Introduction

Soybean (Glycine max [L.] Merr.) is a crop with substantial economic value, which represented 61 percent of the world’s oilseed production, soybean oil could be used into food products such as margarine, salad dressings and cooking oils, and industrial products such as plastics and biodiesel fuel (http://www.soystats.com) (Wilson 2004, 2008). Thus, development of increasing the soybean oil content is important for commercial production.

Soybean oil content is quantitatively inherited in soybeans and controlled by multiple genes by classical genetic studies (Wilcox 1985). As the development of molecular technology and quantitative genetics, numerous of soybean oil content quantitative trait loci (QTLs) have been reported over the past two decades, and there were over 188 QTLs listed in the USDA Soybean Genome Database (SoyBase, http://www.soybase.org). The mapped populations primarily included segregation populations, recombinant inbred lines (RILs) and backcross populations, and the main mapping methods included the analyses of variance (ANOVAs; Lander et al. 1987), interval mapping (IM; Lander and Botstein 1989; Haley and Knott 1992; Jansen 1993), composite interval mapping (CIM; Rodolphe and Lefort 1993; Zeng 1994), and multiple interval mapping (MIM; Kao et al. 1999). Among the published QTLs, some of them showed ‘hot regions’ that were identified four or more times at the same or similar intervals in different studies, and these regions include Gm05: 35.2–40.8 Mb, Gm09: 40.3–46.8 Mb, Gm12: 34.1–40.6 Mb, Gm14: 33.8–49.2 Mb, Gm15: 0.8–13.9 Mb, Gm18: 51.6–59.8 Mb, Gm19: 32.9–48.0 Mb and Gm20: 23.5–34.6 Mb (http://www.soybase.org).

Epistatic interaction effect between genes and their interaction with environments were two important genetic factors, which make a substantial contribution to the variation observed in complex traits (Carlborg and Haley 2004; Würschum et al. 2013). Most traits controlled by major genes are less affected by these two factors, while most quantitative traits are greatly affected by either one of them or both (Xu and Crouch 2008). Till now, many researches have been conducted to identify the QTL × environmental effects in various crops by these methods in recent years, such as tomato (Eshed and Zamir 1996; Chapman et al. 2012), rice (Liu et al. 2007; Wang et al. 2012b), corn (Wang and Guo 2010; Yang et al. 2015), soybean (Korir et al. 2011; Zhang et al. 2015; Qi et al. 2016), wheat (Mohan et al. 2009; Deng et al. 2015), common bean (González et al. 2015), cotton (Jia et al. 2014), chickpea (Bajaj et al. 2015), rapeseed (Xu et al. 2015), peanut (Wang et al. 2015), rapeseed (Luo et al. 2014), asparagus bean (Xu et al. 2013) and groundnut (Ravi et al. 2011). These studies showed that QTL × environment interactions greatly influenced the genetic architecture. Thus, it is meaningful to study the genetic characteristics of soybean oil content across different environments.

The identification of genes/QTLs and the development of higher-density molecular genetic maps have facilitated the study of complex quantitative traits, and the analysis of genes have influenced the related traits into individual Mendelian factors (Xing et al. 2012). Till now, few reports were reported focus on mapping soybean oil content QTLs under 25 environments or more based on a high-density genetic maps. Therefore, in this study, a RIL population was sequenced previously by specific-length amplified fragment sequencing (SLAF-seq) and a high density genetic map has been constructed (Qi et al. 2014a, b), with the population was tested for more than 10 years and employing multiple mapping methods to (1) identify the major QTLs of oil content and environment specific QTL, (2) mining the related genes of major QTLs, and (3) analyze epistatic QTL pairs across years (or environments) to better understand the implications of these QTLs when breeding for soybean oil content.

