MERIDIAN

Hackathon winner. A privacy-first pharmacogenomics system that turns a governed genome-processing run into source-bound medicine guidance.

Genome in → gene results → medicine guidance → daily instructions → constrained AI check → sources.

PharmCAT is the bioinformatics engine underneath the product. It is not the product screen and the normal user does not upload a PharmCAT report. The normal input is a single-person GRCh38 VCF or VCF.GZ. Importing an existing PharmCAT Reporter JSON remains available under Other ways to start for expert inspection only.

How it works

1. The browser uploads the genome directly to a private, run-specific Cloud Storage session. Raw DNA does not pass through Vercel or the medical model.
2. A private Cloud Run worker checks the file, runs the pinned official PharmCAT pipeline and seals the input hash, command, software versions, coverage and output hashes in a run manifest.
3. Deterministic code reads the exact PharmCAT gene calls and matched CPIC annotations. It does not ask a model to invent a gene, medicine or dose.
4. The app groups those source results into plain English: discuss another medicine, review the dose, or no gene-based starting change.
5. After a medicine is selected, the app shows only drug-specific daily facts supported by a pinned prescribing-information record. A routine match is disabled until the exact product and form are confirmed, and it uses only answers the person actually supplied.
6. MedGemma may review a completed, session-owned run for conflicts, evidence gaps, lifestyle constraints and useful clinician questions. It returns fact IDs and typed actions, not medical prose. Invalid or invented references are rejected.

Genetics can change exposure and dosing. It does not reliably predict which antidepressant will improve depression. The system does not diagnose, prescribe, declare a medicine safe, or tell a patient to start, stop or change treatment.

What is real and repeatable

• Patient results are never hardcoded and no failure is replaced by a demo result.
• Raw VCF processing uses the official PharmCAT tool in a pinned wrapper image.
• Missing positions are not treated as reference calls.
• CYP2D6 is withheld unless a validated structural and copy-number-aware outside call is added.
• Medicine rows come from the uploaded run's PharmCAT annotations, not a fixed shortlist.
• Clinical rules are versioned source releases with exact source IDs, revisions and digests.
• An official PharmCAT example is available for UI inspection, but it is clearly separate from a private genome run and cannot use the medical model.
• A failed upload, PharmCAT run, source check or model response remains a visible failure.

Data used

Data	Purpose	Sent to MedGemma?
Raw GRCh38 VCF/VCF.GZ	PharmCAT calling only	Never
PharmCAT Reporter output and coverage	Gene and medicine results	Derived facts only
Exact PharmCAT/CPIC annotation	Dose and medicine-choice guidance	Source-bound fact only
Current medicines	Deterministic enzyme and interaction context	Canonical names only
Selected medicine and answered routine fields	Drug-specific daily-life check	Yes
Direct identifiers	Not required	Never

The current daily-life release uses pinned US Structured Product Label records. The Australian medicine list in this repository is scope data only; Australian PI/CMI, ARTG and PBS evidence must be reconciled before an Australian clinical release.

Exact demo uploads

Use the two public PharmCAT example artefacts for different checks:

Both are byte-for-byte pinned copies of official PharmCAT Example 1.

File	Upload location	What it validates
public/samples/pharmcat-example.vcf	DNA → Choose DNA file	Official Example 1: one GRCh38 sample with all-reference PGx sites. Runs the real private PharmCAT path. CYP2D6 is correctly withheld because this upload has no validated structural/copy-number outside call.
public/samples/pharmcat-example.report.json	DNA → Other ways to start → Import PharmCAT report	Official published result containing a separate outside CYP2D6 call. Tests deterministic Reporter JSON parsing and the validation UI, but does not rerun PharmCAT or prove upstream coverage.

Both files are downloadable from the matching upload screen. After choosing a file, enter recognised current medicines or select I take none. The built-in Use published example path is a third, no-upload UI demo and is kept separate from a governed genome run.

Pinned file digests:

VCF     f45cc947fa0a38f47307ae5f2cc6e71bf5afffa6f1310b043b962c472db76438
Report  affc3223bfaf9176b71e62b5d8926815079228b7f5abb7265bc41fb3e5adf898

Run the app

npm ci
npm test
npm run build
npm run dev

The published example works in the local Vite app. A real raw-genome run also needs the private Google Cloud services and same-origin Vercel routes described in DEPLOYMENT.md.

Backend checks:

npm test --prefix services/pharmcat-control
(cd services/pharmcat-worker && go test ./...)

Refresh the FDA source release explicitly:

node scripts/sync-clinical-sources.mjs

That command fails if a pinned record or reviewed evidence phrase is no longer present. Updated JSON must be reviewed and committed; runtime results never depend on a live FDA API response.

Deployment status

Part	Repository	Live service
Simple six-step app	Implemented	Deploy to Vercel
Private upload/control API	Implemented	Deploy to Cloud Run + Vercel
Pinned PharmCAT worker	Implemented	Build and deploy as a Cloud Run Job
Deterministic clinical engine	Implemented	Runs with the app/API
Constrained MedGemma gateway	Implemented	IAM-protected Vertex endpoint still required
Australian evidence release	Not complete	Must not be presented as complete

The model endpoint and Google Cloud workload are deliberately not created by running this repo: they use billable infrastructure and require operator identity, access and region decisions.

See ARCHITECTURE.md for the system in plain language and DEPLOYMENT.md for the Vercel and Google Cloud setup.