AD-GPT: Large Language Models in Alzheimer's Disease

TL;DR

AD-GPT addresses the challenge of reliable domain-specific information retrieval for Alzheimer's disease by integrating OMIM and GTEx data into four specialized corpora and a stacked Llama3/BERT architecture. The approach yields superior precision and reliability across four AD-related tasks (genetic information retrieval, gene-brain region, gene-AD, brain region-AD), outperforming a range of state-of-the-art LLMs. The system is deployed as a self-contained Docker package with a GUI and uses QLoRA fine-tuning to achieve efficient multi-task inference. The work highlights the importance of domain adaptation, structured data integration, and planned enhancements such as RAG, CoT, MoE, and RL to keep domain knowledge current and reduce hallucinations.

Abstract

Large language models (LLMs) have emerged as powerful tools for medical information retrieval, yet their accuracy and depth remain limited in specialized domains such as Alzheimer's disease (AD), a growing global health challenge. To address this gap, we introduce AD-GPT, a domain-specific generative pre-trained transformer designed to enhance the retrieval and analysis of AD-related genetic and neurobiological information. AD-GPT integrates diverse biomedical data sources, including potential AD-associated genes, molecular genetic information, and key gene variants linked to brain regions. We develop a stacked LLM architecture combining Llama3 and BERT, optimized for four critical tasks in AD research: (1) genetic information retrieval, (2) gene-brain region relationship assessment, (3) gene-AD relationship analysis, and (4) brain region-AD relationship mapping. Comparative evaluations against state-of-the-art LLMs demonstrate AD-GPT's superior precision and reliability across these tasks, underscoring its potential as a robust and specialized AI tool for advancing AD research and biomarker discovery.

Figures (6)

• Figure 1: Overview of the AD-GPT workflow. AD-GPT integrates multimodal data from publicly available databases, including OMIM and the GTEx Project, to systematically extract AD-related information. This information is curated into four specialized textual corpora, each designed to support distinct research tasks. The system builds upon pre-trained Llama and BERT models, further fine-tuned to enhance performance in domain-specific applications. Model efficacy is benchmarked against state-of-the-art language models using diverse evaluation metrics. To facilitate accessibility and usability, the entire framework is encapsulated within a Docker container and equipped with an interactive GUI, enabling seamless deployment for AD research.
• Figure 2: Performance comparison of different language models across four tasks. (a) Accuracy in Task 1 for answering gene-attribute-related questions. (b) Precision, recall, and F1 score in Task 2 for identifying gene–brain region relationships. (c) Average precision and relevance scores (0–5) in Task 3 rated by experts. (d) Average precision and relevance scores (0–5) in Task 4 rated by experts.
• Figure 3: Performance improvements of AD-GPT following QLoRA fine-tuning on Tasks 3 and 4. The QLoRA fine-tuning process substantially shifted the distributions of precision and relevance scores for both tasks. Two-sided paired $t$-tests were conducted to evaluate the statistical significance of these metric improvements. Cohen’s d and 95% confidence intervals are reported to quantify the effect sizes.
• Figure 4: Representative questions randomly selected from Tasks 1-4 and the responses generated by AD-GPT and its counterparts. AD-GPT consistently demonstrates more accurate, specific, and evidence-backed answers across a range of genomic and disease-association queries.
• Figure 5: Integration of genetic data from GTEx and OMIM. GTEx provides gene location and QTL data, while OMIM offers gene descriptions and disease associations for AD research.