Materials and methods

QTL mapping population and experimental design

A total of 147 F2:10–F2:22 populations derived from RILs were advanced by single-seed descent from crosses between the American cultivars ‘Charleston’ (♀) (Cooper et al. 1995) and Chinese lines ‘Dongnong594’ (♂) (Developed by Northeast Agricultural University, Harbin, Heilongjiang, China). The RIL population was constructed by methods from Chen et al. (2005). A total of 147 RILs were planted in Harbin (HRB; longitude 126°38′E, latitude 45°45′N) from 2002 to 2014; Hongxinglong (HXL; longitude 129°55′–134°35′E, latitude 45°35′–47°17′N) from 2007 to 2013; and Jiamusi (JMS; longitude 130°21′E, latitude 46°49′N) from 2007 to 2013.

The plants were arranged in a randomized complete block experimental design involving single-row plots (1 m length and 0.5 m width), with two and three replicates in 2002–2007 and 2008–2012, respectively. The timing and frequency of cultural management procedures used for the trials were in accordance with normal production practices for the respective environments.

QTL validation population

A set of 194 wild soybean CSSLs population (Suinong 14, a cultivated variety in northeast of China, as the recurrent parent; and ZYD00006, a wild soybean variety, as the donor parent) had been constructed in previously study (Xin et al. 2016), which including 190 substituted segment covered 82.55% genome of wild soybean. The CSSLs and their parents were grown in a randomized complete block design with three replications in Harbin (HRB; longitude 126°38′E, latitude 45°45′N) in 2014 year. The other field management was the same as QTL mapping population.

Measurement of soybean oil content

Seeds from five randomly harvested plants were taken per plot for each line to analysis of seed oil content. Seed oil phenotypic data was analyzed using the FOSS Infratec TM1241 Grain Analyzer (FOSS Tecator AB, Hoeganaes, Sweden) at the 13% moisture basis. The seed oil content for each RIL line was averaged from the data of two or three plots replicates.

Map construction

The linkage map of this RIL population was previously constructed by specific-length amplified fragment sequencing (SLAF-seq) (Qi et al. 2014a, b). The high-density genetic map included 5308 markers on 20 linkage groups and was 2,655.68 cM in length, with an average distance of 0.5 cM between adjacent markers.

QTL mapping

Windows QTL Cartographer Ver. 2.5 (Wang et al. 2012a) was employed to independently analyze the QTL under each environment using a CIM model. IciMapping V 3.3 (http://www.isbreeding.net/software/?type=detail&id=12) was also used for conjoint analyses of QTLs to identify the major loci. These two models were used with a mapping step of 2.0 cM, an LOD significance threshold determined by 1000 permutation tests, and significance set at the P = 0.05 level. The naming of QTL was followed according to QTL-nomenclature described by McCouch (1997). All of the favorable alleles were derived from the parent ‘Charleston’ as indicated by the positive values of the additive effect and the parent ‘Dongnong 594’, as indicated by the negative values of the additive effects.

The MCIM model of QTL Network V2.0 (Yang et al. 2007; Yang et al. 2008) was used to analyze the additive QTL effects, additive × additive epistatic QTL pairs, additive QTL × environment effects, and epistatic QTL pair × environment interactions for seed oil content data of all 25 environments. The critical F value of the MCIM model was calculated with 10,000 permutation tests. The QTL effects were estimated using the Markov chain Monte Carlo method with 20,000 Gibbs sampler iterations and candidate interval selection and putative QTL detection, and the QTL effects were calculated with an experiment-wise type-I error under α = 0.001 (Wang et al. 1994; Yang et al. 2007; Xing et al. 2012).

