PharmKGPT

PharmKGPT: Accelerating Delirium and Alzheimer's Research through AI-Driven Knowledge Discovery

To combat the issue of large language model "hallucinations" in specialized fields like medicine, this platform builds a bioinformatics-focused knowledge Q&A system for specific diseases such as Delirium and Alzheimer's Disease, ensuring responses are derived from curated scientific knowledge rather than generalized training data.

Key Features

• Builds Specialized Knowledge Bases: Automatically extracts entities (Genes, Proteins, Pathways, Processes, Metabolites) and their relationships from targeted literature (e.g., analyzing 153,000+ PubMed abstracts relevant to Delirium and Alzheimer's Disease) to create custom Knowledge Graphs (KGs) focused on these specific neurological conditions.

• Advanced Q&A System: Employs an innovative hybrid Retrieval-Augmented Generation (RAG) approach, combining semantic search with Knowledge Graph queries, to provide high-precision answers to complex biological questions related to Delirium and Alzheimer's Disease.

• Visual Insights: Presents visualized, relevant subgraph snippets alongside answers to enhance understanding and transparency of connections within the knowledge domain.

• Goal: To empower researchers to efficiently explore disease mechanisms for Delirium and Alzheimer's Disease, identify potential drug targets, and accelerate the development of effective therapies.

Table of Contents

1. Prerequisites
   • 1.1 Install Ollama & Models
   • 1.2 Download Reranker Model
   • 1.3 Prepare Data
2. Setup & Run
   • 2.2 Local (Conda)
3. Test
4. Data Version History
5. KG Construction Source Code

1. Prerequisites

1.1 Install Ollama & Models

Install ollama.

# Pull required models
ollama pull deepseek-r1:32b
ollama pull nomic-embed-text:latest
ollama pull gemma3:27b
ollama pull Qwen2.5:0.5b

# Start the Ollama server in the background
ollama serve &

1.2 Download Reranker Model

# Download helper script (if you don't have it)
wget https://hf-mirror.com/hfd/hfd.sh
chmod a+x hfd.sh
export HF_ENDPOINT=https://hf-mirror.com
./hfd.sh BAAI/bge-reranker-large

# Extract the model archive (assuming it downloads as a tar.gz)
tar -zxvf bge-reranker-large.tar.gz

Adjust paths and download methods if necessary.

1.3 Prepare Data

The data archive supporting the study is publicly available through Figshare.

After downloading the archive:

tar -zxvf Data_v7_0509.tar.gz

2. Setup & Run

Choose either Docker or Local setup.

2.1 Local (Conda)

1. Create Environment & Install Dependencies:

   conda create -n rag python=3.10 -y
   conda activate rag
   
   pip install -r requirements.txt
   
   pip install llama-index-llms-gemini==0.4.13
   pip install itext2kg==0.0.7

2. Configure Environment Variables: Create a .env file in the project root with:

   # Path to the downloaded BGE reranker model directory
   RERANKER_PATH=/path/to/your/bge-reranker-large
   
   # Path to the root directory of the dataset you want to use
   # Example: /path/to/your/Data_v7_0509 or /path/to/your/Data_test_v4
   DATA_ROOT=/path/to/your/Data

3. Run the Application Servers: You need to start two components: the backend API server and the web interface builder.

   • Start the Backend API Server:

     # Runs the Flask/FastAPI application in the background, logging output to app.log
     nohup python app.py > app.log 2>&1 &

     Note: The API server typically runs on http://localhost:5000.

   • Start the Web Interface:

     # Runs the Gradio/Streamlit web interface builder
     python builder.py

     Note: The web interface will be accessible by default at http://127.0.0.1:7860/.

4. Querying the System: You can interact with the system in two main ways:

   • Via Web Interface: Open your browser and navigate to http://127.0.0.1:7860/.
   • Via API (Terminal): Send a POST request to the backend API endpoint:

     curl -X POST http://localhost:5000/api/query \
          -H "Content-Type: application/json" \
          -d '{"question": "How does kynurenic acid contribute to delirium?"}' \
          -N

   (Optional: Mention run.py here if it provides a distinct, important way to interact, e.g., "For single-script execution or specific tasks, you might use run.py [add brief explanation if needed].")

KG construction source code

See kg_construction/README.md for the construction source, dependencies, sample input and run commands.

Running Tests

To run the automated tests:

1. Ensure Test Data: Make sure you have the test dataset (e.g., Data_test_v4) downloaded and extracted.
2. Configure Environment: Verify that the DATA_ROOT variable in your .env file points to the test dataset directory (or ensure it's correctly mounted if using Docker).
3. Execute Pytest:

   pytest test/test_QA.py