GenoBridge: a sample-size-adaptive machine-learning pipeline for genomic prediction and predictability-gated mixed-model GWAS

Genomic prediction requires selection of an appropriate statistical model and hyperparameters, along with the marker encoding that each model expects. These choices are trait dependent because the genetic architecture is generally not known before the analysis. In practice, these decisions are often made on ad hoc basis and assessed using only a single cross-validation split. As a result, the reported accuracies tend to be optimistic and are challenging to compare across different traits, studies, and species. Currently, tools are not available to integrate model selection into a hierarchical resampling scheme to yield per-trait accuracy and explicit uncertainty estimates. GenoBridge, a pipeline that is developed to test multiple models for each trait and reports prediction accuracy with confidence intervals. GenoBridge, a pipeline that assigns analytical parameters based on sample size without manual tuning, screens traits for heritable signals by cross-validated prediction, and tests the traits that pass kinship-based mixed-model association. GenoBridge provides cross-validated R 2 alongside correlation, ensuring that only traits with predictive signal that generalize across folds enter association analysis, protecting against misleading accuracy when correlations are used alone. On five datasets spanning four species and a nearly 80-fold range in sample size ( n = 136 to 10,729), GenoBridge reached prediction accuracy within 0.02 to 0.05 of GBLUP and BRR without any user-specified model, and outperformed LASSO and DeepGS in all traits. The mixed model held the genomic inflation factor between 0.85 and 1.03 across all datasets, indicating well-calibrated test statistics. Predictability-gated association successfully identified previously established loci, including FLC , FT , and DOG1 in Arabidopsis; Waxy and GW5 in rice; and Glu-D1 in wheat, in which two independent dough-strength traits mapped to the same marker. Prediction models decouple accuracy and association signal; traits with similar predictability differed by more than 60 orders of magnitude in association significance and depended on genetic architecture rather than prediction performance. GenoBridge delivers calibrated association testing and gene annotation on the same run. GenoBridge achieves prediction accuracy comparable to established genomic prediction methods without manual model selection, while simultaneously providing calibrated association testing and gene annotation within a unified pipeline. GenoBridge is a highly useful and easy-to-use ML-based tool that eliminates manual optimization. GenoBridge is open source and available at https://github.com/nps-genomics/GenoBridge .