Results

Phenotypic variation in the RIL population in 25 environments

Phenotypic value of seed oil content of RILs population in 25 environments across 13 years was shown in Table 1. The differences between the two parents were significant in most of the 25 environments, and Mean values of ‘Charleston’ and ‘Dongnong 594’ were 20.71 and 20.74%, respectively. The range of variation was narrow for ‘Charleston’ (19.03–21.89%) but was wider for ‘Dongnong 594’ (18.96–22.22%). The phenotypic data of the RIL populations also showed wide ranges of variation in the 25 environments. Some environments in this population had higher and others had lower mean of the trait as compared to the parents. Most skewness and kurtosis values were <1.00. In HRB location, the oil coefficient of variation ranged from 0.40 to 1.44; in HXL location, the oil content of RILs ranged from 0.36 to 1.05; in JMS location, the ranged from 0.39 to 1.35. An ANOVA analysis was employed to detect the year and location interaction effects of the phenotypic data, and most of the effects showed a significant interaction (P = 0.05).

Table 1 Phenotype of oil content of the population and parents in 25 environments across 14 years
Full size table

Overall, the phenotypic performance of the RIL populations was continuously distributed and relatively consistent in 25 environments, suggesting that the segregations of this trait fit a normal distribution model.

Main effects of the stable additive QTL analysis

An environment-universal QTL (QTLU) was defined as a QTL that was associated with the trait in at least one macro-environment that corresponded to the combination of year and location (Palomeque et al. 2009; Mao et al. 2013), in this research, the main-effect QTLs associated with the oil content were identified in more than two environments or detected by two methods. Seven of them (QOil3-2, QOil4-1, QOil13-5, QOil17-5, QOil17-7, QOil17-8 and QOil19-2) were identified on chromosomes 3, 4, 13, 17 and 19 by the CIM method. The LOD ranged from 2.70 to 6.22 with significance threshold determined by 1000 permutation test (P < 0.05), QOil17-8 was identified in three environments (2012HRB, 2012JMS and 2013HRB). QOil17-6 was mapped on chromosome 17 by the ICIM method in both 2013HRB and 2014HRB, and the main-effect QTL could explain approximately 10% of the phenotypic variation for the seed oil content. Four main effects QTL (QOil6-1, QOil15-1, QOil18-1 and QOil20-6) were identified by both CIM and ICIM methods and the LOD ranged from 2.67 to 3.60. QOil20-6 on chromosome 20 was identified in 2008HRB by two methods and also detected at the similar genome region in 2008HRB and 2010HXL by CIM method (Table 2), so above 12 QTLs was defined as main effect QTLs for soybean oil content in this research.

Table 2 Main-effect QTL mapping based on multiple environments
Full size table

Environment-specific QTL (QTLSP) was defined to be ass1ociated with the oil content that was significant only in one environment (Palomeque et al. 2009; Mao et al. 2013). In HRB location across 13 years, thirty-six QTLs were detected and the LOD ranged from 2.51 to 4.52. Among these QTLs, four environment-universal QTLs were found and also contributed to main effect QTLs, QOil15-1, QOil17-6 and QOil17-8 could explain approximately 10% of the phenotypic contribution for the seed oil content; thirty-two environment-specific QTLs were detected which located on Gm02, Gm 03, Gm 04, Gm 05, Gm 08, Gm10, Gm 11, Gm 12, Gm 13, Gm 14, Gm 15, Gm 16, Gm 17, Gm 18, Gm 19 and Gm20, and seven of them also contributed to main effect QTLs. In JMS location across 6 years, fourteen QTLs were detected and the LOD ranged from 2.51 to 3.60. Among these QTLs, one environment-universal QTL was found and also contributed to main effect QTLs, it could explain 10% of the phenotypic contribution for the seed oil content, thirteen environment-specific QTL were detected which located on Gm 02, Gm 06, Gm 08, Gm 012, Gm13, Gm 17, Gm 18 and Gm19, and four of them also contributed to main effect QTLs. In HXL location across 6 years, twenty-seven QTLs were detected and the LOD ranged from 2.57 to 6.22. Among these QTLs, no environment-universal QTL were found, all these QTLs were environment-specific QTLs which located on Gm 02, Gm 03, Gm 04, Gm 05, Gm 07, Gm 09, Gm10, Gm 11, Gm 13, Gm 15, Gm 16, Gm 17, Gm 18 and Gm20, and five of them also contributed to main effect QTLs, QOil4-1 and QOil3-5 could explain 11.78 and 16.16% of the phenotypic contribution for the seed oil content (Supplement Table 1).

