FAVES V4: A Production-Deployed Compliance and Liability Screening System for Drug Discovery at 122M-Compound Scale
Aligned to HHS/ONC FAVES and NIST AI 600-1. 362/362 compounds passing post-fix; ~5 ms per molecule live, 122,407,635 documents cached.
Abstract
FAVES V4 is a compliance and liability screening system for drug discovery, aligned to two complementary federal frameworks: the U.S. Department of Health and Human Services / Office of the National Coordinator FAVES rubric for trustworthy AI in healthcare and life sciences, and the NIST AI Risk Management Framework Generative AI Profile (NIST AI 600-1), which names chemical and biological design tools as a distinct CBRN risk class.
The five letters — Fair, Appropriate, Valid, Effective, Safe — are operational properties demonstrated against published evidence in production, with named action-ID anchors drawn from NIST AI 600-1 throughout Section 4.
A single-API screening call returns:
- Regulatory status — DEA, FDA, CWC, EPA, EU REACH
- Structural alerts — 1,585 SMARTS across PAINS, Brenk, NIH, ZINC, and 7 ChEMBL sub-catalogs via RDKit FilterCatalog
- Pharmacokinetic classification — BOILED-Egg
- Prior-art disclosure — InChIKey identity resolution against PubChem and a local 122M-compound database
- Per-field provenance metadata
The system is deployed in production with both live RDKit computation (~5 ms per molecule) and a precomputed database cache (122,407,635 documents).
Results
| Test | Outcome |
|---|---|
| V3 regulatory benchmark | 362/362 compounds passing (post-fix) |
| V4 field completeness | Passing predetermined acceptance criteria |
| BOILED-Egg implementation | Passing |
| Prior-art resolution | Passing |
| External cross-RDKit-version reproducibility | Passing |
| Live latency | ~5 ms per molecule |
| Cached corpus | 122,407,635 documents |
The initial 341/342 pre-fix execution surfaced a single defect — an incorrect cortisol SMILES in the deployed FDA whitelist that omitted the 11β-hydroxyl — root-caused, corrected, redeployed, and re-validated through the documented change-control SOP. Both numbers are reported transparently: the defect surfaced because validation was scoped to the full 102-compound regulatory benchmark rather than an undersized fixture, and the change-control loop functioning visibly in production is a stronger claim than a never-exercised one.
Governance bundle
The five FAVES properties are operationalized through a complete governance bundle:
- Intended Use Statement (Appropriate)
- Risk-based component classification mapped to GAMP and CSA (Valid)
- End-to-end validation protocol with predetermined acceptance criteria (Valid)
- Three-layer audit trail with two non-skippable layers (Safe)
- Change-control SOP with a worked rollback example (Safe)
To our knowledge, this is the first cheminformatics screening system whose full HHS/ONC- and NIST AI 600-1-aligned governance scaffolding is described in operational detail, with each of the five FAVES properties demonstrated against public evidence. The benchmark dataset, validation reports, and latency measurements are released for reproducibility; full operational versions of the seven NMCP-FAVES governance documents are available from the corresponding author.