DECIDER genetics knowledge graph and decision support

This pipeline creates a BioCypher knowledge graph from synthetic data after the example of the DECIDER cohort. This knowledge graph is then connected to a BioChatter instance that allows querying the graph and other information (papers in a vector database, information from web APIs such as OncoKB) via natural language. We provide two applications, one based on BioChatter Light, the other on BioChatter Next.

You can find more detailed information in a vignette on our website and see hosted demonstrations of both applications at the following links:

• Light: https://decider-light.biochatter.org

• Next: https://decider-next.biochatter.org

🐳 Run using Docker

[!IMPORTANT] You need to have Docker installed on your machine to run the following commands. Please go to Docker for instructions.

{bash} git clone https://github.com/biocypher/decider-genetics.git cd decider-genetics docker compose up -d

This will build the KG in the Docker container and start a Neo4j instance as well as a BioChatter Light web app instance configured to only show the KG query interface (for more info see this vignette). The Neo4j instance can be accessed at localhost:7474 and the BioChatter Light instance at localhost:8501.

For the BioChatter Next variant, you can run the following command:

{bash} docker compose -f docker-compose-next.yml up -d

This will similarly build a KG, but also a Milvus instance for the vector database, and instances of the biochatter-server REST API service and the BioChatter Next app, which can be accessed at localhost:3000.

[!IMPORTANT] For using OpenAI GPT as the language model, you will have to provide your API key as an environment variable (OPENAI_API_KEY) in your environment. You can do this using an export command (export OPENAI_API_KEY=sk-...) or by adding it to your bash profile; you could also provide it to Docker using an env file. We use GPT-3.5-turbo as the default model.

Questions

The knowledge graph contains information about patients, genes, variants, drugs, pathways, and clinical data. The schema of the KG can be seen below the query interface as a JSON object. You can ask questions in natural language, such as:

• How many patients do we have, and what are their names?

• What was patient1's response to previous treatment, and which treatment did they receive?

• Which patients have hr deficiency but have not received parp inhibitors?

• Does patient4 have a sequence variant in a gene that is druggable with evidence level "1"? Which drug? Return unique values.

• Is there a patient with overlapping variants compared to patient4?

• Which genes are these overlapping sequences in?

These are only few of infinitely many possible questions, and some may not result in a valid query. The BioChatter Light interface allows manual modification and rerunning of the query for prototyping and debugging.

Vectorstore RAG and OncoKB API questions

For convenience, here are some questions to get you started with the RAG and API calling features:

• (OncoKB API) What is the consequence of the TP53 R273C mutation in high grade serous ovarian cancer?

• (OncoKB API) Are there reports of of the functional fusion of CD47 and ROS1 in ovarian cancer?

• (OncoKB API) What is the consequence of of the functional fusion of CD47 and ROS1?

• (OncoKB API) What is the therapeutic relevance of the BRAF V600E mutation in high grade serous ovarian cancer?

• (RAG) Have there been reports of TP53 being therapeutically relevant in HGSC?

• (RAG) Do pro-inflammatory cytokines play a role in the progression of ovarian cancer?

• (RAG) Does RAD51C play a role in ovarian cancer?

⚙️ Local Installation

You can run the KG build locally using a virtual environment.

{bash} git clone https://github.com/biocypher/decider-genetics.git cd decider-genetics poetry install poetry run python create_knowledge_graph.py