Methodology
Where MolForge's reference data comes from, how the pages are built from it, how AI is used and checked, and how often the data is refreshed.
Data sources
Every data page links the exact upstream records it was built from (DrugCentral drug card, ChEMBL id, UniProt accession, PubChem CID, FDA label search…). The datasets:
- DrugCentral (University of New Mexico) (CC BY-SA 4.0) — Drug structures, names and synonyms, indications and contraindications (OMOP/SNOMED), molecular targets and mechanism-of-action links, pharmacologic classes, ATC codes, regulatory approvals, and FDA FAERS adverse-event disproportionality signals. Last loaded: , release 2024-11-01.
- ChEMBL (EMBL-EBI) (CC BY-SA 3.0) — Measured bioactivities (binding affinities, pChEMBL values) linking compounds to protein targets.
- UniProtKB / Swiss-Prot (CC BY 4.0) — Protein names, genes, organisms, function text, sequences, keywords and subcellular locations. Last loaded: , release unknown.
- PubChem (NCBI / NIH) (Public domain) — Compound identifiers (CID, InChI/InChIKey, SMILES, IUPAC names, synonyms), computed properties and parent/salt/tautomer relationships.
- openFDA / FDA Adverse Event Reporting System (FAERS) (Public domain) — Spontaneous adverse-event reports; we show the disproportionality signals DrugCentral derives from them. A signal is a statistical association, not proof of causation.
- AlphaFold Protein Structure Database (EMBL-EBI / Google DeepMind) (CC BY 4.0) — Predicted 3D protein structures and per-residue confidence (pLDDT). Last loaded: , release v6.
- Open Targets Platform (CC0 1.0) — Target–disease association scores and target tractability. Last loaded: , release 24.09.
- Reactome (CC BY 4.0) — Pathway membership of proteins.
- FDA Orange Book (Public domain) — Patent and exclusivity listings for approved drug products.
- PharmGKB and CPIC (CC BY-SA 4.0) — Pharmacogenomic drug-label annotations and CPIC dosing-guideline links.
“Last loaded” dates are read live from our ingest logs; a source without a date has not been logged by the current ingest pipeline. Attribution and licences are preserved per source.
How pages are built
- Upstream releases are downloaded and loaded into MolForge's own PostgreSQL (AlloyDB) database by versioned ingest jobs; cross-references use stable identifiers (InChIKey, UniProt accession, PubChem CID, DrugCentral id).
- Pages are rendered on the server from that database, so the data is in the HTML (no JavaScript needed to read it) and every page links its sources.
- Computed chemistry (Lipinski rule-of-five checks, QED, synthetic accessibility) uses RDKit. ADMET values are in-silico model predictions, labelled as such, not measurements.
- Only curated entities are offered for search indexing: drug-linked compounds, SwissProt-reviewed proteins
with a druggability, disease or structure record, and DrugCentral drugs. Other generated pages are marked
noindex.
How AI text is generated and checked
- Drug-guide articles: drafted by Google Gemini (2.5 Flash-Lite) from a seed of name, SMILES, class and category, returned as schema-validated JSON, then automatically cleaned of meta-commentary and filler. They are marked as AI-drafted and are not clinically reviewed.
- Insight panels: Google Gemini (2.5 Flash) receives only the page's structured data and is instructed not to introduce facts outside it and to state when data is sparse; output is schema-validated JSON and cached. If generation fails, a deterministic summary is shown instead.
- Structured sections (indications, targets, approvals, adverse-event signals, identifiers) are never AI-generated: they are rendered directly from the source records.
Update cadence
Datasets are re-ingested when their publishers release new versions. Each page states the date its own record was last refreshed; we do not bump dates without a data change.
Known limitations
- Adverse-event signals come from spontaneous reports and show association, not causation or frequency.
- Indications reflect DrugCentral's curation and may lag current labels; check the official label.
- AI-drafted guide text can contain errors — please report them.