Bio-Vector Orbit - Discovery Engine v2.0

A Vector-Powered Discovery Engine that transforms fragmented biological data (papers, DNA sequences, and chemical molecules) into a unified, searchable intelligence layer using NCBI, Qdrant, FastEmbed, and React.

---

🚀 Quick Start with Docker (Recommended for Judges)

Prerequisites: Docker and Docker Compose installed.

1. Start the Application

docker-compose up

That's it! The entire stack will start automatically.

2. Access the Application

• Frontend Dashboard: http://localhost:3000
• Backend API: http://localhost:8000/docs (Interactive Swagger UI)
• Qdrant Dashboard: http://localhost:6333/dashboard

3. Try It Out

1. Open http://localhost:3000
2. Choose Your Discovery Mode:
   • Text: Search PubMed articles (e.g., "Alzheimer's treatments")
   • Protein: Visualize human protein structures from PDB (e.g., "Insulin")
   • Molecule: Discover chemical compounds via SMILES latent search (e.g., "CCO")
3. Watch as the system retrieves data from NCBI, RCSB PDB, or your custom SMILES VAE model.

4. Stop the Application

docker-compose down

---

✨ Features

Core Capabilities

• 🧬 NCBI PubMed Integration: Real-time retrieval of the latest biomedical literature.
• 🧬 PDB Protein Discovery: Immersive structural search across the RCSB Protein Data Bank.
• 🧪 SMILES Chemical Search: Latent space discovery using a custom VAE model for molecular similarity.
• 🔍 Unified Semantic Search: Bge-small-en-v1.5 embeddings power both literature and structural queries.
• 🧱 Advanced Visualizer: PDBe-molstar integration with illustrative lighting and futuristic HUD.
• 📊 Multimodal Results: Hybrid UI renders specialized cards for papers, proteins, and molecules.

Technical Highlights

• Quad-Collection Architecture: Separate vector spaces for Text, Proteins, and Chemical Latent vectors.
• HNSW Indexing: High-precision semantic clustering for sub-second retrieval.
• Unified Embedding Logic: Consistent BGE-Small vectorization across all biological entities.
• Zero-Setup Docker: Full orchestration handles model pre-loading and DB health checks.

---

🏗️ Architecture

┌─────────────┐      ┌──────────────┐      ┌─────────────┐
│   React     │─────▶│   FastAPI    │─────▶│   Qdrant    │
│  Frontend   │      │   Backend    │      │  Vector DB  │
│  (Port 80)  │      │  (Port 8000) │      │ (Port 6333) │
└─────────────┘      └──────────────┘      └─────────────┘
                            │
            ┌───────────────┴───────────────┐
            │                               │
            ▼                               ▼
     ┌──────────────┐                ┌──────────────┐
     │ NCBI PubMed  │                │   RCSB PDB   │
     │   (Biopython)│                │   (Structure)│
     └──────────────┘                └──────────────┘

Technology Stack

• Frontend: React 19 + Vite + TypeScript + TailwindCSS
• Backend: Python 3.11 + FastAPI + Uvicorn
• AI Models: FastEmbed (BGE-Small) + Custom Keras VAE (SMILES)
• Vector Database: Qdrant (Hybrid Cloud/Local)
• Data Sources: NCBI, RCSB PDB, PubChem
• 3D Visualization: PDBe-molstar

---

📁 Project Structure

Vectros-in-orbit2.0/
├── backend/
│   ├── main.py              # Central search router & endpoints
│   ├── qdrant_db.py         # Multi-collection client logic
│   ├── ingest_proteins.py   # PDB structure scraper & ingestion
│   ├── smiles.py            # Latent molecular search logic
│   ├── Blog_simple...h5     # Chemical discovery VAE model
│   ├── Dockerfile           # Neural backend container
│   └── requirements.txt     # Unified dependency list
├── components/
│   ├── ProteinViewer.tsx    # Illustrative 3D visualization HUD
│   ├── ProteinCard.tsx      # Structure-focused result card
│   ├── MoleculeCard.tsx     # Chemical formula search card
│   └── ResultCard.tsx       # Standard PubMed article card
├── App.tsx                  # Dual-mode state & hybrid UI logic
└── ...

---

🧪 How It Works

1. Discovery Mode: User toggles between Literature, Proteins, and Molecules.
2. Neural Retrieval:
   • Text/Protein: BGE embeddings find semantic matches in PubMed/PDB.
   • Chemical: Keras VAE converts SMILES into latent vectors for similarity search.
3. Hybrid Rendering: Backend returns standardized types, and the UI adapts the card style (PubMed vs. PDB vs. SMILES).
4. Immersive Viewing: Proteins are loaded into a high-precision 3D viewport with illustrative depth and real-time metadata HUD.

---

---

🐳 Running with Docker

The application is fully containerized for a seamless experience across all platforms.

# Build and start all services
docker-compose up --build

# Shutdown the application
docker-compose down

---

🎯 Use Cases

• Drug Discovery: Find papers on specific protein targets or compounds
• Literature Review: Semantic search across biomedical abstracts
• Hypothesis Generation: Discover connections between biological concepts
• Education: Explore biological topics with visual 3D structures

---

📝 License

MIT License - See LICENSE file for details

---

👥 Author

Built for the Vectors in Orbit Hackathon by Karray Aziz

---

🙏 Acknowledgments

• NCBI PubMed: For providing open access to biomedical literature
• Qdrant: For the powerful vector database solution
• FastEmbed: For efficient, lightweight embeddings
• iCn3D: For 3D molecular visualization

---

Questions? Open an issue or contact the maintainer.

Enjoy exploring the biology discovery engine! 🧬🚀