Water buffalo ( Bubalus bubalis ) is an important source of milk in Egypt and China. Buffalo milk is well known for its high milk qualities. Compared with cow milk, it has higher contents of fat, protein, lactose, and minerals and is more suitable for the manufacture of various dairy products ( Michelizzi et al., 2010 ). The Egyptian buffalo (EGB) and the indigenous Chinese buffalo (CHB) belong to two different buffalo subspecies, the river buffalo and the swamp buffalo, respectively. They differ significantly in milk yield, approximately 2200-2400 kg/year for the riverine buffalo, and 500-700 kg/year for the swamp buffalo ( Shi et al., 2012 ). Along with the increasing demand for high-quality buffalo milk, increasingly more buffalo-breeding programs are aiming for buffaloes with improved milk performance, especially those with increased milk yield.

Marker-assisted selection is a useful approach to assist animal breeding. Recently, candidate genes harboring single nucleotide polymorphisms (SNP) significantly associated with milk fat content ( Li et al., 2018c ; Gu et al., 2019 ), protein percentage ( Manzoor et al., 2018 , 2020 ), and fatty acid composition ( Cosenza et al., 2017 , 2018 ) have been identified on different buffalo chromosomes. As for milk yield, one of the most important economic traits in buffalo industry, associated polymorphisms have been identified at the whole genome level ( Wu et al., 2013 ; Venturini et al., 2014 ; de Camargo et al., 2015 ; Menon et al., 2016 ; El-Halawany et al., 2017 ; Iamartino et al., 2017 ; da Costa Barros et al., 2018 ; Du et al., 2019 ). These studies have made great contributions to the progress of various breeding programs in buffalo. The continuous development of molecular technology enables researchers to comprehensively study the genome of buffalo and, thus, provides more useful information for buffalo breeding.

With the development of next-generation sequencing techniques, information for the whole buffalo genome sequence is now available at https://www.ncbi.nlm.nih.gov/assembly/GCA_003121395.1 . Restriction site-associated DNA (RAD) sequencing (RAD-Seq) is one of the next-generation sequencing techniques that have been deployed for detection of large number of SNP quickly and inexpensively ( Peterson et al., 2012 ). To our best knowledge, the RAD approach has not been applied in buffaloes to identify sequence polymorphisms.

In this study, we performed double digest (dd) RAD-Seq to detect genome-wide SNP in EGB and CHB populations with extreme difference in milk yield. Unlike the Egyptian buffalo, which has a relatively high milk yield, the Chinese buffalo, Haizi buffalo, is traditionally used for labor in farmlands with very low milk yield that is barely enough for the consumption of the calf. It is one of the most famous local buffalo breeds in China for its strong adaptability to hard environments and tasty meat. However, with the development of agriculture mechanization, Haizi buffalo is no longer needed in agricultural farming and was at the edge of extinction with only 1,132 head of live buffaloes in 2006 ( Cheng et al., 2008 ). The marker-assisted selection programs, aiming at improving the milk yield trait in Haizi buffalo and thus facilitating its conversion from draft-purpose only to meat-and-milk use, provide a practical way for the conservation of this unique buffalo breed.

The objectives of our present study included the screening of SNP that show significant differences in allelic frequencies between these two populations, categorization of functional genes associated with these SNP by gene ontology (GO), and pathway analyses to further understand their potential values as candidate genes closely associated with milk yield trait in buffaloes.

Blood samples from 45 two-year-old female buffaloes (25 and 20 individuals from the Chinese Haizi swamp buffalo and Egyptian river buffalo, respectively) were collected. The geographic sites were Yancheng, Jiangsu Province, China (33°50' N, 120°22' E) and Cairo, Egypt (30°03' N, 31°58' E), respectively.

Genomic DNA (gDNA) was extracted from the whole blood using the DNeasy Blood & Tissue Kit (Qiagen, Hilden, Germany) according to manufacturer's instructions. The quality of gDNA was verified on 1% agarose gel and then quantified using a spectrophotometer (Nanodrop 2000, Thermo Fisher Scientific, Waltham, MA, USA).

