Genomic characterization of a global Corylus avellana L. collection: insights into hazelnut ancestry and genetic determinants of production traits Authors: Monique Salardi-Jost, Svenja Mager, Gabriele Dini, Mercy Wairimu Macharia, Mercè Rovira, Valerio Cristofori, Mario Enrico Pé, Andrea Cavallero, Claudio Todeschini, Giuseppe Genova, Matteo Dell’Acqua Published: 2025 ABSTRACT BackgroundEuropean Hazelnut (Corylus avellana L.) is a high-value, understudied crop that exhibits a large diversity of traits. Commercial production relies on a small number of globally grown varieties. Advancing hazelnut breeding requires a comprehensive understanding of its genetic diversity and its association with valuable traits. Here, we used genotyping-by-sequencing to sequence 316 hazelnut varieties, retaining 282 samples and 44,757 single nucleotide polymorphisms (SNPs) after quality control. This collection was used to study genetic diversity, identify genomic loci under selection, and describe genome-wide association with nut traits, i.e. nut size, perimeter and caliber.ResultsWe uncovered three genetic clusters and five ancestral populations, with Eurasian varieties (Türkiye, Azerbaijan, and Georgia) being the most distinct and European and U.S. populations showing higher admixture. We identified four loci under positive selection in Eurasian varieties, including one within Cav06g02620 (probable calcium uniporter) on chromosome 6. We also characterized the nut morphology of 151 genotyped varieties, leading to the identification of ten SNPs associated with nut traits.ConclusionsWhile functional validation of candidate genes and the use of a larger population are important for future analysis, our findings provide a basis for exploring the genetics of quantitative traits in hazelnut, supporting breeding and conservation efforts.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12870-025-07596-2. Backgound The European Hazelnut (Corylus avellana L.) is a diploid (2n = 2x = 22), perennial, monoecious species, characterized by the presence of separate staminate and pistillate flowers on the same plant. Its genome size is approximately 378 Mb [1]. As a wind-pollinated species, C. avellana relies on airborne pollen for fertilization, exhibiting sporophytic self-incompatibility (SSI) controlled by a single polymorphic locus [2] and requires pollen from a genetically distinct plant for fruit production. Regarding its cultivation, C. avellana is frequently clonally propagated to maintain desirable phenotypic traits. Propagation typically occurs via suckers (basal shoots), which the plant naturally produces in abundance. C. avellana has a long juvenile phase, with nut production beginning after 3 to 5 years and reaching peak productivity between 7 and 10 years [1]. These biological characteristics pose significant challenges for breeding, as the expression of key agronomic traits may take up several years to be observable. The domestication of C. avellana occurred independently in at least six regions: the British Islands, central Europe, Spain, Italy, the Black Sea region, and Iran [3–6]. Within the Mediterranean Basin, germplasm groups have been identified in Tarragona (Spain), Campania (Italy), and in the Black Sea Region (Türkiye), representing three putative centers of domestication [7]. Secondary gene pools emerged in other regions, such as Latium and Liguria in Italy and Asturias in Spain, likely as a result of local domestication from wild forms and/or hybridization with introduced cultivated varieties [8]. C. avellana has the fifth largest harvested area among nut crops worldwide [9]. It is widespread across Europe and extends into western Asia, including Türkiye and Georgia [10]. Despite its broad native range, optimal production areas are limited by specific climatic conditions, as the species thrives in temperate climates with cool winters and mild summers [11]. Recently, hazelnut production and harvesting areas have expanded outside its historical growing area, into the United States of America, Chile, and some Eastern European countries [8]. Despite its broad cultivation range, global hazelnut production still relies on only about 20 widely grown varieties and an additional 30 genotypes that are currently considered attractive for breeding [12]. In Türkiye, the world’s largest hazelnut producer, 18 out of 20 varieties dominate production, with the majority coming from Tombul, Palaz, Foşa, and Çakıldak [13, 14]. Similarly, in Europe, production is centered around few cultivars, including Negret in Spain, and Tonda Gentile delle Langhe, Tonda Romana, and Tonda di Giffoni in Italy [8]. The reliance on few broadly cultivated cultivars, and the narrow breeding focus on highly selected genetic backgrounds, exposes hazelnut production to multiple threats. Limited genetic diversity translates to limited resilience to new and emerging biotic and abiotic stresses that may jeopardize yield, including those deriving from the climate crisis [15]. Concurrently, the limited exploration of genetic diversity is a missed opportunity for progress towards new and superior hazelnut varieties. The population structure existing in hazelnut orchards is typically weak, largely due to the practice of clonal propagation coupled with cross-pollination, including from wild relatives, which boosts diversity and heterozygosity [16]. About a dozen public hazelnut breeding programs have been globally operational since mid-20th century, with varying success. However more recently, several of them have been discontinued [8], limiting the development of new cultivars and reducing genetic improvement efforts. To effectively bring more diversity, and especially more desirable diversity into hazelnut breeding pipelines, it is necessary to explore the scope and potential for improvement of its genetic diversity. Several studies used DNA markers to characterize the genetic diversity in C. avellana collections [6, 17, 18]. More recently, the genome sequencing of the reference lines Tombul [1], Tonda Gentile delle Langhe [19] and Jefferson [20] has opened the possibility to effectively use genotyping-by-sequencing to derive single nucleotide polymorphisms (SNPs) that are useful to characterize large germplasm collections. SNP based studies confirmed that recent cultivars are highly similar and that Turkish materials are unique [21], with Georgian varieties constituting a distinct group [22]. These studies found evidence of admixture both between wild and domesticated individuals and among different genetic clusters of cultivated varieties across the world [23]. Although these studies contributed to describing the vast diversity present in hazelnut, a large potential remains in understanding which portion of this diversity has a relevance in hazelnut adaptation and improvement. Genotyping-by-sequencing (GBS) has been shown to efficiently yield informative SNPs suitable for diversity analyses in hazelnut, validating its use in SNP-based identification and diversity scans in this crop [24]. Limited studies explored the connection between hazelnut genetic diversity and phenotypic traits. The first quantitative trait loci (QTL) mapping in hazelnut was conducted in a population of 163 F1 progenies derived from a cross between Tonda Gentile delle Langhe and Merveille de Bollwiller [25]. Other studies using natural and artificial populations composed of materials of interest identified QTL for a number of traits including production [26] and