Candidate gene mining of the main-effect QTLs

By comparison of the marker physical position of the main-effect QTLs and sequencing soybean genome (Williams 82, version 289), 1489 candidate genes were screened out from the gene annotation data set of soybean (http://phytozome.jgi.doe.gov/pz/portal.html#!info?alias=Org_Gmax), (Supplemental Table 2). Based on the blast with Kyoto Encyclopedia of Genes and Genomes website (KEGG, www.kegg.jp), eleven of them showed the relationship with seed oil content and/or fatty acid biosynthesis and metabolism based on the gene ontology and annotation information, including one gene in QOil3-2, seven genes in QOil4-1, two genes in QOil6-1 and one gene in QOil13-5 (Supplemental Table 3). These candidate genes functions have already been studied in the previously research. Glyma03g18661 was screened from QOil3-2 and predicted as Beta-ketoacyl synthase related gene, which was a key enzyme for fatty acid synthesis (Garwin et al. 1980). Glyma04g02271 was screened from QOil4-1 and predicted as sterol methyltransferase 2, this gene involved in the steroid biosynthesis (Shi et al.1996), researchers found it may influenced on fatty acid indirectly (Rakotomanga et al. 2005; Merah et al. 2012). Glyma04g05110 and Glyma04g05131 were screened from QOil4-1 and both of them were predicted as fatty acid desaturase (FAD) gene, which contribute to seed linolenic acid levels (Bilyeu et al. 2003). Glyma04g03590 was screened from QOil4-1 and predicted as phospholipid glycerol acyltransferase, which involve in the ether lipid/glycerophospholipid metabolism process, it also verified influence the fatty acid synthesis (Heath et al. 1994). Glyma04g03740 was screened from QOil4-1 and predicted as allene oxide synthase, which involved in the process of alpha-linolenic acid metabolism and specialized for the metabolism of fatty acid hydro-peroxides (Song et al. 1993). Glyma04g06110 was screened from QOil4-1 and predicted as 3-ketoacyl-CoA synthase 10 (KCS10), which may encodes a fatty acid elongase 3-ketoacyl-CoA synthase and influence the fatty acid indirectly (Todd et al. 1999). Glyma04g04961 was screened from QOil4-1 and predicted as alpha beta-hydrolases superfamily protein and this gene involved in the process of glycerophospholipid metabolism, this kind of enzymes were collected in involved in ESTHER database (Lenfant et al. 2013). Glyma06g17625 and Glyma06g17640 were screened from QOil6-1 and both of them were predicted as fatty acyl-ACP thioesterases B (FatB), which encode a group of enzymes displayed high specificity towards 12:0-ACP which correlates with the composition of the oil present in the seeds of that species (Salas and Ohlrogge 2002). Glyma13g28860 was screened from QOil13-5 and was predicted as cytochrome P450, family 710, subfamily A, polypeptide 1 which involved in the process of steroid biosynthesis, and this kind of gene influence the fatty acid indirectly (Muerhoff et al. 1992) All of the details on the gene annotation and expression information are shown in Supplemental Table 3.

