Liu-Hy/GenoTEX

The benchmark comprises 1,384 gene-trait association problems with 41.5 GB of input data (911 datasets) and 237,907 lines of expert-annotated analysis code covering dataset selection, preprocessing, and statistical inference. Problems span both unconditional analysis and conditional scenarios (accounting for factors like age or co-existing traits), with inputs sourced from GEO and TCGA public databases. The framework enables evaluation of LLM agents on standardized computational genomics workflows, with reference implementations and ground-truth results provided for validation.