Integrating multi-model GWAS prior information enhances genomic prediction of cold tolerance traits in Populus simonii

Abstract

Genomic selection (GS) represents a transformative strategy for accelerating the breeding of cold tolerance in forest trees and other perennial species with long generation intervals. Although integrating genetic loci by genome-wide association studies (GWAS) can enhance prediction accuracy, this potential is frequently constrained by the inconsistency of loci detected across statistical models. Here, we developed a robust GS optimization strategy based on a multi-model GWAS framework using 849 accessions from a half-sib population of Populus simonii. We identified a total of 93 significant loci, among which 29 were co-detected by at least two models, including 8 high-confidence loci consistently detected across all three models. Incorporating these loci as fixed effects in GBLUP improved prediction accuracies ranging from 0.25 to 0.60. Notably, this strategy improved prediction accuracy by up to 60% for complex traits such as superoxide dismutase (SOD) activity, thereby alleviating a key limitation of standard GBLUP in capturing major-effect QTLs. Furthermore, we identified and preliminarily characterized the pleiotropic candidate gene PsiNDHM. Overexpression of PsiNDHM mitigated oxidative damage in poplar under cold stress. Together, these results indicate that leveraging consensus loci from multi-model GWAS offers an effective approach for optimizing GS, providing a methodological framework for precision molecular breeding in species with complex genetic architectures, particularly forest trees.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12870-026-09446-1.

Introduction

Global climate change has intensified the frequency of extreme low-temperature events, posing a severe threat to the stability and productivity of forest ecosystems. Poplar is one of the most widely distributed fast-growing tree species in the Northern hemisphere and holds immense economic and ecological value [1], necessitating the improvement of cold tolerance as a priority for climatic adaptation. However, conventional breeding based on phenotypic selection is impeded by low efficiency, primarily due to long generation intervals, high genomic heterozygosity, and the complex genetic architecture underlying cold tolerance [2]. To meet the urgent demand for stress resistant varieties, accelerating the breeding process of poplar has become increasingly critical.

Genome-wide association studies (GWAS) are powerful tools for mapping genomic variants via historical recombination events [3]. GWAS has emerged leading strategy for dissecting the genetic architecture of complex traits in forest trees and has been extensively applied to elucidate the genetic basis of traits in diverse species, including Eucalyptus [4], Picea abies [5], Ginkgo biloba [6], and Populus [7]. However, the efficacy of GWAS in dissecting cold tolerance, a typical quantitative trait, remains severely constrained by the limitations of individual statistical models [8]. Traditional single-locus models struggle with inherent statistical trade-offs. The General Linear Model (GLM) is prone to inflating false positives by neglecting population structure [9], whereas the Mixed Linear Model (MLM), despite incorporating a kinship matrix to control for confounding factors [10], frequently suffers from over-correction. This issue is particularly pronounced when the target trait is highly correlated with population structure, causing true biological signals to be masked as background noise and leading to an excess of false negatives [11, 12]. Although multi-locus models such as FarmCPU and BLINK have been developed to balance false positives and negatives via iterative algorithms [13, 14], the distinct statistical assumptions underlying these methods often yield discrepant results for the same trait. Such algorithm-specific biases hinder the robust detection of intermediate-effect loci, exacerbating the “missing heritability” phenomenon in forest genetics [15] and impeding a comprehensive understanding of the complex regulatory networks governing cold tolerance.

Genomic selection integrates genome-wide marker effects to predict individual breeding values [16] and holds immense potential for shortening breeding cycles and enhancing selection efficiency in forest trees [17]. The widely used GBLUP model operates on the “infinitesimal model” assumption, positing that all marker effects follow a distribution with equal and small variances [18]. However, this assumption diverges from the actual genetic architecture of complex traits like cold tolerance, which are typically governed by a mixed inheritance pattern comprising few major-effect genes and numerous minor-effect genes [19]. Under such conditions, the uniform treatment of all markers by the GBLUP model inevitably dilutes the contribution of major genes [20], thereby imposing a ceiling on predictive accuracy. Previous studies have attempted to improve prediction accuracy by integrating significant loci identified via GWAS as fixed effects into prediction models [7, 21]. Nevertheless, the application of this strategy in forest trees faces substantial challenges, as its effectiveness is strictly contingent upon the reliability of the prior information. Most existing studies rely blindly on a single GWAS model to screen for fixed-effect loci [22, 23]. As noted above, single models can introduce noise (false positives) or miss true signals (false negatives) due to statistical biases. Forcing erroneous prior information as fixed effects not only fails to improve accuracy but can also introduce severe model bias, leading to a deterioration in predictive performance [24]. Consequently, developing a robust feature selection strategy to precisely extract biologically meaningful loci from a noisy background is a remains critical bottleneck in optimizing GS models for forest trees.

To mitigate the uncertainty inherent in single model feature selection, we propose a novel genomic prediction strategy anchored in a multi-model GWAS. Distinct GWAS models (e.g., MLM, FarmCPU, and BLINK) offer complementary statistical advantages. Multi-locus models enhance detection power, whereas single-locus models provide robust control over the genetic background. By integrating significant loci identified via cross-validation across these models as fixed effects, and utilizing the remaining markers to construct a G-matrix for the polygenic background, we aim to maximize the capture of major QTLs while maintaining a precise fit for minor-effect genes [25]. Furthermore, leveraging high-density variation maps from large populations in conjunction with transcriptome data will enhance the precision of signal identification, thereby providing reliable prior information for the GS model [26–28].

In this study, we conducted a comprehensive genetic dissection and genomic prediction of six key cold tolerance-related physiological traits using 849 P. simonii genotypes from a half-sib population. The specific objectives of our research were threefold. (1) To elucidate the genetic regulatory basis of cold tolerance traits using multiple GWAS models, which led to the identification of a key pleiotropic gene, PsiNDHM. (2) To systematically evaluate the impact of training population size and marker density on genomic prediction. (3) To propose a novel fixed-effect GS strategy based on multi-model framework. This strategy significantly enhanced prediction accuracy for complex traits, particularly antioxidant enzyme activities, thus providing a methodological advance for molecular breeding in forest trees and other non-model organisms.

Materials and methods

Phenotypic data collection

This study evaluated 849 genotypes representing 30 half-sib families of P. simonii. Seeds from diverse natural populations were initially collected in 2022 across northern China, spanning the Inner Mongolia Autonomous Region (Kulun, Naiman, and Zhalute Banners, and Keerqin District in Tongliao), as well as Liaoning (Zhangwu, Fuxin), Jilin (Tongyu, Baicheng), and Shanxi (Datong) provinces. Comprehensive geographical coordinates for all 30 families are provided in Table S1. Following seed collection, seedlings were germinated and cultivated throughout 2023.

