BioWorkflow: Retrieving comprehensive bioinformatics workflows from publications.

Area of Science:

• Bioinformatics
• Computational Biology
• Scientific Literature Mining

Background:

• Reconstructing bioinformatics workflows from publications is crucial for scientific analysis but is hindered by dispersed information across various document formats.
• Manual reconstruction is time-consuming and prone to errors, while existing systems lack the full-text, multimodal capabilities for complete workflow extraction.

Purpose of the Study:

• To introduce BioWorkflow, a novel large language model (LLM)-based framework designed for automated, end-to-end reconstruction of bioinformatics workflows from scientific literature.
• To improve the accuracy, completeness, and reproducibility of bioinformatics pipeline extraction and reuse.

Main Methods:

• BioWorkflow employs a retrieval-augmented framework that parses PDFs to create a unified index of text, tables, and figures.
• It uses hierarchical query decomposition with dynamic reformulation and iterative, context-aware retrieval to assemble directed workflows.
• The framework links predicted elements to cited evidence and performs automated consistency checks to ensure traceability and minimize hallucinations.

Main Results:

• BioWorkflow recovers approximately 80% of workflow steps, a significant improvement over existing tools (~20%).
• It demonstrates a >20% improvement in reproducibility, completeness, and accuracy compared to strong LLM baselines.
• Curation time is reduced to 3-5 minutes per paper, enabling faster reuse of published pipelines.

Conclusions:

• BioWorkflow effectively automates the extraction of detailed bioinformatics workflows from scientific publications.
• The framework enhances the reliability and efficiency of reusing published computational pipelines, fostering scientific reproducibility.