🧬 Drug Discovery AI

Multi-agent precision medicine pipeline for novel drug discovery — generating molecules, predicting selectivity, matching clinical trials, and persisting discoveries in under 60 seconds.

---

🚀 What This Is

Drug Discovery AI is a 18-agent pipeline that takes a gene mutation query (e.g. EGFR T790M) and:

1. Parses the mutation with LLM + regex fallback
2. Fetches literature (PubMed), proteins (UniProt), structures (RCSB), compounds (PubChem) in parallel
3. Downloads PDB structures and detects binding pockets (fpocket / centroid fallback)
4. Generates novel molecules via RDKit scaffold hopping + bioisostere SMARTS mutations
5. Docks molecules (Gnina → Vina → AI hash fallback)
6. Dual-docks vs off-target proteins to compute selectivity ratio — the key differentiator
7. Screens ADMET (Lipinski + PAINS + toxicophore images)
8. Optimizes leads with scaffold hopping, bioisostere replacement, and fragment growing — building an evolution tree
9. Forecasts resistance mutations with LLM
10. Matches active clinical trials via ClinicalTrials.gov API v2
11. Builds a knowledge graph and reasoning trace
12. Saves discoveries to Neon PostgreSQL and exposes a full REST API

---

📁 Project Structure

drug-discovery-ai/
├── backend/ # Python 3.11 + FastAPI + 18 agents
│ ├── agents/ # 18 pipeline agents
│ ├── pipeline/ # LangGraph state + orchestrator
│ ├── utils/ # LLM router, DB, ADMET, molecule utils
│ ├── routers/ # FastAPI route handlers
│ ├── data/ # JSON data files (curated profiles, resistance, etc.)
│ └── evaluation/ # Benchmark runner
│
├── frontend/ # Next.js 16 + TypeScript + Tailwind v4
│ └── app/
│ ├── components/ # All UI components (analysis, landing, settings)
│ ├── hooks/ # useSSEStream, useAnalysis, useDiscoveries, useTheme
│ ├── lib/ # api.ts, types.ts, utils.ts, theme.ts
│ ├── analysis/ # [sessionId] analysis page (10 tabs)
│ ├── discoveries/ # Discovery library browser
│ └── settings/ # Theme customizer, pipeline config, API key checker
│
├── .github/
│ └── workflows/ # backend-ci.yml + frontend-ci.yml
├── start.sh # Unix quick-start
├── start.bat # Windows quick-start
└── README.md

---

⚡ Quick Start

Prerequisites

• Python 3.11+
• Node.js 20+
• (Optional) AutoDock Vina or Gnina for real docking
• (Optional) fpocket for pocket detection

1. Backend

cd backend
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
# Fill in at least GROQ_API_KEY (free at console.groq.com)
uvicorn main:app --reload --port 8000

2. Frontend

cd frontend
npm install
cp .env.local.example .env.local
npm run dev

Open http://localhost:3000.

---

🔑 Environment Variables

Backend (backend/.env)

Variable	Required	Description
OPENAI_API_KEY	Optional	GPT-4o-mini (primary LLM)
GROQ_API_KEY	Recommended	Llama 3.3 70B — free at console.groq.com
TOGETHER_API_KEY	Optional	Mistral 7B fallback
NCBI_API_KEY	Optional	Higher PubMed rate limits
LANGCHAIN_API_KEY	Optional	LangSmith observability
DATABASE_URL	Optional	Neon PostgreSQL connection string
AUTO_SAVE_DISCOVERIES	Optional	Set true to auto-persist every run

Frontend (frontend/.env.local)

NEXT_PUBLIC_API_URL=http://localhost:8000
NEXT_PUBLIC_APP_NAME="Drug Discovery AI"

