Developing an Integrated Rare Disease Bioinformatics Resource to Determine Phenotype to Genotype Correlations

National Center For Advancing Translational Sciences

Investigators

Abstract

The establishment of the web application was realized through a partnership with the Advanced Biomedical Computational Science (ABCS) group at the Frederick National Laboratory for Cancer Research, National Cancer Institute (NCI). Through continuous development and refinement, we have integrated browsing for rare diseases with known associations to genes. The disease and gene catalogs include names, symbols, synonyms, and links to rare disease data sources. Variant level details are presented in tabular formats, as well as 2D and 3D visualizations on protein structure. The resource also incorporated a "Literature AI" feature, with results from mining rare disease and associated gene mentions in MEDLINE using transformer models. This year, we focused on extracting valuable genotype-phenotype details embedded in literature using Large Language Model (LLM) systems. Comparative performance evaluations were conducted between various LLM architectures to assess their effectiveness in variant and phenotype extraction from scientific literature. The code architecture was transitioned from prototypes to production-ready modular systems integrated with our current database. Parallel efforts in the manual curation of Farber Disease are being used as a reference dataset to validate LLM-derived results. Additionally, website improvements have been implemented to address feedback from users and incorporate newly generated data and functionalities, including a fully automated pipeline for our "Literature AI" module to accurately retrieve the most up-to-date publications.