Genome-Wide Association Study and QTL Mapping Reveal Genomic Loci Associated with Fusarium Ear Rot Resistance in Tropical Maize Germplasm

Fusarium ear rot (FER) incited by Fusarium verticillioides is a major disease of maize that reduces grain quality globally. Host resistance is the most suitable strategy for managing the disease. We report the results of genome-wide association study (GWAS) to detect alleles associated with increased resistance to FER in a set of 818 tropical maize inbred lines evaluated in three environments. Association tests performed using 43,424 single-nucleotide polymorphic (SNPs) markers identified 45 SNPs and 15 haplotypes that were significantly associated with FER resistance. Each associated SNP locus had relatively small additive effects on disease resistance and accounted for 1–4% of trait variation. These SNPs and haplotypes were located within or adjacent to 38 candidate genes, 21 of which were candidate genes associated with plant tolerance to stresses, including disease resistance. Linkage mapping in four biparental populations to validate GWAS results identified 15 quantitative trait loci (QTL) associated with F. verticillioides resistance. Integration of GWAS and QTL to the maize physical map showed eight colocated loci on chromosomes 2, 3, 4, 5, 9, and 10. QTL on chromosomes 2 and 9 are new. These results reveal that FER resistance is a complex trait that is conditioned by multiple genes with minor effects. The value of selection on identified markers for improving FER resistance is limited; rather, selection to combine small effect resistance alleles combined with genomic selection for polygenic background for both the target and general adaptation traits might be fruitful for increasing FER resistance in maize.

Fusarium verticillioides can survive in soil, healthy seed, and plant residue, and infection of maize can be initiated from seedborne or airborne inoculum as well as systemic infection from the soil through roots to kernels (Morales-Rodríguez et al. 2007). Because of the high rate of maize production for subsistence in many developing countries, the solution to the problems of FER and fumonisin contamination is not to strengthen regulations, but rather to reduce fungal infection and mycotoxin levels in grain. The best strategy for controlling FER and reducing incidence of fumonisin contamination is the development and deployment of maize varieties with genetic resistance. Preharvest host resistance is economical to famers, leaves no harmful residue in food or the environment, and is compatible with other control measures. This strategy requires a clear understanding of the genetics of resistance, and the identification of alleles significantly contributing to reduced F. verticillioides infection and colonization, and fumonisin production (Mukanga et al. 2010).
Resistance to FER is quantitatively inherited and additive, dominant, and additive by dominant effects are important (Boling and Grogan 1965). Mapping studies using biparental populations have shown that resistance to FER is controlled by minor genes with relatively small effects that vary between environments and are not consistent between populations (Mesterházy et al. 2012). Robertson-Hoyt et al. (2006) and Bolduan et al. (2009) reported genotypic correlations between FER resistance and fumonisin accumulation of 0.87 in North Carolina and 0.92 in Germany, respectively, indicating that visual selection of FER resistance should be effective in simultaneously reducing fumonisin contamination. Although genetic variation for resistance to FER exists among maize inbred lines and hybrids, there is no evidence of complete resistance to either FER or fumonisin contamination in maize (Clements and Kleinschmidt 2003;Clements et al. 2004). The search for novel resistance genes against F. verticillioides is a very important activity in the quest to find a lasting solution to FER problems in maize production. Identification of specific allelic variants that confer improved resistance would permit maize breeders to select for recombinant chromosomes in backcross progeny that have desired target resistance allele sequences in coupling phase with the favorable elite polygenic background, facilitating the improvement of disease resistance without decreasing agronomic performance.
Several studies have identified quantitative trait loci (QTL) associated with resistance to F. verticillioides and subsequent reduced fumonisin accumulation (Robertson-Hoyt et al. 2006;Bolduan et al. 2009). For example, linkage-based mapping studies using F 2:3 populations derived from two resistant parents and a common susceptible parent identified nine and seven QTL associated with F. verticillioides resistance, and three of the QTL were common across the two populations (Pérez-Brito et al. 2001). In another study with two populations sharing a common resistant parent, a common QTL was detected on chromosome 4; this QTL was validated in an independent near isogenic line population (Li et al. 2011;Chen et al. 2012). Other QTL mapping studies have also revealed many QTL for F. verticillioides resistance that are stable across environments (Robertson-Hoyt et al. 2006;Ding et al. 2008). Using the GWAS method, seven SNPs were identified for FER resistance based on a diverse inbred line population comprised of 1687 maize inbred lines (Zila et al. 2013(Zila et al. , 2014. These studies revealed the presence of genetic variation for FER and the potential for identifying and deploying molecular markers for improving FER resistance in maize. GWAS has shown great potential for detecting QTL with high resolution in diverse germplasm (Buntjer et al. 2005). In Arabdopsis thaliana, GWAS was conducted using 213,497 SNPs and 473 accessions to reveal climate-sensitive quantitative trait loci (Li et al. 2010). In maize, GWAS has successfully been used to identify several casual genomic loci for different traits (Weng et al. 2011;Wang et al. 2012b;Liu et al. 2014;Samayoa et al. 2015). However, GWAS also has shortcomings, such as detection of false positives due to presence of population structure; fortunately, several advanced statistical methods have been developed to reduce the false positive rate (Andersen et al. 2005;Yu et al. 2006a;Larsson et al. 2013). Compared to traditional linkage-based analyses, association mapping offers higher mapping resolution while eliminating the time and cost associated with developing synthetic mapping populations (Flint-Garcia et al. 2005;Yu et al. 2006b). On the other hand, linkage mapping generates low rates of false positive results, which offset the limitation of so few alleles in offspring populations (Jiang and Zeng 1995;Ding et al. 2015b). Combining GWAS and linkage mapping could exploit the complementary strengths of both approaches to identify casual loci (Fulker et al. 1999;Pedergnana et al. 2014;Motte et al. 2014).
In this study, we used GWAS to identify genomic regions associated with FER resistance in tropical maize germplasm populations that were evaluated across three environments in Mexico. GWAS-identified genomic regions were validated through linkage mapping using four biparental populations. Furthermore, we identified a set of tropical maize inbred lines with high levels of FER resistance that can be used to improve FER in maize breeding programs.

