Rice breeders can stop picking the closest relatives: 25 designed lines match about 100 random ones for grain yield, each design under a second
Synopsis
Which lines to phenotype has largely been decided by habit or by picking those most related to the targets. TSOpt instead designs training sets by exact rank-one updates of the prediction error variance; on 328 IRRI elite rice lines it raised predictive ability over random sampling by 0.097, with a design taking about 0.1 s and running roughly 900 times faster than STPGA. Twenty-five designed lines predicted grain yield as well as about 103 random lines, days to flowering as 76 and plant height as 167; its sparse-testing allocations also beat balanced random allocation when a quarter of line-by-season cells were tested.
Interpretation
Training-set design becomes an exact computation under a second, so a breeder can redesign for each trait, target set and budget before planting. STPGA and TSDFGS previously approximated CDmean on a few principal components of the relationship matrix, and TrainSel evaluated the full criterion repeatedly inside a stochastic search, taking seconds to minutes per design. On the 328 IRRI elite lines the median design time was 0.11 s for TSOpt and 101 s for STPGA, about 884 times longer; on six public panels TSOpt took 0.03 s against 40 to 120 s for the other two tools.
A designed training set reaches the same predictive ability with far fewer field plots, and the saving is largest for small training sets. Random sampling is the usual practice, and picking the lines with the highest mean relationship to the targets is a common rule of thumb that matched random sampling within seasons. Over 20 random splits, three traits and four training-set sizes, TSOpt raised predictive ability over random sampling by 0.097 (95% confidence interval 0.045 to 0.150) and led in 19 of 20 splits; 25 TSOpt lines predicted grain yield as well as about 103 random lines (bootstrap interval 73 to 127), days to flowering as 76 (64 to 99) and plant height as 167 (115 to 220).
The same engine extends design from lines to line-by-season cells, deciding where plots go when they are scarce. Sparse testing previously relied on random allocation balanced per line, or on overlap designs that test some lines in every season. With a quarter of line-by-season cells tested over four seasons in 2011 to 2012, TSOpt's allocations predicted untested cells 0.024 better than balanced random allocation and led in all 20 replicates; scored with eight-season correlations the design never saw, the gain shrank by about a fifth but remained 0.019.
Perspective
The results come from late-stage replicated yield trials of IRRI's irrigated rice breeding programme at one site over four years, covering 328 elite lines with both genotypes and phenotypes and the traits grain yield, days to flowering and plant height. The sparse-testing analysis uses the 321 lines tested in all four seasons of 2011 to 2012. Hybrid, testcross, multi-trait and major-gene designs are exact under their models but were not evaluated on real data. The stated next steps are forward validation across several programmes, crops and cohorts, larger multi-location networks, and criteria aimed at the covariance between testing and future growing seasons.
The forward and sparse-testing results rest on a single new cohort and a single set of seasons, so replicate spread describes method stability rather than variation between cohorts or programmes. Scored with eight-season correlations the design never saw, the advantage at a quarter of the cells shrank by about a fifth and disappeared when half the cells were tested. On public panels the robust criterion was ahead of a fixed heritability of 0.5 by 0.009, not significant, and its value showed when the assumed heritability was too low. The TrainSel comparison uses the sets its authors published rather than running the software. Phenotypes were analysed in two stages through season BLUEs, and a one-stage analysis might change absolute accuracies.