Fire blight QTL analysis in a multi-family apple population identifies a reduced-susceptibility allele in ‘Honeycrisp’

Breeding apple cultivars with resistance offers a potential solution to fire blight, a damaging bacterial disease caused by Erwinia amylovora. Most resistance alleles at quantitative trait loci (QTLs) were previously characterized in diverse Malus germplasm with poor fruit quality, which reduces breeding utility. This study utilized a pedigree-based QTL analysis approach to elucidate the genetic basis of resistance/susceptibility to fire blight from multiple genetic sources in germplasm relevant to U.S. apple breeding programs. Twenty-seven important breeding parents (IBPs) were represented by 314 offspring from 32 full-sib families, with ‘Honeycrisp’ being the most highly represented IBP. Analyzing resistance/susceptibility data from a two-year replicated field inoculation study and previously curated genome-wide single nucleotide polymorphism data, QTLs were consistently mapped on chromosomes (Chrs.) 6, 7, and 15. These QTLs together explained ~28% of phenotypic variation. The Chr. 6 and Chr. 15 QTLs colocalized with previously reported QTLs, while the Chr. 7 QTL is possibly novel. ‘Honeycrisp’ inherited a rare reduced-susceptibility allele at the Chr. 6 QTL from its grandparent ‘Frostbite’. The highly resistant IBP ‘Enterprise’ had at least one putative reduced-susceptibility allele at all three QTLs. In general, lower susceptibility was observed for individuals with higher numbers of reduced-susceptibility alleles across QTLs. This study highlighted QTL mapping and allele characterization of resistance/susceptibility to fire blight in complex pedigree-connected apple breeding germplasm. Knowledge gained will enable more informed parental selection and development of trait-predictive DNA tests for pyramiding favorable alleles and selection of superior apple cultivars with resistance to fire blight.


Introduction
The development of apple (Malus domestica Borkh.) cultivars with genetic resistance offer a potentially sustainable solution to fire blight, a damaging bacterial disease caused by Erwinia amylovora. This pathogen can cause severe structural tree damage and death, resulting in substantial economic losses 1 . Apple production systems have become more vulnerable to fire blight due to widespread planting of susceptible apple cultivars, adoption of high-density planting systems, and lack of control methods effective against all disease phases 1 . Resistance to fire blight is an important goal of many apple breeding programs 2,3 .
Phenotypic evaluation of the degree of resistance or susceptibility to fire blight can be difficult due to quantitative host resistance, differential host × E. amylovora strain interactions, strong environmental influences on disease incidence, and effects of tree vigor on susceptibility [4][5][6] . Segregation for resistance/susceptibility has been observed in susceptible × susceptible families 7 . As a result, phenotypic information alone is not predictive of performance and thus is inadequate to guide breeding decisions targeting resistance to fire blight.
Current understanding of the genetic basis of resistance/ susceptibility to fire blight of germplasm relevant to U.S. apple breeding programs is insufficient for the immediate development of DNA tests. Relevant sources and effects of QTL alleles need to be determined prior to the development and deployment of trait-predictive DNA tests 25 . Pedigree-based QTL analysis (PBA) 26,27 enables the integrated analysis of multiple small pedigree-connected fullsib families. As populations of such germplasm are typical of populations routinely generated and evaluated in breeding programs, detected QTLs are simultaneously validated for breeding relevance 27 . PBA has several further advantages compared to biparental QTL studies, including increased statistical power, representation of wider genetic diversity, and the ability to determine ancestral sources of QTL alleles 27,28 . PBA has been successfully applied in apple for numerous traits 26,[29][30][31][32][33][34] , including resistance/ susceptibility to fire blight 20 . Key experimental resources are available to characterize genetic factors that influence complex traits in apple, including representative pedigreeconnected germplasm sets 28 , high-quality genome-wide single nucleotide polymorphism (SNP) data 35,36 , and FlexQTL TM software (www.flexqtl.nl) 26,37-39 . This study aimed to elucidate the genetic basis of resistance/susceptibility to fire blight in a large set of pedigree-connected germplasm relevant to U.S. apple breeding programs. The objectives were to (1) detect QTLs associated with resistance/susceptibility to fire blight; (2) characterize effects of alleles at stable QTLs represented by multi-SNP haplotypes; and (3) determine frequencies of putative reduced-and increasedsusceptibility alleles among apple cultivars, important breeding parents (IBPs), and progenitors. It was hypothesized that multiple QTLs associated with resistance/ susceptibility to fire blight would be detected.

