GPBench: A benchmark framework for genomic prediction

Genomic prediction (GP) has emerged as a powerful data mining strategy for disease risk assessment and molecular breeding. GP leverages genome-wide genetic variations and phenotypic information through a computational model to estimate genetic potential or disease risk. However, the field of GP still faces several major challenges: (i) dataset bias and insufficient representativeness; (ii) the lack of standardized evaluation protocols and a systematic benchmarking framework; (iii) heavy reliance on empirical expertise for experimental result interpretation, which hinders large-scale and systematic analysis. These issues make it difficult to conduct systematic cross-species evaluation and comparison of GP methods, and their authentic applications. To address these challenges, we established a standardized and extensible GP benchmark framework-GPBench. We first collected and harmonized public datasets from 16 species, covering animals, plants, and microorganisms, comprising 57 traits, 40,821 samples, and 493,821 single-nucleotide polymorphisms. Then, using the standardized datasets, the framework evaluates 21 GP models and provides a unified assessment across four dimensions: effectiveness, sensitivity, robustness, and scalability. Furthermore, we developed a large language model-based intelligent analysis system to facilitate the interpretation of GP experiments and model recommendations. GPBench provides the community with a fair, unified, and scalable platform (https://www.sdu-idea.cn/GPBench/) and an open-source library for comparing and optimizing GP methods, ultimately accelerating the application of GP in life sciences and biomedical research.