MATERIALS AND METHODS
Germplasm materials and experimental design A collection of 940 elite tropical maize inbred lines was assembled from International Maize and Wheat Improvement Center (CIM-MYT) maize breeding programs located in Zimbabwe, Kenya, Colombia, and Mexico, and from the physiology, pathology, and entomology programs was evaluated for disease resistance (Semagn et al. 2012;Wen et al. 2011). One elite maize inbred line, CML155, was used as a resistant check. This line had previously been identified as highly resistant to FER following multiple years of visual evaluation under field conditions in CIMMYT's experimental station of Agua Fria (AF), Mexico. Four biparental-derived populations that included a doubled haploid (DH) population composed of 201 lines derived from crossing CML495 (resistant) to LA POSTA n Correlation below the diagonal is phenotypic correlation coefficient; correlation above the diagonal is genotypic correlation coefficient. Env, environments; SD, standard deviation; CV, coefficient of variation; s 2 G , genetic variance; s 2 GE , genotype-environment interactions variance. ÃÃ Significant at P = 0.01.
The GWAS panel of 940 inbred maize lines was screened for FER resistance in two locations: CIMMYT's experimental station of AF, located in the state of Puebla in Mexico [longitude 97°389W; latitude 20°289N; elevation 100-110 masl (meters above sea level)] in 2010 and 2011 (AF10 and AF11); and CIMMYT's experimental station of Tlaltizapan (TL) located in the state of Morelos, Mexico (longitude 99°79W; latitude 18°419N; elevation 940 masl) in 2011 (TL11). Entries were divided into four sets on the basis of maturity. Sets were randomized within the field and each set was blocked using an a-lattice design and replicated three times. Twenty seeds were planted in 2-m row plots, with 0.2 m between plants in a row and 0.75 m between rows. Two seeds were planted per hill and later thinned to a single plant to give a total of 10 plants per plot.

