Development and application of a 10 K low-density whole-genome SNP array for chickens

Abstract

Genomic selection (GS) accelerates genetic improvement in poultry, but the high cost of high-density SNP arrays limits its large-scale application. Here, we developed and validated a cost-effective 10 K SNP array for chickens based on 55 K array data from 58,496 individuals across 16 commercial lines. The genome-wide panel comprises 10,000 SNPs using a multi-step strategy that integrated the fixation index (F_ST), parentage-informative markers, and linkage disequilibrium (LD) pruning. Population genetic analyses showed that the 10 K array effectively captured major genetic diversity and clearly distinguished among groups. Genomic relationship estimates derived from the 10 K and 55 K arrays were highly concordant (r = 0.992), and genomic estimated breeding values (GEBVs) for seven traits showed strong correlations between the two platforms (r ≈ 0.97). Imputation from the 10 K to the 55 K array largely restored predictive accuracy to high-density levels, especially when the reference panel contained full-sibling individuals. These results indicate that the 10 K SNP array achieves an effective balance between cost and performance, and that a low-density plus imputation strategy is effective when a high-quality reference panel is available.

Introduction

Since Professor Theo Meuwissen first introduced the concept of genomic selection (GS) in 2001 (Meuwissen et al., 2001b), it has been widely applied in animal breeding due to its considerably higher accuracy compared with conventional best linear unbiased prediction (BLUP), leading to shorter generation intervals and accelerated genetic improvement (Schaeffer, 2006; Arirangan et al., 2026). However, the high cost of high-density SNP arrays remains a major constraint for GS (Florian et al., 2020; Qu et al., 2024), especially in poultry, which are characterized by large population sizes and relatively low individual economic value (Tu et al., 2025). Recent efforts in chicken breeding have explored new low-density SNP genotyping panels specifically adapted to local and commercial chicken populations, demonstrating both genomic diversity capture and practical feasibility for genomic prediction (Wang et al., 2025).

As a core tool for GS in poultry breeding, the first chicken SNP genotyping array, a 3 K panel containing 3,072 SNPs, was developed by the U.S. National Poultry Research Center in 2005 (Muir et al., 2008). This was followed by the development of a genome-wide 60 K BeadChip by Wageningen University (Groenen et al., 2011), which greatly expanded genomic coverage but was primarily intended for research use. Subsequently, the first commercial SNP array, the Affymetrix 600 K SNP array, was developed by the University of Edinburgh in collaboration with Aviagen (Andreas et al., 2013). In China, the 55 K SNP array independently developed by the Institute of Animal Science, Chinese Academy of Agricultural Sciences, provides a more suitable platform for the genetic improvement of local chicken breeds (Liu et al., 2019).

With the rapid development of genomic breeding technologies, SNP genotyping platforms have gradually transitioned from research to industrial applications (Hayes et al., 2009; Zhang et al., 2025a). In practical breeding, the scalability of SNP arrays depends not only on marker density and genome coverage but also on cost, operational convenience, and compatibility with target breeds (Liu et al., 2020). Therefore, effectively reducing the implementation cost of genomic selection while maintaining prediction accuracy has become a critical challenge in poultry breeding systems (Wolc et al., 2016; Tu et al., 2024).

One promising solution is the use of low-density SNP arrays in combination with genotype imputation. For breeding companies and practitioners, a key concern is whether using a low-density array affects genetic variance explained and, consequently, selection rankings (Florian et al., 2020; Wang et al., 2025). Therefore, our study not only developed a low-density SNP array but also systematically evaluated its performance. Our objective was to determine whether the 10 K array can provide an effective balance between genotyping cost and prediction accuracy, thereby offering a practical and economically viable tool for routine genomic selection in poultry breeding programs.

Materials and methods

Experimental animals and study populations

A total of 58,496 chickens from 16 commercial broiler lines belonging to six breeding companies were collected for the development of the 10 K SNP array (Supplementary Table S1). These lines included both specialised fast-growing white-feathered broiler lines and yellow-feathered broiler lines, which represent the core parental populations widely used in the commercial broiler production system in China. Each line constitutes a relatively independent breeding population with a well-documented breeding background and a stable genetic structure.

To evaluate the applicability and robustness of the 10 K SNP array across populations with diverse genetic backgrounds, two independent validation populations were additionally established. The first validation population comprised 291 individuals from eight indigenous Chinese chicken breeds, representing genetically diverse populations. The second validation population consisted of 9,560 individuals from independent white-feathered broiler lines. These validation populations were used exclusively for downstream analyses, including population structure inference, kinship assessment, genotype imputation, and evaluation of genomic prediction performance.