disease resistance traits [27] but still there is no study capturing the whole diversity of global hazelnut collections to pursue the identification of QTL markers of breeding relevance. In this study, we characterized a global collection of C. avellana representative of its major areas of cultivation. We genotyped it describing the uniqueness of Eurasian hazelnuts, and pinpointing candidate loci that are under directional selection, shedding light on the possible molecular mechanisms driving the adaptation of hazelnut populations. We then characterized nut phenotypical diversity, measuring nut size, perimeter, and caliber. Using a genome-wide association approach, we reported genetic loci that are associated to these traits and proposed candidate genes. Materials and methods Plant material We started from 316 leaf samples, each coming from an individual plant, including 13 wild types, one C. colurna, one C. colurna x C. avellana hybrid and 301 C. avellana samples. A formal identification of the plant material has been performed by Prof. Matteo Dell’Acqua. A voucher specimen could not be deposited in any available herbarium since all plant material has been destroyed during the processing. Plants were classified as wild types when they were found to be isolated from cultivated orchards and showed no evidence of agronomic management. Leaves were sourced from five collections: the collections of Ferrero Farming Turkey (Türkiye), the National Hazelnut Germplasm of the Institute of Agrifood Research and Technology (Catalonia, Spain), HCo AgriGeorgia (Tbilisi, Georgia), Tuscia University (Viterbo, Italy) and SAGEA Centro di Saggio s.r.l. (Castagnito, Italy). The collections are representative of hazelnut genotypes from 15 countries (Spain, Italy, Türkiye, Greece, the United States, Germany, France, Albania, Romania, Belgium, England, the United Kingdom, Georgia, Azerbaijan, and Former Yugoslavia) plus samples with unknown geographical provenance as well as breeding material and wild types. Geographical provenance reflects the regions where the varieties originally evolved or have been traditionally cultivated, thereby representing their genetic background. In contrast, the orchard country refers to the location where the trees were cultivated, reflecting the environmental conditions under which they were grown. The samples included in this study were cultivated in four countries: Italy, Spain, Türkiye, and Georgia, and no permissions were required to collect samples from any of the locations. Furthermore, our collection also includes nuts harvested in the growing season of 2022 from 151 of these 316 individual plants. These nuts were produced by trees cultivated in orchards in Italy (30 plants), Spain (80 plants) and Türkiye (41 plants). All samples and passport data are reported in Additional file 1. Upon arrival at the laboratory, leaf samples were stored at −80 °C until DNA extraction, while nut samples were stored at 4 °C until phenotyping. Leaf tissue from samples collected in orchards in Georgia and Türkiye was partially desiccated and shipped without ice, whereas leaf tissue from all other samples was fresh and shipped on ice. Genotyping and variant filtering Genomic DNA was extracted from leaves using the Cetyltrimethylammonium bromide (CTAB) method [28]. DNA quality was verified by electrophoresis on 1% agarose gel and using a NanoDrop ND-1000 spectrophotometer. DNA was then shipped to IGATech Technology Services (Udine, Italy) for genotyping-by-sequencing using the double-digestion restriction site associated DNA markers (ddRAD-seq) method [29] using a custom protocol based on the restriction enzymes NspI and MboI endonucleases. Libraries were sequenced in paired-end mode on a NovaSeq 6000 instrument (Illumina, San Diego, CA). Raw Illumina reads were demultiplexed using the process_radtags utility included in Stacks v.2.61 [30] and were aligned to the Tombul reference genome CavTom2PMs-1.0 (https://www.ncbi.nlm.nih.gov/assembly/GCA_901000735.2) with Burrows-Wheeler-Aligner v.0.7.12 (BWA) using the maximal exact matches (MEMs) algorithm [31] with default parameters and selection of uniquely aligned reads (reads with a mapping quality > 4) (Additional file 1). Alignments were sorted and indexed with SAMtools v.1.20 [32] and variants were exported with the Stacks v2.61 [30] programs gstacks and populations, respectively, specifying the options –R = 0.75 and --max-obs-het = 0.8. Variants were further filtered using VCFtools v.0.1.16 [33] calculating the mean and variance of the loci depth (ldepth.mean). Loci were filtered to maintain only those with a minimum depth of 15X (a conservative threshold) and a maximum depth of 68X (the mean of the variance multiplied by 2). Subsequently, (i) samples with more than 20% missing variants were removed from the dataset and (ii) sites with less than 90% genotypic information and Minor Allele Frequency (MAF) < 0.05 were also removed. This led to the creation of a set of SNP markers henceforth called the complete dataset. For population genetics studies, samples and variants were further filtered to decrease redundancy that may hamper ancestry reconstruction according to the following steps. First, a pairwise distance matrix was calculated by Tassel 5.0 [34] on homozygous SNPs to identify pairs of identical genotypes (i.e. clones). Genotypes were considered clones when they had less than 0.001 difference between each other; the clone with the highest proportion of missing alleles was removed retaining only one representative genotype. Then, SNP markers were filtered for linkage disequilibrium (LD) at a threshold of r²=0.5 using the indep-pairwise function in PLINK v1.90 [35], employing a 150-variant window moving in increments of 5 variants. These filtering steps resulted in a SNP set referred to as the reduced dataset. Genetic diversity and ancestry analyses Genetic diversity and ancestry analyses used the reduced dataset. A neighbour-joining (NJ) tree was generated computing the genetic distances using bitwise distance and performing 2,000 bootstrap resampling on R/poppr v.2.9.6 [36], and the resulting tree was visualized using R/ggtree v.3.12.0 [37]. Principal Component Analysis (PCA) and Discriminant Analysis of Principal Components (DAPC) were performed using R/adegenet v.2.1.10 [38], and the optimal number of clusters (K) for the DAPC was identified using the function find.cluster(). The Bayesian Information Criterion (BIC) statistics were calculated at increasing values of K from 1 to 15 to measure the goodness of fit at each K using all Principal Components (PCs) and default settings. Mean pairwise fixation index (FST) statistics were calculated using the program populations from Stacks v2.61 [30]. To do that, individuals were grouped based on the three clusters (K = 3) inferred by the DAPC analysis. The reduced dataset was used as input, and populations was run with the --vcf, --fstats, and --popmap options enabled, with the population map specifying the cluster membership of each individual. The program populations from Stacks determines private alleles as alleles occurring only in a single population [39, 40]. For each population supplied, at each variable SNP at a locus, the observed heterozygosity (H) and the nucleotide diversity π are calculated as the following:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\pi\:=1-\:\frac{\sum\:\left(\frac{{\:\eta\:}_{i}}{2}\right)}{\frac{\eta\:}{2}}.