NeuraGraph: A Visual Workflow Framework for Reproducible Biomedical Text Mining with LLM-Agents

Summary Biomedical text mining increasingly demands hybrid pipelines integrating traditional NLP toolkits and LLM-based agents, yet building reproducible and modular workflows remains challenging. We present NeuraGraph, a lightweight open-source Python framework for reproducible biomedical text mining workflows integrating traditional NLP components and LLM-based agents. NeuraGraph provides hierarchical workflow composition through reusable subgraphs, built-in batch experimentation with standardized evaluation, and an LLM-powered reporting and refinement engine for workflow optimization. We evaluated NeuraGraph on named entity recognition and chemical-induced disease relation extraction using the BioCreative V CDR benchmark. An optimized relation extraction workflow improved Micro-F1 from 0.656 to 0.673 and Macro-F1 from 0.642 to 0.666 through reporting-guided refinement. Compared with functionally equivalent custom implementations, NeuraGraph reduced implementation effort from 366 to 135 source code lines for relation extraction workflows while maintaining comparable extraction performance. The framework facilitates systematic comparison, reproducible evaluation, and rapid development of hybrid biomedical text mining pipelines. Availability and Implementation NeuraGraph is implemented in Python 3.12+ and released under the MIT license, with all dependencies listed in a requirements.txt file for straightforward installation. Source code and documentation are available at https://github.com/tyrone1979/neuragraph. The software runs natively on Linux, macOS, and Windows. Supplementary information Supplementary data are available at Bioinformatics online.