All experimental chickens were reared, vaccinated, and managed according to standard commercial broiler management practices. Blood samples were collected via venipuncture, and all procedures were approved by the Animal Welfare and Ethics Committee of the Institute of Animal Science, Chinese Academy of Agricultural Sciences (IAS-CAAS, Beijing, China).

Phenotype collection

Phenotypic records were collected for traits commonly used in commercial chicken breeding programs. Body weight was measured for all individuals at 28 and 40 days post-hatching using electronic scales, with individual identification numbers recorded. Feed intake was measured continuously from 28 to 40 days of age. Feed conversion ratio (FCR) was calculated as the ratio of feed intake to body weight gain during the same period, using the formula: FCR = (feed intake) / (body weight gain). Daily egg production was monitored at a fixed time from the onset of laying using a barcode scanning system. Based on these records, cumulative egg number at 43 weeks (43W_EN) and 49 weeks (49W_EN) was calculated from the onset of laying to the corresponding week of age. Age at first egg (AFE) and egg weight at 43 weeks (43W_EW), measured using an electronic scale, were also recorded.

Prior to analysis, all phenotypic records were cross-checked against individual identification numbers to eliminate recording or matching errors. For each trait, outliers exceeding ±3 standard deviations from the population mean were removed. The quality-controlled phenotypic data were subsequently used for genetic parameter estimation and genomic prediction analyses (Valerio-Hernández et al., 2023). Descriptive statistics of all phenotypic traits used for subsequent analyses, including sample size, mean, standard deviation, and range, are provided in (Supplementary Table S2).

55 K SNP array genotyping

Genomic DNA was extracted from whole blood samples using a magnetic bead–based method, following standard molecular biology protocols (Berensmeier, 2006; Sambrook and Russell, 2006). Prior to genotyping, DNA integrity and concentration were assessed, and only samples meeting routine quality requirements without evident degradation were retained for downstream analyses.

Genotyping was performed using a chicken 55 K SNP array developed by the Institute of Animal Science, Chinese Academy of Agricultural Sciences (IAS-CAAS). All procedures were conducted according to the Illumina Infinium HD Assay Ultra standard protocol. Arrays were scanned using the HiScan system (Illumina, USA), and genotype calling and initial quality control were carried out using GenomeStudio software (Illumina, USA). Genotyping services were provided by Beijing Compson Biotech Co., Ltd.

Design principles and construction of the 10 K SNP array

The 10 K SNP array was designed as a cost-effective genotyping tool for commercial chicken breeding programmes. During array development, particular attention was given to its applicability across populations with diverse genetic backgrounds and its suitability for downstream analyses, including population structure inference, pedigree verification, genotype imputation, and genomic selection.

Genomic locus conversion

All individuals were genotyped using the 55 K SNP array developed by the Institute of Animal Science, Chinese Academy of Agricultural Sciences (IAS-CAAS). As the 55 K array is based on the 6a reference genome, FASTA files corresponding to GRCg6a and GRCg7b were downloaded from the Ensembl database (https://www.ensembl.org/Gallus_gallus/Info/Index). A conversion chain file from GRCg6a to GRCg7b was generated using minimap2 v2.28 (https://github.com/lh3/minimap2) and transanno v0.3.0 (https://github.com/informationsea/transanno), and SNP loci were converted with liftOver v1.0.0 (http://hgdownload.soe.ucsc.edu/admin/exe/linux.x86_64/liftOver) (Kuhn et al., 2013; Li, 2018).

Identification of the first SNP set based on population differentiation

Pairwise genetic relatedness among individuals was estimated using the –genome option in PLINK v1.9. Individuals with a PI_HAT value greater than 0.25 were excluded prior to SNP selection. Fixation indices (F_ST) were subsequently calculated among populations using VCFtools v0.1.13 (Danecek et al., 2011). A one-to-many comparison scheme was applied, and SNPs were ranked according to their F_ST values. SNPs with higher F_ST values were retained for subsequent panel construction.

Identification of the second SNP set based on pedigree validation

Quality control was performed in PLINK with the following criteria: call rate ≥ 0.95, Hardy–Weinberg equilibrium test p-value ≥ 1 × 10⁻⁶, and minor allele frequency (MAF) ≥ 0.30. For each SNP, MAF stability was evaluated using bootstrap resampling, and the standard error and lower bound of the 95% confidence interval were estimated (Chatterjee and Bose, 2005). SNPs were further filtered based on these lower confidence bounds. Only loci retained across all populations were kept for subsequent analyses. Linkage disequilibrium (LD) pruning was performed at multiple thresholds to reduce marker redundancy while retaining genome-wide representation. Pedigree validation accuracy was evaluated by both accuracy and error rates, where accuracy was calculated as Accuracy = ((correctly assigned sires / total sires) + (correctly assigned dams / total dams))/2, and the error rate was defined as the number of individuals with incorrectly assigned parents divided by the total number of individuals.

