similar_to is fingerprint-derived, related_drug is a catalog drug-class link, and conjugated_with is a payload–linker–ADC conjugate relationship.

CNS seed proposals only mutate SMILES with the fixed rules below, then validate with RDKit. This is not a generative model.

• cl-to-f — Cl→F halogen walk
• f-to-cl — F→Cl halogen walk
• ome-to-oh — OMe→OH demethylation
• oh-to-ome — OH→OMe methylation
• nme-expand — N-H→N-Me (amine methyl)
• aryl-h-to-f — Aryl-H→F (c1ccccc1 walk)

Exploration metrics

On-screen High / Medium / Low labels and numeric scores are an Exploration Score. They are not drug quality, efficacy, clinical probability, or investment potential.

Internal exploration metric based on catalog and structural signals. Not a prediction of efficacy, safety, potency, or development success.

Documented criteria

• Core: family-center flag, scaffold/payload/linker role, SMILES, smilesSource, InChIKey, approved/clinical/SAR tags, family size, derived-child count, relation count, notes
• Derivative: drug/derivative/adc role, SMILES, smilesSource, approved/clinical tags, Morgan similarity to center, family size, relation count

Normalization

• Additive rule score, then capped at 100
• Relative percentile in the tab pool: top 12% High (X-S), 12–62% Medium (X-A/X-B), remainder Low (X-C). Pools under 5 compounds stay Medium

Limitations

• Scores follow catalog completeness and tag coverage.
• The similarity term is a structural fingerprint, not a potency model.
• Approved/clinical tags are catalog tags, not a live regulatory API.
• Computed proximity (similarity, clusters, exploration scores) is not presented as a literature fact.