QTL detection
In both years, QTLs were detected on Chrs. 6, 7, and 15 with positive (Bayes Factor; BF; 2lnBF 10 > 2) to strong (BF > 5) evidence (Table 1 and Supplementary Table S1; Supplementary Figs. S1, S2, S3, and S4). QTL modes varied by less than 5 cM across years, with ranges of 49-53 cM, 24-28 cM, and 68-72 cM for the Chr. 6, 7, and 15 QTLs, respectively (Table 1 and Supplementary Table  S1). There was positive evidence for a second QTL on Chr. 7 in 2016 (located at 52-66 cM, 28.98-33.23 Mbp) but not in 2017. In both years, there was strong to decisive (BF > 10) evidence for at least one QTL on Chr. 8 and  (Table 1 and  Supplementary Table S1). The two possible Chr. 8 QTL intervals in 2017 were 22-40 cM (9.4-22.0 Mbp) and 52-64 cM (27.5-30.9 Mbp). In 2016, there was strong evidence for a QTL on Chr. 16 but no evidence in 2017 (Table 1). There was no evidence for any QTLs in the rest of the genome in either year (data not shown). Use of adjusted SLB BLUPs estimated across years as phenotypic values did not provide further insights as results were similar to results of QTL analyses with within-year adjusted SLB BLUPs as phenotypic values (results not presented). A dominance model did not provide further insights as there was no evidence for dominance effects or additional QTLs (data not presented).  Table 1) and contained 423 functionally annotated genes, including 74 (18%) genes putatively involved in responses to diseases and biotic stresses (Supplementary Table S2). Of the 369 annotated genes in the common 11.7-Mbp QTL interval on Chr. 7, annotations for 81 genes (22%) indicated involvement in responses to diseases and biotic stresses (  Table S2).
The reduced-susceptibility haplotype of 'Honeycrisp', 6E, was traced to 'Frostbite' (Fig. 2   Honeycrisp's pedigree was repeated as an inset figure to clearly depict inheritance of reduced-susceptibility haplotype 6E and increased-susceptibility haplotype 6R. For a given individual, haplotype listing order was determined based on phasing information. Unknown parents are designated by UP_xxx. "-" refers to other progenitor haplotypes not segregating in full-sib families or unresolved haplotypes. Colorized haplotypes were those with significant effects (p < 0.05) in both 2016 and 2017. For better visibility of complex lineages, uneven vertical scaling between pedigree rows was used inherited 7P (n = 8) typically had low-susceptibility responses with an adjusted SLB BLUP mean of 0.19 across years (Fig. 1 and Supplementary Tables S7, S8). Offspring that inherited 7B (n = 38) or 7T (n = 7) demonstrated moderate to high susceptibility responses, with adjusted SLB BLUP means across years of 0.60 and 0.66, respectively (Fig. 1).

Discussion
QTLs on Chrs. 6, 7, and 15 were consistently detected and characterized across pedigree-connected breeding families derived from 27 IBPs. A rare reducedsusceptibility allele at the Chr. 6 QTL was identified in 'Honeycrisp' 40 , an IBP and economically important U.S. apple cultivar. The new genetic information obtained of estimated haplotype effects has potential utility in targeting reduced susceptibility to fire blight in apple breeding.