$$\end{document} where ni is the count of allele i in the population, and n is the sample size of all alleles in the population at that locus [41, 42]. Wright’s inbreeding coefficient (FIS) measures the degree of difference between observed and expected heterozygosity within subpopulations [43] of the individual relative to the population:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{IS\:=\:\frac{\pi\:\:-\:H}{\pi\:}\:.}$$\end{document} ADMIXTURE v.1.3.0 [44] was employed to further define the ancestry in the population testing values ranging from 1 to 20, using the default termination criterion and implementing a fivefold cross-validation procedure. Parameter standard errors were calculated using 2,000 bootstrap replicates. The correlation of the residual differences between the true genotypes and those predicted by the model was assessed using EvalAdmix v.0.95 [45]. To interpret the ADMIXTURE results on a map the historical provenances of varieties were pooled into six larger areas, namely Eurasia (Azerbaijan, Georgia, and Türkiye), Northern Europe (France, Germany, and England), Central Europe (Greece, Romania, and Former Yugoslavia), Italy, Spain, or United States using R/mapmixture v.1.1.3 [46]. Statistics about the genetic variation within and between populations were called by using fstats in Stacks v2.61 [30]. The LD decay was calculated by thinning the complete dataset using PLINK v.1.9 [35], removing variants at random until only 10,000 variants remained. Then, LD decay was estimated as function of physical distance according to the Hill and Weir Eq. [47] on base R v. 4.4.2 [48]. In addition to the ancestry analysis, we checked for migration events and introgression. We used TreeMix v1.13 [49] with 0–5 migration events and 5 replications each and using -k = 25, -sed and -se options. The R package ggplot v3.5.2 was used to plot 6 trees from TreeMix results, one tree for each value of m and choosing the best of 5 iterations. To get the optimal value for m, the OptM v0.1.9 package [50] in R was used, applying the Evanno method. Identification of potential candidate loci under natural selection Candidate variants under selection were identified calculating the FST between the distinct DAPC clusters with BayeScan v2.1 [51]. Cluster 1 and 3 were combined to gain insights into the distinction of cluster 2 (mainly Eurasian varieties) and the other varieties. Input file conversion was performed with PGDSpider v2.1.1.5 [52]. Only sites with a qvalue < 0.05 were considered as significant. Positive values of alpha were interpreted as diversifying selection and negative values as balancing or purifying selection. Candidate genes were explored with two methods. First, the closest gene to each outlier locus was identified with BEDTools v2.31.0 [53] and its probable function was determined via a protein blast against Arabidopsis thaliana genes. In cases where the most significant result corresponded to a hypothetical protein, the next best match with a functional annotation was provided. Second, genes within the LD decay distance upstream and downstream were derived, and a gene ontology (GO) enrichment analysis was performed on them with R/topGO v.2.56.0 [54]. Phenotyping of nut traits and genome-wide association study We had access to nuts from a subset of the genotyped varieties, and for each accession, we randomly selected 10 nuts and photographed them in a shaded chamber with constant light conditions and constant focal distance with a Sony Alpha NEX-3 N camera (36 mm focus lens, f/5.6 aperture, 1/60s shutter speed, ISO AUTO). Images were processed using ImageJ v. 1.53 [55] to extract size, perimeter, minimum and maximum caliber of each individual nut. The nut size represented the total space enclosed within a nut, while the perimeter measured the length of the nut. The maximum caliber referred to the longest distance between any two points along the nut’s outline, while the minimum caliber represented the shortest distance between any two points along the nut’s outline. Nut samples were pooled into the same larger groups used to plot ADMIXTURE results on a map, adding the only sample from Albania into the Eastern Europe group, and their diversity was explored with R v. 4.4.2 [48]. A Shapiro-Wilk test was conducted to check for normality on R v. 4.4.2 [48], and a Levene test was conducted to assess homoscedasticity of variances on R/car v. 3.1.3 [56]. Using R/ggpubr v. 0.6.0 [57], when a given trait did not pass at least one of these tests, a Kruskal-Wallis test was conducted. Conversely, a one-way ANOVA was conducted when both assumptions were met. Pearson correlation coefficients between nut traits were calculated using R/PerformanceAnalytics v. 2.0.4 [58]. Nut traits were combined with genotypic data in the complete dataset to conduct a genome-wide association study (GWAS) seeking associations between SNPs and measured traits on the set of 151 varieties that had been both genotyped and phenotyped. The GWAS was performed using the Fixed and Random model Circulating Probability Unification (FarmCPU) [59] method as implemented in R/GAPIT v.3 [60]. Three genetic PCs estimated from the LD-pruned dataset were included in the model to account for population structure, and qq plots were assessed for model fit. The Bonferroni correction was applied to establish significance at a nominal p-value of 0.05. When relevant, a less stringent threshold based on false discovery rate (FDR) correction was displayed in the Manhattan plots. To assess potential inflation of test statistics due to population structure or cryptic relatedness, the inflation factor (λ) was calculated for each nut trait. To do that, p-values were converted to chi-squared statistics with one degree of freedom. The inflation factor (λ) was then obtained by dividing the median of the observed chi-squared statistics by the expected median under the null distribution. The phenotypic variance explained (PVE) for each significantly associated SNP was calculated following the equation described by Giordani et al. 2022 [61] for the variance of a QTL (Vqtl).\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{V}_{qtl}=2freq\times\:\left(1-freq\right)\times\:effect2$$\end{document} Here, freq is the allele frequency and effect is the estimated additive effect of the allele from the FarmCPU model. Then, PVE was calculated as the proportion of this variance divided by the total phenotypic variance of each trait. To evaluate marker informativeness, we also calculated the Polymorphism Information Content (PIC) for each significantly associated SNP. For biallelic loci, PIC was computed as:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:PIC\:=1\:-\:\left({p}^{2}+{q}^{2}\right)-2{p}^{2}{q}^{2}$$\end{document} where p and q represent the frequencies of the major and minor alleles, respectively. SNPs with PIC < 0.10 are considered to have limited discriminative power, as they tend to be either too rare or too common in a population. Candidate genes were characterized with the same method used for outlier loci. Results Genotyping-by-sequencing provided a large dataset of variants in a broad hazelnut collection We genotyped 316 leaf samples (314 C. avellana, one C. colurna and one C. colurna x C. avellana hybrid) with geographical provenance from 15 different countries (Fig. 1). Fig. 1Geographical provenance of C. avellana samples genotyped in this study. In addition to the 258 samples shown on the map, 13 wild types, 12 breeding materials, and 33 samples with unknown geographical