Formal taxonomic identification of the plant materials was performed. Sample collection was explicitly authorized by the respective local forestry bureaus. All field studies and experimental procedures strictly adhered to relevant institutional, national, and international guidelines and legislation. No voucher specimens were deposited in a publicly available herbarium, as the experimental materials were derived from widely recognized natural populations. However, our study utilized live plant materials that are currently established and uniformly maintained in the experimental plantation at the Tongliao Institute of Forestry and Grassland (Inner Mongolia, China).

A field trial was established at this site via clonal afforestation in April 2024. The plantation was arranged in a randomized complete block design, with six ramets per genotype planted at a 50 × 50 cm spacing.

Phenotypic measurements were conducted on 849 P. simonii genotypes at four time points in 2024. The timing was based on temperature fluctuations recorded in Tongliao (Table S2). The sampling periods were September 19 (T1), September 30 (T2), October 14 (T3), and October 21 (T4). The measured physiological and biochemical parameters included chlorophyll content (Chl), relative electrical conductivity (EC), superoxide dismutase activity (SOD), total antioxidant capacity (T-AOC), soluble sugar content (Sugar), and reduced glutathione (GSH). For all measurements, the mature leaves from the 5th to 7th whorls exhibiting uniform growth were selected from each genotype. Chlorophyll content was measured using a handheld chlorophyll meter (Model LD-YC, Shandong Lainde Intelligent Technology Co., Ltd.). Three uniformly developed mature leaves were selected per plant. Each leaf was measured three times. The average was recorded as the Chl value for that individual. For relative electrical conductivity, four mature leaves were randomly sampled per plant. Leaf discs of 0.5 cm diameter were excised, avoiding main veins. Eight discs per sample were submerged in deionized water and shaken gently to measure initial conductivity (R1). Samples were then heated in a water bath for 20 min, and after cooling to room temperature, final conductivity (R2) was measured. EC was calculated using the formula EC = (R1/R2) × 100%. For the remaining biochemical assays, four mature leaves per plant were collected. They were immediately frozen in liquid nitrogen and stored at − 80 °C. After thorough homogenization, the contents of SOD, T-AOC, Sugar, and GSH were determined using commercial kits (Beijing Solarbio Science & Technology Co., Ltd.), following the manufacturer’s instructions.

Whole-genome resequencing and SNP variant detection

Genomic DNA was extracted from leaf samples of each genotype by Shijiazhuang Boread Biotechnology Co., Ltd. utilizing a magnetic bead-based plant DNA extraction kit. DNA quality was assessed by multiple approaches: purity was evaluated using a NanoDrop spectrophotometer (OD₂₆₀/OD₂₈₀). Concentration was quantified using a Qubit fluorometer. Integrity was verified via 1% agarose gel electrophoresis at 120 V for 45 min. Whole-genome resequencing libraries were constructed and sequenced on the BGI DNBSEQ-T7 platform using a paired-end 150 bp (PE150) sequencing strategy at a target sequencing depth of 10×. After quality control of raw reads, clean reads were aligned to the P. simonii reference genome using BWA-MEM [29]. This reference genome was assembled and provided by the research group of Professor Xiyang Zhao at Jilin Agricultural University. SNP calling was conducted using the standard sentieon workflow [30, 31]. The resulting raw variant set was then filtered using VCFtools and PLINK with the following criteria: missing rate ≤ 1%, minor allele frequency (MAF) ≥ 0.05, per-sample sequencing depth (DP) ≥ 3, and population mean depth ≥ 10. All insertions and deletions (InDels) were excluded. Missing alleles were imputed using Beagle v5.4 [32]. Following these quality control steps, 55,522 high-quality SNP markers were retained for subsequent analysis. The chromosomal distribution of SNPs was visualized using the CMplot package in R with a 0.5 Mb sliding window [33].

Statistical analysis and heritability estimation

Best linear unbiased predictors (BLUPs) for each trait across all individuals were derived by fitting a linear mixed model using the lme4 package in R [34]. All subsequent analyses were based on these BLUP values (Eq. 1). Descriptive statistics were computed for the phenotypic data across four sampling stages (T1, T2, T3, and T4), including the maximum, minimum, standard deviation (SD), coefficient of variation (CoV), kurtosis, and skewness. Pairwise correlation coefficients among the six traits were calculated using the cor.test function in R.

1

where y_ijk is the phenotypic measurement for genotype i in block j from family k; µ is the overall mean; G_i is the genotypic random effect G∼N(0,Iσ²_g) B_j is the fixed block effect; Fam_k is the random family effect Fam∼N(0,Iσ²_f); and ε_ijk is the residual ε∼N(0,Iσ²_e).

Narrow-sense heritability (h²) was estimated using a mixed linear model (MLM) incorporating whole-genome SNPs. A one-step approach was adopted to partition genetic and environmental variance by directly modeling raw phenotypic data from all biological replicates, using the rrBLUP package in R [35]. The kinship matrix (K) was constructed using the A.mat function in rrBLUP [36]. Narrow-sense heritability is as follows (Eq. 2):

2

RNA-seq analysis

Uniform four-week-old P. simonii tissue culture seedlings were selected and subjected to cold stress at 4 °C in a climate chamber for 0 (control), 12, and 24 h. Three biological replicates were included for each treatment condition. At each time point, mature leaves were collected and immediately frozen in liquid nitrogen. RNA-seq library construction and sequencing were performed by Beijing Biomarker Technologies Co., Ltd. on the Illumina platform. Differentially expressed genes (DEGs) were identified using the following criteria: fold change ≥ 2 and false discovery rate (FDR) < 0.01.

Genome-wide association studies

Following quality control, GWAS was performed using the GAPIT package (v3.4, https://github.com/jiabowang/GAPIT) in R, based on BLUP values and the filtered SNPs markers described above. To ensure robust detection and comprehensive capture of genetic variation, four distinct statistical models were employed: the General Linear Model (GLM), Mixed Linear Model (MLM), BLINK, and FarmCPU. This multi-model approach was adopted to leverage the strengths of these algorithms, thereby enhancing the reliability of association signal detection. Based on significance thresholds applied in recent GWAS studies of poplar and related forest tree species, a LOD score of > 4 (equivalent to P < 1 × 10^− 4) was adopted as the significance threshold for candidate locus identification [7]. This threshold was considered appropriate given the polygenic nature of complex quantitative traits in forest trees, the characteristically low linkage disequilibrium decay in outcrossing tree populations, and the limited statistical power of the present study, thereby optimizing the balance between Type I and Type II errors. Loci surpassing this threshold are referred to as suggestive association loci throughout this manuscript. Finally, Manhattan plots and quantile-quantile (Q-Q) plots were generated using the CMplot package to visualize association signals and the distribution of P-values [33].