Bi-allelic QTL model fit
FlexQTL TM software applies a bi-allelic QTL model that assumes the presence of two functional alleles (Q and q) at a QTL, where Q is associated with high (e.g., high susceptibility) and q with low (e.g., low susceptibility) phenotypic values. In this study, Q/q allele assignments by FlexQTL TM software for several IBPs were largely inconclusive likely due to the presence of multiple functional alleles with differing quantitative effects (Supplementary Table S3). A similar conclusion was made by Verma et al. 34 ; two QTLs associated with fruit acidity in a pedigree-connected apple germplasm set were detected. Verma et al. 34 reported that the presence of a third, intermediate-effect allele might have hindered accurate estimation of parental QTL genotypes at detected QTLs. The presence of multiple functional alleles in this study could explain the detection of other inconsistent QTLs (e.g., on Chr. 16). Under-representation of alleles might have also made it challenging for FlexQTL™ software to estimate some parental QTL genotypes. Lower observed fire blight severity levels in 2017 compared to 2016 in the dataset used here 7,23 could have contributed to inconsistencies in Q/q allele assignments across years due to less distinct phenotypic contrasts between functional allele groups. Higher numbers of reduced-susceptibility haplotypes across QTLs associated with lower susceptibility to fire blight. Distributions, means, and 95% confidence intervals of adjusted shoot length blight best linear unbiased predictions (SLB BLUPs) across years for 314 apple offspring with increased-susceptibility haplotypes present and absent grouped by a number of reduced-susceptibility haplotypes at a combination of three QTLs on chromosomes 6, 7, and 15. Offspring with haplotype 6R, 7B, 7T, and/or 15A were considered to have increased-susceptibility haplotypes present; reduced-susceptibility haplotypes were 6E, 6N, 6O, 7P, 15J, and 15L (corresponding with designations in Fig. 1). Offspring with neutral-effect haplotypes (i.e., haplotypes exhibiting non-significant effects in 2016 and/or 2017) at all three QTLs were considered to have no increased-and reduced-susceptibility haplotypes (gray bar). Different mean separation letters represent the least significant differences with a Bonferroni p adjustment (α < 0.05) Resistance/susceptibility alleles were defined via analyses of SNP haplotypes within QTL intervals. Haplotypes that did not have consistently significant effects could be neutral-effect alleles or were under-represented. Because the approach used to determine effects of individual haplotypes (within-year one-way ANOVAs) examined haplotypes individually, small sample sizes and/or effects of haplotypes at other QTLs might have led to coincidental associations; however, it provided a framework for identifying haplotypes of interest.

Identities of chromosomes 6, 7, and 15 QTLs
The Chr. 6 10 , and Khan et al. 16 also Colored haplotypes had significant effects (p < 0.05) among offspring in the pedigree-connected reference germplasm set in 2016 and 2017. Green colored haplotypes are putatively associated with reduced susceptibility whereas gold-colored haplotypes are putatively associated with increased susceptibility. Haplotypes denoted by italics were imputed a Important breeding parent (IBP) or progenitor represented in the pedigree-connected germplasm set used for QTL analyses (Fig. 2 This is the first report of a QTL associated with the degree of susceptibility to fire blight in 'Honeycrisp'derived families. FlexQTL TM software's estimate that 'Honeycrisp' was heterozygous for the Chr. 6 QTL supported the haplotype analysis findings. Found exclusively in 'Honeycrisp' and its progenitors, haplotype 6E was significantly associated with reduced susceptibility albeit of moderate effect compared with 7P and 15L ( Fig. 1 and Supplementary Table S5). Haplotype 6E was traced to 'Frostbite', a cultivar previously classified as highly resistant (HR; Fig. 2) 23 . To our knowledge, haplotype 6E is a rare reduced-susceptibility allele and appears to be unique to germplasm derived from 'Frostbite' (Fig. 2). Previously reported MS to HR levels of cultivars Frostbite, Honeycrisp, Keepsake, and WA 38 might be partially explained by the presence of 6E (Table 2) 23 . In addition to inheriting 6E from 'Honeycrisp', 'WA 38' 41 inherited reduced-susceptibility haplotype 6N from 'Enterprise', which could explain 'WA 38's lower susceptibility classification compared to 'Honeycrisp' 23 . The presence of increased-susceptibility haplotype 6R, which segregated in families derived from 'Honeycrisp', 'Minnewashta', and 'Sansa', might partially explain the previously reported HS to MS levels of several cultivars ( Table 2). The extended haplotype sharing across the Chr. 6 QTL between 'Honeycrisp' and 'Minnewashta' indicated that haplotype 6R in these two cultivars possibly originated from a common unknown ancestor. Selection for haplotype 6E and against 6R might be a useful approach to developing breeding populations with lower susceptibility to fire blight.
Novel putative fire blight QTL on chromosome 7 Significant Chr. 7 haplotypes (7B, 7P, 7T) had relatively low representation or their effects might have been confounded by haplotypes at other QTLs. Detection of the Chr. 7 QTL was likely due to segregation of the increased-susceptibility haplotype 7B in a 'Pinova' × NY 03 full-sib family with 22 offspring. Because reduced-susceptibility haplotype 7P was only present with reduced-susceptibility haplotypes at other QTLs (Supplementary Table S11), the association between haplotype 7P and reduced susceptibility might have been coincidental. Increasedsusceptibility haplotype 7T should be considered putative because it was only represented by seven offspring. Conclusions about the effects of reduced-susceptibility haplotype 7P and increased-susceptibility haplotypes 7B and 7T were limited by small family sizes, variability between years for some haplotypes, and putative effects of haplotypes at other QTLs. Therefore, the Chr. 7 QTL should be considered putative and needs to be validated in future studies.