FER inoculations and evaluation
The experiments were artificially inoculated with a local toxigenic F. verticillioides isolate using the nail punch/sponge technique (Drepper and Renfro 1990), $7 d after flowering. A single-spore isolate of F. verticillioides was increased on sterile maize kernels, incubated for 14 d at 25°. After incubation, the spores were harvested, and concentration estimated using a hemocytometer and adjusted to 5 · 10 6 spores ml 21 in sterile distilled water with 0.2 ml/l Tween-20 surfactant (polyoxyethylene 20-sorbitan monolaurate). The primary ear of each plant in a plot was inoculated using a nail punch/sponge inoculation method with a suspension that contained 5 · 10 6 spores ml 21 about 7 d after flowering. The same inoculation method was used for both the GWAS panel and QTL mapping population.

Genotypic data
Total DNA was extracted from young leaves using the cetyltrimethylammonium bromide method (CIMMYT 2005), and DNA quality, purity and quantity for each sample was checked using gel-electrophoresis and spectrophotometer (NanoDrop ND8000, Thermo Scientific). A total of 854 maize inbred lines with good quality DNA were genotyped using an Illumina MaizeSNP50 BeadChip which contained 56,110 SNP markers (Ganal et al. 2011). The SNP genotyping was performed on an Illumina Infinium SNP genotyping platform at Cornell University Life Sciences Core Laboratories Center using the protocol developed by the Illumina Company. The genotypic data summary (allele frequency, heterozygous rate, and missing rate) were calculated by PLINK v1.07 software (Purcell et al. 2007).
The four biparental populations used for linkage mapping were genotyped by low density markers from the Kompetitive Allele Specific PCR (KASP) genotyping system of LGC Company (http://www. lgcgroup.com/) (Semagn et al. 2014). A total of 1250 SNPs were screened to identify markers polymorphic between the two parental lines. Of the polymorphic SNP markers, 200 were selected and used to genotype the entire population. Markers with allele frequency between 0.4 and 0.6 for both DH and F2:3 populations were included in the analysis.

Statistical analyses
Descriptive statistics (such as mean, range, skewness, and kurtosis) and correlations of phenotypic data were conducted in Excel 2010. Genetic correlation, and best linear unbiased estimates (BLUEs) were calculated using SAS (SAS Institute 2011) with multiple environments traits analysis package (META) which can be found on CIM-MYT Dataverse (http://hdl.handle.net/11529/10217) (Vargas et al. 2013). For the single environment BLUE, a mixed linear model was performed including line as a fixed effect, days to silking as a fixed linear covariate, and replication and block within replication as random effects. In the combined experimental analysis, each combination of location and year was considered an environment, with a mixed linear model including line as a fixed effect, days to silking (DTS) as a fixed linear covariate, and year, line · environment interaction, replication within environment, and block within replication as random effects.
n Table 2 Phenotypic (below the diagonal) and genetic (above the diagonal) correlation coefficient between FER resistance and agronomic traits where Pheno was phenotypic data; Env was environments, which was the combination of location and year; Rep was replication; Block was block in a-lattice design; Entry was the inbred lines used in this study.
Variance components were estimated using VarCorr function after fitting the linear mixed model (lmer) with the REML option in R software (R Core Team 2015). The single environment repeatability (H 2 ) was estimated using the following formula (Knapp et al. 1985): Broad-sense heritability (H 2 ) was estimated using the formula below (Knapp et al. 1985): where s 2 G is genetic variance, s 2 GE is genotype · environment interactions variance, s e 2 is error variance, E is the number of environments, and r is the number of replications in each environment.

Association analysis
A subset of 2000 SNP markers were randomly selected from 10,736 SNPs that remained after removing SNPs with missing values .10%; minor allele frequency of ,10%; and physical position interval ,50 kb. This subset of SNP markers was used for STRUCTURE analysis (Yu et al. 2009). The population structure was determined using an admixture model with correlated allele frequency in software STRUCTURE v2.3.3 (Pritchard et al. 2000). A burn-in of 10,000 iterations followed by 100,000 Monte Carlo Markov Chain replicates was conducted to test k values (number of subpopulations) in the range of 2-9. Each k was replicated four times, and most lines were assigned into clusters with a probability .0.6 (Falush et al. 2003).
Principal Component Analysis (PCA) was conducted in Eigensoft V3.0 software Patterson et al. 2006). Genetic distancebased neighbor-joining (NJ) analysis and a genetic kinship matrix were conducted using TASSEL V3 (Bradbury et al. 2007) and the tree visualized using FigTree v1.3.1 (Rambaut and Drummond 2009). Linkage disequilibrium (LD) measured as D9 was calculated using TASSEL software (Bradbury et al. 2007). Haplotype was built using the LD-based method as described by Gabriel et al. (2002), and SNPs are considered to be in the same haplotype or in "strong LD" if the one-side upper 95% confidence bound on D9 was .0.98 and the lower bound was above 0.7, and was calculated using PLINK v1.07 software (Purcell et al. 2007).
A mixed linear model that included BLUEs, marker, kinship matrix (K), and PCA was conducted using TASSEL software (Bradbury et al. 2007). Haplotype generated by PLINK and haplotype genotypes were used to conduct association mapping using the mixed linear model with PCA and Kinship in TASSEL software.

