Radiology Dataset Database

A pipeline for automatically discovering, extracting, and structuring radiology datasets from the literature.

• data/: contains the final dataset table (e.g., radiology_db.csv)
• notebooks/: tutorials and exploratory data analysis notebooks
• scripts/: scripts for running the database building pipeline
• radiology_dataset_db/: source code for querying PubMed, extracting dataset metadata, and building the database
• tests/: pytest-based testing suite

⚙️ Installation

1. Install and run vLLM (for local LLM inference)

conda create -n vllm python=3.10 -y
conda activate vllm
pip install vllm

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen2.5-7B-Instruct \
  --port 8001 \
  --enforce-eager \
  --enable-auto-tool-choice \
  --tool-call-parser hermes

2. Set up the radiology dataset database environment

git clone git@github.com:pachterlab/radiology_dataset_db.git
cd radiology_dataset_db
conda create -n radiology_dataset_db python=3.10 -y
conda activate radiology_dataset_db
pip install -e .

3. Rename .env_sample to .env and fill in your Entrez email and API key (optional but recommended for higher rate limits)

🚀 Usage

Modify .env and radiology_dataset_db/config.py as needed to customize PubMed query/LLM settings. Runtime defaults like modality/output paths are now configured via CLI args in scripts/build_db.py. Then run:

python scripts/build_db.py --database-modality MODALITY

Currently supported modalities:

• Radiology: --database-modality radiology
• Single-cell RNA-seq: --database-modality scrnaseq
• Bulk genomics (e.g., bulk RNA-seq, WGS/WXS): --database-modality bulk_genomics
• Spatial transcriptomics: --database-modality spatial_transcriptomics

To enable parallel extraction, increase --num-threads (for example --num-threads 8).

To add more modalities (e.g., genomics, pathology):

1. Define new pubmed query and extraction instructions in radiology_dataset_db/config.py
2. Implement new dataset schema class and extraction function in radiology_dataset_db/extract_MODALITY_dataset_information_llm.py
3. Import and call the new extraction function in scripts/build_db.py and add a conditional to check the modality type
4. Optionally, update .github/workflows/update_dbs.yml to run the pipeline for the new modality on a schedule
5. Optionally, add some ground truth papers to tests/conftest.py, and add to get_modality_info in tests/test_llm_output.py to check that the new extraction function is working as expected All instructions are notaded in the code with comments like #* add additional extraction instructions and functions for other modalities here, e.g. genomics, pathology, etc

Example codex prompt used to add a scRNA-seq dataset schema and extraction function:

Pleas write a module very similar to extract_radiology_dataset_information_llm.py called extract_scrnaseq_dataset_information_llm.py that looks for scRNA-seq/snRNA-seq data. It should look for name, num_patients, sequencing_technology (eg 10X, SMARTSEQ, Parse, etc), disease, species, tissue, cell/nuclei. Also have fields for paper_title, paper_link, paper_year etc that get populated afterwards. Add instructions in config.py, and add an extra condition to build_db.py (areas to edit are marked by "#*"). Add integration test structure in test_llm_output.py and add a placeholderground truth paper to conftest.py to test the new extraction function.

Testing

Just unit tests:

pytest

Just integration tests:

pytest -m integration (not all tests need to pass because LLM has some randomness, but most should pass consistently)

All tests:

pytest -m ""