An AI-Powered Trisomy 21 Research Assistant

Area of Science:

• Genetics and Bioinformatics
• Artificial Intelligence in Medicine
• Down Syndrome Research

Background:

• Down syndrome (trisomy 21) is linked to numerous health issues, with a rapidly growing body of over 34,000 publications.
• General AI models struggle with the specificity required for scientific literature retrieval.
• Retrieval-augmented generation (RAG) enhances AI reliability by linking outputs to source texts, but standard methods don't prioritize experimental data.

Purpose of the Study:

• To develop a section-aware RAG system, the T21 Research Assistant, to focus on primary experimental evidence in Down syndrome research.
• To improve the accuracy and reliability of information retrieval from the extensive Down syndrome literature.
• To provide researchers with timely and evidence-based answers grounded in experimental results.

Main Methods:

• Developed a section-aware RAG system prioritizing "Results" sections of manuscripts.
• Utilized a curated dataset of 1,789 open-access Down syndrome publications from PubMed Central.
• Implemented a multistage pipeline: query validation, retrieval, reranking, synthesis, and citation verification using NVIDIA Nemotron models.

Main Results:

• The T21 Research Assistant demonstrated strong performance in expert-curated question evaluations.
• Achieved a BERTScore F1 of 0.712 and recall of 0.758.
• Outperformed or matched leading proprietary and open-source models in accuracy and reliability.

Conclusions:

• The T21 Research Assistant effectively grounds AI responses in primary experimental evidence from Down syndrome research.
• This section-aware RAG system offers a significant advancement in navigating complex scientific literature.
• The tool is accessible for researchers at https://bioinformatics.cuanschutz.edu/t21-res-assi/.