Identification of the third SNP set based on even distribution

Candidate SNPs were further filtered by retaining loci with MAF ≥ 0.10 and subsequently merged across populations. SNP density along chromosomes was assessed based on physical coordinates, and loci were preferentially selected from genomic regions with lower marker density. Additional loci were included on chromosomes 16 and 31. SNPs were selected to achieve a uniform distribution across autosomes 1–31 and the Z chromosome.

Integration of the three SNP datasets

The final 10 K SNP panel was generated by integrating three SNP sets derived from pedigree validation, population differentiation, and genome-wide distribution criteria. When overlapping or competing loci were identified among datasets, SNPs were retained according to a predefined selection order. Priority was first given to pedigree-informative SNPs used for parentage verification, followed by SNPs with high population differentiation, and finally by SNPs selected to improve genome-wide uniformity.

Functional annotation

The final panel of 10,000 SNPs was functionally annotated using the Ensembl Variant Effect Predictor (VEP) (https://www.ensembl.org/Tools/VEP) based on the GRCg7b reference genome. SNPs were classified into functional categories including intronic, upstream/downstream, intergenic, synonymous, missense, splice-site, untranslated regions (UTRs), and regulatory regions (McLaren et al., 2016).

Genotype quality control and imputation

Initial genotype quality control was performed using PLINK v1.9 to ensure data reliability. SNPs with MAF < 0.05 or a genotype call rate < 0.90 were excluded from subsequent analyses. Variants that failed genotyping were removed using the –snps-only just-acgt option. The 10 K SNP array data were then imputed to the 55 K level, with both the reference and target panels derived from the same populations. Imputation was conducted using Beagle v5.2 under default parameters (Browning et al., 2018). Imputation performance was evaluated primarily at the genomic prediction level, focusing on recovery of prediction accuracy and variance explained.

Population structure and linkage disequilibrium analysis

By assessing the population structure across all study cohorts, the array's capacity to capture genetic associations is validated. Population structure was analyzed across all breeds using principal component analysis implemented in PLINK v1.9. The first two principal components were visualized with the ggplot2 package in R (Ginestet, 2011). Ancestry inference was further conducted with ADMIXTURE v1.3.0 (Alexander et al., 2009). LD decay was quantified and visualized using PopLDdecay with a maximum distance of 500 kb (Zhang et al., 2018). Analyses were performed for the pooled population comprising all individuals to provide an overall estimate of linkage disequilibrium across the studied populations.

Estimation of genetic parameters and breeding values

Genomic breeding values (GEBVs) and heritability estimates were calculated using the ASReml software. These analyses confirmed the suitability of the 10 K SNP array for genetic evaluation and genomic selection across diverse chicken populations.

Genetic parameters were estimated using a single-trait animal model based on the restricted maximum likelihood (REML) method. The statistical model applied was as follows:

$$y=\text{Xb}+\mathrm{Z}\alpha +e$$ (1)

where $y$ represents the vector of phenotypic values for each trait, $X$ and $Z$ are the incidence matrices of fixed and additive genetic effects, respectively; $b$ denotes the vector of fixed effects, including generation and batch; $\alpha$ is the vector of additive genetic effects; and $e$ is the vector of random residual effects.

The variance–covariance structure of the random effects was defined as:

$$var\left[\begin{array}{c}a\\ e\end{array}\right]=\lfloor \begin{array}{cc}G{\sigma }_a^2& 0\\ 0& I{\sigma }_e^2\end{array}\rfloor$$ (2)

where ${\sigma }_a^2$ represents the additive genetic variance, ${\sigma }_{\mathrm{e}}^2$ the residual environmental variance, $G$ the genomic relationship matrix constructed from SNP genotypes, and $I$ the identity matrix.

Heritability was calculated as:

$$h^2=\frac{{\sigma }_a^2}{{\sigma }_a^2+{\sigma }_e^2}$$ (3)

The genomic relationship matrix G was constructed using the method of VanRaden (VanRaden, 2008). Heritability estimates and genomic estimated breeding values were computed using the ASReml v4.1 package in R.

Genomic selection evaluation and cross-validation

To evaluate the GS performance of different SNP datasets, a five-fold cross-validation scheme was implemented. Genomic estimated breeding values (GEBVs) were predicted using the same statistical model implemented in the R package sommer across all SNP datasets, ensuring comparability among panels. Fixed and random effects were kept identical across analyses, with only the genotype data differing among SNP datasets.