Candidate gene screening and haplotype analysis

The candidate genomic regions were delineated as the intervals spanning 50 kb upstream and downstream of each significant SNPs identified by GWAS. The coding sequences (CDS) were extracted using TBtools software. The comprehensive functional annotation was conducted via the BMKCloud platform, integrating NR, GO, KEGG, and Pfam databases. The core candidate genes were prioritized by integrating these functional annotations with DEGs identified under cold stress.

Local association analysis was subsequently performed on the genomic regions harboring core candidate genes. SNPs in strong LD (R² > 0.9) with significant loci and with P < 0.001 were selected as high-quality markers for further analysis. Missense mutations within the CDS were prioritized for haplotype construction. Haplotype effect plots and local Manhattan plots were generated using the IntGenicPlot package in R to dissect the effects of allelic variation on phenotypic performance [37].

Reverse transcription quantitative PCR (RT-qPCR) analysis

Based on the previous GWAS results, the candidate gene PsiNDHM, significantly associated with Sugar was identified. The CDS of PsiNDHM was amplified from P. simonii cDNA by PCR and cloned into the pMDC32 vector under the control of the CaMV 35 S promoter (pMDC32-PsiNDHM) [38]. In parallel, an empty pMDC32 vector (negative control) was constructed as a negative control. Both vectors were separately transferred into Agrobacterium tumefaciens strain GV3101 for Agrobacterium-mediated transient transformation of four-week-old 84 K poplar seedlings. Transient transformation efficiency was confirmed by evaluating the relative expression of PsiNDHM via RT-qPCR. Rather than using stable independent lines, independent Agrobacterium infiltration events on distinct individual plants were performed.

Two days after transformation, all plants were exposed to 4℃ for 0 h (control) and 48 h in a climate chamber, after which mature leaves were harvested for gene expression analysis and physiological measurements. Total RNA was extracted using a Plant Total RNA Extraction Kit (Beijing Tiangen Biotech Co., Ltd., China). One microgram of RNA was reverse-transcribed to synthesize cDNA using a PrimeScript™ First-Strand cDNA Synthesis Kit (Beijing Takara Biomedical Technology Co., Ltd., China). RT-qPCR was carried out using a LightCycler^® 480 Real-Time PCR System (Roche Applied Science, Penzberg, Germany) and TB Green^® Premix Ex Taq™ (TaKaRa, Cat. No. RR420A). The relative gene expression level was calculated using the 2^−ΔΔCT method. All primer sequences used in this study are listed in Table S5. The PagActin gene was used as an internal reference to normalize expression data. The contents of H₂O₂ and malondialdehyde (MDA) were determined using commercial assay kits (Solarbio Science & Technology Co., Ltd, Beijing, China) following the manufacturer’s instructions.

Genomic prediction

Genomic prediction was performed using the genomic best linear unbiased prediction (GBLUP) model [36]. The statistical model is specified as follows (Eq. 3):

3

Where y is the vector of phenotypic values (BLUPs). 1 is a vector of ones and µ represents the overall mean. Z is the incidence matrix linking individuals to their genetic values (an identity matrix in this study). g is the vector of genomic estimated breeding values (GEBVs), assumed to follow g ~ N (0, Gσ²_g), where G is the genomic relationship matrix constructed using SNP marker [36], and σ²_g is the additive genetic variance. e is the vector of residuals with e ~ N (0, Iσ²_e), where I is the identity matrix and σ²_e is the residual variance. The BLUP values used as phenotypic input for GBLUP were estimated using the lme4 package, with block effects modeled as random effects and thereby accounted for prior to genomic prediction. Consequently, only the overall mean was retained as a fixed effect in the GBLUP model.

Prediction accuracy was evaluated using a ten-fold cross-validation procedure. The 849 genotypes were randomly partitioned into ten equally sized folds, in each iteration nine folds constituted the training population and the remaining fold served as the validation population. Prediction accuracy was quantified as the Pearson correlation coefficient (r) between predicted genomic estimated breeding values (GEBVs) and observed BLUP values. To ensure robustness, the entire cross-validation procedure was repeated five times. All analyses were conducted using the GAPIT package in R.

To systematically investigate the key factors influencing GBLUP prediction accuracy, we evaluated two sets of parameters spanning multiple gradients: marker density and training population size. Thirteen marker density gradients were established, comprising 50, 200, 555, 1666, 2,776, 3,887, 5,552, 11,104, 16,657, 27,761, 38,865, 49,970, and 55,522 SNPs. Six training population size gradients were evaluated, consisting of 85, 255, 425, 595, 764, and 849 individuals.

Optimization strategies for GS based on GWAS prior information

Three optimization strategies were developed by incorporating GWAS results as prior information into genomic prediction models. Each strategy was designed with a distinct analytical objective, necessitating a tailored data usage framework. Strategy I was designed as an exploratory analysis to characterize the relationship between marker density and prediction accuracy across a range of P-value thresholds. GWAS was performed on the complete dataset of 849 individuals prior to cross-validation, and the resulting SNP subsets were used to construct genomic relationship matrices (G-matrices) for GBLUP. Although this design involves overlap between feature selection and model evaluation, no phenotypic information from validation individuals entered the prediction model, and the G-matrix was derived exclusively from genotypic data. This approach is therefore considered appropriate for the exploratory objectives of Strategy I, and absolute accuracy estimates should be interpreted accordingly. For Strategies II and III, GWAS-derived information was directly integrated into prediction models as marker weights or fixed effects, respectively. A strict nested cross-validation framework was therefore implemented to prevent data leakage. In each fold, training and validation sets were partitioned prior to any association analysis. GWAS was then independently conducted within the training population. All subsequent model parameterization steps, including P-value threshold determination, fixed-effect locus identification, and marker filtering, were performed exclusively using training set data. The validation set was completely withheld from all stages of feature selection and model parameterization, ensuring unbiased estimation of predictive performance.