The ddRAD-Seq library was constructed by staff in Personal Biotechnology Co., Ltd (Shanghai, China) according to Peterson et al. (2012) . Enzymes used in this study were purchased from the New England Biolabs (NEB), Beverly, MA, USA. Briefly, 500 ng of gDNA was double-digested with Hind III (5'- A|AGCTT -3') and Bfa I (5'- C|TAG -3') at 37 °C for 3 h in a reaction volume of 20 μL. The digested gDNA was purified using Agencourt AMPure XP beads (Beckman Coulter, Brea, CA, USA). Then, in a 50-μL ligation system, adapter P1 (containing a unique 7-bp barcode sequence and a Hind III restriction site overhang) and adapter P2 (containing a Bfa I overhang) were ligated to the digested gDNA by using T4 DNA ligase. Ligation was performed at 22 °C for 60 min, followed by an inactivation step of 65 °C for 10 min. The resulting samples were purified by following the standard AMPure XP bead protocol to remove unligated adapters and adapter-adapter ligation products.

After the purification step, the fragments were screened using the Pippin Prep system (Sage Science, Beverly, MA, USA) to obtain 200-400 bp fragments. Enrichment of PCR of the library was performed using a Phusion High-Fidelity PCR Kit (NEB, Beverly, MA, USA) in a 20-μL PCR reaction volume containing about 20 ng of the size-selected DNA as the template and 2 μM each of P1 and P2 adapter primers. The PCR conditions were: initial denaturation at 98 °C for 30 s, then 14 cycles of 98 °C for 15 s, 65 °C for 30 s, and 72 °C for 30 s, followed by a final extension step at 72 °C for 5 min. The PCR samples were purified again with AMPure XP beads and checked on an Agilent Bioanalyzer (Agilent Technologies Inc., Santa Clara, CA, USA) to confirm the size distribution of fragments, and then quantified using a Qubit 3.0 fluorometer (Life Invitrogen, USA). Samples were then combined in equimolar ratios and sequenced on the platform of Illumina HiSeq 2500.

According to the 7-bp barcode sequences assigned to each sample in the ddRAD-Seq library, raw Illumina reads were de-multiplexed, which allowed the matching of each sequence read to a single sample. High-quality (HQ) cleansed reads were obtained by following these filtering criteria: trimming out the barcode sequences as well as the adapter sequences; removing reads with more than 50% bases having Phred quality scores lower than 20; and eliminating the pair-end sequences shorter than 50 bp. Then, the HQ reads were mapped to the reference genome ( https://www.ncbi.nlm.nih.gov/assembly/GCA_003121395.1 ).

The Burrows-Wheeler Aligner (BWA) tool (v0.7.12) was used to align cleansed reads against the reference buffalo genome (GCA_003121395.1 UOA_WB_1) with the demo BWA mem settings. Variant calling was performed according to Zhu et al. (2018) . The SNP were further screened by using vcftools based on these criteria: minor allele frequency (considering 45 samples together) ≥0.01; minimum depth of coverage for each sample ≥2; SNP missing rate across all samples (computed per population) ≤0.7; minimum Phred quality score for each bases ≥10; and max alleles/min alleles ≤2.

The SNP, successfully genotyped in all the 45 samples (missing genotype rate = 0) with a minimum allele frequency of 0.01 in both CHB and EGB, were retained for further bioinformatics analysis. Differences in allelic frequencies between CHB and EGB were calculated for all the polymorphic loci using the Population Differentiation option of GenePop version 4.2 ( http://genepop.curtin.edu.au/ ). Significant differences were established at P<0.05. Gene ontology and pathway analyses were performed by the PANTHER classification system (v.14.0) ( Mi et al., 2019 ). We used the “Gene List Analysis” tool. Briefly, the list of genes (official gene symbols) was uploaded, and the Bos taurus reference was selected as the reference list. For four functional categories (GO biological process complete, GO molecular function complete, GO cellular component complete, and PANTHER pathways), Fisher's Exact test (calculating False Discovery Rate, FDR) was applied to determine whether there was a statistical over- or under-representation of genes/proteins in the input list relative to the reference list ( Mi and Thomas, 2009 ). The interactions between proteins encoded by the promising candidate genes were predicted using the software Strings ( http://string-db.org/ ).