QTL mapping in biparental populations
Linkage maps were constructed using IciMapping v3.2 with Kosambi method for map distance calculation (Kosambi 1944;Wang et al. 2012a). The total map length for POP1 (DH population) was 1260 cM and included 166 SNPs and the average marker interval was 8.83 cM; the map length of POP2 was 991 cM and included 154 SNPs and the average marker interval was 8.93 cM. Linkage maps were not constructed for POP3 and POP4 as the number of retained markers was small (118 for POP3 and 93 for POP4). The Inclusive Composite Interval Mapping (ICIM) method in IciMapping v3.2 was used for QTL mapping Wang et al. 2012a). ICIM retains all the advantages of composite interval mapping (CIM) over interval mapping and avoids the possible increase of sampling variance and the complicated background marker selection process that are in CIM (Li et al. 2007. The step of ICIM was set to 1 cM, and the LOD threshold was set to 2.5. The total proportion of phenotypic variance explained by the detected QTL was calculated by fitting all significant SNPs simultaneously in a linear model to obtain R 2 adj . The proportion of the genotypic variance explained by all QTL was calculated as the ratio of pG = R 2 adj /h 2 (Gowda et al. 2015). The Single Marker Analysis (SMA) method in IciMapping V3.1 software was used for POP3 and POP4 QTL mapping, since the number of polymorphic markers were not enough for linkage map constriction. BioMercator V3.0 software (Arcade et al. 2004) was used to integrate significant markers to the maize physical map of the B73 reference genome (B73 RefGen_v1). The physical positions and sequence of SNP markers were obtained from the Illumina public ftp site (ftp://ussd-ftp.illumina. com/Whole%20Genome%20Genotyping%20Files/Archived_non-Human_Products/Maize_SNP50/).
Based on GWAS results, the sequences flanking SNP markers significantly associated with FER resistance were used to perform BLAST searches against the "B73" RefGen_v2 (MGSC) (http://blast. maizegdb.org/home.php?a=BLAST_UI) to obtain the physical position of significant SNPs.

Data availability
The original genotype and phenotype of the GWAS population are available in File S1 and File S2 and the original data of the four biparental populations are available in File S3 (POP1), File S4 (POP2), File S5 (POP3), and File S6 (POP4).