Strategy I. Marker selection based on P-value gradients

GWAS was first performed on all 849 individuals using the best-performing association model identified from the association analysis, and all SNPs were ranked according to their GWAS P-values. Ten SNP subsets were then defined based on the following P-value thresholds: 0.01, 0.03, 0.05, 0.07, 0.1, 0.2, 0.3, 0.4, 0.5, and 1.0, where the subset at P < 1.0 retained all available SNPs and served as the baseline. No prior biological annotation or functional information was incorporated into this selection process. Each SNP subsets was subsequently used to construct G-matrix, which was then applied in the GBLUP model for genomic prediction. Predictive accuracy for each subset was evaluated through ten-fold cross-validation across all 849 individuals. This strategy aimed to systematically evaluate how P-value-based filtering influences genomic prediction accuracy and to identify the optimal marker set for genomic selection in this population.

Strategy II. Fixed-effect strategy using significant loci from a single model

This strategy employs a “fixed effects + random effects” hybrid framework. The best-performing single GWAS model was selected first, identified as the model accounting for the maximum proportion of phenotypic variance for the target trait. Significant SNPs identified by this model were incorporated into the GBLUP model as fixed effects, and the remaining non-significant SNPs were used to construct the G-matrix and modeled as random effects. This strategy aims to enhance the capture of large-effect loci by explicitly weighting major QTLs as fixed effects. The extended model is specified as follows (Eq. 4):

4

where y is the n × 1 vector of phenotypic values (BLUPs). X is the n × m genotype matrix of significant SNPs (where m is the number of significant SNPs) and b is the m × 1 vector of fixed effect coefficients for these loci. Z is the n × n design matrix for random effects (an identity matrix). g is the n × 1 vector of genomic breeding values (random effects) estimated from the remaining non-significant markers, following the distribution g ~ N (0, Gσ²_g). e is the residual vector following e ~ N (0, Iσ²_e).

Strategy III. Fixed-effect strategy using joint significant loci from multiple models

To comprehensively capture the genetic architecture of complex traits, this strategy integrates detection results from three complementary GWAS: MLM, BLINK, and FarmCPU. These models offer complementary statistical strengths: multi-locus models (e.g., FarmCPU and BLINK) demonstrate superior power in detecting minor-effect loci, whereas single-locus models (e.g., MLM) provide robust control over the genetic background. Significant SNPs identified across these models were pooled to form a comprehensive set of candidate loci, all of which were incorporated into the GBLUP model as fixed effects, the remaining genome-wide markers were used to construct the G-matrix to model the polygenic background as random effects. This strategy aims to maximize extraction of genetic information from diverse statistical frameworks, thereby improving the estimation of both major and minor genetic effects.

Results

Phenotypic variation and trait correlations under cold stress

Phenotypic distributions of the six traits (EC, Chl, SOD, T-AOC, Sugar, and GSH) across the four sampling periods (T1–T4) were approximately normal, as assessed by frequency histograms (Fig. 1). Based on the Kolmogorov-Smirnov test, only chlorophyll content at the third time point (P < 0.05) and Sugar at the first time point followed a normal distribution, whereas the remaining traits deviated from normality or skewed distributions (P < 0.05) (Table S3). Considerable phenotypic variation was observed across all traits, providing a broad genetic basis for downstream analysis. The magnitude of variation differed substantially among traits. EC and Chl remained relatively stable, with coefficients of variation (CoV) consistently below 20% across all sampling periods (Chl: 13.52%–16.81%, EC: 14.01%–19.88%). In contrast, defense-related traits (GSH, SOD, T-AOC, and Sugar) exhibited considerably higher variability, with CoV exceeding 30%. Notably, GSH showed the greatest variation, with CoV exceeding 50% across all four time points (51.80%–58.76%), highlighting substantial genetic diversity in antioxidant regulation among individuals.

Fig. 1.

Descriptive statistics and pairwise correlations of six cold-tolerance traits across four cold-stress stages in P. simonii. (a–d) Phenotypic data from T1, T2, T3, and T4, respectively. Each panel displays. (i) marginal histograms showing the distribution of each trait (EC, Chl, T-AOC, SOD, Sugar, and GSH). (ii) scatterplots with Pearson correlation coefficients and associated significance levels (***P < 0.001, **P < 0.01, *P < 0.05) for all pairwise trait combinations

Physiological responses shifted distinctly as cold stress intensified from T1 to T4. Mean EC progressively increased, reflecting cumulative membrane damage. Conversely, mean values of Chl and defense metabolites (GSH, SOD, T-AOC, and Sugar) generally declined, indicating that progressive membrane damage was paralleled by reductions in both photosynthetic capacity and antioxidant defense.

Correlation analysis revealed strong synergistic interactions within the defense system (Fig. 1). Significant positive correlations were consistently observed among the antioxidant traits (SOD, GSH, and T-AOC) and Sugar. In contrast, Chl showed weak or non-significant associations with the other five traits (Fig. 1), suggesting that photosynthetic maintenance and biochemical defense are likely regulated by independent genetic pathways.

Multi-model GWAS detection

To identify genetic variations associated with cold tolerance, GWAS was performed on all six traits using four statistical models. Principal component analysis (PCA) revealed that the first three principal components (PC1–PC3) collectively explained 17.38% of the total genetic variation, and the 849 accessions were broadly clustered into seven groups (Figure S3). The GLM model showed inflated test statistics and poor control of type I error across all traits (Figure S1b), indicating its limited reliability for candidate locus prioritization. A total of 93 significant loci were identified across three models (MLM, BLINK, and FarmCPU), distributed across multiple chromosomes. Notably, chromosomes 1, 8, and 17 showed significant enrichment, harboring 20, 15, and 15 loci respectively (Figure S1a, Figure S1c), suggesting the presence of genomic hotspot regions for cold tolerance.

The number of significant loci varied among models and traits (Fig. 2e). Multi-locus models (FarmCPU and BLINK) exhibited superior statistical power compared to the single-locus models (MLM). FarmCPU identified the highest number of loci (57) (Fig. 2a), followed by BLINK (45). At the trait level, Chl was associated with the most loci (21), while GSH and T-AOC were associated with the fewest (10). Of the 93 loci, 29 were co-detected by at least two models, indicating high robustness and 8 high-confidence loci were consistently detected across all three models and were prioritized for subsequent functional validation and as fixed effects in genomic prediction.

Fig. 2.