Raw sequence data obtained from the present study were deposited in the NCBI Sequence Read Archive under the BioProject number of PRJNA554744 (SRA accession: SRR9831104- SRR9831148).

In the present study, the Illumina HiSeq sequencing of the 45 ddRAD libraries generated a total of 833.68 million raw reads corresponding to about 118 Gb of sequence data. The average number of raw reads per sample was 18,526,242, ranging from 13,094,116 to 25,602,402 with a median of 19,117,758 and standard deviation (SD) of 3,029,678. After trimming and filtering, we obtained approximately 797.23 million (95.63%) HQ reads, of which, 99.24 and 97.10% of the bases had a Phred quality score of 20 (Q20) and 30 (Q30), respectively. The average number of HQ reads per sample was 17,716,223, ranging from 12,532,968 to 23,686,928 with a median of 17,850,416 and SD of 2,984,099. High-quality reads were then aligned with the reference buffalo genome, and about 99.93% of them were successfully mapped. The average sequencing depth was 0.86X. Summary of the ddRAD-Seq results were shown in Table 1 .

A total of 110,129 and 150,535 putative SNP were identified in Egyptian and Chinese buffaloes, respectively. The distribution of these SNP across the chromosomes was similar within the two populations ( Figure 1 ). Most of them were distributed in the intergenic and intronic regions ( Figure 2 ). The flow chart for SNP detection and bioinformatics analysis were explained in Figure 3 .

Altogether, 608 synonymous and 541 non-synonymous mutations were detected on the exons of functional genes. Of all these 1149 SNP, we focused only on high-confidence SNP with no missing data among all the 45 samples, having a minimum reads coverage depth of 2 and showing significant differences (P<0.05) in the allelic frequencies between CHB and EGB, which yielded 15 SNP distributing on 12 chromosomes and resulting in non-synonymous amino acid mutations ( Table 2 ).

For these 15 loci, the average reads coverage depth was 15X for EGB (ranging from 9 to 22X) and 20X for CHB (ranging from 13 to 28X). Except for the two SNP on MTMR10 and TACC2 , most polymorphisms were displayed only in one population. Furthermore, three mutations harbored by RASGRP1 (K7E), OTUD7A (H20Q), and TACC2 (E1523K) caused alterations between acidic and alkaline amino acids. These SNP are especially worthy of further validation in more populations.

Over the last decade, accumulated studies have identified candidate genes on buffalo chromosome 1, 2, 3, 8, 12, 14, 15, 22, and 23 closely related to milk yield by either genome wide association study or candidate gene approach. In the present study, we identified some novel SNP located on these candidate genes ( Table 3 ), which might provide potential evidence for their roles in milk production.

For further analysis, we retained only high-confidence SNP with no missing data among all the 45 samples, having at least a minimum coverage depth of 2, and a minimum allele frequency of 0.01 in both EGB and CHB populations. The final data set contained 1490 SNP with average reads coverage of 17.66X ( Table 4 ). The majority of them were located in intronic (47.92%) and intergenic (47.25%) regions, followed by 1.74% of them in the 3’UTR and 1.21% in exons. The rest were located in the upstream (0.74%), downstream (0.67%), and 5’UTR (0.47%), respectively. Of all the 1490 SNP, 886 SNP loci associated with 658 functional genes were of particular interest to us due to their significant differences in allelic frequencies between EGB and CHB (P<0.05).

Gene ontology analysis of these 658 genes with the PANTHER classification system (v.14.0) revealed the top five biological processes (BP), molecular functions (MF), and cellular components (CC) in which these functional genes were involved ( Figure 4 ). The fold enrichment for BP, MF, and CC ranged from 3.61 to 9.81, 5.21 to 7.09, and 2.69 to 3.99, respectively.

