Build with the MultiEndpointTox API

Real physics-based docking against 7 curated protein targets (hERG, CYP3A4, CYP2D6, CYP2C9, PPARγ, androgen receptor, ERα — real RCSB PDB structures). Disabled by default even once deployed (docking.enabled: false) — requires optional vina/meeko dependencies not in the main requirements.txt.

• GET /docking/status
• GET /docking/targets
• POST /dock
• POST /dock/batch
• GET /dock/{target}
• POST /dock/enhanced
• GET /dock/endpoint/{endpoint}/enhanced
• POST /predict/ensemble
• POST /descriptors/3d
• POST /pharmacophore/features
• POST /pharmacophore/compare

Every route returns HTTP 503 with an explanation — never a fabricated affinity or pose — when docking is disabled or Vina isn't installed.

Real Phase I/II metabolite prediction via SyGMa (Ridder & Wagener, ChemMedChem 2008, doi:10.1002/cmdc.200700312) — 138 Phase I + 27 Phase II published, literature-derived reaction rules. Returns the full reaction graph, real likelihood scores, 2D structure SVGs, and structural-alert screening of every metabolite.

• GET /metabolism/status
• POST /metabolism/predict
• GET /metabolism/predict/{smiles}

Structural alerts on metabolites are a mechanistic hypothesis, not a validated toxicity prediction — labeled as such in every response.

Real, on-disk per-endpoint validation data — model version history with SHA256/training timestamp/git commit, real cross-validation fold statistics, and a live applicability-domain (AD) staleness health check. Nothing here is computed or estimated by the API.

• GET /validation
• GET /validation/{endpoint}
• GET /validation/ad-health

The AD health check currently reports 4 of 7 endpoints as STALE (hepatotox, ames, skin_sens, cytotox) — a real, documented, in-progress incident, not hidden. See the Validation page's Applicability Domain Status panel.

Real CPIC-sourced gene-drug evidence — 57 level A/A-B pairs across 16 genes, live-fetched from api.cpicpgx.org with real guideline URLs and PMIDs. No patient genotype input exists in this API; every phenotype is an explicit scenario. Anything outside the curated evidence set is reported 'Not currently supported', never guessed.

• POST /pharmacogenomics/predict
• POST /pharmacogenomics/score
• POST /pharmacogenomics/report
• POST /pharmacogenomics/adjust-toxicity
• GET /pharmacogenomics/genes
• GET /pharmacogenomics/version

/adjust-toxicity calls the real, unmodified toxicity model and applies an illustrative PSI adjustment on top — the model itself is never touched.

curl -X POST https://api.bionexusdiscovery.com/pharmacogenomics/predict \
  -H "Content-Type: application/json" \
  -d '{"smiles": "CN1CCC23C=CC(O)C4Oc5c(O)ccc(c25)C1C34", "genes": ["CYP2D6"]}'
# codeine — real CYP2D6 prodrug-activation evidence (CPIC level A, PMID 22205192):
# Poor Metabolizer -> decreased toxicity risk (less morphine formed)
# Ultrarapid Metabolizer -> increased toxicity risk (excess morphine —
# the real mechanistic basis for the FDA's black-box warning)

Publication-scale batch screening — SMILES paste, CSV, TSV, SDF, or ZIP of SDF files. Chunked, resumable, cancellable async jobs (up to 10,000 compounds/job) with real per-compound provenance and CSV/Excel/JSON/PDF export. No model outputs changed — every compound scored via the same pipeline as POST /predict/multi.

• POST /batch/jobs
• GET /batch/jobs/{id}
• GET /batch/jobs/{id}/items
• GET /batch/jobs/{id}/summary
• POST /batch/jobs/{id}/resume
• POST /batch/jobs/{id}/retry-failed
• POST /batch/jobs/{id}/cancel
• GET /batch/jobs/{id}/export?format=csv|xlsx|json|pdf

Rate-limited to 10 job submissions/minute — physics-real compute triggers, not a cached lookup.

curl -X POST https://api.bionexusdiscovery.com/batch/jobs \
  -F "smiles_text=CCO
CC(=O)Nc1ccc(O)cc1,paracetamol" \
  -F "endpoints=hepatotox,herg" \
  -F "chunk_size=50"
# -> {"id": "...", "status": "pending", "total": 2, ...}
# then poll: curl https://api.bionexusdiscovery.com/batch/jobs/{id}