Slash pine breeding moves past phenotype alone: training-fold marker preselection lifts growth-trait prediction by 11.9%

Genomic prediction in slash pine now picks markers and models trait by trait instead of applying one configuration to the whole tree. Such breeding had relied mainly on field phenotypes and BLUP breeding values, while marker-assisted selection delivered modest gains for polygenic traits. Across 240 trees, 53,229 transcriptome SNPs and 100 held-out folds from 10 repeats of 10-fold cross-validation, no single model led on all twelve traits: DBH reached the highest mean prediction ability of 0.493 with XGBoost at 300 SNPs, while MOE reached 0.157 with GBLUP using all markers.

Keeping association analysis inside each training fold lets marker ranking use only training-tree information. Ranking markers on all individuals before cross-validation would let validation individuals enter marker selection and inflate apparent accuracy. Each outer fold held 216 training and 24 held-out trees, with GWAS run independently per training set and fold-specific principal components and a kinship matrix as covariates; after p-value preselection, mean prediction ability across model-density combinations rose from 0.242 to 0.271 for growth traits and from 0.214 to 0.219 for wood quality, while oleoresin traits stayed at 0.101 under both schemes.

QTN covariates built from PVE thresholds or cross-fold recurrence frequency added measurable value for only a few traits. A no-QTN baseline served as the control under the same trait, model, marker density and fold structure. Nine of twelve traits retained the no-QTN baseline; the frequency strategy exceeded it only for Crown (ΔPA = 0.0123), beta-pinene (ΔPA = 0.0015) and levopimaric acid (ΔPA = 0.0072), and PVE thresholds often admitted thousands of covariates against 216 training trees per fold.

On a common phenotype-BLUP benchmark, GS-selected trees retained most same-cycle index gain and showed an annualized advantage under assumed shorter cycles. Both selection routes were scored on the same phenotype-BLUP index, avoiding direct comparison across different breeding-value shrinkage scales. From Top 1 to Top 30, GS-selected sets retained 80.3%–95.1% of same-cycle benchmark gain; under assumed 15-year conventional and 7-year GS cycles, the annualized GS-to-phenotype ratio was 2.00 at Top 1, 1.74 at Top 10 and 1.96 at Top 30, spanning 1.72–2.04 across Top 1–30.

The conclusions rest on one trial site near Hangzhou, 240 slash pine trees from 33 open-pollinated families and a single 2019 phenotyping occasion, with 53,229 transcriptome-derived SNPs not anchored to chromosomes. The intended readers are researchers and engineers working on forest-tree genomic selection and breeding-scheme design; the source states no next experimental step.

The authors note that configuration selection and performance comparison used the same held-out folds, so selected configurations may be optimistic and are not an unbiased estimate from an additional independent outer validation; random individual-level folds can place relatives in both training and validation, so prediction ability partly reflects within-population relatedness. They also state that the annualized advantage depends on the assumed 8-year reduction in breeding-cycle length, and that the study did not measure realized progeny response or verify a 7-year operational cycle.