Mechanistic Discovery in Complex Diseases through Multi-Modal AI and Knowledge Graphs

Understanding the molecular mechanisms underlying complex diseases remains a central challenge in biomedical research. Disorders such as neurodegenerative and inflammatory diseases result from interactions across multiple biological scales, making causal interpretation difficult despite the rapid expansion of experimental and clinical data. In recent years, large volumes of heterogeneous evidence—from scientific literature and curated databases to omics and imaging—have become available. However, this knowledge remains fragmented across formats and modalities. While manual curation provides high-quality mechanistic resources, it cannot keep pace with the rate of discovery, leading to incomplete and quickly outdated representations of disease biology. This dissertation develops automated and scalable approaches to transform heterogeneous biomedical evidence into continuously updated, computable knowledge representations. The work focuses on artificial intelligence (AI)–based methods for constructing, expanding, and evaluating biomedical knowledge graphs (KGs) that encode causal relationships and support systematic, mechanism-centered analysis of complex diseases. To enable scalable updates of mechanistic KGs, a semi-automated natural language processing (NLP) framework based on a fine-tuned BERT model was developed. Using tau phosphorylation in Alzheimer's disease as a case study, automated relation extraction combined with expert validation allowed the integration of newly reported regulators while maintaining high curation quality. The approach was further extended to additional applications, including expansion of the Heme KG to improve coverage of the Heme-TLR4 inflammatory axis, and construction of a comorbidity graph linking COVID-19 and neurodegenerative disorders, which revealed shared pathogenic mechanisms. The dissertation further benchmarks large language models (LLMs) for biomedical triple extraction in Biological Expression Language (BEL). Although LLMs generate fluent mechanistic statements, systematic evaluation reveals limitations, including hallucinated relations and output instability, underscoring the need for structured evaluation and human-in-the-loop validation. Finally, automated knowledge construction is extended beyond text by extracting mechanistic information from biomedical figures, demonstrating that visual data often contain causal relationships not explicitly described in the accompanying text. Overall, this work develops AI-assisted approaches for transforming fragmented biomedical evidence into structured, evolving KGs and supports mechanistically grounded discovery in complex disease research.