GS performance was evaluated using four complementary metrics, all calculated based on the test datasets only:

$$Accuracy=\frac{cor\left(GEBV,y\right)}{\sqrt{h^2}}$$ (4)

where GEBV denotes the predicted genomic breeding value, y denotes the corrected phenotype in the test set, and $h^2$ is the estimated heritability of the trait.

$$Rankcorrelation=\rho \left(\text{GEBV},\mathrm{y}\right)$$ (5)

where $\rho$is Spearman’s rank correlation coefficient, reflecting the consistency of individual ranking.

$$\text{Unbiasedness}=b\text{from}y=a+b·\text{GEBV}+\epsilon$$ (6)

where b is the regression coefficient of corrected phenotypes on predicted GEBVs; a value close to 1 indicates an unbiased prediction.

$$\text{MSE}=\frac{1}{n}\sum _{i=1}^n{\left(y_i-GEB{V}_i\right)}^2$$ (7)

where n is the number of individuals in the test set.

All reported GS evaluation metrics were derived exclusively from the test datasets to avoid overestimation of prediction performance.

Results

Development and composition of the 10 K SNP array

A low-density 10 K SNP array designed for multi-line applications was developed using 55 K SNP genotypes from 58,496 chickens representing 16 commercial broiler lines (Fig. 1). After harmonization to the GRCg7b reference genome and quality control, 46,807 high-quality SNPs were retained, of which 19,302 remained polymorphic across all lines. Principal component analysis (PCA) based on this marker set revealed clear genetic separation among commercial lines, indicating that the dataset adequately captured inter-line genetic differentiation (Fig. 2A).

Fig. 1.

Technical pipeline for the development and application of the 10 K SNP array.

Fig. 2.

Evaluation of key steps in the development of the 10 K SNP array. (A) Principal component analysis (PCA) illustrating the genetic structure of the 16 chicken breeds used for array development. (B) Distribution of MAF, with the dashed line indicating the lower bound of the 95% confidence interval estimated by bootstrapping, used to filter out loci with unstable frequency estimates. (C) Linkage disequilibrium decay curve estimated from the pooled population. (D) Parentage assignment accuracy assessment showing that the accuracy (red line) increased rapidly with the number of SNPs and reached saturation at approximately 300 loci.

The final 10 K SNP array integrates three functionally complementary marker categories. A total of 299 SNPs exhibiting high levels of genetic differentiation among lines were included to enhance population discrimination. To support pedigree verification, SNPs were further evaluated for allele frequency stability and cross-population consistency (Fig. 2B), together with linkage disequilibrium characteristics (Fig. 2C). Parentage assignment accuracy increased rapidly with marker number and exceeded 99% across all populations when approximately 300 SNPs were used (Fig. 2D); based on this result, 330 SNPs were selected as pedigree-informative markers.

In addition, 12,468 SNPs showing stable allele frequencies across multiple lines were incorporated as broadly applicable markers for cross-population analyses and quality control (Table 1). After integration of all marker categories, candidate SNPs were further optimized for chromosomal distribution to improve genome-wide coverage uniformity. Ultimately, 10,000 SNPs were selected to constitute the final array (Supplementary Table S3). Despite its low marker density, the 10 K SNP array achieves a balanced design that combines population genetic resolution, pedigree verification capability, and multi-line robustness, providing a practical genotyping platform for large-scale genetic analyses and breeding applications.

Table 1.

Number of samples and SNPs retained after quality control in each chicken line.

Breed	Sample size	Pihat > 0.25	SNPs removed after quality control	SNPs retained after quality control
A	4,968	161	12,570	36,573
B	3,625	137	10,555	38,588
C	2,592	86	14,939	34,204
D	4,054	66	19,357	29,786
E	3,413	270	8,109	41,034
F	3,345	288	6,709	42,434
G	5,604	401	8,537	40,606
H	6,006	295	6,995	42,148
I	150	94	5,662	43,481
J	100	61	5,654	43,489
K	100	68	5,667	43,476
L	100	59	5,657	43,486
M	8,967	184	10,625	38,518
N	7,597	184	6,521	42,622
O	6,341	299	7,585	41,558
P	1,534	237	6,187	42,956

Characteristics of the 10 K SNP array

The final 10 K SNP array was systematically evaluated in terms of genome-wide coverage, spatial marker distribution, and functional composition to assess its suitability for population genetic analyses, kinship inference, and genomic selection applications.

Across the genome, SNP markers were evenly distributed across all 31 autosomes and the Z chromosome, with the number of markers per chromosome showing a strong positive correlation with chromosome length (Fig. 3A). No large marker-depleted regions were observed, indicating that the array achieves uniform genome-wide coverage, consistent with the design expectations of a low-density SNP array for whole-genome analyses.

Fig. 3.

