Phenotyping part of the target family improves within-family genomic prediction in rice nested association mapping families

Abstract Genomic prediction uses DNA markers to help rice breeders choose promising lines before costly yield trials. A new breeding family usually lacks its own yield records, so its lines are predicted from other families, often less accurately. With limited field plots, breeders can measure yield in some lines of the family being selected, called the target family, or adopt a more flexible model, which needs no extra plots. It is unclear which helps more, whether they combine, and whether better prediction selects higher-yielding lines. Here, two-site data for 15 nested association mapping families sharing one parent were reanalyzed with 2000 training records per target family, so gains reflect which lines were measured, not how many. In a four-strategy comparison added after the original analysis, replacing 100 other-family records with records from 5–15% of the target family raised average within-family predictive ability, the correlation between predicted and recorded yield of held-out lines, from 0.16–0.20 to 0.19–0.26 for both linear ridge regression and a nonlinear Gaussian-kernel model, with gains in 11–14 of 15 families. Switching to the Gaussian model alone, tuned for selected-line yield, changed it by only 0.005–0.011, although the two gains did not differ clearly after multiple-comparison correction; combining both gave the highest predictive ability. At 10% selection, lines chosen with target-family records yielded more on average, with 95% intervals above zero only at Shanghai; the originally planned comparison of records versus model was positive but not conclusive. In a multiparent population predicted from these families, its own records lowered selected-line yield at one site, with mixed effects at the other. In an elite International Rice Research Institute cohort, adding 25 current-season records raised average predictive ability of both models in every season, while selected-line yield varied. In these related families, phenotyping part of a new family before ranking the rest improved within-family prediction for both models. The results support using ridge regression trained with target-family records as the benchmark for flexible models, with strategies compared by both predictive ability and selected-line yield; predictions and code are publicly released.