Increased-susceptibility haplotype at chromosome 15 QTL prevalent among IBPs
Increased-susceptibility haplotype 15A was prevalent among IBPs and cultivars, which was likely due to the presence of 'Golden Delicious' in the pedigrees of many cultivars (Fig. 2). For example, haplotype 15A in 14 of 19 IBPs was inherited from 'Grimes Golden' through 'Golden Delicious' (Fig. 2 and Supplementary Table S9). The presence of haplotype 15A might partially explain HS to MS classifications of several cultivars (Table 2) 23 . The prevalence of haplotype 15A among IBPs indicates that it will likely segregate among offspring in many breeding families. Future selection against haplotype 15A may eliminate individuals with a higher potential to be susceptible.

Highly resistant cultivar 'Enterprise' had four reducedsusceptibility haplotypes
The identification of four reduced-susceptibility haplotypes in 'Enterprise' was not surprising as 'Enterprise' has previously demonstrated strong levels of resistance to fire blight 20,23 . Because FlexQTL TM software estimated that 'Enterprise' was homozygous qq at the three QTLs, it was expected that all Chrs. 6, 7, and 15 QTL haplotypes inherited from 'Enterprise' would be significantly associated with reduced susceptibility. However, SNP haplotype analyses within the QTL intervals revealed that four (6N, 6O, 7P, 15L) of Enterprise's six haplotypes had significant effects, which indicated that 'Enterprise' was homozygous (qq) for the Chr. 6 QTL and heterozygous (Qq) for the Chr. 7 and Chr. 15 QTLs. Low representation (n = 19 offspring) of 'Enterprise's haplotypes might have hampered accurate estimation of haplotype effects. Therefore, reduced-susceptibility haplotypes traced to 'Enterprise' should be considered putative.

Interactions at and among detected QTLs were not purely additive
Conclusions regarding combined QTL effects were limited by lack or under-representation of some genotype classes at and across QTLs. Under-representation of genotype classes is a common limitation of multi-locus studies in pedigree-connected germplasm. The approach used to examine combined QTL effects was like the Qallele dosage model used by Verma et al. 34 . Although conclusions about the effects of specific compound QTL genotypes were limited, this approach increased statistical power by eliminating the need to examine specific allelic combinations separately.
Non-additive interactions among fire blight QTLs have been previously reported 20 and were observed in this study. Under a purely additive model, adjusted SLB BLUP means were expected to be significantly lower with each additional reduced-susceptibility haplotype regardless of whether increased-susceptibility haplotypes were present or absent. Because additional reduced-susceptibility haplotypes did not always correspond to significantly lower susceptibility (Fig. 3), there might have been dominance effects at a QTL and/or epistatic interactions among QTLs. However, empirical examination of dominance effects and/ or epistatic interactions was not possible in this study.

Reduced-susceptibility haplotypes were present in several susceptible cultivars
The presence of putative reduced-susceptibility alleles in susceptible cultivars was not surprising and suggested epistatic interactions among multiple QTLs. Resistance alleles have been previously identified in susceptible individuals. For example, 'Enterprise's resistance allele at a major Chr. 7 QTL was traced by van de Weg et al. 20 to its MS progenitor, 'Cox's Orange Pippin'. In this study, multiple HS to MS cultivars (e.g., Gala, Hudson, Jonathan, Melrose, Northern Spy, Yellow Newtown) had reducedsusceptibility alleles, which indicated that parental susceptibility classifications might not be predictive of offspring susceptibility levels due to segregation at multiple additives and/or epistatic QTLs. The development of trait-predictive DNA tests for detected QTLs might enable a more informed selection of parents for generating breeding populations with low susceptibility to fire blight.