Overview of the 10K SNP array features. (A) Distribution of SNP marker density within 1-Mb windows across each chromosome. (B) Genomic annotation categories of the 10 K SNP array: intronic variants account for 43.7%, upstream variants 24.2%, downstream variants 21.1%, and other functional categories (including coding, UTR, regulatory, and splicing variants) constitute the remaining ∼11%. (C) Number of SNPs per chromosome plotted against chromosome length. Chromosome lengths are shown on the right axis (based on GRCg7b), and SNP counts on the left axis.

Functional annotation revealed that the majority of SNPs were located in non-coding regions, in line with the design principle of genome-wide arrays optimized for linkage disequilibrium–based coverage rather than direct detection of functional mutations. Specifically, intronic variants accounted for 43.70% of all markers, followed by upstream (24.20%), downstream (21.10%), and intergenic variants (7.20%) (Fig. 3B). Only a small proportion of SNPs (2.30%) were located within coding regions, with synonymous substitutions predominating, indicating that the panel is largely composed of markers with broad information content rather than potential functional bias.

Further analysis showed that within ±20 kb windows surrounding each SNP, a total of 14,288 genes, 37,218 transcripts, and 1,104 regulatory elements were covered (Fig. 3C). This indicates that most markers are capable of capturing potential functional and regulatory variation through linkage disequilibrium. Taken together, these results demonstrate that the 10 K SNP array achieves a balanced trade-off between genome-wide coverage, functional representativeness, and marker informativeness.

Application of the 10 K SNP array in population genetic diversity

Principal component analysis was performed on a diversity panel comprising both commercial and indigenous chicken populations to systematically evaluate the ability of the 10 K SNP array to resolve population genetic structure.

PCA based on the 10 K SNP panel revealed pronounced and stable genetic differentiation among major chicken population groups. Fast-growing white-feathered broiler lines formed a highly compact and distinct cluster, clearly separated from indigenous breeds and specialized yellow-feathered chickens (Fig. 4A). The indigenous Beijing-You chicken also showed a well-defined and separable clustering pattern, indicating that even at a relatively low marker density, the 10 K SNP array provides sufficient resolution to discriminate genetically differentiated populations.

Fig. 4.

Population structure analysis. (A) PCA plot of one white-feathered chicken breed (B) and eight indigenous chicken breeds: Beijing You Chicken (BJY), Chahua Chicken (CH), Dagu Chicken (DG), Daweishan mini Chicken (DWS), Jixing Yellow Chicken (JXH), Piao Chicken (P), Wuding Chicken (WD), and Tibetan Chicken (ZJ). (B) PCA plot of 16 breeds including yellow-feathered and white-feathered chickens; W_ labels represent seven white-feathered chicken breeds, and Y_ labels represent nine yellow-feathered breeds. (C) PCA plot of nine yellow-feathered chicken breeds. (D) PCA plot of seven white-feathered chicken breeds. (E) Kinship analysis of nine breeds, assuming five ancestral populations (K = 5).

When the analysis was restricted to commercial populations, PCA consistently separated white-feathered and yellow-feathered chickens into two major genetic clusters across analytical scales (Fig. 4B). White-feathered lines exhibited tighter clustering and clearer separation among lines, reflecting long-term intensive artificial selection, closed breeding systems, and a consequently narrower genetic background. In contrast, yellow-feathered lines displayed a more overlapping distribution, consistent with their more complex breeding histories and higher levels of inter-population gene flow.

Further within-group PCA analyses revealed finer-scale population stratification. Nine specialized yellow-feathered lines were partitioned into four distinct genetic clusters, whereas seven white-feathered lines formed five clearly separable clusters (Fig. 4C–D). SNP-based relatedness analyses further supported these groupings and showed high concordance with inferred ancestry proportions (Fig. 4E). Notably, these genetic structure patterns were highly consistent with known breeding origins, pedigree relationships, and production records.

Overall, these results demonstrate that, despite its relatively modest marker density, the 10 K SNP array robustly captures genetic structure at both inter- and intra-population levels across diverse chicken populations. The resolution provided by this panel is sufficient for applications such as population stratification control and breed discrimination.

Application of the 10 K SNP array in kinship assessment

To evaluate the effectiveness of the 10 K SNP array for pedigree verification and relatedness-based breeding applications, genomic relatedness estimates derived from the 10 K panel were systematically compared with those obtained from the 55 K high-density SNP array and from recorded pedigree information.

Among 100 individuals genotyped using both the 10 K and 55 K arrays, genomic relatedness coefficients estimated from the two platforms showed a very high level of concordance, with an extremely strong correlation (r = 0.992, 95% confidence interval: 0.991–0.994; Fig. 5A). The mean difference between the two estimates was negligible, and no evident systematic bias was observed across the range of relatedness values relevant to breeding practice (Fig. 5B), indicating that reducing marker density to 10 K did not substantially compromise the accuracy of genomic relatedness estimation.