provenance were also genotyped, totaling 316 samples. Black circle sizes correspond to number of samples, with bigger circles meaning more samples, ranging from a minimum of 1 to a maximum of 97 samples Geographical provenance of C. avellana samples genotyped in this study. In addition to the 258 samples shown on the map, 13 wild types, 12 breeding materials, and 33 samples with unknown geographical provenance were also genotyped, totaling 316 samples. Black circle sizes correspond to number of samples, with bigger circles meaning more samples, ranging from a minimum of 1 to a maximum of 97 samples DdRAD-seq genotyping generated an average of 5,307,805 reads for each of the 316 leaf samples. On average, 78.07% (4,235,226) reads per sample mapped on the Tombul reference genome v.1.0 (GCF_901000735.1_CavTom2PMs-1.0.fna). Since we sequenced a reduced representation library, the mean genomic coverage in the whole dataset is 1.50X; however, samples present varying coverage, with a standard deviation of 24.5X on average (Additional file 1). The high standard deviation reflects the differences in conditions of the leaves used for DNA extraction. Since leaf samples from orchards in Georgia and Türkiye were shipped under less-than-ideal conditions, this factor may have contributed to slightly lower mapping rates in some cases. At the same time, the number of mapped reads is sufficient to include these samples in downstream analyses. After quality filtering, including removal of samples with poor quality, the complete dataset consisted of 282 varieties and 44,757 SNPs. The reduced dataset was obtained by both removing 141 clones and markers in linkage via LD pruning, resulting in 141 varieties and 16,378 SNPs retained. High genetic diversity, low fixation and a genetically distinct population are present in the global C. avellana collection The NJ tree of the reduced dataset shows that hazelnut diversity can be partially attributed to geography. For readability, Fig. 2a shows a simplified visualization (variety names, bootstrap values, and branch lengths omitted) while the fully annotated tree can be found in Additional file 2. Varieties with shared geographic provenance tend to form clusters. For example, Turkish and Spanish samples each form their own separate clusters. 9 out of the 10 wild type varieties form a separate cluster, together with one Italian (Tonda Gentile delle Langhe) and one Spanish variety (Pirineu). The probable origin of samples with unknown provenance can be inferred by their clustering nearby varieties of known origin. Varieties with the same variety name are not always clustering together as expected, such as Sivri and Sivri A (Additional file 2). The DAPC best summarizes hazelnut diversity with three genetic clusters, as indicated by BIC statistics (Fig. 2b). DAPC cluster 2 (mainly hazelnut varieties from Türkiye, Georgia and Azerbaijan) is separated from DAPC clusters 1 (mainly Italian and varieties with unknown provenance) and 3 (mainly Spanish varieties) (Fig. 2c, Additional file 1). While DAPC cluster 2 is distinctly separated from the other two, clusters 1 and 3 exhibit some overlap. Fig. 2Population diversity and structure analysis. Varieties are colored according to their geographical provenance. A NJ-tree rooted on the Tombul variety with 2000 bootstrap resampling. B BIC values for different values of DAPC clusters, indicating that the optimal number of clusters is 3. C PCA of the SNPs. The 3 genetic clusters identified through DAPC are indicated by different shapes Population diversity and structure analysis. Varieties are colored according to their geographical provenance. A NJ-tree rooted on the Tombul variety with 2000 bootstrap resampling. B BIC values for different values of DAPC clusters, indicating that the optimal number of clusters is 3. C PCA of the SNPs. The 3 genetic clusters identified through DAPC are indicated by different shapes While the nucleotide diversity is relatively high in all three clusters (between 0.23 and 0.29 nucleotide differences per site) and similar values are observed for heterozygosity (ranging from 0.22 to 0.24), cluster 2 shows a much higher number of private alleles (155) than cluster 1 [35] or cluster 3 (0) (Table 1). Cluster 3, however, stands out for the comparably higher inbreeding coefficient (FIS), 0.19, as compared to the value of ~ 0.07 in both clusters 1 and 2. Table 1Statistics about the genetic variation within the three DAPC clustersCluster1Cluster2Cluster3Samples in cluster673242Private alleles351550Mean observed heterozygosity0.240.220.24Mean value of π (Nucleotide diversity)0.250.230.29Mean measure of FIS (inbreeding coefficient)0.070.070.19 Table 1Statistics about the genetic variation within the three DAPC clustersCluster1Cluster2Cluster3Samples in cluster673242Private alleles351550Mean observed heterozygosity0.240.220.24Mean value of π (Nucleotide diversity)0.250.230.29Mean measure of FIS (inbreeding coefficient)0.070.070.19 Statistics about the genetic variation within the three DAPC clusters Pairwise comparison of FST was estimated between all three DAPC clusters as well as between cluster 2 and the combined clusters 1 and 3 (cluster1_3) (Table 2). The smallest FST (~ 0.03) is observed between cluster 1 and 3, which also shows overlap in the DAPC and lower numbers of private alleles. Slightly higher genetic differentiation can be seen when comparing cluster 2 to either cluster 1 or 3. In general, the genetic structure is low, as confirmed by the FST of only ~ 0.04 between the combined cluster 1_3 and the more distinct cluster 2. Table 2Pairwise FST values between the three DAPC clusters as well as between cluster 2 and the combined clusters 1 and 3Cluster2Cluster3Cluster1Cluster1_3Cluster2-0.050.050.04Cluster3--0.03- Table 2Pairwise FST values between the three DAPC clusters as well as between cluster 2 and the combined clusters 1 and 3Cluster2Cluster3Cluster1Cluster1_3Cluster2-0.050.050.04Cluster3--0.03- Pairwise FST values between the three DAPC clusters as well as between cluster 2 and the combined clusters 1 and 3 To further deepen the understanding of the similarities and differences between hazelnut clusters, the samples’ ancestry was investigated. According to this analysis, five genetic ancestral populations best explain the structure in the dataset (Fig. 3a, Additional file 3). The results confirm an overall high level of admixture in most samples. However, hazelnut varieties coming from Azerbaijan, Georgia, and Türkiye, which are grouped in DAPC cluster 2 (Fig. 2b), exhibit considerably lower admixture levels (Fig. 3b). The samples from other provenances show higher levels of admixture, though each geographical area has one main ancestral cluster, especially Italian samples (ancestral cluster 1), Spanish samples (ancestral cluster 2) and Central European samples (ancestral cluster 5). Fig. 3Genetic admixture analysis results. A Q-estimates bar plot of the hazelnut panel for ancestral clusters = 5. Individuals are represented by a bar along the x-axis, ordered by their origin. Different colors represent the different ancestral clusters. The height of each colored bar indicates the proportion of the genome of each sample corresponding to the respective ancestral cluster (y-axis). B The average genomic proportions of each ancestral cluster, derived from ADMIXTURE analysis, are pooled into loosely defined geographical areas. For each pie chart, the number of individuals is indicated in