Identification of epistatic effect QTLs and epistatic QTL interactions with the environment

With the analyzing the oil content data set of all 25 environments by QTLnetwork software, four epistatic QTL pairs were identified. One epistatic QTL pair, which was located on Gm01 (70.5 cM) and Gm18 (0.2 cM), showed additive epistatic effects and interactions with the environment (AAE), the phenotypic variation explained by the epistatic QTLs was 7.41%, and the phenotypic variation explained by epistasis × environment interactions was 15.26%. Furthermore, this QTL pair showed negative values for epistatic effects in 2004HRB, 2010HRB, 2011HRB, 2014HRB, 2007JMS and 2010JMS environment. The second epistatic QTL pair, located on Gm03 (5.1 cM) with the marker interval of Mark859393-Mark855624 and Gm15 (0.1 cM) with the marker interval of Mark18771-Mark23202, showed additive epistatic effects and interactions with the environment. The phenotypic variation explained by the epistatic QTLs was 7.01%, and the phenotypic variation explained by the additive epistasis × environment interactions was 14.55%, and this QTL pair presented positive values for the epistatic effects and positive epistatic effects in the 2003HRB, 2006HRB 2007HRB and 2010HRB environments, although it showed the negative values of 2012HRB; moreover, it also showed epistatic interaction in HRB environment rather than the other location. The third epistatic QTLs pairs were located on Gm04 (0.1 cM) with the marker interval of Mark771298-Mark756945and Gm08 (30.1 cM) with the marker interval of Mark1312817-Mark1334259, showed additive epistatic effects and interactions with the environment. The phenotypic variation explained by the epistatic QTL was 8.16%, and the phenotypic variation explained by the epistasis × environment interactions was 18.24%. This QTL pair had negatie epistatic effects in the 2008JMS, 2009JMS and 2012JMS environments, although it had a positive epistatic effect in 2011JMS. It showed epistatic interaction in JMS environment rather than the other location. The forth epistatic QTLs pairs were located on Gm06 (30.2 cM) with the marker interval of Mark450142-Mark493782 and Gm07 (0.1 cM) with the marker interval of Mark582982-Mark532996, showed additive epistatic effects and interactions with the environment. The phenotypic variation explained by the epistatic QTL was 10.17%, and the phenotypic variation explained by the epistasis × environment interactions was 22.05%. This QTL pair had negatie epistatic effects in the 2007HXL, 2009 HXL and 2012HXL environments, although it had a positive epistatic effect in 2011HXL, it also showed epistatic interaction in HXL environment rather than the other location (Table 3). Furthermore, one of QTL pair contributed over 20% of the phenotypic variation, these four QTL pairs contributed around 70% of phenotypic variation interaction with environments in total.

Table 3 Estimated epistatic effects (AA) and environmental (AAE) interaction of QTLs for soybean seed oil content across all environments
Full size table

QTL validation

To identify the effect of QTLs mapping results, a CSSLs population was used. We screened out the CSSLs with main effect QTL fragment, environment-specific QTL fragment and non-QTL fragment. Distribution of CSSLs seed oil content was showed in Fig. 1. The seed oil content of the CSSLs with main effect QTL fragment was range from 21 to 22.76%. CSSL-158 with the QOil3-2 showed the highest oil content (22.76%) in CSSLs. CSSL-77 brings the QOil17-5, QOil17-6, QOil17-7 and QOil17-8 together, which the seed oil content was 21.73%. CSSL-52 with QOil6-1 and CSSL-91 with QOil13-5 showed the seed oil content were 22.15 and 22.33%, respectively. The seed oil content of CSSLs with environment-specific QTL fragment were range from 20.7 to 22.38%. The CSSLs with main effect QTL fragment and most environment-specific QTL fragment showed the significant higher than the Suinong14 (the recurrent parent of CSSLs, average of seed oil content was 20.20%) and the CSSLs without any QTL fragment (seed oil content of non-QTL fragment CSSLs was average at 20.20%).