Fig. 5.

Correlation of breeding value estimation between 55 K and 10 K SNP arrays. (A) Scatter plot of genomic kinship coefficients estimated using the 10 K SNP array versus the 55 K SNP array. The correlation coefficient R = 0.992 (95% CI: 0.991–0.994) indicates high consistency. (B) Bland–Altman analysis comparing kinship coefficients estimated from the 55 K and 10 K SNP arrays. The x-axis represents the mean of the two estimates, and the y-axis represents the difference. The solid blue line indicates the mean bias (bias = 0.0021), and the dashed red lines indicate the limits of agreement (limits = [–0.0285, 0.0326]). No systematic trend is observed within the 0.0–0.5 range, demonstrating high interchangeability of the two methods. (C) Pearson correlation coefficients evaluating the linear correlation between breeding values estimated by the two arrays, as well as the correlation between estimated breeding values and phenotypes. (D) Spearman rank correlation coefficients assessing the consistency of breeding value rankings across the population.

Further comparison between genomic relatedness matrices and pedigree-derived relatedness matrices revealed inconsistencies attributable to pedigree recording errors. Genomic relatedness inferred from the 10 K SNP array enabled clear discrimination between true and incorrectly assigned parents, and pedigree corrections based on the 10 K array were fully consistent with those obtained using the 55 K SNP array (Table 2). These results demonstrate that the 10 K SNP array provides sufficient resolution for reliable pedigree verification and relatedness inference.

Table 2.

Kinship coefficients estimated from different data types.

IID1_IID2	55K_Kinship Coefficient	10K_Kinship Coefficient	PED_Kinship Coefficient
K10p01000003_K10p01000005	0.5058	0.5006	0.5
K10p01000003_K10p01000007	0.2448	0.2447	0.25
K10p01000003_K10p01000008	0.0372	0.0376	0.25
K10p01000003_K10p01000009	0.1857	0.1841	0.25
K10p01000003_K10p01000012	0.0401	0.0533	0.25
K10p01000003_K10p01000019	0	0	0.25
K10p01000005_K10p01000009	0.2749	0.2838	0.25
K10p01000005_K10p01000012	0.0191	0.0245	0.25
K10p01000005_K10p01000016	0.2968	0.3025	0.25
K10p01000005_K10p01000019	0	0	0.25
K10p01000007_K10p01000008	0.2956	0.2694	0.5
K10p01000007_K10p01000009	0.4576	0.4529	0.5
K10p01000007_K10p01000012	0.2739	0.2759	0.5
K10p01000007_K10p01000016	0.2923	0.2919	0.25
K10p01000007_K10p01000019	0	0	0.25
K10p01000007_K10p01000022	0.2959	0.3	0.25
K10p01000007_K10p01008110	0.1642	0.1807	0
K10p01000007_K10p01008321	0.1852	0.1954	0
K10p01000008_K10p01000012	0.4724	0.4772	0.5
...	...	...	...

Genomic relatedness estimates derived from the 10 K SNP array are highly consistent with those obtained from a high-density SNP platform, supporting the use of the 10 K array as a cost-effective and reliable tool for routine pedigree validation and relatedness monitoring in commercial breeding programs.

Application of the 10 K SNP array in genomic selection

The performance of the 10 K SNP array for genomic selection was evaluated by comparing genomic estimated breeding values derived from the 10 K panel with those obtained from the 55 K high-density SNP array, with particular emphasis on prediction accuracy, ranking consistency, and the impact of genotype imputation.

Across 9,500 individuals and seven economically important traits, GEBVs estimated using the 10 K SNP array showed consistently high concordance with those derived from the 55 K array, with correlation coefficients of approximately 0.97 for all traits, indicating that the overall genetic signal captured by the low-density panel was largely preserved.

When prediction accuracy was assessed as the correlation between phenotypes and GEBVs, the 10 K SNP array exhibited an average reduction of approximately 0.03 relative to the 55 K array (Fig. 5C). This reduction was consistent across traits with different heritabilities and genetic architectures, suggesting a limited and stable loss of information attributable to reduced marker density.

Given that relative ranking of individuals is often more critical than absolute GEBV values in practical breeding programs, rank correlations between GEBVs derived from the two arrays were further examined. The 10 K and 55 K SNP arrays showed high ranking consistency across all traits, with Spearman correlation coefficients ranging from 0.96 to 0.97 and only a marginal decrease (∼0.01) compared with the high-density array (Fig. 5D). These results indicate that the 10 K SNP array provides sufficiently robust ranking information for selection decisions.