Genetic architecture of cold-tolerance traits in P. simonii revealed by GWAS. (a) Manhattan plots of GWAS for six physiological traits. The dashed red horizontal dashed line indicates the significance threshold. (b–d) Local association and haplotype analysis of significant SNPs associated with Chl (b), EC (c), and Sugar (d), showing linkage disequilibrium (LD) structure among associated SNPs and phenotypic differences among haplotype groups. Group differences were assessed using the Kruskal-Wallis test (multiple groups) (pairwise comparisons). Asterisks indicate the following levels of statistical significance. *P < 0.05, **P < 0.01, ***P < 0.001. (e) Venn diagram illustrating number and overlap and number of significant loci identified by the three GWAS models across the six traits

Quantile-quantile (Q-Q) plots were further examined to assess model fit (Figure S1). For all three models, observed P-values for the vast majority of SNPs aligned closely with expected values under the null hypothesis in the low-significance region, indicating effective control of confounding factors such as population structure and kinship. In the high-significance tail, observed P-values deviated markedly from the null distribution, suggesting the presence of genuine genetic association signals.

Candidate genes and allelic effects at key loci

A total of 100 candidate genes were identified within a 50 kb genomic windows flanking significant SNP loci. Functional annotation against the NR, GO, and KEGG databases assigned putative functions to 93 of these genes (Table S4). By integrating these annotations with differentially expressed genes (DEGs) identified under cold stress, 14 high-confidence candidate genes were prioritized (Table 1). These genes are primarily involved in phytohormone signal transduction, secondary metabolism, redox homeostasis, and cellular organization. Notably, a highly significant SNP associated with Sugar content (P = 3.64 × 10⁻⁵) was identified on chromosome 5. This locus resides in the 3’ flanking region of the gene Posim05aG0172700 (Figure S2). Functional annotation revealed that this gene encodes the NAD(P) H-quinone oxidoreductase subunit M (NDHM), a core subunit of the chloroplast NADH dehydrogenase-like (NDH) complex. We therefore designated this gene PsiNDHM.

Table 1.

Annotation and association statistics of consensus significant loci identified by multi-model GWAS for cold tolerance related traits in P. simonii

GeneID	Name	Model	Trait	Marker	PVE	MAF	P-value	Annotation
Posim01aG0173400	IAA16	MLM	Chl	chr01a_15800801	0.019	0.112	5.93E-05	Auxin-responsive protein IAA16-like
Posim02aG0001700	POLY	FarmCPU	T-AOC	chr02a_141246	0.019	0.243	4.94E-05	Poly(ADP-ribose) polymerase
Posim04aG0167000	SAUR24	BLINK	T-AOC	chr04a_17765803	0.018	0.161	7.32E-05	Auxin-responsive protein SAUR24-like
Posim04aG0193000	WAK2	FarmCPU	Chl	chr04a_20101456	0.027	0.09	9.87E-07	Wall-associated receptor kinase 2-like
		BLINK	SOD	chr04a_20104873	0.028	0.06	7.24E-07
		MLM			0.022
Posim04aG0228500	HPCA1	FarmCPU	Chl	chr04a_23130123	0.02	0.057	3.15E-05	Leucine-rich repeat receptor protein kinase HPCA1 isoform X4
		BLINK	EC	chr04a_23129176	0.034	0.423	4.01E-08
		MLM			0.025
Posim05aG0172700	NDHM	FarmCPU	Sugar	chr05a_18327183	0.02	0.405	3.64E-05	NAD(P)H-quinone oxidoreductase subunit M
Posim06aG0068500	DDC	BLINK	EC	chr06a_5851123	0.019	0.095	4.93E-05	D-dopachrome decarboxylase
		MLM			0.019
Posim07aG0038900	PLP2	BLINK	EC	chr07a_3513852	0.037	0.058	1.20E-08	Patatin-like protein 2 isoform X2
Posim07aG0108600	LECRK1	BLINK	SOD	chr07a_13508352	0.022	0.473	1.31E-05	G-type lectin S-receptor-like serine/threonine-protein kinase LECRK1
		FarmCPU	GSH	chr07a_13508531	0.018	0.063	7.33E-05
Posim08aG0214500	cyp726A27	BLINK	Sugar	chr08a_21682040	0.034	0.369	4.09E-08	Cytochrome P450 726A27-like
		FarmCPU			0.027
		MLM			0.019
Posim09aG0077200	PR1	FarmCPU	Chl	chr09a_8917032	0.028	0.323	8.19E-07	Pathogenesis-related protein 1
Posim11aG0026400	ANR	BLINK	T-AOC	chr11a_2632502	0.038	0.121	6.25E-09	Anthocyanidin reductase ((2 S)-flavan-3-ol-forming)
		FarmCPU	T-AOC		0.026
Posim12aG0022600	Lr10	BLINK	SOD	chr12a_2427636	0.036	0.05	1.43E-08	Rust resistance kinase Lr10-like isoform X5
		MLM			0.018
Posim15aG0025000	FLS2	BLINK	EC	chr15a_1835867	0.02	0.08	3.69E-05	LRR receptor-like serine/threonine-protein kinase FLS2 isoform X2
		FarmCPU	Chl	chr15a_1836521	0.027	0.077	8.58E-07
		MLM			0.018

Local association and pleiotropy analysis of the PsiNDHM region

To further dissect the genetic effects of this locus, a fine-scale local association analysis was conducted on the genomic region harboring PsiNDHM. The results revealed evidence of pleiotropy, with significant signals detected simultaneously for Chl, EC, and Sugar, while no significant associations were found for the other three traits. This suggests that the natural variation at this locus may coordinately influence photosynthetic maintenance, membrane integrity, and osmotic adjustment.

Haplotype analysis based on the significant SNPs elucidated the phenotypic consequences of allelic variation at this locus (Fig. 2b–d). For Chl, three significant SNPs (chr05a_18322596, chr05a_18322630 and chr05a_18322631) located in the 3’-UTR defined three major haplotypes (Hap1 ‘GGCCAA’, Hap2 ‘GGTTGG’, and Hap3 ‘TTCCAA’). One-way ANOVA indicated significant differences in chlorophyll content among haplotypes (P < 0.05). Individuals carrying Hap3 ‘TTCCAA’ exhibited the highest chlorophyll retention under cold stress, suggesting that this haplotype confers enhanced photosynthetic stability (Fig. 2b). For EC, 14 haplotypes were constructed using six highly linked loci, a comparison of the two major haplotypes (frequency > 10%) revealed highly significant differences in EC values (P < 0.001) (Fig. 2c). Individuals carrying Hap2 ‘GGTTTTAAGGCC’ displayed significantly lower relative electrical conductivity, indicating superior protection of cell membrane integrity. For Sugar, a significant locus (chr05a_18327183) was identified within the CDS, the two haplotypes defined by this locus (Hap1 ‘GG’ and Hap2 ‘TT’) showed significant differences in Sugar (P < 0.01) (Fig. 2d). Individuals carrying Hap1 ‘GG’ accumulated significantly higher levels of Sugar, potentially enhancing cold tolerance through osmotic adjustment. Collectively, these findings suggest that PsiNDHM represents a promising pleiotropic candidate gene, whose natural variation is associated with coordinated differences in photosynthetic efficiency, membrane integrity, and osmolyte accumulation under cold stress in P. simonii, warranting further functional characterization.

