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

Abstract

Automated extraction of molecular interactions from scientific literature has outpaced the development of systematic methods for integrating this information with curated models and knowledge graphs. Here we present VIOLIN (Versatile Interaction Organizing to Leverage Information in Networks), a configurable, attribute-aware reconciliation framework that formally compares newly extracted interaction lists against structured baseline graphs. VIOLIN classifies each interaction as a corroboration, contradiction, flagged case, or extension, and supports configurable attribute inclusion strategies and mismatch semantics to adjust reconciliation strictness. We evaluate VIOLIN using interaction lists generated by two traditional NLP systems (REACH, INDRA) and two large language models (GPT-4.1, Llama 3) across multiple literature corpora and structurally distinct baseline graphs. Across all conditions, reconciliation outcomes were stable and interpretable, with extensions dominating and corroboration-contradiction balance reflecting intrinsic structural relationships between baseline graphs and extracted evidence. Sensitivity analyses demonstrate that attribute inclusion and classification scheme selection shift category boundaries predictably. Benchmark evaluations confirm high algorithmic correctness and alignment with expert curation. VIOLIN is publicly available as a Python package and through web-based interface (https://nmzlab.github.io/Tools-UI).