RESULTS
Phenotypic data analysis of GWAS panel Significant phenotypic variation for FER was observed in both the Agua Fria and Tlaltizapan experiments. The mean ear rot severity ranged from 0 to 87% with an overall mean of 22.96% in TL11, from 0 to 47% with an overall mean of 7.76% in AF11, and from 0 to 61% with an overall mean of 9.6% in AF10. Disease severity was higher in Tlatizapan than Agua Fria, possibly revealing differences in aggressiveness of F. verticilioides strains used. In the combined analysis, mean ear rot ranged from 0 to 74% with an overall mean of 16.03% (Table 1). The distribution of FER scoring in individual and combined environments was close to normal with a skew toward the lower level of infection ( Figure S1). Reflect kurtosis analysis revealed that ear rot resistance was continuously distributed, revealing the quantitative nature of F. verticillioides resistance ( Figure S1 and Table 1). Both genotypic components of variance (s 2 G ) and genotype · environment interaction (s 2 GE ) were significant (P , 0.01), from the combined ANOVA analysis, and s 2 G was also significant in the three single environments analysis. The repeatability (H 2 ) of FER scores was generally high, ranging from 0.89 in TL11 to 0.71 and 0.68 in AF11 and AF10, respectively. In combined analysis, the broadsense heritability (H 2 ) of the trials was 0.66, indicating that F. verticillioides resistance was controlled by genetic factors and that the data could confidently be used for accurate mapping of F. verticillioides resistance genes.
Genetic and phenotypic correlation between ear rot aspect, ear rot score, and PIA were significant, ranging from r = 0.90 to 0.98 (Table S2). Therefore, subsequent data and GWAS analyses were conducted using PIA as a FER parameter. Genetic and phenotypic correlations between environments were highly significant, and the phenotypic correlation between combined mean and mean of the three single environments was significant (Table 1). Low but significant correlations were observed between FER (PIA) and DTS (Table 2). However, a moderate genetic correlation (r = 0.38) was observed between FER resistance and stem lodging. This is expected as F. verticillioides can grow within the maize plant as an endophyte, and can become pathogenic and incite stalk rot when conditions become stressful to the plant.
Response of 940 maize inbred lines to FER revealed several lines that consistently had mean disease severity scores ,5% across the three environments. Analysis of combined phenotypic data from the different environments identified 63 maize inbred lines that were highly (PIA ,5%) resistant to FER (Table S3). These tropical inbred lines can immediately be used as a source of FER resistance in breeding programs.
Phenotypic data analysis of QTL mapping populations Significant phenotypic variation for FER was observed for the four biparental populations (Table S5). For all populations, genotypic components of variance (s 2 G ) were significant (P , 0.01) from the single environment ANOVA analysis. For combined ANOVA, both genotypic components of variance (s 2 G ) and genotype by environment interaction (s 2 GE ) were significant (P , 0.01) for POP1 and POP2, revealing that F. verticillioides populations in the two environments might have been different. In combined analysis, the broad-sense heritability (H 2 ) of the trials was 0.74 for POP2 and 0.52 for POP1. The repeatability was generally high for each single environment, for example, the repeatability of POP1 in the TL12A environment was 0.73 and in AF12A was 0.69. Those results indicate the data could confidently be used for QTL mapping.
Genotypic characterization of GWAS panel A total of 56,110 SNPs were generated for 854 maize inbred lines using the Illumina maize SNP50 BeadChip. The number of SNP markers per chromosome ranged from 3965 SNPs on chromosome 10 to 8625 SNPs on chromosome 1 (Figure 1). The average SNP missing value was 7.0% and 2112 SNPs (3.76%) had a missing value .40%. Of the 56,110 SNPs, 14.6% had a MAF (minor allelic frequency) ,0.05, while 55.8% had a MAF .20%. Most of the markers (96.3%) had a heterozygous rate ,2.5%, and only 0.01% had a heterozygosity .40%. After eliminating SNP markers with a missing value .40% and MAF ,5%, a total of 43,424 SNPs were retained for GWAS.
From the 940 maize inbred lines evaluated against FER, 818 lines were included in GWAS analysis, after removing lines with .20% heterozygosity and those with .20% missing SNP markers ( Figure  S2). Population structure estimated using 2000 random SNPs and the software STRUCTURE v2.3.3 divided the inbred lines into three subgroups ( Figure 2). Using k = 3, 97.3% of the maize inbred lines were assigned to three groups, and only 6.8% of the lines were assigned into mixed population (Figure 2). The largest subgroup (blue color in Figure  2 of the K = 3) was composed on germplasm coming from different breeding programs of CIMMYT, including the lowland breeding program, physiology, pathology, and programs in Africa. Most inbred lines in the second subgroup (red color in Figure 2 of K = 3) comprised of germplasm derived from CIMMYT's drought tolerant population LaPostaSeq. The third subgroup (olive green color in Figure 2 of K = 3) contained germplasm mainly from CIMMYT's lowland breeding program. NJ tree constructed using 43,424 SNP markers and 818 maize inbred lines clustered the lines into three major groups ( Figure S3), and the grouping was confirmed using PCA analysis ( Figure S4). The dK result from STRUCTURE analysis and the absolute difference of eigenvalue between PCs indicted there were three major subgroups in the GWAS panel (Figure 2 and Figure S5). The results obtained following STRUCTUE, PCA, and NJ tree cluster analyses were consistent; therefore, the first three PCAs were used as a covariate in the mixed linear model in GWAS analysis. n Association mapping for FER resistance GWAS analysis using combined phenotypic data identified 45 SNPs that were significantly associated with FER resistance with P value ,10 23 ( Figure 3A). The markers were distributed on all chromosomes except chromosome 7; and the number of SNPs per chromosome ranged from 1 on chromosome 8 to 14 on chromosome 10. The most significant SNP was located on chromosome 10 (PZE-110022154) with the lowest P value (P , 5 · 10 25 ) and it explained 2.06% of the phenotypic variation. The second SNP with lowest P value was located on chromosome 5 and it also explained 2.06% of the phenotypic variation. Detailed information of 45 SNPs significantly associated with FER resistance is provided in Table 3. Genome-wide Manhattan plots for single environment analysis are attached ( Figure S6). Quantile-quantile plots showed that population structure was controlled well by the mixed linear model ( Figure S7).
Haplotype built based on LD as described by Gabriel et al. (2002) resulted in 7063 haplotypes (Table S4). The maximum number of markers per haplotype was 19, the minimum 2, and the average number was 2.96 SNPs per haplotype. Haplotype-based GWAS in a mixed linear model identified 15 haplotypes that were significantly associated with FER resistance and these were distributed in bin 2.05, 5.03/5.04, 7.02, 8.03/8.04, and 10.03 ( Figure 3B). Haplotype analysis increased the power of marker detection; for example, haplotype 5076 on chromosome/bin 7.02 that was significantly associated with FER resistance (P value = 4.45 · 10 27 ) was not detected in single marker GWAS analysis (Table  4). However, some markers were detected by both single marker and haplotype-based GWAS analysis. Haplotype 4168 on chromosome 5 accounted for 3.1% of variation for FER and the two markers PZE-105116484 and PZE-105116502 associated with this haplotype explained 2.1 and 1.8% of phenotypic variation for FER, respectively (Table 4).