Panther pathways analysis showed that the most significantly over-represented pathways were the ionotropic glutamate receptor pathway (P = 1.18E-09, FDR = 1.94E-07), including eight genes ( CACNA1A, GRIA2, GRIA4, GRIK3 , GRIK4, GRIK5, GRM3 , and SHANK2 ); the endothelin signaling pathway (P = 1.56E-04, FDR = 6.40E-03), including 10 genes ( ADCY2, ADCY5, EDNRA, PIK3R1 , PIK3R6, PLCB1, PLCB4, PRKCE , PRKG1, PRKX ); and the gonadotropin-releasing hormone receptor pathway (P = 2.91E-05, FDR = 2.38E-03), including 14 genes ( ANXA5, CACNA1C, CREB1, FST , INHBA, INHBB, MAP3K1, NFATC2 , PBX1, PIK3R1, PLCB1, PRKCE , SCG2, SMAD2 ). The fold enrichment of three pathways was 7.00, 3.83, and 2.49, respectively. A total of 29 genes ( Figure 5 ) were involved in these three pathways.

Further analysis revealed that nine ( ADCY5, CACNA1A, CREB1, INHBA , INHBB, PIK3R1, PLCB1, PRKCE , and SMAD2 ) out of these 29 genes were significantly enriched in such biological processes as the regulation of hormone secretion, reproductive process, as well as ovulation ( Figure 6 ). These genes, associated with high-confidence SNP showing significant differences in allelic frequencies between CHB and EGB, involved in the most significantly over-represented pathways, significantly enriched in the biological processes closely related to the activities of lactation, were considered to be the most interesting candidate genes for milk yield trait in buffaloes.

Restriction site-associated DNA sequencing is a fast and useful technique to generate large numbers of SNP. Analyses based on SNP developed by RAD sequencing usually focused on high-confidence SNP only and removed those SNP showing low depth of coverage and low SNP calling rate among samples. The criteria set for screening of high-confidence SNP varied among documents. For example, a minimum coverage depth of 5 and more than 90% of the samples being successfully genotyped were adopted in a recent study ( Gao et al., 2019 ). However, in another study, the minimum coverage depth was set to be 3, and 70% of the samples were required to have data to process a SNP ( Hayashi et al., 2017 ). In the current study, a total of 886 high-confidence SNP were obtained. For each of them, the minimum coverage depth was 2, and the average reads coverage was 17.66X. All SNP were successfully genotyped in all the 45 samples with a minimum allele frequency of 0.01 and showed significant differences in allelic frequencies between CHB and EGB.

Functional genes associated with these high-confidence SNP were significantly over-represented in three pathways, which contained 29 genes with 41 related SNP. Of these, nine genes ( ADCY5, CACNA1A, CREB1, INHBA , INHBB, PIK3R1, PLCB1, PRKCE , and SMAD2 ) were considered the most promising candidate genes for milk yield trait due to their modulating roles in hormonal regulation of the lactation cycle.

