Pedigree and phenotype records used for this study were provided by our lab. The pedigree contains 26,539 animals from 7 generations, including 14,226 Yorkshire and 12,313 Landrace animals. There were 12,661 and 10,635 records of birth weight for Yorkshire and Landrace piglets, respectively. After excluding disqualified records (missing birth date or abnormal records), 10,267 and 8,919 records Yorkshire and Landrace piglets were included, respectively. A total of 674 purebred sows (453 Yorkshire, 221 Landrace) born between 2014 and 2016 were selected for RAD-seq. After eliminating abnormal values (deviated from the third quartile), 668 high quality records were analyzed.

Genomic DNA was isolated from the ear tissue of pigs; the double-digest restriction enzyme associated DNA sequencing method (RAD-seq) was performed using the methods of Andolfatto et al. (2011) with appropriate modifications. Briefly, the DNA concentration of all samples was normalized to 50 ng/pr in 96-well plates, and digested with FastDigestTaq I- MspI (Thermo Fisher Scientific) in 30 μL volume containing 20 μL DNA (1 μg). An anneal adapter (10 μM) was ligated to the digestion products by T4 DNA ligase with 23 TaqI-Ms. Then, 24 ligation products were pooled together to form one library with 15 μL per sample. Agencourt^® AMPure^® XP Reagent was used for library size-selection. The PCR system contained 50 ng size-selection products, 25 μL KAPA HiFi HotStart ReadyMix (kapasystem), and 10 pmol primers. PCR products were purified by Agencourt^® AMPure^® XP Reagent. The final library quality (concentration and fragment size distribution) was determined by a Qubit 2.0 Fluorometer (Thermo Fisher Scientific) and BiopticQsep100 DNA Fragment Analyzer (Bioptic), respectively. Every four library products (96 different barcodes) were mixed together in equal parts which a total weight at 170 ng. The cycling system contained 48 μL library mix, 1 × T4 DNA ligase buffer, 0.5 μL T4 DNA ligase (600 U/μL), and 100 pmol Splint Oligo, were reactions at 37°C and fragment size distribution were determined by a Qubit 2.0 Fluorometer (Thermo Fisher Scientific) and Bioptic Qsep100 DNA Fragment Analyze sample volume of Agencourt^® AMPure^® XP Reagent. Finally, the purified cyclizing libraries were sequenced with a BGI-seq500 platform (PE100).

Sequenced paired-end reads for each sow were identified by barcode and aligned against the Sscrofa reference genome (version Sscrofa 11.1)¹ using the Burrows-Wheeler Aligner (version 0.7.12) software (Li and Durbin, 2009). SAMtools (version 0.1.19) (Li et al., 2009) was used to generate the consensus sequence for each sow and prepare input data for SNP calling with the Genome Analysis ToolKit (version 3.4) (McKenna et al., 2010). Raw SNPs with sequencing depth greater than 2,500 or less than 50 were removed, as SNP with extreme sequencing depth is most likely caused by a repeat DNA sequence or alignment error. The SNPs underwent quality control (QC) in which those with a call rate > 0.5, minor allele frequency (MAF) > 0.05, and p-value > 10^–6 for the Hardy-Weinberg equilibrium test were kept, resulting in 140,948 SNPs. The missing genotypes were imputed with Beagle software (Browning and Browning, 2007), and the SNPs were filtered again with the above QC criteria. Finally, 139,634 high quality SNPs were retained for subsequent analysis.

These individuals were also genotyped with a Geneseek Porcine 50K SNP Chip (Neogen, Lincoln, NE, United States), which contained 50,697 SNPs across autosomes and sex chromosomes. QC of the SNPs was conducted using PLINK (version 1.07) (Purcell et al., 2007). The SNPs with MAF > 0.05, call rate > 0.97, and individual call rate > 0.95 were retained. Furthermore, we removed SNPs that were not mapped to the Sscrofa 11.1 genome, leaving 45,180 SNPs. The missing genotypes were imputed with Beagle software and underwent QC with the above QC criteria. Finally, 45,175 high quality SNPs were included.

