Standardizing Benchmarks for Plant Genomic Prediction

Genomic prediction (GP) is now widely used in crop improvement, but the rapid expansion of GP methods has outpaced the development of standardized evaluation practices. Reported gains from deep learning and other complex models are difficult to interpret when studies use different datasets, baselines, validation schemes, metrics, tuning budgets, and reporting practices. In this Perspective, we argue that plant GP requires a shift from model-centric performance claims toward standardized, scenario-aware, and application-oriented benchmarking. We highlight four sources of complexity that shape model performance: species and genetic-background diversity, trait architecture, population structure, and genotype-by-environment interaction. We then review current plant GP resources and draw lessons from benchmarking efforts in gene regulatory network inference, single-cell model assessment, and protein structure prediction. On this basis, we propose a nine-step workflow covering curated datasets, harmonized preprocessing and metadata, marker-density scenarios, required baselines, controlled hyperparameter tuning, fixed validation splits, multi-dimensional evaluation metrics, reproducibility reporting, and breeder-facing recommendations. Such benchmarks would make reported gains easier to verify, reduce selective reporting, support cost-aware deployment, and transform plant GP from a collection of fragmented performance claims into a reproducible, comparable, and practically deployable framework for modern breeding.