QTL mapping of FER resistance
Five QTL were detected in the DH population (POP1); two on chromosome 1 and one each on chromosomes 2, 3, and 5 (Table 5). The QTL on chromosome 2 accounted for 15.41% of the total phenotypic variation observed for FER in this population; while the QTL on chromosome 5 explained 13.56% of the phenotypic variation (Table 5). Combined, the five QTL detected in POP1 explained 49% of the total phenotypic variance observed for FER. For POP2, an F 2:3 population, six QTL were detected that together accounted for 25% of the observed phenotypic variation. The QTL on chromosome 1 accounted for 11.36% of the phenotypic variation for FER resistance. The QTL in bin 4.03/04 explained 9.27% of the phenotypic variance, while that in bin 10.03 accounted for 7.82% of the phenotypic variance (Table 5). SMA was used for QTL mapping for POP3 and POP4 as few polymorphic markers were detected in these populations. For POP3, six markers were significantly associated with FER resistance and these were distributed in three regions of chromosome 5; bins 5.03, 5.04, and 5.05 (Table 5). The phenotypic variation for FER explained by these markers ranged from 4.56 to 6.73%, revealing that these were minor QTL. The SNP in bin 5.04 had the greatest effect, explaining 6.73% of the observed phenotypic variance for FER. For POP4, four markers were associated with FER resistance and these were in bin 2.04, 2.06, and 2.07 (Table 5). The phenotypic variation explained by these markers ranged from 12.56 to 15.84% and the SNP in bin 2.07 had the largest effect, explaining 15.84% of the phenotypic variance for FER resistance (Table 5).
Forty-five single SNP markers and 15 haplotypes identified through GWAS, together with 15 QTL identified through linkage mapping were integrated onto a maize physical map using the software BioMereator V3.0 (Sosnowski et al. 2012). The map generated by the software was convenient for visualizing QTL and significant SNPs together. Eight common loci were identified on six chromosomes; on chromosome/bin 2.04, 3.06, 4.04, 4.08, 5.03, 5.04, 9.01, and 10.03 (Figure 4). The QTL on chromosome 2 in bin 2.04 was detected in two biparental populations as well as single marker GWAS. The chromosome 5 (bin 5.04) locus was detected in one biparental population and by both single marker and haplotype GWAS. The locus on chromosome 10 bin 10.3 contained 14 significant SNP markers, one haplotype and one QTL.