parentheses. Geographical grouping is meant for readability on the map but does not correspond to clusters as defined in the analysis. Samples belonging to each group are reported in Additional file 1 Genetic admixture analysis results. A Q-estimates bar plot of the hazelnut panel for ancestral clusters = 5. Individuals are represented by a bar along the x-axis, ordered by their origin. Different colors represent the different ancestral clusters. The height of each colored bar indicates the proportion of the genome of each sample corresponding to the respective ancestral cluster (y-axis). B The average genomic proportions of each ancestral cluster, derived from ADMIXTURE analysis, are pooled into loosely defined geographical areas. For each pie chart, the number of individuals is indicated in parentheses. Geographical grouping is meant for readability on the map but does not correspond to clusters as defined in the analysis. Samples belonging to each group are reported in Additional file 1 LD decay was calculated to define the search window for candidate gene identification around SNPs potentially under selection in Eurasian varieties, as well as those significantly associated with nut traits. Moderate differences in LD structure are observed between chromosomes. For instance, chromosome 2 and chromosome 5 exhibit broader regions with elevated r² values, particularly in the first ~ 10 Mb of chromosome 2, and between 15 and 30 Mb on chromosome 5, suggesting local reductions in recombination or recent selection events. In contrast, chromosomes 9 and 11 show generally lower LD across their entire length (Fig. 4, left panel). Regarding the genome-wide LD decay, r² decreases sharply within the first 0.5–1.5 Mb and approaches background levels beyond 2 Mb. This pattern reflects a rapid LD decay typical of outcrossing perennial species, such as hazelnut (Fig. 4, right panel). Fig. 4Linkage disequilibrium (LD) in the C. avellana genome. The left panel shows LD evolution across 11 chromosomes, with the lines representing r2 values as rolling windows along physical genomic positions. The right panel represents LD decay as a function of physical distance of SNP markers, expressed in million base pairs (Mb). Lines are color-coded by chromosome Linkage disequilibrium (LD) in the C. avellana genome. The left panel shows LD evolution across 11 chromosomes, with the lines representing r2 values as rolling windows along physical genomic positions. The right panel represents LD decay as a function of physical distance of SNP markers, expressed in million base pairs (Mb). Lines are color-coded by chromosome To determine the genomic distance used in candidate gene searches, the LD decay threshold was defined as the distance at which r² dropped to half of its maximum observed value on each chromosome. These values, calculated per chromosome, range from 145,457 bp on chromosome 11 to 449,368 bp on chromosome 4, and are summarized in Table 3. Table 3Chromosome-specific LD decay distances in the Corylus Avellana genome. The physical distance (in base pairs) at which LD decays to half of its maximum value (r²) for each of the 11 chromosomes is reported. These LD decay distances were used to define the search windows upstream and downstream of significantly associated SNPs for candidate gene identificationChromosomeLD decay (bp)Half LD decay (bp)1292,166146,0832343,899171,9503437,219218,6104449,368224,6845369,741184,8716255,188127,5947350,131175,0668172,63786,3199175,26487,63210195,46297,73111145,45772,729 Table 3Chromosome-specific LD decay distances in the Corylus Avellana genome. The physical distance (in base pairs) at which LD decays to half of its maximum value (r²) for each of the 11 chromosomes is reported. These LD decay distances were used to define the search windows upstream and downstream of significantly associated SNPs for candidate gene identificationChromosomeLD decay (bp)Half LD decay (bp)1292,166146,0832343,899171,9503437,219218,6104449,368224,6845369,741184,8716255,188127,5947350,131175,0668172,63786,3199175,26487,63210195,46297,73111145,45772,729 Chromosome-specific LD decay distances in the Corylus Avellana genome. The physical distance (in base pairs) at which LD decays to half of its maximum value (r²) for each of the 11 chromosomes is reported. These LD decay distances were used to define the search windows upstream and downstream of significantly associated SNPs for candidate gene identification Residuals and optM plots (Additional file 4) based on TreeMix results indicate an m value of 2 as optimal reflection of the true number of migration events. The TreeMix trees (Fig. 5) suggest gene flow especially among Eurasian samples as well as among samples from England, France and Germany which also show some overlap with Italian and Spanish samples. Considering m = 2, the corresponding tree (Fig. 5) indicates gene flow from Former Yugoslavia towards Mediterranean and Wild Type populations as well as a tendency of gene flow from southern into northern Europe populations. Fig. 5Maximum-likelihood population trees inferred by TreeMix with different numbers of migration edges (m = 0–5). Each panel shows the best-fit tree topology (best of 5 iterations) with migration edges represented as arrows. The direction of arrows indicates gene flow, and the color scale reflects migration weight, with warmer colors representing higher inferred migration proportions Maximum-likelihood population trees inferred by TreeMix with different numbers of migration edges (m = 0–5). Each panel shows the best-fit tree topology (best of 5 iterations) with migration edges represented as arrows. The direction of arrows indicates gene flow, and the color scale reflects migration weight, with warmer colors representing higher inferred migration proportions A set of SNPs were revealed to be under strong diversifying selection between Eurasian hazelnuts and the rest of the collection A genome scan was performed to identify outlier SNP loci that constitute candidates to be under selection based on differences in allele frequencies between Eurasian and the other varieties. Four outliers under diversifying selection (positive alpha value) on four chromosomes [1, 6, 7, 11] were found (Fig. 6, Additional file 5). The most significant outlier, and the only outlier with a qvalue < 0.01 (0.0008), is on chromosome 6, with an alpha value of 1.76 and a FST value of 0.34. Fig. 6Manhattan plot of alpha values for identification of putative candidate loci under natural selection among populations formed by samples belonging to cluster 2 and samples belonging to the combined clusters 1 and 3 of the DAPC analysis. X-axis shows the genomic position on the 11 chromosomes, while the y-axis reports the alpha values. Orange and blue dots are the non-significant loci, while brown dots represent the significant outlier candidates (qvalue < 0.05) Manhattan plot of alpha values for identification of putative candidate loci under natural selection among populations formed by samples belonging to cluster 2 and samples belonging to the combined clusters 1 and 3 of the DAPC analysis. X-axis shows the genomic position on the 11 chromosomes, while the y-axis reports the alpha values. Orange and blue dots are the non-significant loci, while brown dots represent the significant outlier candidates (qvalue < 0.05) The closest gene to each outlier was determined (Table 4). The most significant loci for diversifying selection between Eurasian and other samples was