To further enhance the performance of the low-density array, genotype imputation was applied using reference panels constructed from homologous populations. Reference panels of varying sizes were generated using individuals genotyped with the 55 K SNP array, and imputation performance was evaluated for a low-heritability trait (egg number, h² = 0.15) and a moderate-to-high heritability trait (egg weight, h² = 0.40). Following imputation, prediction accuracy based on the 10 K SNP array increased substantially and was comparable to that obtained with the 55 K array. For the two evaluated traits, prediction accuracy varied only modestly across the tested reference panel sizes (Fig. 6A).

Fig. 6.

Effects of genotype imputation using reference panels of varying proportions and sampling strategies on breeding value accuracy and heritability estimates. (A) Prediction accuracy of genomic estimated breeding values (GEBVs) for egg number under different proportions of the 55 K SNP reference panel and two sampling strategies (Random and Contain). (B) Estimated heritability (h²) for egg number under varying reference panel proportions and the two sampling strategies. (C) Prediction accuracy of genomic estimated breeding values (GEBVs) for egg weight under different proportions of the 55 K SNP reference panel and the two sampling strategies (Random and Contain). (D) Estimated heritability (h²) for egg weight under varying reference panel proportions and the two sampling strategies.

Pedigree structure further influenced imputation performance. When full-sib individuals were included in the reference panel, imputed GEBVs were nearly identical to those derived from the 55 K SNP array. In contrast, imputation based solely on half-sib relationships resulted in a slight reduction in prediction accuracy (0.003–0.012) (Fig. 6A, C), highlighting the importance of close familial relationships for optimal imputation (Supplementary table S4).

In terms of captured genetic variance, heritability estimates derived from the 10 K SNP array alone were approximately 0.04 lower than those obtained from the 55 K array. This loss was fully recovered after genotype imputation, with heritability estimates closely matching those from the high-density array regardless of sampling strategy (Fig. 6B, D).

Collectively, these results demonstrate that although the use of the 10 K SNP array alone leads to a small and predictable reduction in genomic prediction accuracy, this limitation can be effectively mitigated through genotype imputation. The combined application of the 10 K SNP array and imputation achieves performance comparable to that of high-density genotyping, providing a cost-effective and scalable solution for genomic selection in large commercial chicken breeding populations.

Discussion

Genomic selection, as a revolutionary breeding technology, has been widely applied in the breeding practices of many species (Meuwissen et al., 2001a; Wiggans et al., 2011; Tan et al., 2023). However, in practical poultry breeding programs, achieving an appropriate balance between genotyping cost and genomic prediction performance represents a central challenge for the large-scale implementation of genomic selection (Florian et al., 2020). We addressed this issue by systematically evaluating the performance of a low-density chicken SNP array across multiple populations.

In this study, based on 58,496 SNPs derived from the 55 K SNP array across 16 lines of yellow-feathered and white-feathered chickens, we designed a 10 K low-density SNP array through the selection of population-specific and parentage-informative markers. The performance of this array was systematically evaluated in multiple scenarios, including population structure analysis, parentage verification, relationship estimation, and genomic selection. Although strategies for SNP panel design and marker selection differ across studies (Liu et al., 2021; Wang et al., 2025; Zhang et al., 2025b), our results regarding the application of low-density panels and imputation are consistent with previous findings (Jiang et al., 2022). Moreover, our study provides general design considerations and practical recommendations for low-density SNP panel construction and application.

VEP annotation revealed that the majority of variants included in the 10 K SNP array are enriched in non-coding regions (intronic, upstream/downstream, and intergenic regions), with coding variants accounting for only ∼2.3%. Compared with high-density arrays or panels constructed from whole-genome resequencing, our SNP selection strategy prioritized broad population applicability and maximization of population genetic information. This approach is commonly adopted and has been demonstrated to be effective in low-density panel design (Andreas et al., 2013; Herry et al., 2018). Furthermore, PCA and STRUCTURE analyses based on the 10 K SNP array successfully distinguished commercial white-feathered chickens from several indigenous chicken breeds, with the commercial white-feathered lines forming more compact and easily distinguishable clusters, while yellow-feathered lines displayed more dispersed patterns and weaker separation. These findings are consistent with previous population structure studies using higher-density SNP arrays (55K/WGS), which demonstrated that population stratification revealed by high-density panels can often be replicated using carefully designed low-density panels, provided that the SNPs are evenly distributed across chromosomes and include markers informative for population differentiation (Tan et al., 2024).

We reported that genomic relationship estimates based on the 10 K SNP array showed a correlation greater than 0.99 with those obtained from the high-density 55 K SNP array across 100 individuals. In addition, parentage assignment achieved an accuracy exceeding 99% using 330 parentage-informative SNPs. Previous studies in other species have demonstrated that approximately 200–400 well-designed SNPs are sufficient to achieve high-confidence parentage assignment in most populations, suggesting that the number of markers included in our array is consistent with recommended ranges (Gebrehiwot et al., 2021).