Fig. 1
figure 1

Distribution of CSSLs population oil content. Dark blue represent the CSSL with main effect QTLs, white blue and red represent the border of CSSL with environment-specific QTLs, yellow represent the recurrent parent (Suinong14) of CSSLs population, green represent the CSSLs without any QTL fragment. (Color figure online)

Full size image

Discussion

High-density genetic map could identify more recombination events in a population and will increases accuracy of QTL mapping (Xie et al. 2010). In soybean, many genetic linkage maps have been published based on restriction fragment length polymorphism (RFLP) markers (Keim et al. 1990), RFLP, isozyme, morphological, and biochemical markers (Shoemaker and Olson 1993; Lark et al. 1993), simple sequence repeat (SSR) markers (Cregan et al. 1999) and genetic map of integrated SSR, RFLP, random amplified polymorphic DNA (RAPD), amplified fragment length polymorphisms (AFLP), etc. markers (Song et al. 2004). With the advances in genome sequencing technology, few high-density genetic map have been constructed, Choi et al. (2007) constructed a soybean transcript map with single nucleotide polymorphisms (SNPs). Hyten et al. (2010) built a high-density integrated genetic linkage map“Universal Soy Linkage Panel” (USLP 1.0) with 1536 SNPs. Lee et al. (2015) constructed a high-density genetic map with SNPs based on the Infinium BARCSoySNP6 K BeadChip array. Song et al. (2016) have constructed a high resolution genetic linkage maps based on new soybean genome sequence assembly data Glyma1.01. In this study, a high-density soybean genetic map was used, and it was created by SLAF-seq that included 5308 markers on 20 linkage groups and was 2,655.68 cM in length, with an average distance of 0.5 cM between adjacent markers (Qi et al. 2014a, b), furthermore, the RILs were crossed between an American cultivars ‘Charleston’ and a Chinese lines ‘Dongnong594’ as distant hybridization, Charleston was semi-dwarf lines and selected from HC74-634RE × HC78-676 (Cooper et al. 1995), and ‘Dongnong594’ was selected from Dongnong42 serials, Northeast Agricultural University in China. In order to reduce the errors caused by the environment, the RILs were planted in HRB (from 2002 to 2014), JMS (from 2007 to 2013) and HXL (from 2007 to 2013), and each of the locations or years was different from a statistical perspective. As described by Jansen et al. (1995), the accuracy of the QTL position and effects can be evaluated if the data are collected in different locations and years from a statistical perspective.

Twelve main-effect QTLs were identified, they contributed the seed oil content significantly. Comparison of the result with previously study with the same population and similar data set based on the SSR marker linkage map (Qi et al. 2014a, b), there were no overlap for the main effect QTL with previously research, only few matches with environment specific QTL, but the QTLs detected in this study showed much accuracy with significantly much markers and shorter average maker distance. Furthermore, six of them had a good match with those detected in previously research. QOil4-1 showed similar position and bigger interval than the consensus QTL detected by Qi et al. (2011) and a main effect QTL detected by Wang et al. (2014a, b), although the interval was bigger than the result of meta-analysis, but the peak position was closed. QOil6-1 was identified by both the CIM and ICIM methods and it mapped with a smaller confidence interval at the central part of a QTL, which was identified under 4 environments by Hyten et al. (2004). QOil13-5 located on Gm03 from 3.1 to 3.3 Mb had been identified as the three QTLs Seed oil 37-7, Seed oil 37-8 and Seed oil 38-4 by Eskandari et al. (2013) and Rossi et al. (2013), respectively. QOil15-1 showed a smaller genome region as detected by Li et al. (2011). QOil19-2 was detected as a hot region by Mansur et al. (1996), Orf et al. 1999, Hyten et al. (2004), Qi et al. (2011). QOil20-6 was identified by both the CIM an ICIM methods in 2008HRB and had the same QTL peak, which were also identified by Diers et al. (1992), Sebolt et al. (2000), Tajuddin et al. (2003), Shibata et al. (2008), Qi et al. (2011), Specht et al. (2001), Wang et al. (2014a, b) in the classic population, this genomic region was also a very hot region. As the soybean genome of Williams 82 Glyma1.01 was sequenced completely and published by Schmutz et al. (2010), which make the identification of QTLs or genes much easier. Based on the release of sequence data and bioinformatics tools, 1489 candidate genes were screened from the main-effect QTLs, and eleven of them showed the relationship with seed oil content and/or fatty acid biosynthesis and metabolism based on the gene ontology and annotation information and their homologous or similar genes have been researched for the functional analysis or validation in the other plant. All of the gene annotation details and expression information are shown in Supplemental Table 3. To validate the effect of main effect QTLs, a wild soybean CSSLs population was used. We found seed oil content of CSSLs with main effect QTL and environment-specific QTL were range from 21 to 22.76% and 20.7 to 22.38%, respectively. Which were significant higher than the recurrent parent of CSSLs and the CSSLs without any QTL fragments. The consistency with validation population and previously research indicated that these QTL regions were “hot regions” that should be studied in further research on genes and marker-assisted selection (MAS) and the candidate genes could be the further research target.