It is known that milk production is under the control of various hormones secreted by the neuroendocrine systems, which include the reproductive (such as estrogen, progesterone, prolactin, and oxytocin), metabolic (such as growth hormone, corticosteroids, thyroid hormones, and insulin), as well as mammary (such as leptin) hormones ( Neville et al., 2002 ; Svennersten-Sjaunja and Olsson, 2005 ; Crowley, 2015 ). Of all these nine genes, INHBA, INHBB , and SMAD2 were significantly enriched in the TGF-β (transforming growth factor β) signaling pathway (P = 3.2E-3). The inhibin beta A subunit (INHBA) and inhibin beta B subunit (INHBB) are components of activins and inhibins, which belong to the TGF-β superfamily and are regulators for the synthesis and secretion of the pituitary follicle-stimulating hormone (FSH). Activins act as the stimulator. Inhibins, on the other hand, neutralize activins’ activities by binding to them ( Bilezikjian et al., 2006 ). They act with receptor-activated Smads (including Smad2) and co-mediator Smad (Smad4) to transfer the signal from cell surface to the nucleus and regulate the transcription of a variety of genes involved in follicular development, growth of oocytes ( Knight and Glister, 2006 ; Xing et al., 2014 ), and embryo differentiation ( Zhang et al., 2015 ). ADCY5, CREB1 , and PLCB1 were significantly enriched in the pathways of insulin secretion (P = 3.1E-3) and thyroid hormone synthesis (P = 2.2E-3). These three genes, together with PIK3R1, PRKCE , and CACNA1A , were also enriched in the pathways of estrogen signaling (P = 1.1E-4), aldosterone sythesis and secretion (P = 5.9E-5), and cholinergic synapse (P = 2.9E-6), respectively.

Due to their critical roles in hormone-related biological processes, association studies regarding intragenic SNP harbored by these functional genes and reproductive traits in farm animals have been conducted. For example, polymorphic loci in INHBA and INHBB were found to have significant effects on sperm quality and fertility in boars ( Lin et al., 2006 ); intronic INHBA SNP was reported to be associated with fertility of stallions ( Giesecke et al., 2010 ) and sperm quality in Chinese Holstein bulls ( Sang et al., 2011 ); INHBB SNP (at 3’UTR) were related to litter size in sheep ( Chu et al., 2011 ); SMAD2 was considered an important candidate for total number born in swine ( Wang et al., 2018 ). However, association studies regarding their sequencing variations and milk performance in buffaloes have not been documented yet. Based on the facts that these nine genes played various roles in the regulation of lactation-associated hormones, together with our findings that they were associated with high-confidence SNP showing significantly different allelic frequencies between CHB and EGB, we suggested that they were worthy of further investigations as candidate genes having influences on milk yield trait in buffaloes.

We also identified eight high-confidence SNP associated with five functional genes ( DIAPH3, FSTL4, GMDS, KCNMA1 , and SLC44A5 ), which have been previously documented as candidate genes for reproductive traits in buffaloes ( Wu et al., 2013 ; de Camargo et al., 2015 ; Li et al., 2018a ). They were all significantly different in the distribution of allele frequencies between EGB and CHB ( Table 5 ). Five SNP were intergenic, such as SNP associated with DIAPH3, GMDS , and KCNMA1 , which have been reported to be closely associated with calving interval and age at third calving in buffaloes. The remaining three SNP were intronic, with one and two SNP located within FSTL4 and SLC44A5 , respectively. In a previous study that combined GWAS (genome-wide association study) and RNA-seq of follicular granulosa cells, FSTL4 was the nearest functional gene associated with a SNP closely related to age at second calving and age at third calving in Italian Mediterranean buffaloes ( Li et al., 2018b ). In the current study, a novel intronic SNP located on FSTL4 was identified with significantly different allelic frequency distribution between CHB and EGB. Our results further confirmed that these functional genes, closely associated with genetic variants in two divergent populations with extreme phenotype in terms of milk yield, may explain the variances underlying the reproductive behavior in EGB and CHB.

In the present study, we used EGB and CHB, two phenotypically-divergent populations, to exploit genetic variants with potential effects on milk yield in buffaloes. Based on the strategy of selecting animals with extreme target trait for genotyping, previous association studies between DNA markers and interested traits have proved this to be an effective experimental design to identify candidate genes associated with target traits in other animals ( Fontanesi et al., 2012a , b ; Liu et al., 2018b ). Together with the application of ddRAD-Seq, an efficient and cost-effective approach for SNP detection, genomic variants identified in this study provided additional insights into candidate genes affecting milk yield in buffaloes.

Our results provide potential genetic variances for the selection of milk yield trait in buffalo. We suggest nine genes which are involved in the hormonal regulation of lactation process as promising candidate genes worthy of further investigations for favorable alleles closely related to milk yield trait.