Genomic selection has swept through plant breeding over the past two decades because it estimates the genetic value of lines from markers alone, cutting costs and outpacing phenotypic selection for quantitative traits like yield. But the standard implementation — truncation selection, in which breeders simply pick the top-ranked lines by their genomic estimate of the genotypic value and mate them at random — carries a hidden cost. Compared with phenotypic selection, it strips out genetic diversity faster, and that erosion hinders long-term success. Breeders who optimize only for the next generation may be quietly mortgaging the ones after that.

Researchers have proposed a bewildering zoo of alternatives. Weighted genomic selection places extra emphasis on rare alleles with positive effects. The optimal haploid value and the expected maximum haploid breeding value both shift attention from the individual to the gametes it can produce, asking what theoretical best line could emerge from a candidate’s chromosomes. Cross selection indices such as the usefulness criterion, which combines the expected progeny mean with the expected segregation variance, and the best progeny value, which considers the expected gap between a cross’s mean and its theoretically best offspring, let breeders rank crosses rather than parents. Optimal contribution selection and optimal cross selection go further, searching for entire sets of crosses that maximize gain under constraints on inbreeding. Software packages like SimpleMating, Genomic Mating, and AlphaMate now automate the optimization. The problem is that nobody knew which of these tools to use, in which generation, for a breeding program that only runs a few cycles.

That question matters because most of these methods were designed with long-term programs in mind. They typically deliver their advantages after five or more generations, while producing lower gain than truncation selection in the first generations — exactly the generations a breeding program needs to refinance itself. A team led by Emanuel Grosch and Carola Zenke-Philippi focused specifically on a short-term recurrent genomic selection program spanning five generations, from the selection of parental inbred lines through three crossing cycles to the production of final inbred lines. Their goal was to find out whether the specific strengths of different measures could be exploited by using them at particular stages of the scheme rather than committing to one method throughout.

The study drew on two real datasets: 224 oat inbred lines derived from factorial crosses among 15 founder genotypes, evaluated across multiple German locations and years, and 293 wheat doubled-haploid lines from crosses between elite varieties and resistance donor lines carrying exotic alleles. Marker effects for yield were estimated with ridge regression best linear unbiased prediction. From these starting points, the team simulated 19 different breeding strategies over 300 runs each, evaluating every strategy by two criteria: the mean genomic estimate of the genotypic value of the final 216 inbred lines, and their mean genetic diversity, measured by expected heterozygosity and the number of fixed alleles. The reference strategy was truncation selection for the genomic estimate of the genotypic value with random assignment of mating pairs — the simple, widely used approach in both plant and animal breeding.

The toolbox the researchers deployed was broad. For selecting individuals, they used the genomic estimate of the genotypic value itself and the expected maximum haploid genotypic value, obtained by simulating doubled-haploid progeny of each candidate. For ranking crosses, they tested the best progeny value, the maximum genomic estimate of the genotypic value — the value of the best doubled-haploid line a cross could theoretically produce if it inherited the best allele at every locus — the usefulness criterion, the expected genotypic value of grand offspring, and genetic Rogers’ distance. Cross-planning methods ranged from random mating and mating best with best, through the Cockerham-Burrows scheme that crosses the superior half of the selected fraction with the inferior half, to greedy top-cross selection, heuristic set optimization inspired by optimal cross selection, and the SimpleMating package with a relatedness threshold.

The first result was sobering: no strategy relying on a single selection criterion or cross selection index throughout managed to beat the reference on both gain and diversity simultaneously. Mating best with best produced the highest selection gain of all 19 strategies — and the lowest diversity in both datasets. The usefulness-based strategies delivered higher gain but lower diversity than the reference, an effect stronger for the SimpleMating implementation than for set optimization. The expected maximum haploid value and the grand-offspring metric performed comparably to the reference, consistent with earlier findings that these measures only pay off after five or more generations. Meanwhile, the maximum genomic estimate used across all steps retained the highest diversity of any single-index strategy, at the cost of gain that was slightly higher in oat and slightly lower in wheat.