Resistance donor
Developing host resistance is the preferred strategy for managing FER, especially for smallholder farmers across the tropics, who largely produce maize for their own consumption, and often lack resources to adopt other control strategies. However, effective use of this strategy requires identification of sources of resistance that are stable and effective across environments. We evaluated 940 maize inbred lines in three environments and identified 63 inbred lines that were highly resistant to F. verticillioides. These sources of FER resistance complement a few that have been reported in tropical germplasm (Pérez-Brito et al. 2001;Small et al. 2012). The broad-sense heritability (H 2 = 0.66) was high, revealing that FER resistance was genetically controlled, thus, significant improvements for FER resistance can be achieved through breeding. Furthermore, the 63 inbred lines resistant to FER constitute a valuable tool for understanding the genetic basis and architecture of FER resistance in tropical maize germplasm. These lines should be evaluated in multiple environments to confirm stability of FER resistance.

QTL for FER resistance
Forty-five SNPs and 15 haplotypes associated with FER resistance were identified through single marker and haplotype-based GWAS and 15 QTL were identified through linkage mapping in four biparental populations. Using the software BioMereator V3.0, eight loci, containing significant markers from GWAS and linkage mapping were identified ( Figure 4). Six loci on chromosomes/bin 3. 06, 4.04, 4.08, 5.03, 5.04, and 10.03 are in regions that have previously been reported (Chen et al. 2012;Ding et al. 2008;Li et al. 2011;Pérez-Brito et al. 2001;Robertson-Hoyt et al. 2006;Zhang et al. 2007), while two loci, on chromosomes/ bin 2.04 and 9.01 are new loci, identified in this study. Two of the loci on chromosomes 4.04 and 9.01 are in regions containing genes encoding putative proteins of unknown function, while six loci are in regions that have been associated with stress tolerance, including FER resis-tance. Results from this study concur with previous reports (Boling and Grogan 1965;Zila et al. 2014) that FER resistance is a complex trait conditioned by multiple genes with minor effects. The loci on chromosome 5.04 contained two significant SNPs, one haplotype and a QTL detected through linkage mapping. This chromosome region has previously been reported in three independent QTL mapping studies (Pérez-Brito et al. 2001;Robertson-Hoyt et al. 2006;Ding et al. 2008). Candidate gene analysis revealed that this QTL was in a region containing a putative protein encoding a glucose/ribitol dehydrogenase protein that catalyzes the oxidation of D-glucose to D-b-gluconolactone using NAD or NADP as a coenzyme in the cell development. This gene belongs to a subset of short-chain dehydrogenase and reductase family of genes which are involved in different biochemical processes including pathogen toxin reduction (Meeley et al. 1992;Moummou et al. 2012).
The locus on the long arm of chromosome 4 (bin 4.08), detected through both GWAS and linkage mapping, has previously been reported (Li et al. 2011;Chen et al. 2012). Markers within this locus localized to a putative protein of unknown function from maize. However, blastp analysis revealed that it had high homology to Arabidopsis 2OG-Fe (II) oxide reductase, a gene that is involved in regulating giberellic acid and abscisic acid biosynthesis, which are involved in plant tolerance to stress, including disease resistance (van Damme et al. 2008;Han and Zhu 2011). Furthermore, chromosome 4.08 is a hot spot region for disease resistance in maize and has been found to harbor resistance QTL to eight maize diseases (Wisser et al. 2006). This would be a good target for developing markers to simultaneously introgress multiple disease resistance genes.
The chromosome 10.03 locus containing 13 SNPs and one QTL is located in a region conditioning resistance to multiple maize disease, including rp1 and rp5 that confers resistance to common rust (Wisser et al. 2006). The candidate with the lowest P value in this region encoded an MADS-box transcription factor (Parenicová et al. 2003). MADS-box family genes are involved in controlling major aspects of plant n development, including embryo and seed development (Gramzow and Theissen 2010), and may increase seed vigor and subsequently increase tolerance to diseases. Other important resistance loci identified in this study included chromosomes 3.06 and 5.03. These two loci have previously been reported associated with resistance to FER in two QTL mapping studies (Robertson-Hoyt et al. 2006;Ding et al. 2008). The locus on bin 3.06 encoded a dense granule Gra7 protein and the bin 5.03 locus was a putative protein of an unknown function. In addition, two new loci were identified, on chromosome 2.04 and 9.01. The chromosome 2.04 locus was associated with a plant peroxidase gene that is involved in cell wall fortification (Kolattukudy et al. 1992). The chromosome 9.01 locus encoded a protein of unknown function. Although many SNPs localized to genic regions, the currently limited understanding of pathways contributing to FER resistance restricts our ability to precisely predict what genes might be involved in resistance to this complex disease. However, information from this study provides a basis for further research into elucidating the genetic architecture and pathways leading to FER resistance in maize.