PsiNDHM overexpression mitigates oxidative damage in poplar under cold stress

To investigate the role of PsiNDHM in cold tolerance, transient overexpression of PsiNDHM was achieved via Agrobacterium-mediated infiltration in 84 K poplar. RT-qPCR analysis confirmed approximately 13-fold higher PsiNDHM expression in overexpression lines relative to empty vector controls (CK) (Fig. 3a). Under control conditions, the H₂O₂ and MDA contents did not different significantly between overexpression lines and CK. However, overexpression lines exhibited significantly lower H₂O₂ and MDA levels than CK after cold stress (Fig. 3b and c), indicating that overexpression of PsiNDHM attenuates cold-induced reactive oxygen species accumulation and membrane lipid peroxidation.

Fig. 3.

PsiNDHM overexpression reduces oxidative damage in poplar under cold stress. (a) Relative expression levels of PsiNDHM in overexpression lines and empty vector controls (CK). (b–c) Contents of H₂O₂ and MDA in overexpression lines and CK under control and cold-stress conditions. Value for each CK group were normalized to 100%. Data represent the mean ± SD of three independent biological replicates. Asterisks indicate the following levels of statistical significance: *P < 0.05, **P < 0.01, ***P < 0.001

Impact of marker density on genomic prediction in P. simonii

The impact of marker density (50 to 55,522 SNPs) on prediction accuracy was evaluated using the GBLUP model (Fig. 4a). A distinctive two-stage pattern of “rapid increase followed by plateau saturation” was observed as marker number increased. In the low-density range (< 5,552 SNPs), prediction accuracy was highly sensitive to marker number. Specifically, the prediction accuracies for SOD, Sugar, and Chl at 5,552 SNPs increased by 244.9%, 154.5%, and 142.3%, respectively, compared to those at 50 SNPs. Beyond the inflection point of 5,552 SNPs, however, further gains diminished and became trait-specific. SOD, GSH, Sugar, and T-AOC plateaued first, with additional increases in marker density yielding accuracy improvements of less than 8%. In contrast, Chl and EC maintained a slight upward trend, peaking at 27,761 and 38,865 SNPs and reaching maximum accuracies of 0.371 and 0.502, respectively, representing increases of 21.8% and 10.8% relative to values at the inflection point. Overall, EC demonstrated the highest prediction accuracy (r = 0.50), while SOD showed the lowest (r = 0.28). These results suggest that a core set of approximately 5,500 markers is sufficient to capture the majority of genetic variation for most traits, although higher marker densities may further improve prediction accuracy for certain complex traits.

Fig. 4.

Evaluation of genomic prediction accuracy across varying training population sizes, marker densities, and GS optimization strategies. (a) Effects of training population size and marker density on the prediction accuracy of the six cold tolerance traits. (b) Prediction accuracy of the three GS strategies (Strategy I, II, and III) across a range of P-value thresholds. (c) Pairwise accuracy comparison among the three strategies and the relative improvement (%) of Strategy III over the baseline GBLUP model

Impact of training population size on genomic prediction

The influence of training population size on genomic prediction accuracy was systematically evaluated using the GBLUP model (Fig. 4a). Although increasing training population size generally improved predictive performance, considerable heterogeneity was observed among traits in their responses to sample size. Chl, GSH, and T-AOC demonstrated the greatest sensitivity to training population size, with prediction accuracies increasing by 174%, 125%, and 178%, respectively, as the population expanded from 85 to 849 individuals, ultimately reaching maximum prediction accuracies of 0.362, 0.448, and 0.362. By contrast, EC exhibited early saturation, achieving near-plateau accuracy at a training population of 255 individuals (r = 0.434), beyond which further increases in training population size yielded only marginal improvements, a slight decline was observed at the largest population size (r = 0.499 at TP = 849). SOD maintained consistently low prediction accuracy across all population sizes examined (r = 0.177–0.261), with no discernible upward trend, suggesting a more complex underlying genetic architecture. Sugar displayed a distinct trajectory, characterized by markedly low accuracy at the smallest training population (TP = 85, r = 0.123), a pronounced increase at TP = 255 (r = 0.300), and subsequent stabilization at larger population sizes (r = 0.373–0.446). Collectively, these results indicate that maximizing training population size represents an effective strategy for improving genomic prediction accuracy for most traits, with the notable exception of SOD, whose limited responsiveness to increased sample size likely reflects the highly polygenic nature of antioxidant enzyme regulation.

Impact of fixed effects on GS prediction accuracy

The predictive performance of the three modeling strategies (Strategy I, II, and III) was systematically compared across a range of P-value thresholds (Fig. 4b). Overall, Strategy III, which integrates consensus fixed effects derived from multiple GWAS models, consistently achieved the highest prediction accuracy across the six traits at all threshold levels. Strategy III yielded substantial improvements over both Strategy I and II. Under the full-marker condition, prediction accuracies for EC, Chl, T-AOC, SOD, Sugar, and GSH reached 0.578, 0.427, 0.475, 0.412, 0.564, and 0.533, respectively, representing average increase of 14.8% over Strategy I and 6.9% over Strategy II. Notably, SOD showed the greatest response to strategy optimization, with prediction accuracy increasing from 0.258 (Strategy I) to 0.412 (Strategy III), a 59.7% improvement (Fig. 4c). Furthermore, the relationship between prediction accuracy and SNP filtering thresholds exhibited strategy-dependence patterns. In Strategy III, prediction accuracies for all traits showed a monotonic increas as the P-value threshold was relaxed from 0.01 to 1, indicating that incorporating a greater proportion of background markers consistently enhances predictive performance. In contrast, Strategy I displayed non-monotonic fluctuations for certain traits (e.g., SOD and GSH), with a distinct accuracy trough in the moderate significance range (P = 0.03–0.05). The magnitude of improvement varied among traits. While EC and Sugar maintained consistently smooth upward trajectories across all three strategies, the performance gap between Strategy II and III for Chl and GSH widened markedly at more permissive thresholds (P > 0.3), demonstrating the advantage of the multi-model strategy in capturing complementary genetic signals across the genome.

