zERExtractor: An Automated Platform for Enzyme-Catalyzed Reaction Data Extraction from Scientific Literature.

Area of Science:

• Biotechnology
• Computational Enzymology
• Bioinformatics

Background:

• The exponential growth in enzyme reaction literature presents a significant challenge for database curation.
• Vast amounts of enzyme-substrate-condition data remain unstructured, limiting their use in deep learning (DL) models for enzyme activity prediction.
• Existing DL-based data extraction methods often lack fidelity checks and continuous evolution capabilities.

Purpose of the Study:

• To develop an accurate and extensible platform, zERExtractor (Zelixir's Enzyme Reaction Data Extractor), for extracting enzyme-catalyzed reaction data from scientific publications.
• To create a unified multimodal information extraction framework capable of processing molecular reaction diagrams, tables, and text.
• To integrate enzymatic reaction descriptors into structured storage for improved DL-driven modeling.

Main Methods:

• Implementation of a human-in-the-loop pipeline utilizing fine-tuned large language models (LLMs) and DL.
• Integration of data fidelity validation by experts and active learning for continuous model evolution.
• Development of a multimodal information extraction framework covering molecular diagrams, tables, and text.

Main Results:

• zERExtractor achieved 89.9% accuracy in table recognition and over 98% accuracy in molecular image recognition on synthetic datasets.
• The platform outperformed the strongest baseline by more than 2% on synthetic data.
• Consistent performance above 95% accuracy was maintained on realistic benchmarks.

Conclusions:

• zERExtractor effectively bridges the data gap in enzyme reaction information.
• The platform provides a scalable framework for accurate multimodal data extraction.
• This advancement supports DL-driven enzyme modeling, computational enzymology, and biotechnology applications.