In the evaluation of 9,500 individuals across seven traits, we observed that direct use of the 10 K SNP array for GEBVs prediction yielded a Pearson correlation of ∼0.97 with GEBVs derived from the high-density array. Compared with the 55 K SNP array, the correlation between GEBVs and phenotypes decreased by an average of ∼0.03, while rank correlations decreased by only ∼0.01 (remaining within 0.96–0.97). Many studies using low-density (LD) panels followed by imputation have reported similar findings: direct use of LD panels results in a modest reduction in predictive ability, but the construction of suitable high-density reference panels followed by imputation can largely restore predictive accuracy. For instance, imputation from 7K–75 K LD panels to high-density arrays in Nelore cattle achieved high genotype concordance and acceptable recovery of predictive accuracy (Carvalheiro et al., 2014). In livestock species including cattle, pigs, and chickens, LD panel plus imputation strategies have been demonstrated as a cost-effective and broadly applicable approach. Our findings (∼0.03 accuracy loss, almost fully restored by imputation) are consistent with these results (Bolormaa et al., 2015; Grossi et al., 2018), and recovery was most effective when the reference panel contained closely related individuals (e.g., full-sibs).

Although the performance of the 10 K SNP array is robust, several areas warrant improvement. First, the sample sizes for some lines were relatively small, and further expansion will be needed to validate the robustness of imputation and marker representativeness. Second, SNPs from small chromosomes (e.g., 32, 33, W) were excluded from this design. If future studies focus on traits linked to these chromosomes, targeted supplementation of SNPs or additional markers should be considered. Third, numerous studies have shown that imputation performance is highly dependent on reference panel composition and software choice. BEAGLE, widely used due to its computational efficiency and ability to perform both phasing and imputation simultaneously, does not always guarantee the highest imputation accuracy. In fact, its accuracy may decline when the reference sample size is limited. Therefore, the choice of imputation software should be tailored to the specific study design and available resources (Marchini and Howie, 2010).

For practical chicken breeding, we recommend a mixed strategy combining the 10 K SNP array with imputation as the preferred approach. When a high-density reference panel can be established, this strategy allows substantial reduction of genotyping costs while restoring GEBVs accuracy to levels comparable to those obtained with medium- or high-density arrays. In the absence of such resources, we recommend adopting within-family imputation strategies (full-sib or half-sib based). Overall, the 10 K SNP array offers a practical and cost-effective balance between genotyping cost and performance, making it well suited for large-scale routine breeding and population management programs.

Conclusions

The 10 K SNP array combines cost-effectiveness with robust capabilities for characterizing population structure, estimating kinship, and verifying parentage. When coupled with an appropriate high-density reference panel and imputation strategies, it achieves genomic prediction performance comparable to that of medium- and high-density arrays. This array provides a practical solution for large-scale, routine poultry genotyping and breeding applications. By enabling genomic estimated breeding value evaluation at scale, it facilitates early assessment of key production traits, supports the selection of superior individuals, shortens the generation interval, and accelerates genetic improvement.

Ethical approval

The study protocol was approved by the Ethics Review Committee of the Institute of Animal Science (IAS) of the Chinese Academy of Agricultural Sciences (CAAS) (reference no. IAS2019-44) and conducted in strict accordance with the Regulations for the Administration of Affairs Concerning Experimental Animals established by the Chinese Ministry of Science and Technology.

Funding

This study was funded by grants from the National Key Research and Development Program of China (2022YFD1301600), the China Agriculture Research System (CARS-41) and Special Fund for Basic Scientific Research Business Expenses of the Chinese Academy of Agricultural Sciences (Y2025YC55).

Data and model availability statement

The datasets used in the current study have not been deposited in an official repository. All the models used in the study are presented in the Material and Methods section of this paper. The data and models are available from the corresponding author upon reasonable request.

Conflict of Interest

All authors disclosed no relevant relationships.

CRediT authorship contribution statement

Xiaochun Ma: Writing – original draft, Visualization, Methodology, Formal analysis, Conceptualization. Yanru Chen: Visualization, Methodology, Data curation. Minghui Wang: Visualization, Formal analysis, Data curation. Lu Bai: Formal analysis, Data curation. Maiqing Zheng: Software, Resources, Investigation. Guiping Zhao: Resources, Project administration, Funding acquisition, Conceptualization. Jie Wen: Supervision, Project administration, Funding acquisition. Ranran Liu: Writing – review & editing, Validation, Supervision, Project administration, Funding acquisition, Conceptualization.

Disclosures

None.