We sequenced the whole genome of 28 boars, the ancestors of 453 Yorkshire sows (unpublished), with an average sequence depth ∼19× (ranged from 17.06× to 22.24×). After genome alignment with Burrows-Wheeler Aligner and SNP calling with the Genome Analysis Toolkit, 17,017,067 raw SNPs were detected. These SNPs were filtered using the Genome Analysis Toolkit with parameters “QUAL < 30 || QD < 2.0 || FS > 60.0 || MQ < 40.0 || MQRankSum < −12.5 || ReadPosRankSum < −8.0,” and using PLINK with MAF < 0.05 and p-value < 10^–6 for the Hardy-Weinberg equilibrium test. We removed 761,590 additional SNPs with missing genotypes across the 28 boars, leaving 11,668,346 high quality SNPs, which were taken as the reference panel for imputation.

The SNPs determined by RAD-seq (140,948 SNPs) and SNP chip (45,180 SNPs) were merged to produce a high density SNP set for sequence level imputation. After removing 427 duplicate SNPs from both SNP sets, 185,701 SNPs remained. We performed sequence level imputation with Beagle by taking the whole genome sequencing data of 28 Yorkshire boars (described above) and 20 Landrace pigs (downloaded from https://figshare.com/articles/data2019_tar_gz/9505259). After QC (MAF < 0.05 and p-value < 10^–6), we obtained 9,012,073 overlapping SNP markers for the two breeds and imputed the RAD_ chip SNPs of the Yorkshire and Landrace pigs to a genome-wide level.

Both pedigree and RAD_SNP information were used to build a kinship matrix among individuals to estimate the variance components of birth weight. The mixed linear model for this estimation was:

where Y is the phenotype vector, b is a fixed effects vector, i.e., herd-year-season, sex (only in pedigree-based estimation), breed (2 breeds in SNP-based and 6 strains in pedigree-based estimation) and birth parity, u is a vector of additive genetic effects following the multinormal distribution: u ∼ N (0, Aσa2) and ∼ N (0, Gσa2), respectively in pedigree and RAD_SNP based estimations, where A is the pedigree relationship matrix and G is the genomic relationship matrix constructed based on SNPs as described in VanRaden (2008). p is a material effects vector: p ∼ N (0, Iσp2) and e is a residuals vector: e ∼ N (0, Iσe2), and I is an identity matrix. σa2, σp2, and σe2 are the additive genetic, material genetic, and residual variances, respectively. X, Z₁, and Z₂ are the incidence matrices for b, u, and p, respectively. The variance components were estimated using the average information restricted maximum likelihood procedure in DMU software (version 6, release 5.2²). Heritability of birth weight was estimated as:

The standard error of heritability was obtained as Klei and Tsuruta (2008) described.

The mixed model including a random polygenic effect can be expressed as:

where Y is the phenotype vector, which is corrected with estimated breeding values and fixed effects (only residuals left), and estimated breeding values are evaluated with the average information restricted maximum likelihood procedure in DMU; b is the estimator of fixed effects including breed, g is the SNP substitution effect and a is the vector of random additive genetic effects following the multinormal distribution a ∼ N (0, Gσa2), in which G is the genomic relationship matrix that is constructed based on SNPs as described in VanRaden (2008), and σa2 is the polygenetic additive variance. X, Z, and M are the incidence matrices for b, a, and g, respectively. e is a vector of residual errors with a distribution of N (0, Iσe2). All single-marker GWAS analyses were conducted using the EMMAX software (Kang et al., 2010). Based on the Bonferroni correction, the genome-wide significant threshold was P < 1/N, where N is the number of informative SNPs.

The BayesFM-MCMC package (Fang and Georges, 2016) was used to finely map causative variants, in which the threshold for SNP clustering was set as r² = 0.5; the length of the Markov chain was 510,000 with the first 10,000 discarded (burn-in period). The threshold to declare significance was set at 1.1 × 10^–5, which was determined from 0.05 divided by the number of SNPs in the GWAS region. We corrected the phenotypes by subtracting the corresponding breeding values and fixed effects, where the breeding values were estimated via the DMU package.

