Multi-Agent Chatbot for Explainable Genetic Risk Analysis : Extending XPRS with Multi-Source Evidence Retrieval for Verifiable Risk Communication = Multi-Agent Chatbot for Explainable Genetic Risk Analysis

다국어 초록 (Multilingual Abstract)

Polygenic risk scores (PRS) are increasingly used to quantify genetic susceptibility to complex diseases, yet their clinical utility remains limited by poor interpretability. Recent advances such as eXplainable Polygenic Risk Score (XPRS) improve model transparency by decomposing PRS into gene-level contributions; however, interpreting these outputs still requires substantial domain expertise. This thesis presents a multi-agent, evidence-grounded chatbot designed to as a digital genetic counselor by generating clear, scientifically supported explanations of patient-specific XPRS results.
Using breast cancer susceptibility as a primary case study, the proposed system integrates three complementary evidence sources: (1) a vector-indexed knowledge base for structured retrieval of gene-specific mechanistic context, (2) the Semantic Scholar API to retrieve peer-reviewed scientific literature, and (3) the Open Targets platform to verify gene-disease associations. A multi-agent workflow comprising retrieval, generation, and evaluation components anchors explanation generation to validated sources and minimizes hallucinated citations. The system produces structured summaries that combine gene-level contribution scores with verifiable scientific references.
The prototype was evaluated through a pilot mixed-methods user study involving quantitative usability surveys and qualitative feedback from participants familiar with genomic data and polygenic risk concepts. The results indicate that integrating curated genetic databases with controlled LLM-based reasoning supports more interpretable and evidentially reliable communication of polygenic risk.

목차 (Table of Contents)

• Introduction 3
• Related Work 5
• Methods 8
• Results 17
• Discussion and Conclusion 20
• References 23
• Appendix 24