AD-GPT: large language models in Alzheimer's disease.

BACKGROUND

Alzheimer's disease (AD) research produces extensive genomic and clinical data, yet general large language models (LLMs) often generate inaccurate or superficial outputs. We introduce AD-GPT, a domain-specific framework for reliable information retrieval and synthesis of AD-related knowledge.

METHODS

We integrated curated genomic resources, including cis-eQTL and sQTL data across 13 brain regions from GTEx, genomic location information from NCBI, and gene function annotations from OMIM, together with approximately 150,000 AD-related publications from NCBI's PubMed. AD-GPT adopts a retrieval-augmented generation (RAG) workflow with task-specific database partitioning, a BERT-based query router, and fine-tuned Llama models, augmented with router and context verifiers to validate task assignment and evidence relevance, supporting three tasks: genetic information retrieval, association study reasoning, and general AD-related knowledge synthesis.

RESULTS

AD-GPT consistently outperformed strong baseline LLMs in evidence-grounded evaluation metrics across all tasks, including factual consistency, citation validity, and instruction-level faithfulness. Task-specific retrieval and stacked routing improved evidence grounding and substantially reduced hallucination in complex AD-related queries.

CONCLUSION

AD-GPT harmonizes curated genomic databases with biomedical literature, offering a scalable and accurate informatics tool to advance AD research.