SnpEff (version 4.3t) (Cingolani et al., 2012) was used to annotate the function of SNPs, in which the genome sequence and the genomic annotation databases (.gff) were required. The Sscrofa11.1 genome were downloaded from the National Center for Biotechnology Information³ and the genomic annotation file (.gff) was downloaded from the web ftp://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/000/003/025/GCF_000003025.6_Sscrofa11.1/GCF_000003025.6_Sscrofa11.1_genomic.gff.gz.

We obtained 139,634 SNPs from RAD-seq and only 45,175 SNPs from SNP chip analysis. First, we compared the allele frequencies (AF) of SNPs garnered from both genotyping platforms (Figure 1A). Compared with SNP chip analysis, RAD-seq more frequently found SNPs with lower AF. Specifically, the likelihood of RAD-seq finding SNPs with AF < 0.1 was nearly 0.3, almost two times higher than that of SNP chip analysis (∼0.1). We also compared the distance between adjacent SNPs determined by the two genotyping methods (Figure 1B). The adjacent SNPs found by RAD-seq were much closer together than those found with SNP chip analysis, suggesting that RAD-seq is more informative and may be helpful to detect causative genes. Finally, we determined the overlapping SNPs between the two SNP sets, and surprisingly found only 427 SNP overlaps.

We estimated heritability prior to the association study to fully understand how much birth weight is inherited. We used pedigree information and genome SNPs to estimate heritability. There were 14,226 and 12,313 individuals in the pedigree, and 10,267 and 8,919 records of birth weight for Yorkshire and Landrace, respectively. Genome-wide SNP information was used to build kinship among individuals and heritability was estimated as 0.094 ± 0.065. Then, once again using the pedigree, we estimated heritability in Yorkshire and Landrace pigs at 0.162 ± 0.026 and 0.131 ± 0.025, respectively (see Table 1), which are closer to previous reports than the heritability found when genome-wide SNP information was used.

Next, we performed an association study for genome-wide SNPs based on a mixed model that accounted for population kinship (see section “Materials and Methods”). SNP sets from RAD-seq and SNP chip analysis were merged together, with two signals on chromosome 1 and 4 exceeding the threshold (Figure 2A). The positions of the lead SNPs for the two regions were chr1: 97,745,041 and chr4: 112,031,589, respectively; the MAF of the lead SNPs were 0.24 and 0.34 and they explained 6.36% and 4.25% of the phenotypic variance, respectively. We then focused on the two GWAS regions surrounding the lead SNPs, which are determined as the surrounding 1∼2 Mb region around the lead SNP. To confirm the two GWAS signals, we performed separate GWAS for the RAD-seq and SNP chip datasets. The region on chromosome 4 was determined to be significant for the RAD-seq dataset but not for the SNP chip dataset; where the reverse was true for the region on chromosome 1 (Figures 2C,E). Despite only reaching significance in one dataset, the –logP values of both regions peak in both datasets, confirming the reliability of the GWAS signals. To check for false positives caused by population stratification, we closely examined the theoretical and observed p-values with a Q-Q plot⁴. The -logP values are well fit by a linear regression against theoretical -logP values (Figures 2B,D,F), suggesting that population stratification has been well corrected for, although, it is important to note that two breed populations were simultaneously investigated.

