PetriDish

Pharmacogenomic + autoimmune decision-support infrastructure, calibrated to the Indian genome.

Live · Case study · Architecture · Contributing

---

Why this exists

Global drug-dosing tables were built from Western genomes. Indian patients have different variant frequencies, different HLA distributions, and different drug-response profiles.

Gap	Reality
CYP2C19*2 loss-of-function	23% in South Asians · 15% globally → clopidogrel routinely mis-dosed
HLA-B*15:02	9% in South Asians · 4% globally → SJS/TEN carbamazepine mortality 10–30%
32% of Indian variants	Absent from gnomAD → returned as "variant of uncertain significance"
RA diagnostic delay	4–7 years in India vs 1–2 years in the West → irreversible joint damage

What it does

Ingest a VCF (or 23andMe / CSV), run through the deterministic orchestration router, and return one auditable clinical report — with every score traceable to a curated rule.

Everything in the stack ships with a trace_id. One click retrieves the full reasoning chain, model versions, source edges, and whether the output is clinical-surface-safe or hard-blocked as a hypothesis.

Live modules

Clinical (surface-safe):

• IndoPGx — 61 CPIC-graded variants · sub-200ms
• IndoG6PD — tribal Odisha primaquine safety
• IndoAutoImmune 2.0 — HLA precision medicine · hard benchmark gate
• CardioRisk India — 6 simulation engines for Indian MI risk
• Co-Branded Report — diagnostic-lab PDF · 6 sections

Infrastructure:

• Orchestration Layer — deterministic router · no LLM in routing path
• Genomic Report Reader — VCF/23andMe/CSV ingest
• Trust & Audit — trace_id · reasoning chain · shared confidence-tier enum

Research (hypothesis-only · schema-quarantined):

• IndoVarClassifier v1.0 — real methods · fully cited · grant-research-ready
• Discovery — GNN link prediction + VUS classification
• Structural Intelligence — ESM + Boltz-2 four-tier funnel
• Generative Hit-Design — real RDKit chemistry + BRICS scaffold-hop + 5 go/no-go gates → CRO handoff

End-to-end proof:

• Case Study · Priya — one patient · every layer · 13 milliseconds

Non-negotiable design principles

1. No LLM in the clinical routing path. Same input → same output, every call.
2. Predictions are structurally separate from facts. :PREDICTED_TARGETS never merges with :TARGETS. Enforced in code, not by convention.
3. Every claim traces to a source. Curated rule, published paper, or explicit verification_required=true.
4. Every module cites its methods. No proxy backend without a documented swap-in target.

Architecture

See ARCHITECTURE.md for the full 5-layer stack.

┌─────────────────────────────────────────────────────────┐
│  Diagnostic lab / hospital / doctor                      │
└─────────────────┬───────────────────────────────────────┘
                  │  VCF · 23andMe · CSV
                  ▼
┌─────────────────────────────────────────────────────────┐
│  1  Ingestion & normalization                            │
│  2  Deterministic orchestration router                   │
│  3  Interpretation modules (clinical-safe)               │
│  4  Hypothesis modules (schema-quarantined)              │
│  5  Trust & Audit (trace_id on every response)           │
└─────────────────┬───────────────────────────────────────┘
                  │  Standardized clinical report envelope
                  ▼
┌─────────────────────────────────────────────────────────┐
│  Cardiologist · compliance officer · wet-lab partner     │
└─────────────────────────────────────────────────────────┘

Tech stack

• Backend · Python 3.11 · FastAPI · Neo4j · RDKit · NetworkX · pydantic
• Frontend · Next.js 14 App Router · TypeScript · React 18
• Data · IndiGen (CSIR-IGIB) · PharmGKB · CPIC · gnomAD v4 · ClinVar · Ensembl VEP · PrimeKG · IMPPAT 2.0 · GenomeIndia · PDB · AlphaFold DB · UniProt · Reactome
• Hosting · Vercel (frontend) · Railway (FastAPI + Neo4j)

Quick start

git clone https://github.com/shailesh2790/petridish.git
cd petridish

# Backend
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
uvicorn api.reason:app --reload --port 8000

# Frontend (separate terminal)
npm install
npm run dev

Open http://localhost:3000.

.env.example lists the environment variables. Neo4j credentials are only needed for graph modules; most modules work with just an empty .env.

Contributing

• CONTRIBUTING.md — how to add a module, code style, PR checklist
• ARCHITECTURE.md — 5-layer stack, module pattern, envelope contract
• CODE_OF_CONDUCT.md — Contributor Covenant

Good first contributions:

• Add a curated Indian variant to api/patient.py (with citation)
• Wire a new module's Next.js proxy route
• Improve a page's UI/UX using src/components/ui.tsx primitives
• Add citations to a module's /citations endpoint
• Fix a broken SMILES in api/real_chemistry.py::APPROVED_AUTOIMMUNE_DRUG_SMILES

Open an issue first for anything larger than a proxy tweak.

Positioning

PetriDish is decision-support infrastructure — not a diagnostic device, not a drug. Every clinical output is triage for a licensed practitioner. Every hypothesis output requires wet-lab or literature cross-check.

License

Apache 2.0

Contact

Shailesh Tripathi · shailesh.tripathi2706@gmail.com · LinkedIn