Functional annotations
Examination of functional annotations provided a list of annotated genes within the Chrs. 6, 7, and 15 QTL intervals, which could be used as a preliminary framework for a candidate gene approach. Further studies would be needed to identify and characterize the causal genes underlying detected QTL intervals.

Detection and characterization of chromosome 8 QTL(s) were hampered
Multiple Chr. 8 QTLs might underlie variation in resistance/susceptibility to fire blight in apple, but the characterization of Chr. 8 QTL(s) detected in this study was limited. Because the location of Chr. 8 QTL(s) could not be precisely determined, Chr. 8 was not targeted for further analyses. Van de Weg et al. 20 Table 1) overlapped with the putative epistatic QTL detected by van de Weg et al. 20 . Detection and characterization of Chr. 8 QTL interval(s) could have been hampered by limited families examined, relatively small family sizes, and/or phenotyping method utilized.

Study limitations
A limitation of this study was reliance on phenotypic data from inoculation with a single E. amylovora strain 7 . E. amylovora strains vary for virulence, which can result in variable fire blight incidence and severity 5,8,42,43 and complicate elucidation of the genetic basis of the degree of susceptibility to fire blight. However, this study provided an experimental framework that can be replicated with different E. amylovora strains.
The large residual variation in the phenotypic dataset leveraged in this study was observed and managed 7 . Because several environmental, host, and pathogen factors impact fire blight incidence and severity 4-6 , large residual variation is common in fire blight field experiments. In this study's germplasm set, Kostick et al. 7 reported that 42-49% of the variation for SLB values was associated with the residual variation. The large experimental variation was managed by using adjusted SLB BLUPs as phenotypic values in QTL analyses.
Small family sizes likely limited the detection of QTLs, characterization of haplotypes, and examination of QTL × QTL interactions. Small family sizes combined with the uneven representation of some IBPs' genomes, referred to as "skewed average allelic representation" 28 , led to under-representation or no representation of some haplotypes and compound QTL genotypes. Due to trait complexity, van de Weg et al. 20 argued that the use of a single large family to dissect resistance/susceptibility to fire blight might be a more useful QTL analysis approach, although only for the few alleles able to segregate in a single family. Inevitably, the representation of some haplotypes was limited; however, haplotypes characterized in this study will enable targeted crosses to be made for future studies to validate haplotype effects. For example, an individual (e.g., 'Keepsake', 'Frostbite', 'WA 38') that is heterozygous for the rare reduced-susceptibility haplotype 6E could be used as a parent in a biparental QTL study.

Breeding implications
The complexity of resistance to fire blight will continue to make breeding for resistance challenging. QTL haplotypes characterized in this study had moderate effects in the pedigree-connected reference germplasm set. A better understanding of alleles at detected QTLs and their interactions is needed before information gained can be used to develop and deploy trait-predictive DNA tests for the Chr. 6, Chr. 7, and Chr. 15 QTLs detected in this study.
Reduced-and increased-susceptibility allele information gained has utility in targeting reduced susceptibility to fire blight in breeding. In the short-term, allele information can be used to inform parental selection for more targeted crosses. Parents that have multiple increased-susceptibility and no reducedsusceptibility alleles at reported QTLs (e.g., 'Ginger Gold', 'Minnewashta', 'Pinova', 'Sansa', and 'Sunrise') should be avoided if strong resistance to fire blight is expected in the next generation. In contrast, cultivars homozygous for reduced-susceptibility alleles at QTLs and/or with reduced-susceptibility alleles at multiple QTLs (e.g., 'Enterprise') could be useful as parents when breeding for reduced-susceptibility to fire blight. DNA-informed parental selection combined with targeted phenotypic seedling selection in the greenhouse might be an effective approach to developing breeding populations with lower incidence and severity of fire blight. Additionally, the Chr. 6 QTL might be a useful target for DNA test development due to the breeding relevance of 'Honeycrisp', high QTL repeatability across years, and phenotypic variation explained by this QTL. Selection for haplotype 6E in 'Honeycrisp'-derived parents (e.g., 'WA 38') or selection against offspring with haplotype 6R might be valuable approaches to developing breeding populations with low susceptibility to fire blight. Although reduced susceptibility is not complete resistance, a successful combination of multiple reduced-susceptibility alleles from different sources will contribute to achieving more durable resistance 8 . In the long-term, breeders should focus on selecting against increased-susceptibility alleles and pyramiding resistance alleles at major genes derived from wild germplasm with reduced-susceptibility alleles to achieve durable resistance to fire blight. Techniques like fast-track breeding [44][45][46] could be used to accelerate introgression and pyramiding of favorable alleles.

