A multi-agent, multi-knowledge-graph biomedical question–answering system built with Google ADK and TRAPI/ReasonerAPI.
This project lets a user ask questions like:
“What causes epilepsy, which genes are involved, what drugs target those genes, and are there clinical trials?”
The system:
- Converts the question into a ReasonerAPI / TRAPI query graph.
- Uses a multi-agent orchestration layer to query multiple biomedical knowledge graphs (KGs):
- Monarch Initiative KG
- Clinical Connections KG
- BigGIM Drug Response KG
- Clinical Trials KG
- Drug Approvals KG
- Wellness Multiomics KG
- Merges and ranks evidence across KGs.
- Uses an LLM to generate a consolidated, provenance-aware explanation.
Biomedical facts are scattered across many specialized KGs and APIs. No single source can fully answer multi-hop questions that span:
- Disease → Gene → Drug → Clinical Trial
- Disease → Phenotype → Gene → Pathway
- Gene → Variant → Drug Response
Clinicians and researchers must manually jump between portals and APIs, which is:
- Time-consuming
- Error-prone
- Hard to reproduce
- Difficult to explain or share as a single coherent answer
LLMs can explain, but LLMs alone hallucinate if not grounded in structured, curated knowledge.
This project builds a KG-aware chat assistant that:
- Accepts natural language biomedical questions
- Generates a TRAPI query graph using ReasonerAPI conventions.
- Queries multiple KGs in parallel via agents/tools:
- Monarch Initiative (disease–gene–phenotype)
- Clinical Connections (causal gene–drug–disease relationships)
- BigGIM (expression / omics / drug response)
- Drug Approvals KG (FDA labels)
- Clinical Trials KG (NCT trials)
- Wellness Multiomics KG (pathways, variants, omics)
- Merges and ranks evidence into a canonical mini-KG slice
- Produces a clear, explainable answer with provenance information
- Chatfront Agent – UI-facing conversation agent
- Query Graph Builder Agent – converts NL question → TRAPI query graph
- Multi-KG Orchestrator Agent – routes TRAPI queries to KPs
- Per-KG Agents/Tools – Monarch, Clinical Connections, BigGIM, Drug Approvals, Clinical Trials, Multiomics
- Evidence Merger Agent – canonicalises nodes/edges, aggregates scores, ranks answers
- Explanation Agent – generates human-readable answer from merged evidence
- Google ADK (Agent Developer Kit) – multi-agent orchestration
- TRAPI / ReasonerAPI – standardised biomedical query/response format
- Biolink Model – semantic categories and predicates
- Python 3.10+, FastAPI – serving the chat API
- httpx – async HTTP calls to KGs
- Optional: Docker, GitHub Actions for CI
git clone https://github.com/yogesh-parte/MultiKG-bioAgent.git
cd multikg-bioAgent
uv venv
source .venv/bin/activate # Windows: .venv\Scripts\Activate.ps1
uv sync
See the docs/ folder for:
- PROJECT_PROPOSAL.md – problem, motivation, high-level solution
- ARCHITECTURE.md – agent design and call graph
- AGENTS.md – per-agent responsibilities and prompts
- KGS_INTEGRATION.md – how each KG is integrated
- TRAPI_EXAMPLES.md – example TRAPI messages and responses
- ROADMAP.md – planned features.
See asset folder for interactive visualization with the end product
- Interactive-Software-Workflow-Concept1.html
- Interactive-Software-Workflow-Concept2.html
- Workflow-MultiKG-BioAgent.html
We gladly welcome issues, ideas, and PRs.
See CONTRIBUTING.md for:
- how to set up your dev environment
- coding style
- how to add a new Knowledge Provider
- how to extend the agent workflow
⸻
This project is licensed under the Apache License 2.0 – see LICENSE.
⸻
This project builds on the work of:
- NCATS Translator and the Translator ReasonerAPI/TRAPI ecosystem
- Monarch Initiative and other biomedical knowledge graph providers
- The open-source and clinical data science communities
- Yogesh PARTE, PhD
- Rohan Routh, MS
- Celdrick Kuta, MS
- Sravanti Tatineni, MS