ARDIA PRECISION HEALTHGoverned AI for healthcare revenue & precision care
Back to Home
Discovery & Research

Governed drug discovery —
grounded, never generated

Discovery-to-trial decisions built on measured data and published rules. Every claim cites a real source — or the system abstains. No generative chemistry, no invented affinities, no black-box scores.

The governed discovery pipeline

The same neuro-symbolic core behind Ardia’s reimbursement engine, pointed upstream at the discovery-to-trial decision layer — deterministic rules + real-database retrieval + guarded reasoning + human-in-the-loop. No new trust model required.

1Built now

Cite-or-Abstain Provenance & Non-Generative Guarantee

Every output runs through TARA’s Scholar tier — plan → retrieve from a named public source → reason → cite — with a source chip on each claim and a tamper-evident trace (query, source, version, timestamp, reviewer). The non-generative rule is enforced in the code path: no measured evidence means the system abstains rather than inventing an answer.

2In build

Companion Diagnostic & Biomarker→Coverage Strategy

For a biomarker-defined program (EGFR, ALK, BRAF, MSI-H/TMB, HLA-B, DPYD) it maps the required test to FDA companion-Dx approvals, NCCN/CPIC guidance, and the concrete Medicare coverage path (MolDX Z-code, applicable LCD/NCD, CPT/PLA codes) — the direct bridge from discovery to the reimbursement business Ardia already owns.

3In design

Governed Target Validation & Prioritization

Ranks candidate targets on real human genetic and genomic evidence from Open Targets (association scores, ClinVar/GWAS support, tissue specificity, tractability) — never intuition or generated hypotheses. Targets with weak or absent human evidence are flagged “insufficient evidence,” not scored optimistically.

4In design

Evidence-Grounded Repurposing & Chemical Starting Points

Retrieves compounds with measured activity against a target from ChEMBL — real IC50/EC50/Ki, mechanism of action, and approved-drug status — rather than inventing molecules. If no measured bioactivity exists it abstains and says so; it never fabricates a structure or an affinity.

5In design

ADMET & Safety-Liability Surfacing

Surfaces measured ADMET and pharmacology (ChEMBL properties, CYP metabolism, PGx and drug–drug-interaction flags) plus published, rule-based structural alerts (PAINS, Brenk, Lipinski). Any predicted value is labelled predicted and visually separated from measured — never a black-box “toxic/safe” verdict.

6In design

Trial Landscape & Feasibility

Pulls the live competitive and feasibility picture from ClinicalTrials.gov — who is running trials for a target/indication, their endpoints, eligibility, enrollment and sponsors — each cited to an NCT id. Supports go/no-go and trial design; shows only real registered trials and makes no efficacy prediction.

What we will not do

Most “AI drug discovery” sells the parts that hallucinate. These are the capabilities we deliberately exclude — because a claim you can’t cite is a claim we won’t ship.

×
De-novo generative chemistry
Invented structures have no measured data behind them and cannot be cited.
×
Docking / affinity scores presented as fact
Unvalidated numbers dressed as measurements.
×
Retrosynthesis auto-planning as validated output
Implies wet-lab reliability Ardia does not have.
×
Black-box QSAR “toxic / safe” verdicts
Structure-only verdicts are unfalsifiable and dangerous in a clinical context.
×
Clinical efficacy prediction (“this drug will work”)
Predicting efficacy is diagnostic; Ardia is explicitly non-diagnostic.
×
Predicted ADMET shown as measured
Blurring predicted and measured is a subtle honesty failure.

EGFR × osimertinib, the governed way

A single fully-grounded example on real public data — every value below was retrieved live from ChEMBL, ClinicalTrials.gov and PubMed and links back to its source. It illustrates the method, not a novel discovery or an efficacy claim.

TargetEGFR — ChEMBL target CHEMBL203, UniProt P00533 (single protein, human) ChEMBL →
DrugOsimertinib (Tagrisso / AZD9291) — ChEMBL CHEMBL3353410, max phase 4, first approval 2015 ChEMBL →
MechanismIrreversible, mutant-selective EGFR inhibitor (measured bioactivity curated in ChEMBL) PMID 24893891 →
DevelopabilityMW 499.62 · ALogP 4.51 · HBD 2 · HBA 8 · PSA 87.55 Ų (ChEMBL calculated) → passes Lipinski Ro5 & Veber ChEMBL →
Active trialsNCT04181060 (osimertinib ± bevacizumab, Ph 3) · NCT07738172 (Ph 3) ClinicalTrials.gov →
⚖ AbstainedOpen Targets EGFR–NSCLC association score was unavailable at query time — the system returned “insufficient evidence” rather than inventing a score.

Molecular descriptors are ChEMBL calculated properties (from structure), clearly separated from measured bioactivity. Illustrative of the governed method on real public data — not a treatment recommendation.

Real databases & published rules — nothing invented

ChEMBL
Measured bioactivity & drug data
Open Targets
Target–disease genetic evidence
PubChem
Open chemistry
ClinicalTrials.gov
Global trial registry
PubMed
Primary literature
DrugBank
Drug / target reference
BindingDB
Measured affinities
openFDA / DailyMed
Labels & adverse events

Rule set (each verified to its primary publication): Lipinski Ro5, Veber, Egan, Ghose, Congreve Ro3, ligand efficiency (LE/LLE), PAINS & Brenk structural alerts, the Bowes 2012 44-target secondary-pharmacology safety panel, and hERG / ICH S7B cardiac-repolarization assessment.

Same governance, upstream

This is a design/roadmap capability for a pre-revenue company — built on the reasoning core, guardrails and cite-or-abstain that are real today. It feeds directly into precision medicine and the TARA governance the whole platform runs on.

Talk to us →