Abstract 3651: AI-derived immune cell knowledge graphs to advance immunotherapy

Abstract Immune cells are critical units in human immunity, yet our current understanding of their roles remains limited. A critical barrier is a lack of human natural language in describing the complex molecular roles of immune cells in diverse contexts. Current descriptors of molecular features such as genes and pathways are limited by biological terms available in the functional annotation databases, which were not created specifically to describe immunity, lacking granularity and completeness. Although richer language exists in biological literature, they require labor-intensive curation. We hypothesize that by applying cutting-edge AI techniques such as large language models (LLMs), we can curate a knowledgebase specific for describing human immunity, consisting of rich molecular-cellular relationships and functional descriptors not available in current databases. Querying this knowledgebase can substantially enrich our language in describing and understanding the complex roles of immune cells, and generate novel, verifiable hypotheses for discovery and development of immunotherapy. To address these challenges, we have constructed immune cell type-specific knowledge graphs (KG) for T cells, NK cells, B cells, and macrophages by parsing 9, 995, 3, 665, 7, 503, and 3, 615 relevant published abstracts since 2020, respectively. We have fine-tuned BioBERT models to perform named entity recognition (NER) on genes, diseases, and cell types, formulated rule-based algorithms to extract pathways and biomedical concepts. We then employed the LLaMA3-70B model to infer relationships between the extracted entities based on abstract using carefully designed prompts. We present the first immune cell type-specific KGs. Using LLMs, we overcome limitations in immune vocabularies, extracting concise biomedical phrases to expand vocabularies and better characterize immune cell biology beyond existing knowledge bases. These enriched vocabularies can annotate gene sets, providing annotations that reflect recent literature. The KGs, represented as directed networks of biomedical entities, support hypothesis generation and experimental design. By querying relationships among genes and concepts, researchers can identify genes linked to specific phenotypes and predict phenotypic changes following interventions such as gene silencing or cytokine stimulation. These KGs enable dynamic queries to explore immune cell complexity, identify therapeutic targets, and design experiments to investigate signaling pathways. To enhance accessibility, we provide a web application for exploring genes and phenotypes in a cell type-specific manner: https://kchen-lab.github.io/immune-knowledgegraph.github.io/. Citation Format: Shan (Chloe) He, Yukun Tan, Vakul Mohanty, Hind Rafei, Katy Rezvani, Ken Chen. AI-derived immune cell knowledge graphs to advance immunotherapy [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 3651.