Answers are grounded in this paper and cite their sources.
Genomic selection (GS) offers a practical route for identifying superior hybrid combinations before extensive field testing, but predictability (PA) may depend jointly on the genomic representation and the prediction algorithm. We used 18 trait–population combinations across five crop hybrid public datasets including maize, wheat, rice, strawberry, and pumpkin. Six predictor sets were compared: all genome-wide markers, 2000 Spearman-selected markers, and eigenvectors of additive (G), dominance (D), and additive-by-additive epistatic (G#G) relationship matrices. Seven conventional, machine-learning (ML) and deep-learning (DL) models were evaluated. PA was assessed by random cross-validation and three parent-aware schemes (T2, T1, T0), in which both, one, or neither parent of a hybrid was represented in training. Across datasets, PA ranged from 0.39 to 0.93. Adding dominance or epistatic information to additive predictors gave little consistent improvement and occasionally reduced PA. Genomic best linear unbiased prediction (GBLUP) showed the most stable performance, whereas ML and DL models varied more; support vector regression and multilayer perceptron excelled for specific traits, and Anchored Genomic NETwork (AGNET) was competitive but did not consistently outperform GBLUP. Under the parent-aware schemes, PA declined from T2 to T0, with tree-based models most affected. Overall, tailoring prediction strategies to specific traits and populations remains essential.
No takes yet. Share an insight, caveat, or question.
Vitale et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: