EnzyExtract

Extract kinetics data from PDFs using LLMs.

Installation

git clone https://github.com/ChemBioHTP/EnzyExtract
cd EnzyExtract
pip install -e .

Furthermore, create a .env file in the project root. Add your OpenAI API key:

OPENAI_API_KEY=...

Quickstart

A preliminary tutorial is available on Google Colab or downloadable here.

Database

For explanations of each header, see docs/enzyextract_headers.md. Reading it is recommended due to possible surprises.

Initial Release

Filename	kcat count	Location	Description
Original, crude	242,115	data/export/TheData_kcat.parquet	As originally described
Deduplicated, crude	218,095	data/export/TheData_unpruned.parquet	With deduplication
EnzyExtractDB_176463	176,463	EnzyExtractDB/EnzyExtractDB_176463.parquet	Deduplicated and filtered

The data is released as CC BY-NC 4.0. (NC is due to Elsevier and Wiley TDM APIs.)

The code is released as MIT.

Usage

See experiments/example/pipeline/ex_step*.py for example scripts. The scripts should be run sequentially, though file paths may need to be adjusted.

Steps:

1. ex_step0_run_preprocessing.py:
   • Handles the preprocessing steps (ResNet, Table Extraction)
   • create a .enzy folder for simplified file management
2. ex_step1_run_tableboth.py
   • Given PDFs and preprocessed data, feed to LLMs using Batch API.
   • File locations should be automatically saved to .enzy/llm_log.tsv.
3. ex_step1b_run_pdf_binaries.py
   • Alternative to step0 and step1: feed PDF binaries directly to Claude.
4. ex_step2_download.py
   • Small script to retrieve batches from Batch APIs.
5. ex_step3_llm_to_df.py
   • Convert the LLM output to parquet files.
6. ex_step5_generate_identifiers.py
   • Optional: Attach sequence identifiers (EC number, UniProt ID, PDB ID, SMILES, PubChem ID) to the data from step3.

Evaluation

See experiments/example/evaluation/ex_step*.py.

1. ex_eval1_compare_dfs.py
   • Evaluate and benchmark LLM data against a trusted dataset.
2. ex_eval2_plot_dfs.py
   • Plot the data from ex_eval1.

Accessions

Enzyme accession pipeline is a WIP. Please see docs/pipeline/enzyme_accessions.md.

For docs on getting substrate IDs, please see docs/pipeline/substrate_accessions.md

Architecture

Citation

If you find EnzyExtract useful, please cite it as below:

@article{wei2025finding,
  title={Finding the dark matter: Large language model-based enzyme kinetic data extractor and its validation},
  author={Wei, Galen and Ran, Xinchun and AI-Abssi, Runeem and Yang, Zhongyue},
  journal={Protein Science},
  volume={34},
  number={9},
  pages={e70251},
  year={2025},
  publisher={Wiley Online Library}
}