All animals in the study belonged to the Norwegian Red cattle breed. For the chromosome wide QTL scan, animals were organized in a granddaughter design consisting of 18 elite sire families with a total of 716 sons and 507,000 granddaughters. To fine-map QTL in the casein region, the animal data was expanded to 31 elite sire families with a total of 1112 sons, ranging from 23 to 70 sons for the smallest and largest families, respectively. The total number of daughters in this analysis was approximately 1.9 million, with an average of 1670 daughters per son. The families were chosen based on sufficiently large family sizes and/or availability of trait data. The pedigree of each animal in the study was traced back as far as known. Daughter yield deviations (DYDs) of the sons were used as performance information in the analyses. The DYDs for milk production traits [protein percentage (P%), protein yield (PY), milk yield (MY), fat percentage (F%) and fat yield (FY)] were available from the national genetic evaluation carried out by GENO Breeding and AI Association, and evaluated using a BLUP animal model 17.

For the initial QTL scan, we used a map consisting of 399 SNPs covering the entire BTA6 18. To fine-map QTL, we constructed a dense marker map consisting of 73 SNPs in and around the casein region on BTA6, covering approximately 750 kb. Fifty-four of the 73 SNPs in the map were detected by PCR resequencing of promoters and exon regions of all four casein genes (CSN1S1, CSN2, CSN1S2 and CSN3), nine SNPs were available from 19, whereas ten SNPs were selected from the Bovine Genome Sequencing Project 20. Physical distances between markers were determined from one single scaffold, NW_001495211, available from the latest assembly of the bovine genome Btau_4.0 20. The average distance between SNPs was 10,462 bp (ranging from 7 to 302,143 bp). A description of the SNPs, including accession numbers in dbSNP, assays for genotyping on the MassARRAY system (Sequenom, San Diego, USA), marker allele frequencies and predicted physical distances between markers can be found in Additional file 1.

A combined linkage and linkage disequilibrium (LDLA) method 5 was used to analyze milk production traits based on the information on markers from the 399-marker map described in 18 and a dense SNP map (73 markers) constructed for the casein region (see Additional file 1). For the midpoint of each marker bracket, the log-likelihood of a model containing the QTL (LogL(G_i)) was calculated as well as a model fitting only background genes (LogL(0)) using the ASREML package 21. Our test statistic, LogL difference, was then calculated as the difference in log-likelihood between the first and the second model. This LogL difference times 2 is equal to the Likelihood Ratio Test-statistic (LRT) of 22. According to Baret and coworkers, the distribution of the LRT under the null hypothesis can be seen as a mixture of two chi square distributions with 0 and 1 degree of freedom (df), respectively. Significance levels for the LRT are then found from a chi square distribution with 1 df but doubling the probability levels 22. Then, to obtain a significance level of 0.0005, the LRT value corresponding to a chi square distribution with 1 df and P = 0.001 is utilized. This LRT value is 10.8, and thus the corresponding LogL difference must be 5.4 or higher to achieve a significance level of 0.0005.

DYDs of the sons were used as performance information in the analyses. The model fitted to the performance information for each trait and each SNP was: DYD_i = μ + s_i + x_ib + a_i + e_i where DYD_i is performance of son i, μ is the overall mean, s_i is a fixed effect of sire of son i, x_i is 0 if son i is homozygous 1 1 (e.g. AA); 1 if son i is heterozygous 1 2 (e.g. AT or TA); or 2 if son i is homozygous 2 2 (e.g. TT), b is the effect of the SNP, a_i is a polygenic effect of son i, and e_i is a residual effect. For each single marker, the log-likelihood of a model containing the SNP effect (LogL(H1)) was calculated as well as a model without this SNP effect (LogL(H0)) using the ASREML package 21. Our test statistic, LogL difference, was then calculated as the difference in LogL between the first and the second model as described above. A SNP effect was regarded significant if the LogL difference exceeded 5.4.

Additionally, multiple SNP association tests were carried out for the most significant markers from the single SNP association test. The tests were implemented by fitting a fixed effect of the SNP in the above-mentioned model and repeating the analyses for the most significant SNPs in turn. Test statistics for the analyses were as described above.

An analysis package, CRIHAP, was developed for determining haplotypic phases and imputing missing genotypes for all individuals (Nome and Lien, unpublished). The programs are based on both linkage and linkage disequilibrium information generated by the CRI-MAP 2.4 23 and PHASE version 2.1 2425 programs. Map information and genotypes for all animals were imported into the Haploview program 26 to calculate LD (r²) between markers.