Haplotype-based GWAS analysis
Because of the rapid LD between markers, haplotype analysis may provide more detection power compared to single marker GWAS and is more practical for breeding (Yan et al. 2011). Four methods are Figure 4 Visualization of all loci associated with FER resistance that were detected in this study using the software BioMereator V3.0. The black ovals represent the location of the eight loci detected by both GWAS and linkage mapping. The numbers on the right of the chromosome indicate the physical position of the chromosome with million base pair as unit.
commonly used to build haplotypes (Gabriel et al. 2002;Yan et al. 2009;Gore et al. 2009;Lu et al. 2010Lu et al. , 2011Ding et al. 2015a): (1) use of a fixed number of markers as a window to slide across the chromosome to build the haplotype; (2) use of a fixed physical distance interval of 10 kb in maize to build the haplotype; (3) use of gene-based physical position to build the haplotype; and (4) use of LD information to put high LD markers together to constitute a haplotype. The marker density in our study was medium to high so we chose the LD-based haplotype build method. Using this approach, haplotype GWAS detected some resistance loci that were not detected by single marker GWAS, whereas the single marker result was reflected in haplotype-based GWAS. This indicates that haplotype-based GWAS has a high marker detection efficiency but requires high density markers to build a haplotype. On-going genotyping by sequencing projects will furnish enough marker density to exploit the advantages of haplotype-based GWAS.
Candidate genes colocalized with associated SNPs SNPs and haplotypes associated with FER resistance were located within or adjacent to 38 putative candidate genes which were obtained from the MaizeGDB (http://www.maizegdb.org/) genome browser based on physical position of significant SNPs, MaizeCyc database version 2.0 (http://maizecyc.maizegdb.org/). The Phytozome database (http:// phytozome.jgi.doe.gov/pz/portal.html) that was used for defining relevant pathways and annotating possible functions of candidate genes (Caspi et al. 2010) could annotate functions to 21 out of the 38 candidate genes (Table 3 and Table 4). Thirteen of the 45 SNPs localized to intergenic regions, 10 were inside exons, nine were located in introns, and nine were located in promoters; five localized to the 39 untranslated region and five to the 59 untranslated region (Table 3). The most significant SNP on chromosome/bin 5.04 was in a region associated with a gene encoding a glucose/ribitol dehydrogenase, a protein that catalyzes the oxidation of D-glucose to D-b-gluconolactone using NAD or NADP as a coenzyme. This gene family is a subset of short-chain dehydrogenases and reductases, involved in pathogen toxin reduction (Meeley et al. 1992;Moummou et al. 2012). These results reveal the complex nature of FER resistance in tropical maize, and indicate that various mechanisms might be involved in conditioning FER resistance, including complex biosynthesis processes, which also might include interactions between multiple metabolic pathways.

Conclusion
This study identified a set of inbred lines that can potentially be used as sources of resistance to develop hybrids with resistance to FER. Further validation of the potential sources of resistance in multiple environments is required, but the small number of inbred lines makes this process cost-effective. Eight loci harboring FER QTL were identified through integrating GWAS and linkage mapping results. Two are new loci while six colocalized to loci that have previously been described (Chen et al. 2012;Ding et al. 2008;Li et al. 2011;Pérez-Brito et al. 2001;Robertson-Hoyt et al. 2006;Zhang et al. 2007). Some SNPs associated with these loci localized to within or close to genes with known function. Candidate gene analyses for significant SNPs provided targets for further research to elucidate mechanisms of FER resistance. Our results confirmed earlier reports that many genes are involved in FER resistance.