prime.sarkome.com

General LLMs lack precise biomedical knowledge. This PrimeKG chat lets users query a unified precision medicine graph in natural language, accelerating insights from genes to phenotypes.

Comment

Inspiration

Drug discovery takes 10-15 years and costs billions, yet efficient data exploration remains difficult. We wanted to replace complex database queries with a simple interface, allowing researchers to "chat" directly with the 129,000 entities in the PrimeKG biomedical knowledge graph.

What it does

PrimeKG Explorer is an AI-powered interface for biomedical research. It uses Google Gemini to translate natural language into structured graph queries, enabling users to find drug repurposing opportunities, discover therapeutic targets, and visualize molecular mechanisms. It includes semantic search and real-time reasoning visualization.

How we built it

We built a React and TypeScript frontend using Vite for performance. The core logic relies on Google Gemini 3.0 and its Function Calling capabilities to select among 7 specialized tools. These tools query a Neo4j database containing the biological knowledge. We used localStorage for client-side session persistence.

Challenges we ran into

  • AI Complexity: Managing 7 different tools and keeping the AI context window efficient was difficult.
  • State Management: Creating a seamless single-page experience without page reloads required a custom session key architecture.
  • Visualization: Extracting and displaying real-time execution steps (traces) from the Gemini stream required a custom aggregation system.

Accomplishments that we're proud of

  • Autonomous Logic: The AI successfully chains over 10 tool calls to answer complex questions like finding inhibitors for specific disorders.
  • UX Quality: We achieved a polished, production-ready interface with light/dark modes and auto-resizing inputs.
  • Zero-Config: The application runs directly in the browser without complex backend setup.

What we learned

We learned how to optimize prompts for Gemini's Function Calling to ensure the model chooses the right API endpoints. We also validated that strict TypeScript typing prevents bugs when integrating AI with structured data APIs.

What's next for prime.sarkome.com

  • Immediate: Add Firebase authentication and cloud history synchronization.
  • Medium-term: Support multi-modal inputs (PDFs/images) and create autonomous agents for complex workflows.
  • Long-term: Integrate Clinical Trial matching and expanded knowledge graphs.

Built With

  • cloud
  • google
  • google-gemini-2.0
  • google-gemini-2.0-flash
  • google-gemini-3.0
  • google-genai-sdk
  • google-genai-sdk-database:-neo4j-(graph-database)-apis:-custom-primekg-api-(hosted-at-https://kg.sarkome.com/docs)
  • neo4j
  • primekg-api
  • python
  • python-frontend-frameworks:-react-18
  • react-18
  • tailwind-css
  • tailwind-css-ai-&-models:-google-gemini-3.0-(flash/pro)
  • typescript
  • vite
Share this project:

Updates