Syntrx

AI-powered personalized medicine from your DNA.

Syntrx parses the raw genetic data file you already have from 23andMe or AncestryDNA and generates a personalized health report across four domains — drug metabolism, nutrient optimization, diet and fitness, and lifestyle-modifiable risk — every recommendation cited to a CPIC clinical guideline, FDA pharmacogenomic label, or peer-reviewed source.

The clinical content is produced by a deterministic CPIC-grounded engine; an optional LLM "narrator" only adds plain-English flavour on top. That separation is what makes Syntrx auditable: the same input produces the same recommendations, and the recommendations can be validated bit-for-bit against published guidelines.

The synthetic test suite achieves 100% concordance against expected CPIC phenotypes across the curated profiles.

Architecture at a glance

                        ┌──────────────┐
   23andMe / Ancestry → │ Parser Agent │ → user genotypes (curated 49+ SNP catalog)
                        └──────────────┘
                               │
                               ▼
                        ┌──────────────┐
                        │ Lookup Agent │ ← ChromaDB (PharmGKB / SNPedia / PubMed)
                        └──────────────┘
                               │
                               ▼
                        ┌──────────────┐
                        │  Synthesis   │ → deterministic engine (CPIC tables)
                        │    Agent     │   + optional LLM narrator
                        └──────────────┘
                               │
                               ▼
                        ┌──────────────┐
                        │ Safety Agent │ → strip diagnostic phrasing,
                        └──────────────┘    add physician disclaimers
                               │
                               ▼
                       Report (JSON / PDF / UI)

Why deterministic + LLM, not pure-LLM?

A pure-LLM pharmacogenomics tool can hallucinate dose changes. A deterministic engine grounded in CPIC tables cannot. Syntrx's recommendation layer is closed-form Python with no LLM in the path; the LLM is only allowed to add a non-prescriptive narrator paragraph, and the Safety Agent strips anything that smells diagnostic. This gives auditable accuracy and a friendly voice.

Project layout

syntrx/
├── backend/            FastAPI + multi-agent pipeline (Python 3.11+)
│   ├── app/
│   │   ├── core/        ← parser, SNP catalog, phenotype caller, recommendation engine
│   │   ├── agents/      ← Parser → Lookup → Synthesis → Safety
│   │   ├── services/    ← LLM provider abstraction, pipeline, PDF, drug-interactions
│   │   ├── knowledge/   ← ChromaDB wrapper for the RAG store
│   │   ├── api/         ← FastAPI routes
│   │   ├── db/          ← SQLAlchemy models
│   │   └── schemas/     ← Pydantic request/response models
│   ├── data/sample/     ← Synthetic 23andMe files with known phenotypes
│   ├── tests/           ← Pytest suite (parser, phenotype, recs, pipeline, API)
│   └── scripts/         ← generate_sample_data, evaluate (CPIC concordance), ingest_pharmgkb
├── frontend/           React + Vite + Tailwind + Recharts
│   └── src/
│       ├── components/  ← ConfidenceBadge, DomainSection, FindingCard, Upload, SummaryStats
│       ├── pages/       ← Home, Report, Interactions
│       └── lib/         ← Typed API client
├── docker-compose.yml  ← Postgres + Redis + backend + frontend
├── .vscode/            ← debug configs + recommended extensions
└── .github/workflows/  ← CI (pytest + frontend build)

Getting started

One-command Docker

cp .env.example .env
docker compose up --build
# Frontend → http://localhost:5173
# API      → http://localhost:8000/docs

Local dev (no Docker)

Backend

cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python -m scripts.generate_sample_data           # synthetic 23andMe files
python -m app.cli analyze data/sample/sample_warfarin_sensitive_pgx_storm.txt
uvicorn app.main:app --reload                    # API on :8000, docs at /docs
pytest -q                                         # full suite
python -m scripts.evaluate                        # CPIC concordance report

Frontend

cd frontend
npm install
npm run dev                                       # Vite dev server on :5173

Configuring the LLM (optional)

Without an API key, Syntrx uses a deterministic mock narrator and the clinical engine still produces a complete report. To turn on a real LLM:

# .env
LLM_PROVIDER=anthropic                # or openai
ANTHROPIC_API_KEY=sk-ant-…
LLM_MODEL=claude-sonnet-4-6

Switch between providers without code changes.

