Genomic selection, first proposed in 2001 by Meuwissen, Hayes and Goddard, revolutionized breeding by predicting the genetic value of plants from dense marker data rather than waiting years for field phenotypes. Instead of identifying a handful of major genes, the method captures the collective action of thousands of small-effect variants scattered across the genome. The standard workhorse is GBLUP, a genomic best linear unbiased prediction model that assumes all markers contribute equally through a relationship matrix. But the approach depends entirely on the quality and composition of the marker set fed into it. Most soybean studies rely on single nucleotide polymorphisms, or SNPs, the single-letter changes that dot the genome in their millions. Whether those SNPs should be chosen at random, or prioritized based on statistical association with the trait of interest, has remained an open and practically important question.

The research team, led by Xiaoyue Zhu, Ruixin Zhang, Changhong Guo and Yongjun Shu, attacked the problem using two public soybean datasets, one covering drought tolerance and the other plant height. These traits were deliberately chosen because they represent opposite ends of the genetic architecture spectrum. Drought tolerance is a highly complex, polygenic trait shaped by many genes of modest effect interacting with the environment, while plant height tends to have a somewhat simpler genetic basis. From whole-genome resequencing data, the researchers identified genome-wide markers of three distinct classes: SNPs, simple sequence repeats known as SSRs or microsatellites, and insertion-type structural variants, abbreviated INS-SVs, which are chunks of DNA hundreds of bases long that have been pasted into new genomic locations.

The methodological core of the study lies in its GWAS-assisted marker prioritization pipeline. Rather than treating every marker as equally informative, the team ran genome-wide association analyses using five different statistical methods: the generalized linear model, or GLM; the mixed linear model, or MLM; FarmCPU, a multi-locus random effect model; fastGWA, a resource-efficient mixed model tool; and BOLT-LMM, another powerful mixed model approach originally developed for human genetics. Each method ranks markers by their statistical association with the trait. The researchers then constructed Top-K marker panels, taking the strongest associated markers at different densities, and pushed these panels through twelve different genomic selection models to see how prediction accuracy responded. This combinatorial design, five GWAS methods times multiple panel sizes times twelve prediction models, produced an unusually comprehensive map of what works and what does not.

The headline finding is that GWAS-prioritized marker panels generally outperformed random 5K SNP panels, demonstrating that feature selection informed by association mapping genuinely adds value. Even more striking, the curated panels achieved accuracy comparable to, or better than, the full-marker SNP GBLUP baseline within the internal validation framework. That last point matters enormously for breeding economics. A full-marker analysis requires genotyping every individual at hundreds of thousands of positions, which is costly at scale. If a carefully chosen panel of a few thousand markers can match the predictive power of the complete set, breeding programs can slash genotyping budgets without sacrificing accuracy. The study identified a marker density of 5,000 markers as a practical sweet spot, balancing prediction accuracy against the number of markers that must be assayed.

The structural variant story is more nuanced and arguably more scientifically interesting. Structural variants, especially insertions, have long been suspected of carrying hidden functional information that SNPs miss. Landmark studies in tomato and Arabidopsis have shown that widespread structural variation affects gene expression and crop improvement in ways that SNP arrays cannot capture. The soybean pan-genome work published in Cell in 2020 reinforced this picture, revealing substantial presence-absence variation between wild and cultivated soybeans. The Harbin team therefore hypothesized that adding INS-SV markers to prediction panels might boost accuracy. The results were trait-dependent: for drought tolerance, some combined-marker panels showed small positive gains in prediction accuracy, measured as delta-r values, while for plant height the gains were negligible. In other words, structural variants are not a universal upgrade, but for certain complex traits they may carry information that SNPs alone do not.

Equally instructive was the behavior of the prediction models themselves. Across the experiments, GBLUP and Bayesian regression models, including members of the so-called Bayesian alphabet such as BayesB, showed relatively stable and reliable performance regardless of marker panel composition. This stability is reassuring for breeders, since it suggests the standard modeling toolkit remains robust even when marker inputs change dramatically. The relative performance of combined-marker panels, however, depended on both the trait being predicted and the GWAS method used for prioritization. A panel that excelled under FarmCPU prioritization might underperform when markers were selected by BOLT-LMM, and vice versa. This method-dependence is a caution against one-size-fits-all recommendations and underscores the need for trait-specific and method-specific validation before deploying any curated panel in a real breeding pipeline.

The study also situates itself within a growing literature on incremental feature selection for genomic prediction. Previous work in Norway spruce showed that preselecting QTL markers enhances genomic selection accuracy, and other research has explored stepwise approaches to marker refinement. The soybean results extend this principle to a major legume crop and, crucially, to a multi-marker-type framework that includes structural variants alongside SNPs and SSRs. The authors frame GWAS-assisted prioritization as an efficient feature-selection strategy, one that leverages association signals already computable from existing data rather than requiring new phenotyping or genotyping investments. For breeding programs in developing regions, where genotyping budgets are tight, this could be a genuinely transformative workflow: run a GWAS once, extract the top markers, and use the compact panel for routine genomic prediction of candidate parents and progeny.

There are, of course, limitations that temper the enthusiasm. The evaluation relied on internal validation within the same datasets, which can inflate accuracy estimates relative to independent validation across populations and environments. Genomic prediction accuracy is notoriously sensitive to the composition of the training population, the heritability of the trait, and the relatedness between training and validation individuals. The authors themselves emphasize that multi-type marker panels should be evaluated against a standard SNP reference on a trait- and method-specific basis, a recommendation that effectively functions as a quality-control protocol for any breeding program considering adoption. Drought tolerance, in particular, is strongly environment-dependent, and prediction models trained on one set of environments may falter elsewhere. The small positive delta-r values observed for some drought tolerance panels, while encouraging, are modest rather than dramatic breakthroughs.

Nevertheless, the significance of this work extends beyond soybean. As sequencing costs continue to fall, the bottleneck in crop genomics is shifting from data generation to data curation, deciding which of the millions of available variants actually deserve a place in the prediction model. This study provides one of the clearest empirical answers to date: association-informed selection of markers, at moderate densities around five thousand, offers a practical and efficient route to accurate genomic prediction, while exotic marker types should be added selectively and validated rigorously. For a crop that supplies roughly a quarter of the world’s vegetable oil and protein, even incremental gains in breeding efficiency translate into real agricultural impact. As climate change intensifies drought pressure on soybean-growing regions from the American Midwest to Northeast China, tools that accelerate the development of resilient varieties are not just scientifically elegant, they are urgently necessary. The Harbin team’s marker-panel roadmap offers breeders a tested template for getting there faster.

Subject of Research: Comparative evaluation of GWAS-prioritized SNP, SSR, and insertion-type structural variant marker panels for genomic prediction of drought tolerance and plant height in soybean

Article Title: Comparative evaluation of GWAS-prioritized SNP, SSR, and insertion-type structural variant marker panels for genomic prediction in soybean

Article References: Zhu, X., Zhang, R., Guo, C., & Shu, Y. (2026). Comparative evaluation of GWAS-prioritized SNP, SSR, and insertion-type structural variant marker panels for genomic prediction in soybean. Theoretical and Applied Genetics, 139(10), Article 280. https://doi.org/10.1007/s00122-026-05397-1

Image Credits: AI Generated

DOI: 10.1007/s00122-026-05397-1

Keywords: soybean, genomic selection, genomic prediction, GWAS, SNP markers, SSR markers, structural variants, drought tolerance, plant height, GBLUP, marker-assisted selection, plant breeding