Haplotype blocks were constructed for the casein loci CSN1S1, CSN2 and CSN1S2 for which we found highly significant brackets or single SNPs associated with protein yield. A script was made to deduce maternal and paternal haplotypes for all individuals and different haplotype blocks using haplotypic phases from the CRIHAP program package. As for the single SNP analyses, DYDs of the sons were used as performance information in the analyses. The model fitted to the DYDs, for each trait and each haplotype, was DYD_i = μ + s_i + x_ib + a_i + e_i where DYD_i is the performance of son i, μ is the overall mean, s_i is a fixed effect of sire of son i, x_i is a row-vector indicating which haplotypes and how many copies are carried by the son; and b is a column indicating the random effects of the haplotypes; a_i is a random polygenic effect of son i, and e_i is a residual effect. The test statistic (LogL difference) was found as previously described for the single SNP association test. Phenotypic standard deviations for protein and milk yield were 36.75 kg and 1137.79 kg, respectively. These deviations were used to scale the haplotype effects into phenotypic standard deviations for each of the traits for a standardised presentation.

Results of the initial QTL scan for milk yield, protein yield, protein percentage, fat yield and fat percentage (LDLA analysis using the 399-marker map) are shown in Figure 1. For details about the markers, see Table S1 in Nilsen et al. 18 or http://cilit.umb.no/maps/. The analysis reveals highly significant results (LogL difference > 5.4, P < 0.0005) mainly in two different regions. Milk yield, protein yield and especially fat and protein percentages show highly significant results in the region between approximately 25 and 45 Mb. This QTL, previously fine-mapped in Norwegian Red cattle 3, is potentially caused by a polymorphism in the ABCG2 gene 45. Additionally, highly significant results were found for milk and protein yields in the casein cluster region at approximately 90 Mb. The results from the initial scan were followed up by LDLA analyses in a high-resolution map constructed for the casein region (73 SNPs) and using an extended number of families. The result of this analysis for protein yield and percentage are shown in Figure 2 (for details about the markers, see Additional file 1). The LogL difference for protein yield was found for the interval between the markers BTA6-02720 and CSN1S1-Prom_175 (LogL difference = 19.5), but several additional significant results appear for numerous marker brackets in CSN2 and CSN1S2. No significant result was found for marker brackets in the CSN3 gene. The interval between CSN1S1_192 and CSN1S1-BMC_17969 was the only one with significant LogL difference for protein percentage (LogL difference = 5.6).

Data was also analysed for association between single SNPs and DYDs for protein yield and milk yield. Highly significant results were found for a number of SNPs in CSN2 and CSN1S2 for both protein yield (PY) and milk yield (MY) (Figure 3 and Figure 4, respectively). SNPs with the highest LogL differences were CSN2-BMC_9215 and CSN2_67 for both traits (LogL difference = 26.4 for PY and 15.7 for MY for both SNPs), in addition to CSN1S2-BMC_17192 for MY (LogL difference = 15.8).

In most cases when fitting an effect of the most significant SNPs in a multiple SNP association test it highly reduced LogL differences for the other SNPs in the region. The most striking results were found for SNPs CSN2-BMC_9215 and CSN2_67. These two SNPs are in complete LD with each other and both removed almost all peaks for other markers in the region. The result for CSN_67 is presented in Figure 5. In accordance with the LDLA results no significant association was found between SNPs in the CSN3 gene and DYDs for PY.

The dense SNP map in the casein region made it possible to construct haplotypes within the casein loci. Such an analysis revealed five haplotypes for CSN1S1, seven haplotypes for CSN2 and six haplotypes for CSN1S2 (Figure 6). LD between pairs of loci varied from complete disequilibrium to almost no disequilibrium, and was much higher between SNPs in CSN2 and CSN1S2 than between SNPs in any other gene (Figure 7). The extent of LD between SNPs within CSN1S1, CSN2 and CSN1S2 allowed us to construct an extended haplotype block covering all three genes, creating 12 haplotypes with a population frequency above 0.9% (Additional file 2).

LogL differences for the four individual casein loci for PY and MY are shown in Table 1. As shown in Figure 8 and Figure 9, respectively, highly significant results were found in the CSN2 and CSN1S2 genes for both PY and MY. Six haplotypes were identified for CSN2. Estimation of the effect of haplotypes within loci on PY and MY revealed two haplotypes that tend to be negative (haplotype 2 and 5) and four haplotypes that tend to be positive (haplotypes 1, 3, 4 and 6) for CSN2 (Figure 8). For CSN1S2, we detected three haplotypes that are negative for both MY and PY (haplotypes 2, 3 and 4) (Figure 9). In contrast, both haplotypes 1 and 5 seem to be positive for both MY and PY. In addition, LogL differences for the extended haplotype block covering CSN1S1-CSN2-CSN1S2 were highly significant for both PY and MY (Table 1). The effects of the 12 haplotypes created for this block are shown in Figure 10. Effects of haplotypes for MY and PY were in the same direction for both traits, with four haplotypes tending to be negative (haplotypes 2, 3, 6 and 7) and eight haplotypes that seem to be positive for both traits.