---

🧪 Tech Stack

Backend

• FastAPI 0.115 + uvicorn + SSE streaming
• LangGraph 0.2 pipeline orchestration
• LangSmith observability (enterprise trace dashboard)
• RDKit — molecule generation, ADMET, depiction
• Multi-LLM fallback: OpenAI GPT-4o-mini → Groq Llama 3.3 70B → Together Mistral 7B → deterministic template
• Neon PostgreSQL via SQLAlchemy asyncio + asyncpg
• External APIs: PubMed, UniProt, RCSB, PubChem, ClinicalTrials.gov

Frontend

• Next.js 16 App Router + TypeScript strict
• Tailwind CSS v4 + amber minimal theme system
• GSAP 3.12 + ScrollTrigger — hero animations, counter reveals
• Framer Motion 11 — agent status stagger, 2D/3D crossfade
• D3.js 7 — knowledge graph force-directed, evolution tree
• Recharts — ADMET radar chart, docking score chart
• shadcn/ui (Radix primitives) — accessible components
• NGL — 3D molecular viewer
• Biome — linting + formatting (no ESLint, no Prettier)

---

🏆 Win Factor Features

Feature	Description
Selectivity Ratio	Dual-docks top leads vs. off-target proteins. ratio = target / off-target. ≥3.0 = High selectivity
Evolution Tree	D3 visualization of how seed molecules transform through scaffold hopping + bioisostere operations
LangSmith Tracing	Enterprise observability — every agent call, token count, latency, auditable in real time
Clinical Trial Matching	Live ClinicalTrials.gov API — links discovery to active patient trials
Resistance Forecasting	LLM predicts which secondary mutations will emerge under treatment pressure
Discovery Database	Neon PostgreSQL — persists all discoveries, browseable in the Discoveries Library

---

🗣️ Demo Script (Hackathon)

1. "Meet Sarah. 52 years old. Lung cancer. EGFR T790M. Erlotinib stopped working."
2. Type EGFR T790M → Launch Analysis
3. Watch 18 agents stream live in the PipelineStatus panel
4. Top Leads tab: "Our AI generated molecules that don't exist in any database. This one binds 3.2× harder to the cancer target than to healthy ABL1 kinase."
5. Evolution Tree: "Here's exactly how the seed molecule was transformed — scaffold hop +1.4 kcal/mol, bioisostere +0.8 kcal/mol."
6. Clinical Trials: "3 active Phase II trials targeting EGFR T790M. Our molecule targets the same pocket."
7. Open LangSmith tab: "Every decision, auditable. Enterprise ready."
8. Click Save Discovery → "Permanently saved to our Neon database."

---

🛠️ Development

Lint & Format

# Backendcd backend
ruff check .
mypy .# Frontendcd frontend
npm run check # Biome check
npm run check:fix # Auto-fix
npm run typecheck # TypeScript

Run Benchmark

cd backend
python -c "from evaluation.benchmark_runner import run_benchmark_casesimport asyncior = asyncio.run(run_benchmark_cases())print(f'Accuracy: {r[\"accuracy\"]*100:.0f}%')"

Verify SelectivityAgent

cd backend
python -c "from agents.SelectivityAgent import SelectivityAgentimport asyncior = asyncio.run(SelectivityAgent().run({ 'docking_results': [{'smiles': 'CC(=O)Nc1ccc(O)cc1', 'binding_energy': -8.5, 'compound_name': 'Test'}], 'mutation_context': {'gene': 'EGFR'}, 'analysis_plan': type('P', (), {'run_selectivity': True})(),}))print('SelectivityAgent OK:', r['selectivity_results'][0]['selectivity_label'])"

---

🌿 Git Branches

Branch	Owner	Scope
main	—	Protected. Merge via PR only
feature/backend-agents	Backend dev	Python agents, pipeline, bioinformatics
feature/frontend-ui	Frontend dev	Next.js pages, components, GSAP
feature/testing-validation	QA dev	Pytest, benchmark, E2E validation

---

📄 License

MIT — built for a hackathon. Not for clinical use. All results are computational predictions only.