Demo flow

1. Drag-drop backend/data/sample/sample_warfarin_sensitive_pgx_storm.txt into the upload card. (Generated by scripts/generate_sample_data.py.)
2. The pipeline runs Parser → Lookup → Synthesis → Safety. The trace appears in the dashboard.
3. The report opens with four collapsible domains. The "Clopidogrel resistance", "Warfarin sensitivity", "DPYD reduced activity" and "Statin myopathy" findings should all be flagged for physician follow-up.
4. Click Download PGx card (PDF) to get a printable one-page summary designed to be handed to a prescriber.
5. Click Drug interaction checker, paste in a few drugs (codeine, plavix, simvastatin, omeprazole, sertraline) and hit Check to see severity-coded gene-drug flags.

Validation

cd backend
python -m scripts.evaluate

                CPIC Concordance
profile                          Gene     Expected      Got                    Match
average_european                 CYP2D6   Normal        Normal Metabolizer       ✓
average_european                 CYP2C19  Normal        Normal Metabolizer       ✓
average_european                 VKORC1   Standard      Standard …               ✓
cyp2d6_poor_metabolizer          CYP2D6   Poor          Poor Metabolizer         ✓
warfarin_sensitive_pgx_storm     CYP2C9   Poor          Poor Metabolizer         ✓
warfarin_sensitive_pgx_storm     VKORC1   High          High …                   ✓
warfarin_sensitive_pgx_storm     SLCO1B1  Poor          Poor Function            ✓
warfarin_sensitive_pgx_storm     DPYD     Intermediate  Intermediate …           ✓
warfarin_sensitive_pgx_storm     CYP2C19  Poor          Poor Metabolizer         ✓
ultrarapid_2c19_caffeine_slow    CYP2C19  Ultrarapid    Ultrarapid Metabolizer   ✓
ultrarapid_2c19_caffeine_slow    CYP1A2   Slow          Slow Metabolizer         ✓
apoe_e4_carrier_with_iron        APOE     ε4            Single ε4 carrier        ✓
apoe_e4_carrier_with_iron        HFE      C282Y/H63D    C282Y/H63D compound …    ✓

Concordance: 13/13 = 100.0%

What Syntrx covers today

• Pharmacogenomics (23 SNPs): CYP2D6, CYP2C19, CYP2C9, CYP3A5, CYP1A2, VKORC1, SLCO1B1, DPYD, TPMT, UGT1A1, IFNL3, G6PD.
• Nutrigenomics (10 SNPs): MTHFR, VDR, FUT2, BCO1, SLC23A1, TCN2, GC.
• Diet / fitness (10 SNPs): APOA2, MCM6/LCT, ACTN3, ADH1B, ALDH2, FTO, TCF7L2, PPARG, TAS2R38, TRPV1.
• Risk awareness (6 SNPs): APOE, HFE, F5 (Factor V Leiden), F2 (prothrombin).

The full curated list lives in backend/app/core/snp_catalog.py.

Regulatory positioning

Syntrx is positioned as a genetic literacy and wellness tool — it does not diagnose disease, prescribe medication, or replace medical advice. The Safety Agent enforces that:

• Every finding that suggests a prescription change carries a "discuss with prescriber" disclaimer in its action list.
• The synthesis prompt is explicit that the LLM may not invent dosages or change clinical claims; the Safety Agent regex-strips any LLM output that drifts into diagnostic phrasing.
• The global disclaimer is embedded in every report payload and on the PDF card.

Roadmap

• PharmGKB bulk ingestion + Pinecone for production RAG
• Authentication (Clerk / Auth0) and per-user report history
• Comparison view: diff two reports (yours vs a family member's)
• Lifestyle-modifiable polygenic risk scores (T2D, CAD, MDD)
• Mobile-first redesign of the report viewer