Discussion

Multi-model integration improves the robustness of GWAS signals

Based on resequencing data from 849 P. simonii individuals and BLUP values for six cold-tolerance traits, this study systematically compared the detection performance of three GWAS models (MLM, FarmCPU, and BLINK). The results revealed substantial heterogeneity in the genetic architecture of cold-tolerance traits and demonstrated that multi-model integration significantly improves the robustness of association signals. The genetic basis of individual traits differed markedly. Chlorophyll content appears to be under oligogenic control, with significant loci concentrated on chromosomes 1 and 5. In contrast, biochemical defense traits such as SOD exhibited a typically polygenic architecture, with up to 42 significant loci distributed across eight chromosomes, each exerting only minor effects. From a methodological standpoint, single-locus models such as MLM showed limited detection power for highly complex traits such as SOD, whereas multi-locus models such as FarmCPU improved power but remained susceptible to algorithm-specific assumptions. Notably, 68.8% of the 93 significant loci were detected by only a single model. To address this, we adopted a multi-model cross-validation strategy, identifying 29 consensus loci supported by at least two models; among these, eight high-confidence loci were consistently identified across all three models. These consensus signals leverage the complementary strengths of individual models to capture both major and minor genetic effects [39], and loci detected consistently across multiple models may represent more statistically robust candidates than those identified by a single model alone, although cross-model agreement alone does not constitute definitive evidence of biological causality. Nonetheless, these high-confidence candidates provide valuable prior information for both functional validation and genomic prediction [40–42].

In summary, we demonstrate that relying on a single GWAS model risks missing key regulatory loci in non-model species such as forest trees, given their rapid LD decay and complex population structures [43, 44]. A multi-model integration strategy leverages the complementary strengths of distinct algorithms, enabling more comprehensive capture of the genetic architecture underlying complex traits, from major QTLs to minor polygenic effects. While multi-model consensus does not guarantee the identification of causal variants, it provides a pragmatic framework for prioritizing candidate loci with greater statistical support. This is particularly critical for long-generation perennial species such as Populus, where genetic transformation and phenotypic validation are exceptionally time-consuming and resource-intensive, making the robustness and completeness of candidate loci identification a prerequisite for precise molecular breeding.

Mechanistic insights into cold tolerance mediated by the key candidate gene PsiNDHM

While GWAS identifies statistical associations, bridging the gap between genetic signals and causal mechanisms requires multi-omics integration. By coupling GWAS findings with cold-stress transcriptomic profiles, we identified 14 candidate genes exhibiting significant differential expression and functional relevance to cold tolerance (Table 1). This convergence of genetic and transcriptomic evidence substantially strengthens the credibility of these candidates. Central to these findings is PsiNDHM on chromosome 5, whose allelic variations was found to be pleiotropically associated with photosynthetic efficiency, membrane integrity, and soluble sugar accumulation. PsiNDHM encodes the NAD(P) H-quinone oxidoreductase subunit M, a core subunit of the chloroplast NADH dehydrogenase-like (NDH) complex that is essential for sustaining redox homeostasis during cyclic electron transport around Photosystem II (PSII) [45, 46]. Given that Sugars function as both osmolytes and ROS signaling molecules [47, 48] and that PSII quantum efficiency reflects photosynthetic integrity under stress [49, 50], our results are consistent with the possibility that PsiNDHM participates in modulating cyclic electron flow to stabilize the PSII redox balance under low temperatures, potentially mitigating photoinhibition-induced ROS accumulation and facilitating soluble sugar synthesis, thereby contributing to cellular osmoprotection. Importantly, transient overexpression assays provided preliminary functional evidence supporting a role for PsiNDHM in cold stress tolerance. Overexpression of PsiNDHM in 84 K poplar significantly reduced H₂O₂ and MDA levels under cold stress (Fig. 3), suggesting that PsiNDHM contributes to attenuating oxidative damage and preserving membrane integrity, although the precise mechanistic basis remains to be fully elucidated. Taken together, these findings are consistent with a regulatory axis integrating photosynthesis, redox homeostasis, and carbon metabolism, each step of which will require direct experimental validation. Beyond PsiNDHM, the identification of additional genes including IAA16 (auxin signaling), WAK2 (cell wall integrity), and LECRK1 (ROS scavenging) [51–53] highlights a multi-layered, multi-layered defense network underpinning cold tolerance in P. simonii.

Enhancing prediction accuracy via multi-model GWAS priors

Tree breeding is impeded by inherent challenges, primarily long generation intervals and the need for extensive field trials. Consequently, there is an urgent need for novel approaches to accelerate genetic gain. Genomic selection has recently emerged as a powerful tool for accelerating forest tree breeding programs [54]. In this study of P. simonii, the GBLUP model yielded prediction accuracies ranging from 0.25 to 0.60 across the six cold tolerance traits. Recent studies in forest tree species have demonstrated that incorporating trait-associated functional loci can significantly improve predictive accuracy [21, 55]. This finding has been corroborated by empirical evidence in crop and horticultural species [22, 56, 57] and supported by simulation studies [58]. Specifically, integrating quantitative trait loci (QTLs) as fixed effects has been improve GS performance [7, 59]. The statistical rationale for treating GWAS-identified loci as fixed effects lies in relaxing the distributional constraint imposed by the GBLUP assumption that all marker effects follow a Gaussian distribution with equal variance, which forces effect estimates to shrink towards zero. While this assumption is robust for highly polygenic traits governed by many minor loci, it systematically underestimates the effects of major QTLs and the heritable phenotypic variance they explain, thereby constraining the upper bound of prediction accuracy [27]. We provide empirical support for this mechanism, Strategy III achieved a 59.7% improvement in prediction accuracy for SOD over Strategy I. By reallocating significant loci from random to fixed effects, the model relaxes the distributional constraints on major QTLs, capturing additional explainable variance beyond the polygenic background.

