Abstract
"This dataset provides a curated benchmark of 50 research queries focused on gene-disease associations, designed to evaluate the capabilities of automated scientific knowledge discovery systems, particularly those using Large Language Models (LLMs). Each query targets a distinct human gene selected from the HUGO Gene Nomenclature Committee (HGNC) database and is framed as a comprehensive research question requiring deep synthesis of information (e.g., \"Provide a comprehensive overview of the diseases and phenotypes associated with the HBB gene...\"). The genes span four categories representing varying levels of information availability and complexity: Common\/Well-Studied, Cancer-Related, Male Fertility-Related, and Rare Diseases. This benchmark enables standardized evaluation of system performance across dimensions like knowledge recall, information depth, and the ability to uncover non-obvious or rare associations, facilitating the development and comparison of advanced scientific search and synthesis tools."