The phenotype data correlation analysis was conducted between oil content and protein content, plant height, seeds length and width, seeds and pods weight per plant, seeds weight per plant, 100-seeds weight etc. agronomic traits of the RIL population for all environments, results showed there were significance negative relations between oil content and protein content in most of environments (Supplemental Table 4, except two environmental data showed minor positive relation), the result were identified with previously important research (Brim and Burton 1979; Brummer et al. 1997; Wilcox 1998; Cober and Voldeng 2000). Hwang et al. (2014) conducted a genome-wide association study of seed protein and oil content in soybean, six SNPs had a significant association with both protein and oil and showed negative relation between protein and oil. However, based on the genetic map and phenotype data under multiple environments of this RIL, main-effect QTLs of seed protein content (Qi et al. 2016) and oil content (this research) have been identified, we found the a main effect QTL located on 24.6–24.8 Mb of Gm20, the positive allele was from “Charleston”, which showed the positive relationship for oil and protein, furthermore, this region also inside in a hot region identified by Bolon et al. (2010), who using Soy GeneChip and transcriptome sequencing technology association analysis identified the candidate genes for soybean protein and oil content. Moreover, there were only six genes in this interval, but no evidence showed they involved in fatty acid biosynthesis or metabolize, it could be a further research point.

Epistasis effects and environmental factors make a substantial contribution to variation in complex traits, for quantitative trait, the phenotype of a given genotype cannot be predicted by the sum of its component single locus effects (Carlborg and Haley 2004). Previously studies showed the mechanism of epistasis of complex traits is much more complicated comparison with the individual gene effects and might decrease the individual QTLs effect (Zhao et al. 2008; Palomeque et al. 2010; Korir et al. 2011). The genetic architecture of a complex trait was influenced by the main gene and also can be largely attributed to epistasis and gene × environment interactions (Allard 1996), additive–additive epistatic effects of QTL pairs and random terms for environmental effects, additive-environmental interaction effects and additive–additive–environmental interaction effects were the key factor (Yang et al. 2005, 2007, 2008). Several recent studies have been conducted to demonstrate epistatis and QTL × environmental interaction effects (Ma et al. 2007; Xing et al. 2012; Wu et al. 2012; Liu et al. 2016). In this study, four QTL pairs were screened, and they could explain approximately 70% of the phenotypic variation, which is a high value because some pairs could explain around 20% of the phenotypic variation of seed oil content. These QTL pairs showed additive epistatic effects and interactions with the environment which detected under serious type-I errors below α = 0.001. Epistasis is common and can cause cryptic genetic variation in quantitative traits and can produce small additive effects and generate a loss of heritability and lack of replication (Mackay et al. 2009; Mackay 2014). The identification of epistatic interactions between genes and/or QTLs is a valuable starting point for a more thorough understanding of these genetic networks (Carlborg and Haley 2004). Future breeding strategies with to increase seed oil content should consider not only the main effect of QTLs, but also exploring epistasis effects more thoroughly and their interactions with the environments.