STRESZCZENIE

This thesis introduces ResearchShop Disease2Gene, a comprehensive system for automated extraction of gene-disease associations from full-text biomedical literature.
Unlike prior systems that typically rely on abstracts or limited text segments, Disease2Gene is distinguished by its full-text processing capabilities and integration of multiple literature retrieval techniques. The system employs intelligent literature retrieval, citation-based prioritisation, and multi-strategy full-text acquisition that prioritises PMC Open Access while also retrieving content from DOI links and publisher websites using Playwright browser automation and Trafilatura extraction.
A two-stage Large Language Model architecture uses Google’s Gemini models: a lightweight model for initial gene discovery and a more capable model for detailed schema-driven attribute extraction.
The central contribution is a comprehensive multi-layer validation framework designed to mitigate LLM hallucinations. The framework brings together a few practical checks: it verifies gene symbols against the HGNC database; it recognises HGVS-style variant descriptions; and it double-checks citations to ensure the quoted text is actually present in the original paper. Taken together, these steps help keep the extracted information accurate and easy to review.