The turning point came when the researchers planned crosses with the maximum genomic estimate early in the scheme. Because this index considers only the best allele at each locus, crossing partners can carry one favorable but up to three unfavorable alleles at a locus, which inflates the diversity of the progeny. Strategies that used this index for cross planning in the inbred-line generation, or additionally in the first generation, produced higher gain than the reference while maintaining equal or higher diversity in both datasets. By contrast, two classic diversity-oriented tactics failed. The Cockerham-Burrows mating scheme lowered gain without raising diversity, and planning crosses by maximum genetic distance between partners could not be recommended at all — in wheat it raised gain but lowered diversity, the opposite of its intent.

The combination strategies, derived heuristically from these observations, delivered the study’s headline finding. Every combination of cross selection indices and cross-planning methods produced higher genetic gain than the reference. The standout was a strategy that planned crosses with the maximum genomic estimate in the first two generations and in the inbred-line generation, then switched to the usefulness criterion with heuristic set optimization: it achieved both higher gain and higher diversity than the reference in both datasets. Strategies combining the maximum genomic estimate with the best progeny value, either consecutively or alternating between generations, retained the most diversity of all combination strategies while still outperforming the reference on gain. The logic is elegant: the maximum genomic estimate preserves heterozygosity early because it ignores allele frequencies, and that banked diversity is then exploited by the usefulness criterion or the best progeny value, which reward crosses whose favorable alleles are actually likely to appear in offspring. The best progeny value balances a high mid-parent value with heterozygosity, peaking at an allele frequency of 0.75 — favoring high frequency but not fixation of a positive allele — whereas the maximum genomic estimate values an allele identically at every frequency except zero.

The differences between the oat and wheat datasets added a practical caveat. In oat, where founder parents were pre-selected for desirable traits, the theoretical indices translated into real gain more reliably. In wheat, where exotic resistance donors introduced alleles not pre-selected for yield, the risk grew that a cross favored by these hypothetical values would fail to deliver, because Mendelian sampling and limited recombination can strip out the favorable alleles the index assumed would combine. The authors also flagged limitations: the simulations assumed a purely additive genetic architecture, family sizes were fixed at ten progeny per cross, and the segregation variance estimates required for the usefulness criterion remain challenging to predict accurately in practice, which may compress the real-world differences between strategies. Still, the core conclusion stands and carries real weight for breeders. Genetic diversity in a short-term program is not just a reservoir for the future — it is variability in the final doubled-haploid or inbred lines, and lines that reach comparable yield through different alleles will most likely differ in other agronomic and resistance traits. That means combination strategies could benefit multi-trait selection even more than single-trait selection. Because the advantages arise from properties inherent to the measures themselves, the researchers expect them to hold across datasets, even if magnitudes shift. And since breeders can compare a handful of tailored strategies on a standard desktop computer, the message is strikingly accessible: the next leap in breeding progress may not require new genotyping platforms or bigger trials, just smarter sequencing of the selection tools already on the shelf.

Subject of Research: Combining selection criteria and cross-planning methods into breeding strategies for short-term recurrent genomic selection in oat and wheat

Article Title: Combining different selection criteria into breeding strategies for short-term recurrent genomic selection

Article References: Grosch, E., Herrmann, M., Schurack, S., Beuch, S., Kurra, V., Mascher, M., Avni, R., Oldach, K., Röhrs, I., Bekele, W., Hanemann, A., Ram Mehta, R., Reinbrecht, C., Serfling, A., Stahl, A., Stucke, M., & Zenke-Philippi, C. (2026). Combining different selection criteria into breeding strategies for short-term recurrent genomic selection. Theoretical and Applied Genetics, 139(11), Article 316. https://doi.org/10.1007/s00122-026-05410-7

Image Credits: AI Generated

DOI: 10.1007/s00122-026-05410-7

Keywords: genomic selection, plant breeding, genetic gain, genetic diversity, oat, wheat, cross selection indices, truncation selection, usefulness criterion, simulation study, recurrent selection, doubled-haploid lines