found to lie directly in the exon of a calcium uniporter. Table 4Protein blast results of the peptide sequence of the Corylus Avellana genes closest to the putative candidate loci under diversifying selection against the genome of Arabidopsis thaliana. The number of candidate genes in half LD decay distance up- and downstream of the candidate loci is reported in the last columnC. avellana – Gene IDGenomic position of geneGenomic position of outlier (distance to gene)A. thaliana – Gene IDA. thaliana - Gene descriptionQuery coverE-valueGenes in LD decay distance Cav01g12730 Chr1,13,637,906–13,644,80313,637,556(− 350 bp)NP_193319.6BTB/POZ domain-containing protein96%0.00E + 0022 Cav06g02620 Chr6,2,081,984–2,084,1972,083,200(0)NP_179959.1calcium uniporter (DUF607)92%6.00E-9729 Cav07g08710 Chr7,7,364,824–7,370,2467,371,795(+ 1549 bp)NP_568355.1tRNA (guanine-N-7) methyltransferase79%7.00E-12043 Cav11g08360 Chr11,12,653,942–12,658,37712,643,567(− 10375 bp)NP_177388.2CCAAT-binding factor66%9.00E-755 Table 4Protein blast results of the peptide sequence of the Corylus Avellana genes closest to the putative candidate loci under diversifying selection against the genome of Arabidopsis thaliana. The number of candidate genes in half LD decay distance up- and downstream of the candidate loci is reported in the last columnC. avellana – Gene IDGenomic position of geneGenomic position of outlier (distance to gene)A. thaliana – Gene IDA. thaliana - Gene descriptionQuery coverE-valueGenes in LD decay distance Cav01g12730 Chr1,13,637,906–13,644,80313,637,556(− 350 bp)NP_193319.6BTB/POZ domain-containing protein96%0.00E + 0022 Cav06g02620 Chr6,2,081,984–2,084,1972,083,200(0)NP_179959.1calcium uniporter (DUF607)92%6.00E-9729 Cav07g08710 Chr7,7,364,824–7,370,2467,371,795(+ 1549 bp)NP_568355.1tRNA (guanine-N-7) methyltransferase79%7.00E-12043 Cav11g08360 Chr11,12,653,942–12,658,37712,643,567(− 10375 bp)NP_177388.2CCAAT-binding factor66%9.00E-755 Protein blast results of the peptide sequence of the Corylus Avellana genes closest to the putative candidate loci under diversifying selection against the genome of Arabidopsis thaliana. The number of candidate genes in half LD decay distance up- and downstream of the candidate loci is reported in the last column Chr1, 13,637,906–13,644,803 13,637,556 (− 350 bp) Chr6, 2,081,984–2,084,197 2,083,200 (0) Chr7, 7,364,824–7,370,246 7,371,795 (+ 1549 bp) Chr11, 12,653,942–12,658,377 12,643,567 (− 10375 bp) A total of 99 genes (Additional file 6) were found in LD decay distance around all candidate loci under diversifying selection. The LD decay values for each chromosome are reported in Table 3. Specifically, the LD decay is 292,166 bp on chromosome 1, 255,188 bp on chromosome 6, 350,131 bp on chromosome 7, and 145,457 bp on chromosome 11. The GO analysis reported that 40 functional categories are overrepresented, with most containing maximally three genes (Additional file 7). The five categories with more than three genes include tubulin and hydrogen peroxide pathways as well as plant ovule development. Diversity in nut traits and genome wide association study A subset of 151 varieties from our complete dataset was characterized for their nut traits, including nut size (mean = 154.41 mm2, SD = 26.42 mm2), perimeter (mean = 51.13 mm, SD = 5.07 mm), maximum caliber (mean = 15.92 mm, SD = 1.73 mm), and minimum caliber (mean = 12.72 mm, SD = 1.35 mm). The distribution of the nut traits within each provenance follows a normal distribution (Shapiro-Wilk test, p > 0.8). According to the Levene test, the variances of nut size (p = 0.001657), maximum caliber (p = 0.0002984), and minimum caliber (p = 0.03035) within each provenance are not homoscedastic, while the variances of nut perimeter (p = 0.3027) are homoscedastic. The ANOVA and Kruskal-Wallis test revealed no significant differences in nut traits based on provenance (Additional file 8). Pearson correlations showed that all traits are significantly positively correlated to each other (Additional file 9). GWAS identified multiple SNPs significantly associated with key nut traits. Three SNPs are associated with average size, two with nut maximum caliber, three with nut minimum caliber, and two to nut perimeter (Fig. 7). The inflation factors (λ) calculated from the FarmCPU models range from 0.73 to 0.87, with values of 0.73 for nut perimeter, 0.75 for minimum caliber, 0.86 for nut average size, and 0.87 for maximum caliber. Fig. 7Nut illustrations, Manhattan and QQ-plots of nut size (A), nut maximum caliber (B), nut minimum caliber (C), and nut perimeter (D). The schematic illustrations on the left provide a visual reference for how each trait was measured. In the Manhattan plots, the X-axis shows the genomic position of the SNPs in the genome, and the Y-axis reports the negative log base 10 of the p-values. The solid green horizontal lines represent the significance threshold based on the Bonferroni correction, which was applied at a p-value of 0.05. When relevant, a dashed horizontal green line shows a less stringent threshold based on false discovery rate (FDR) correction. In the qq-plots, the expected (X-axis) versus the observed (Y-axis) − log10 of association p-values are reported Nut illustrations, Manhattan and QQ-plots of nut size (A), nut maximum caliber (B), nut minimum caliber (C), and nut perimeter (D). The schematic illustrations on the left provide a visual reference for how each trait was measured. In the Manhattan plots, the X-axis shows the genomic position of the SNPs in the genome, and the Y-axis reports the negative log base 10 of the p-values. The solid green horizontal lines represent the significance threshold based on the Bonferroni correction, which was applied at a p-value of 0.05. When relevant, a dashed horizontal green line shows a less stringent threshold based on false discovery rate (FDR) correction. In the qq-plots, the expected (X-axis) versus the observed (Y-axis) − log10 of association p-values are reported The proportion of PVE by the associated SNPs ranged from 5.93% to 23.28%, with the strongest association observed for nut perimeter on chromosome 7 (23.28%), suggesting that the detected loci contribute moderate effects on the studied traits (Table 5). Furthermore, all significantly associated SNPs show PIC values above 0.10, ranging from 0.14 to 0.36 (Table 5), indicating that they possess sufficient discriminative power. Table 5List of SNPs associated with nut traits. Chromosome, physical position, significance level (p-value), phenotypic variance explained (PVE, %), polymorphism information content (PIC), and the interval based on half LD decay used to look for candidate genes surrounding the associated SNPs are shownTraitChromosomePosition (bp)p-valuePVE (%)PICInterval Start (bp)Interval End (bp)Size119,066,4645.88E-076.570.2218,920,38119,212,547Size434,896,4623.09E-075.930.3634,671,77835,121,146Size114,371,0246.61E-077.350.324,298,2964,443,753Maximum caliber37,456,3831.68E-0710.860.247,237,7747,674,993Maximum caliber518,926,3681.96E-1415.400.1618,741,49819,111,239Minimum caliber62,670,0007.68E-1420.620.292,542,4062,797,594Minimum caliber99,387,6307.87E-088.790.199,299,9989,475,262Minimum caliber912,275,3322.08E-088.730.2912,187,70012,362,964Perimeter721,048,0227.05E-1023.280.1420,872,95721,223,088Perimeter9831,7051.12E-066.240.21744,073919,337 Table 5List of SNPs associated with nut traits. Chromosome, physical position, significance level (p-value), phenotypic variance explained (PVE, %), polymorphism information content (PIC), and the interval based on half LD decay used to look for candidate genes surrounding the associated SNPs are shownTraitChromosomePosition (bp)p-valuePVE (%)PICInterval Start (bp)Interval End (bp)Size119,066,4645.88E-076.570.2218,920,38119,212,547Size434,896,4623.09E-075.930.3634,671,77835,121,146Size114,371,0246.61E-077.350.324,298,2964,443,753Maximum caliber37,456,3831.68E-0710.860.247,237,7747,674,993Maximum caliber518,926,3681.96E-1415.400.1618,741,49819,111,239Minimum caliber62,670,0007.68E-1420.620.292,542,4062,797,594Minimum caliber99,387,6307.87E-088.790.199,299,9989,475,262Minimum caliber912,275,3322.08E-088.730.2912,187,70012,362,964Perimeter721,048,0227.05E-1023.280.1420,872,95721,223,088Perimeter9831,7051.12E-066.240.21744,073919,337 List of SNPs associated with nut traits. Chromosome, physical position, significance level (p-value), phenotypic variance explained (PVE, %), polymorphism information content (PIC), and the interval based on half LD decay used to look for candidate genes surrounding the associated SNPs are shown Based on half LD decay up- and downstream the associated SNPs (Table 3), intervals were selected to look for candidate genes surrounding the 10 associated SNPs (Table 5). The closest gene to each of the associated SNPs, their genomic position and their description are shown in Table 6. Table 6Protein blast results of the peptide sequence of the C. avellana genes closest to the SNPs associated to nut traits against the genome of Arabidopsis thaliana. The number of candidate genes in half LD decay distance up- and downstream of the candidate loci is reported in the last columnTraitC. avellana – Gene IDGenomic position of geneGenomic position of associated SNP (distance to gene)A. thaliana – Gene IDA. thaliana - Gene descriptionQuery coverE-valueGenes in LD decay distanceNut maximum caliber Cav03g07620 Chr3, 7,461,550–7,471,5007,456,383 (+ 5167 bp)OAP00664.1 SPK1 94%044Nut maximum caliber Cav05g10360 Chr5, 18,935,016–18,937,12518,926,368 (+ 8648 bp)NP_850380.1myosin-G heavy chain-like protein51%2.00E-5919Nut minimum caliber Cav06g03190 Chr6, 2,676,086–2,676,7662,670,000 (+ 6086 bp)NANo significant similarity foundNANA33Nut minimum caliber Cav09g09750 Chr9, 12,286,994–12,288,65312,275,332 (+ 11662 bp)NP_172346.1Presenilin-199%03Nut minimum caliber Cav09g08540 Chr9, 9,387,663–9,389,2939,387,630 (+ 33 bp)NP_191315.1N-acetyl-l-glutamate kinase98%6.00E-1559Nut perimeter Cav07g15600 Chr7, 21,044,588–21,047,06321,048,022(−3434 bp)NP_188199.1uncharacterized protein AT3G1578058%9.00E-1222Nut perimeter Cav09g01010 Chr9, 833,665–836,738831,705 (+ 1960 bp)OAP11528.1 SCPL51 99%020Nut size Cav01g16120 Chr1, 19,059,680–19,064,11319,066,464(−6784 bp)CAD5317706.1unnamed protein product72%0.057Nut size Cav11g04070 Chr11, 4,378,421–4,381,0484,371,024 (+ 7397 bp)NP_001320169.1VQ motif-containing protein23%9.00E-2213Nut size Cav04g24360 Chr4, 34,899,618–34,902,39334,896,462 (+ 3156 bp)NP_194497.1heat shock protein 2199%2.00E-4262 Table 6Protein blast results of the peptide sequence of the C. avellana genes closest to the SNPs associated to nut traits against the genome of Arabidopsis thaliana. The number of candidate genes in half LD decay distance up- and downstream of the candidate loci is reported in the last columnTraitC. avellana – Gene IDGenomic position of geneGenomic position of associated SNP (distance to gene)A. thaliana – Gene IDA. thaliana - Gene descriptionQuery coverE-valueGenes in LD decay distanceNut maximum caliber Cav03g07620 Chr3, 7,461,550–7,471,5007,456,383 (+ 5167 bp)OAP00664.1 SPK1 94%044Nut maximum caliber Cav05g10360 Chr5, 18,935,016–18,937,12518,926,368 (+ 8648 bp)NP_850380.1myosin-G heavy chain-like protein51%2.00E-5919Nut minimum caliber Cav06g03190 Chr6, 2,676,086–2,676,7662,670,000 (+ 6086 bp)NANo significant similarity foundNANA33Nut minimum caliber Cav09g09750 Chr9, 12,286,994–12,288,65312,275,332 (+ 11662 bp)NP_172346.1Presenilin-199%03Nut minimum caliber Cav09g08540 Chr9, 9,387,663–9,389,2939,387,630 (+ 33 bp)NP_191315.1N-acetyl-l-glutamate kinase98%6.00E-1559Nut perimeter Cav07g15600 Chr7, 21,044,588–21,047,06321,048,022(−3434 bp)NP_188199.1uncharacterized protein AT3G1578058%9.00E-1222Nut perimeter Cav09g01010 Chr9, 833,665–836,738831,705 (+ 1960 bp)OAP11528.1 SCPL51 99%020Nut size Cav01g16120 Chr1, 19,059,680–19,064,11319,066,464(−6784 bp)CAD5317706.1unnamed protein product72%0.057Nut size Cav11g04070 Chr11, 4,378,421–4,381,0484,371,024 (+ 7397 bp)NP_001320169.1VQ motif-containing protein23%9.00E-2213Nut size Cav04g24360 Chr4, 34,899,618–34,902,39334,896,462 (+ 3156 bp)NP_194497.1heat shock protein 2199%2.00E-4262 Protein blast results of the peptide sequence of the C. avellana genes closest to the SNPs associated to nut traits against the genome of Arabidopsis thaliana. The number of candidate genes in half LD decay distance up- and downstream of the candidate loci is reported in the last column 21,048,022 (−3434 bp) 19,066,464 (−6784 bp) The 232 genes found in half LD decay up- and downstream of each associated SNP are reported in Additional file 10. Since all traits are correlated to each other (Additional file 9), a unified GO analysis considering all the genes related to the four traits was conducted. This results in 32 terms with significantly overrepresented GO categories (p < 0.01). In the molecular function category, 19 significantly overrepresented terms are identified, four of which are related to methyltransferase activity. In the biological process category, nine terms were significantly overrepresented. Within this category, we focused our subsequent analysis on the term flavonol biosynthetic process, which, although not the one with the lowest p-value, is the only term for which a clear biological interpretation could be established based on the literature (Additional file 11). Discussion Diversity analysis revealed genetically distinct samples in an otherwise highly diverse and admixed dataset Hazelnut is known for its phenotypic and genetic diversity, which plays a crucial role in its environmental adaptability and its agricultural value. Despite the economic and ecological significance of this species, relatively few studies have comprehensively examined the origin of the diversity within and between hazelnut populations. In this study, we addressed this gap by investigating the genetic diversity of a global collection of C. avellana cultivated varieties and wild types using genome-wide SNP markers. We observed a tendency for correspondence of clusters according to the provenance of the accessions, which is in agreement with other studies describing hazelnut genomic diversity [22]. The probable origin of samples with unknown provenance could be deduced by their clustering with nearby varieties of known origin. The distinct clustering of Turkish and Spanish varieties likely reflects independent domestication histories and limited gene flow between these countries (Fig. 2). This pattern aligns with previous findings that emphasize the existence of regionally differentiated gene pools within cultivated hazelnuts [62]. Only a partial distinction of wild types from hazelnut cultivated varieties could be observed within the clustering alongside Italian and Spanish varieties. Separation of wild types from cultivars was obtained when studying cultivated and wild material sampled in Spain [63], Italy [7], Portugal [64, 65] and Slovenia [66], probably due to domestication and selective