Materials and methods
Germplasm Germplasm in this study, a subset of the U.S. Apple Crop Reference and Breeding Pedigree Sets described by Peace et al. 28 , were represented in triplicate, propagated on M.111 rootstock, in a randomized complete block design field planting. The germplasm, planting establishment, and maintenance were previously described by Kostick et al. 7,23 . The final pedigree-connected germplasm set, described by Kostick et al. 7 Table S12). 'Honeycrisp' was highly represented with 112 direct offspring. As described by Kostick et al. 7 , these 32 families were included in analyses due to sufficient average allelic representation (AAR) of their respective IBPs (≥ 12.5 AAR units) 28 .

Phenotypic data
Phenotypic data used in this study were previously described 7 . In 2016 and 2017, up to 10 independent, actively growing shoots per tree (≤3 trees per individual) were inoculated with E. amylovora 153n 47,48 in a replicated field inoculation study 7 . For each inoculated shoot, the response to fire blight was quantified as the proportion of the current season's shoot length that was blighted (SLB). SLB values for each inoculated shoot were calculated by dividing the length of a necrotic lesion within the current season's growth by the total shoot length. Statistical analyses of offspring SLB data were previously described 7 and are briefly summarized here. Using R version 3.5.2 (R Core Team, Vienna, Austria), SLB data were analyzed across and within years with linear mixed models fit by restricted maximum likelihood (REML) via the lme4 package 49 . Within a year, blocks were considered fixed effects whereas offspring and block × offspring were considered random effects 7 . Normal QQ plots (i.e., sample vs. theoretical quantiles plots) were used to check the normality of random effects and residuals (Supplementary Figs. S5 and S6) 7 . Within a year, SLB best linear unbiased predictions (BLUPs), adjusted by trait means (as in Amyotte et al. 50 ), were used to estimate offspring responses. Across year offspring SLB BLUPs were also estimated (data not shown). High adjusted SLB BLUPs (closer to 1.00) indicated high relative susceptibility to fire blight whereas low values (closer to 0.00) indicated low relative susceptibility. Adjusted SLB BLUPs were used as phenotypic values in QTL analyses and within-year SLB BLUP distributions are summarized here. In both 2016 and 2017, offspring responses spanned the spectrum ranging from highly resistant to highly susceptible (Supplementary Table S12). Adjusted SLB BLUPs ranged from 0.04 to 0.90 with a relatively normal distribution and an average of 0.48 (n = 312 offspring) in 2016 and from 0.08 to 0.97 with an average of 0.48 (n = 314 offspring) in 2017 (Supplementary Table S12). In 2017, offspring responses were slightly skewed towards lower adjusted SLB BLUPs than in 2016 but still followed a relatively normal distribution.

Genotypic data
As part of the USDA-SCRI RosBREED project 51-53 , a genotypic data set of the IBPs, offspring, and available progenitors were obtained from the use of the International RosBREED SNP Consortium apple 8 K SNP array v1 35 after marker calling, filtering, and curation of SNP data by Vanderzande et al. 36 . Genetic positions of the 3855 filtered SNPs were obtained from the map described by Vanderzande et al. 36 .

