Evidence-Grounded Clinical Pharmacogenomics Decision Support: Integrating Fine-Tuned Large Language Models with Hybrid Retrieval-Augmented Generation

Pharmacogenomics (PGx) play a key role in personalized medicine by guiding drug and dosage selection based on genetics. However, the volume and complexity of PGx data hinder clinical decision-making. This research proposes a data-driven clinical decision support framework that combines large language models (LLMs) with hybrid retrieval-augmented generation (RAG) to improve answers to PGx queries. The proposed framework evaluates Meta-LLaMA-3.1-8B-Instruct and Qwen3-8B across various configurations, including base models, Low-Rank Adaptation (LoRA) fine-tuning, and hybrid RAG methods. To build a robust dataset, structured data from the Clinical Pharmacogenetics Implementation Consortium (CPIC) and clinical guideline content from ClinPGx are prepared as JSON Lines (JSONL) resources, with CPIC-derived records used for instruction tuning and structured retrieval and ClinPGx guideline text used as a separate retrieval resource. The hybrid retrieval pipeline pairs lexical filtering with dense semantic similarity via sentence embeddings to maximize factual grounding. Evaluation relies on both automated metrics and human clinical review for correctness, relevance, completeness, and clarity. Results indicate that Meta-LLaMA-3.1-8B-Instruct benefits most consistently from the combined RAG and LoRA configuration, while Qwen3-8B shows more modest action-level classification performance but improved evidence-grounded text-generation quality when retrieval is added. Fine-tuning alone proved insufficient, highlighting the limitations of purely parametric knowledge. This study shows that combining retrieval methods with parameter-efficient fine-tuning enhances LLM reliability in clinical settings. The proposed methodology offers a scalable, trustworthy framework for AI-driven decision support in pharmacogenomics and broader healthcare applications.