However, the application of fixed-effect strategies in non-model species such as forest trees, has been constrained by the reliability of prior information. Inclusion of false-positive GWAS signals as fixed effects introduces disproportionate weights on noise, leading to overfitting and reduced prediction accuracy in in the validation population [24]. Consistent with this, we observed limited gains when relying on a single GWAS model: incorporating FarmCPU-identified loci for Chl improved accuracy by only 7.2%. The Multi-Model consensus strategy (Strategy III) addresses this limitation by integrating cross-validated signals from MLM, BLINK, and FarmCPU, functioning as a robust statistical filtering framework that prioritizes loci supported by multiple independent statistical frameworks, ensuring broader and more reliable representation of the trait’s genetic architecture. Consequently, the superior performance of Strategy III derives from assembling a fixed-effect set that more faithfully reflects the underlying genetic architecture of each trait. This approach not only alleviates the underfitting of major-effect loci inherent in standard GBLUP but also reduces model bias introduced by erroneous priors. Collectively, these findings demonstrate that the quality of prior information, rather than its quantity, is the decisive factor in optimizing GS models [60]. This framework offers a robust and scalable solution for non-model forest tree species characterized by complex genetic backgrounds and rapid linkage disequilibrium decay.

The influence of heritability, marker density, and population size on genomic prediction accuracy

In this study, we examined the interacting effects of marker density, training population size, and heritability on genomic prediction accuracy. While narrow-sense heritability defines the theoretical upper limit of prediction accuracy, the translation of this potential into realized accuracy depends critically on the alignment between statistical model assumptions and the underlying genetic architecture of the target trait [61, 62]. Two distinct response patterns were observed, reflecting the contrasting genetic architectures of different traits.

For the low-heritability traits EC (h² = 0.19), prediction accuracy exhibited continuous dependence on both marker density and training population size, with no discernible saturation plateau. This is consistent with classical quantitative genetics theory: when phenotypic variation is dominated by environmental noise, signals from minor-effect loci are easily obscured. Consequently, accurate genomic prediction of such traits require high-density markers to capture short-range LD and large training populations to mitigate statistical noise introduced by environmental variance [61, 63]. For high-heritability traits such as SOD (h² = 0.55) and T-AOC (h² = 0.54), standard GBLUP (Strategy I) yielded unexpectedly low prediction accuracies (r = 0.26), attributable to model misspecification rather than a biological paradox. Under the infinitesimal model assumed by GBLUP, all marker effects are drawn from a normal distribution with equal and small variances. When applied to traits governed by a mixed genetic architecture involving major-effect loci, this assumption induces excessive shrinkage of large-effect signals, compressing them toward zero and diluting their contributions within the polygenic background.

The identification of the PsiNDHM provides evidence for this mechanism. Although haplotype analysis (Fig. 2) demonstrated that this locus exerts substantial effects on cold-tolerance traits, its contribution was inadequately captured within the standard GBLUP framework, explaining why Strategy III achieved a 59.7% improvement in SOD prediction accuracy by modeling consensus loci as fixed effects. Fundamentally, treating these loci as fixed effects relaxes the distributional constraints imposed on major QTLs, enabling the model to recover their unbiased effect estimates and thereby better recapitulate the mixed genetic architecture of the trait [21]. This corroborates recent findings in Norway spruce in rubber tree (Hevea brasiliensis), suggesting that explicitly modeling major QTLs can substantially overcome the prediction accuracy ceiling of GBLUP for complex traits [24].

In summary, the design of genomic prediction strategies should be tailored to the genetic architecture of the target trait rather than focused solely on increasing data volume. For low-heritability traits, large training populations and high marker densities are critical for accurately estimating the polygenic background. In contrast, for high-heritability traits governed by major loci, expanding data volume alone yields diminishing returns. Instead, the key lies in correcting shrinkage bias on large-effect loci through mixed models such as strategy III.

Conclusion

This study presents a genomic selection optimization framework grounded in multi-model GWAS integration. By leveraging the complementary strengths of multiple GWAS models, the proposed strategy captures the genetic architecture of complex traits more comprehensively, encompassing both major QTLs and minor-effect loci. Incorporating consensus loci as fixed effects alleviates the shrinkage bias inherent in standard GBLUP, substantially improving prediction accuracy for traits with mixed genetic architectures. The biological relevance of this framework was further supported by functional characterization of PsiNDHM. Integrating GWAS, transcriptomic, and transient overexpression evidence, we demonstrated that natural variation at this locus is pleiotropically associated with photosynthetic efficiency, membrane integrity, and osmolyte accumulation, and that its overexpression attenuates cold-induced oxidative damage. Collectively, this work provides a robust methodological framework for precision molecular breeding, with particular utility for accelerating genetic gain in forest trees and other long-generation perennial species.

Abbreviations

EC

Electrical Conductivity

Chl

Chlorophyll

T-AOC

Total Antioxidant Capacity

SOD

Superoxide Dismutase

GSH

Glutathione or Reduced Glutathione

Sugar

Soluble Sugar Content

GLM

General Linear Model

MLM

Mixed Linear Model

FarmCPU

Fixed and random model Circulating Probability Unification

BLINK

Bayesian-information and Linkage-disequilibrium Iteratively Nested Keyway

GS

Genomic Selection

GWAS

Genome-Wide Association Study

SNP

Single Nucleotide Polymorphism

BLUP

Best Linear Unbiased Prediction

CoV

Coefficient of Variation

CV

Cross-Validation

TP

Training population

Strategy I

Marker selection based on P-value gradients

Strategy II

Fixed-effect strategy using significant loci from a single model

Strategy III

Fixed-effect strategy using joint significant loci from multiple models

Authors’ contributions

L.Q.W. designed the study. T.T.S., P.L.L., and H.C.L. performed the data collection and statistical analyses. W.T.Z. and H.C.L. prepared the experimental materials. L.R. and K.P.L. performed the transient expression of PsiNDHM. T.T.S., P.L.L., H.C.L., L.Q.W., and M.M.Z. wrote the manuscript. All authors read and approved the manuscript.

Funding

The present study was supported by the Biological Breeding-National Science and Technology Major Project (2023ZD0407501) and the Fundamental Research Funds of CAF (CAFYBB2022QC001-1).

Data availability

The datasets generated and/or analysed during the current study are available in the NCBI repository, accession number PRJNA1474111. The source code, genotype data, phenotype data, and modified GAPIT functions supporting the findings of this study are publicly available in the GitHub repository: https://github.com/lpleTree/priorGWAS-GS. The genotype matrix is distributed via Git Large File Storage (Git LFS). The GAPIT software (version 3) used for GWAS and genomic prediction is available at http://zzlab.net/GAPIT. All other data generated or analyzed during this study are included in this published article and its supplementary information files.

Declarations

Ethics approval and consent to participate

All experimental research on plants (either cultivated or wild), including the collection of plant materials, was conducted in compliance with relevant institutional, national, and international guidelines and legislation. Permissions for sampling were officially granted by the respective local forestry bureaus.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.