The central problem is one of limited land. A traditional randomized complete block design places every genotype in each block, allowing researchers to separate genetic differences from variation caused by soil, moisture, temperature and other environmental factors. But full replication consumes plots rapidly. If a field can accommodate only a fixed number of plants, repeating every entry means that fewer distinct lines can be tested. That reduces selection intensity—the proportion of candidates that can be compared and chosen for future breeding—and may shrink the training population used to build genomic prediction models. Partially replicated, or p-rep, designs take the opposite approach. A subset of entries is repeated across blocks, while the remainder is observed once per location. The repeated entries provide information about field variation, and genomic relationships allow information from those entries to be transferred statistically to their less-replicated relatives.

That transfer is possible because genomic selection does not treat each plant as an isolated data point. DNA markers, commonly single-nucleotide polymorphisms, are used to construct a genomic relationship matrix describing how genetically similar individuals are. In a mixed model, the matrix acts as a map of expected covariance: if an unreplicated line is closely related to several well-measured lines, the model can use their phenotypes to improve its estimate of the untested line’s breeding value. Breeding value refers primarily to the genetic component that can be passed to offspring, while genotypic value may also include non-additive effects such as dominance or epistasis. The strategy is especially attractive for traits such as yield, which are often highly influenced by the environment and therefore difficult to measure precisely from a small number of plots.

The investigators used empirical marker data from three contrasting crop populations. Their maize simulations represented 780 hybrids generated from a full diallel involving 40 selected parental lines, with the parental material originally drawn from 391 inbred lines genotyped at 244,781 markers. The sunflower data included 224 testcross hybrids characterized by 14,755 markers, while the oat population consisted of 699 inbred lines and 17,288 markers. The researchers divided each population into a phenotyped training set and an unphenotyped test set, then simulated yield-related observations for the training entries. The scenarios covered fields containing two or three blocks and between 50 and 600 plots. Each combination of crop, field size and design was evaluated repeatedly: 30 independent constructions of the field layout were paired with 10 phenotype simulations, producing 300 repetitions for each comparison.

Four layouts were tested. Conventional complete block designs were either randomized or optimized, while p-rep designs were likewise randomized or optimized. In the p-rep treatments, replicated entries represented 20 percent of the block size, a level selected after preliminary testing as the minimum that still supported consistent spatial analysis. Optimization was performed in two stages. First, the researchers selected replicated genotypes with the lowest average genomic relatedness, using a criterion called Avg_GRM_self. This was intended to create a genetically diverse and representative subset rather than repeatedly measuring a cluster of similar lines. Second, they optimized the physical positions of genotypes in the field with CDmean, a criterion based on the coefficient of determination. CDmean uses the predicted error variance of genetic estimates to seek arrangements that maximize expected information, while also accounting for assumed spatial correlation between neighboring plots.

Spatial correlation is a major source of hidden error in field experiments. Plants next to one another may share soil conditions, irrigation patterns or pest exposure, causing their measurements to resemble each other even when their genomes differ. If this pattern is ignored, a genotype placed in a particularly favorable or unfavorable part of the field can appear genetically better or worse than it really is. The study compared simple blocking structures with two-dimensional penalized splines, or 2D P-splines. These splines model gradual or irregular trends across rows and columns at fine spatial resolution. A penalty prevents the fitted surface from chasing random noise, allowing the model to use many potential spatial coefficients without producing an unstable estimate. In effect, the method reconstructs a smooth map of field conditions and removes that environmental pattern from the genetic signal.

The researchers also compared three ways of analyzing the resulting data. A single-stage model estimated environmental effects, spatial effects and genomic relationships simultaneously, making it statistically comprehensive but computationally demanding. An unweighted two-stage model first analyzed field data and then passed estimated genotype values to a genomic prediction step. That approach is faster, but it can lose information because the first-stage estimates are treated too simply. A fully efficient two-stage model attempted to correct this weakness by carrying estimation error variances from the first stage into the second. This allows the genomic analysis to distinguish precise genotype estimates from uncertain ones and can approach the performance of a single-stage analysis while retaining the practical advantages of separating the calculations.

Across the simulations, p-rep designs consistently produced more accurate predictions for genotypes that had not been observed in the field. Their advantage was not unlimited: within the phenotyped training set itself, fully replicated designs generally achieved slightly higher accuracy because every genotype had more direct measurements. But that local advantage was offset by the greater selection intensity enabled by p-rep layouts. In a fixed field, breeders could evaluate more unique lines, increasing the chance of identifying rare individuals with valuable combinations of alleles. Optimizing the p-rep design added about 1 percent accuracy compared with randomized p-rep arrangements. The gain was modest but potentially meaningful when accumulated across thousands of candidates and repeated breeding cycles.

The statistical model mattered almost as much as the field layout. The use of 2D P-splines improved prediction accuracy by approximately 1 to 2 percent when combined with single-stage or fully efficient two-stage models. The results therefore point to a positive interaction rather than a single winning ingredient: strategic replication, optimized placement and an analysis capable of resolving fine-scale spatial variation worked best together. The simulations deliberately included additive effects, non-additive effects, environmental interactions and plot-level spatial interactions, while the fitted models were simpler, reflecting the compromises common in real breeding programs. The authors also note that the work is a simulation study, not a direct field demonstration, and that its predictions depend on assumptions about genetic architecture, heritability and spatial covariance. Even so, the findings offer a compelling blueprint for modern crop improvement. As genomic data make it possible to borrow information across related plants, the future of field testing may depend less on repeating every genotype and more on deciding which plants to repeat, where to place them and how intelligently to interpret the resulting landscape of data.