QTL detection
FlexQTL TM software (www.flexqtl.nl) 26 26 , reliable estimates of trait means, numbers of QTLs, and their variances were not considered achieved until lengths of effective chain size for each were at least 100depending on the run, such convergence required Markov chain lengths of 250,000 or 300,000. An additive, bi-allelic QTL model (Q/q) was used where Q and q are associated with high and low phenotypic values, respectively. A dominance model was also initially tested. Separate QTL analyses were conducted using adjusted SLB BLUPs for each year. Phenotypic data were only included for unselected offspring (n = 312 in 2016 and 314 in 2017) to avoid any biases from previous breeding selection (as in van de Weg et al. 20 and Verma et al. 34 ). To ensure reproducibility, two replicate runs with different starting seed numbers were carried out for each year (as in van de Weg et al. 20 , Verma et al. 34 , and Howard et al. 32,33 ).
QTL significance and stability were determined using the Bayes factor parameter (BF; 2lnBF 10 ) and posterior intensity values. Evidence for a QTL was considered positive, strong, or decisive if BF values were > 2, 5, or 10, respectively 54 . Similar to van de Weg et al. 20 , QTL intervals were summarized in consecutive 2 cM "bins" (chromosomal segments) with BF > 2 and QTL interval boundaries were defined by the furthest left and right cM positions, respectively, of the two outer bins. The mode within a detected QTL interval was considered the most probable QTL position. The proportion of phenotypic variance explained by each significant QTL within a replicate run was estimated by dividing the variance explained by the total phenotypic variance (described by Verma et al. 34 ). The proportions of phenotypic variance explained were only reported for repeatable QTLs. For purposes of describing QTL detection results, a single replicate run from each year is presented in the main text (as in Howard et al. 32 ) and results of individual replicate runs are presented in Supplementary Tables S1 and S3.

Functional annotation
Functional annotations of genes within each stable QTL interval were examined. Physical positions of stable QTL intervals using the GDDH13 v1.1 reference genome 21 were used to determine the list of genes to consider. Available gene functional annotations for these intervals were obtained using JBrowse 55 accessed via the Genome Database for Rosaceae (https://www.rosaceae.org) 56 . Genes with putative roles in disease resistance were particularly sought.

QTL genotyping and haplotype characterization of QTL alleles
QTLs targeted for haplotype characterization had strong evidence (BF > 5) in at least one year and at least positive evidence (BF > 2) in the other year. FlexQTL TM software was used to estimate QTL genotypes (qq, Qq, QQ), where q and Q were associated with low and high adjusted SLB BLUPs, respectively.
Haploblocks, i.e., chromosomal segments in which no recombination was observed among selected material (cultivars, selections) of the germplasm set, were previously defined by Vanderzande et al. 36 . Haploblocks used were chosen based on their proximity to detected QTL modes (peaks). SNP haplotypes of these haploblocks were constructed across the germplasm set using Pedihaplotyper software 57 with SNP marker phasing via FlexQTL TM software. Where necessary, deduction of haplotypes was conducted by examining available progenitor and/or progeny SNP haplotype data. Haplotypes were traced through the pedigree to the furthest ancestor and back through descendants via identity-by-descent (IBD). In contrast, shared haplotypes that could not be traced to a common ancestor/founder were considered identical-by-state (IBS).
Where an ancestor lacked SNP information, haplotypes were deduced according to what it must have contributed to immediate offspring and assigned as homozygous for a haplotype where there was only a single deduced haplotype. Pedigrees were visualized using Pedimap software 58 .
Within-year one-way analysis of variances (ANOVAs; R version 3.5.2; R Core Team, Vienna, Austria) for each haplotype were performed to determine if adjusted SLB BLUPs were statistically different for presence vs. absence of a given haplotype. A haplotype with ANOVA p < 0.05 in 2016 and 2017 was marked as a significant-effect haplotype. Significantly higher and lower adjusted SLB BLUP means indicated increased-and reducedsusceptibility 8 (i.e., relatively high and low susceptibility) haplotypes, respectively. Using R-package 'ggpubr' 59 , adjusted SLB BLUP means, 95% confidence intervals, and distributions across years were determined and plotted for haplotypes with significant effects in both years.
To simultaneously examine the effects of multiple reduced-susceptibility haplotypes, offspring were first grouped by presence/absence of increased-susceptibility haplotypes and then by their number (i.e., 0, 1, 2, ≥ 3) of reduced-susceptibility haplotypes across detected QTLs. Adjusted SLB BLUP means, 95% confidence intervals, and distributions across years were determined and plotted for each of these eight groups of offspring. Statistical mean separation was calculated using the least significant difference with a Bonferroni p adjustment for multiple comparisons via R package 'agricolae' 60 .