To further refine the regions containing causative genes and variants, we performed fine mapping of the GWAS region 1∼2 Mb around the lead SNP. To increase fine mapping accuracy, we utilized as many SNPs as possible by merging the SNPs from both RAD-seq and SNP chip analysis and removing duplicate SNPs. After applying a stringent filter, we obtained 5,226 and 7,184 SNPs in the fine mapping regions of chromosome 1 and 4, respectively. With this high density of SNPs, we were able to impute SNPs at a sequence level. Sequence-level imputation requires a sequence-level reference SNP set. We therefore re-sequenced 28 Yorkshire boars with an average coverage of ∼19x and downloaded the whole genome sequencing data of 20 Landrace pigs. This resulted in 11,668,346 and 18,954,748 sequence-level SNPs for Yorkshire and Landrace pigs, respectively. With these SNPs as a reference panel, we imputed the merged RAD-seq and SNP chip SNPs at a sequence level using Beagle software separately for each breed. Then, we employed BayesFM-MCMC software to narrow down the clusters containing causative variants. BayesFM-MCMC first clusters the SNPs within a GWAS region using a hierarchy clustering algorithm according to r² among SNPs; then it models multiple causal variants by carrying out a Bayesian model selection across the cluster and generates the posterior probability for each SNP within the cluster, by which a credible set of SNPs with >95% posterior probability is constructed. The advantages of BayesFM-MCMC are that (1) it narrows down potential causative variants by indicating causal variants in the SNP set; and (2) it efficiently identifies more than one variant if multiple variants control the investigated trait.

However, because BayesFM-MCMC does not solve a mixed model with polygenic effects, we corrected the phenotype values by using the residuals (see section “Materials and Methods”). First, we conducted a single variant association for the GWAS chromosome region, 1,96,745,041–98,745,041, which produced a sharp peak in this region (Figure 3A). We then employed BayesFM-MCMC to further refine the regions, and one cluster signal with a posterior probability equal to 1 (greater than the threshold 0.5) was identified. To examine which SNPs predominantly explained the posterior probability in this cluster, we plotted the posterior probabilities for each SNP (output from BayesFM-MCMC). Most SNPs have miniscule posterior probabilities and no one SNP gives substantial posterior probability (f.i. greater than 0.5 or 0.2) in the identified cluster (Figures 3B,C). We then employed the 95% credible set defined by BayesFM-MCMC to further refine the causal variants, which contained 71 SNPs across a ∼100 kb region from 96,895,307 to 97,098,059 (see Supplementary Table S1 for detail). This 100 kb region contained the peak identified with the scan of single variants (Figure 3A), confirming the refined 100 kb region was reliable.

Fine mapping of the region on chromosome 4, 111,031,589–113,031,589 (Figure 4B), identified one cluster signal with a posterior probability equal to 1. As before, we plotted the posterior probabilities for each SNP but most SNPs once again had miniscule posterior probabilities (less than or 0.05) (Figure 4C). The 95% credible set of causal variants in chromosome 4 contained 432 SNPs across over a ∼870 kb region from 111,700,218 to 112,569,735 (see Supplementary Table S2 for detail). The peak found in the single-SNP association profile (Figure 4A) is covered by this ∼870 kb region, once again confirming the reliability of BayesFM-MCMC for this purpose. The correlation (r²) among SNPs confirmed that they were highly linked, which explains why the individual posterior probabilities of these SNPs are very small.

The 71 SNPs of interest on chromosome 1 are located in the intergenic region, which lies about 53 kb upstream of SKOR2 and over 317 kb downstream of SMAD2 (Table 2, see Supplementary Table S1 for details). We hypothesize that these variants are likely to have regulatory effects on the two nearby genes.

The 432 highly linked SNPs on chromosome 4 are located within four genes, LOC106510205 (covered by 28 SNPs), LOC106510207 (covered by 26 SNPs), VAV3 (covered by 160 SNPs), and NTNG1 (covered by 218 SNPs, see Supplementary Table S2). Among these SNPs, one is a coding amino acid, seven are located in the 3′ untranslated region and 414 are located in the intron (see Supplementary Table S2 for details). The coding variant is a synonymous variant (c.1136 T > A), localized in gene VAV3. The remaining variants are in non-coding sites distributed in all four genes, suggesting the causal variant may have regulatory effect. We searched for functional genes near the tightly linked region, and thereby included SLC25A24, PRMT6, STXBP3 as candidate genes (Table 2).