VIOLIN: A modular framework for scalable reconciliation of heterogeneous interaction graphs

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Automated extraction of molecular interactions from scientific literature is advancing rapidly.
• Integrating these extracted interactions with curated models and knowledge graphs remains a challenge.

Purpose of the Study:

• To present VIOLIN (Versatile Interaction Organizing to Leverage Information in Networks), a configurable framework for reconciling extracted molecular interaction data with structured knowledge graphs.
• To evaluate VIOLIN's performance using interaction lists from various natural language processing (NLP) systems and large language models (LLMs).

Main Methods:

• VIOLIN formally compares newly extracted interaction lists against structured baseline graphs.
• It classifies interactions into categories: corroboration, contradiction, flagged case, or extension.
• The framework supports configurable attribute inclusion and mismatch semantics for adjustable reconciliation strictness.

Main Results:

• Reconciliation outcomes were stable and interpretable across different NLP systems, LLMs, literature corpora, and baseline graphs.
• Extensions were the dominant classification, with the balance of corroborations and contradictions reflecting underlying graph structures.
• Sensitivity analyses confirmed predictable shifts in classification based on attribute inclusion and scheme selection.

Conclusions:

• VIOLIN provides a robust and configurable method for reconciling extracted molecular interactions with existing knowledge graphs.
• Its high algorithmic correctness and alignment with expert curation demonstrate its utility for biological data integration.
• VIOLIN is available as a Python package and web interface, facilitating its adoption in research.