breeding in cultivated varieties. Sivri and Sivri A, despite having the same varietal name, were shown to be genetically different (Additional file 2), with Sivri grouped in DAPC cluster 2 and Sivri A in DAPC cluster 1 (Fig. 2c). Similar observations were reported in previous studies, in which Çakıldak clones were separated into distinct clusters, highlighting that hazelnut varieties, despite sharing a common name, can encompass genetically heterogeneous groups [67]. In other cases, despite being in the same DAPC cluster, plants with the same variety name were more genetically similar to other varieties than to their own namesakes. For example, one Karafindik was more similar to a Yomra than to another Karafindik, one Camponica was closer to a Tonda di Giffoni than to another Camponica, and a Vermellet was more similar to a Sugranyes than to other Vermellet (Additional file 2). This finding may result from an incorrect assignment of variety names or from cases of homonymy, as previously observed in other studies [68]. Furthermore, we identified 141 clones within our collection, likely resulting from the repeated sampling of the same variety. As these genetically identical individuals do not contribute additional information and may introduce bias in downstream analyses, they were excluded from our reduced dataset. Among the three DAPC clusters that best described the diversity in our collection, DAPC Cluster 1 contained most of the Italian samples together with a part of the Spanish samples, while DAPC Cluster 3 was composed of predominantly Spanish samples. These Clusters showed some overlaps, suggesting shared genetic characteristics between these populations. This is also reflected in the low number (for cluster 3 even absence of) of private alleles in these clusters. The slightly higher nucleotide diversity of cluster 3 indicates a moderately higher diversity among Spanish samples, which are probably getting part of their diversity from alleles shared with other groups. The small number or even lack of private alleles in DAPC Clusters 1 and 3, respectively, indicates that there is a higher gene flow among the samples as opposed to DAPC Cluster 2, where the high number of private alleles indicates a constraint to gene flow. Despite the indications for a moderate gene flow between samples from DAPC Clusters 1 and 3, DAPC Cluster 3 shows a higher FIS value. This hints at a deficit of heterozygotes relative to Hardy–Weinberg expectations and might be due to population sub-structure (Wahlund effect), non-random mating, or localized inbreeding. Even with some overlaps of the clusters and the overall low levels of structure and fixation, the Eurasian samples (DAPC Cluster 2) formed a separate genetic group, marked by a high number of private alleles, highlighting that they have been largely underutilized in breeding programs. The genetic distinction of Eurasian varieties on a broad geographic scale confirmed the findings of recent studies. These include the differentiation between European and Asian Corylus lineages [69], the genetic separation between Turkish and Italian C. avellana varieties [70], the identification of a unique pool of C. avellana genetic resources within Georgian varieties in a global collection [22], and the demonstration that Spanish germplasm from Asturias constitutes a distinct genetic pool compared with varieties from Catalonia, Italy, Türkiye, and the USA [71]. Furthermore, chloroplast-based variant analyses reported a differentiation for European versus Caucasian material [72], although chloroplast DNA captures maternal lineage and may not reflect recent nuclear admixture. Additionally, the outcrossing nature of hazelnut combined with the wind-pollination likely facilitate gene flow between species and populations, further contributing to the observed genetic diversity and admixture. Eurasian varieties harbor unique genetic characteristics This distinctiveness of Eurasian samples raised the question about the underlying factors that contribute to this uniqueness. We identified five ancestral populations partly aligning with proposed domestication events for hazelnut, reported in the British Isles, Central Europe, Spain, Italy, the Black Sea region, and Iran [3, 4]. However, no population retained the unadmixed genomic signature of a single ancestral type. Of all groups, the Eurasian samples came closest to a pure ancestral lineage, as they predominantly shared a single major ancestral component, indicating limited admixture with samples from other regions (Fig. 3). By contrast, most non-Eurasian samples displayed admixture, consistent with higher levels of gene flow. Over time, gene flow between ancestral populations was driven by both natural interbreeding and intentional human-mediated breeding (Fig. 3). Most notably, varieties from Italy and Spain are known for being genetically similar [5, 73]. In our dataset, this was supported by the overlap and very low FST value observed between DAPC Clusters 1 and 3, which predominantly contain Italian and Spanish samples (Fig. 2c, Additional file 2, Table 2). The cultivation of hazelnuts was probably introduced from the eastern Mediterranean basin into Spain and southern Italy by the Romans through a significant exchange of germplasm between Italy and Spain [3] which facilitated the cultivation of hazelnut varieties that thrive in both regions. The lower admixture level in Eurasian samples might be partly attributed to gene flow barriers between geographical regions [3], but probably mainly reflects the largely absent breeding programs that cross Eurasian samples with varieties from other provenances. Eurasian varieties are not commonly cultivated in Europe, partially due to their lack of climate adaptation, but also due to differing breeding priorities. In Türkiye, for instance, the focus has been on improving traits like nut and kernel size, but this has also introduced undesirable characteristics such as empty nuts, which may reduce their appeal in European markets [26]. The lack of significant improvement in flavor, a trait highly valued in European markets, may have made Turkish varieties less attractive to European consumers [26]. These untapped allele pools, however, are particularly interesting for hazelnut improvement. An assessment from seed lots collected from Georgia and Azerbaijan identified 79 individuals resistant and 19 individuals highly tolerant to Eastern Filbert Blight (EFB) [74], highlighting their unexplored potential. Investigating introgression indicated that at least 2 migration events happened. Results from introgression analysis via the TreeMix model with m = 2 seem to confirm the findings from the other analyses. The model (m = 2) suggests that C. avellana populations are structured into two main groups: (i) Mediterranean populations (Italy, Greece, Spain, Wildtype) which cluster closest to the root, and (ii) more northern and western European populations (France, England, Germany, Former Yugoslavia), together with the United States and breeding lines, which appear derived from the Mediterranean gene pool. In addition to these two groups, differentiation and a certain degree of isolation of Eurasian populations can be inferred as well. The two major migration events shown in the tree indicate that genetic material from Former Yugoslavia might have been integrated into the southern/Mediterranean populations while gene flow also occurred from southern populations into northern Europe. However, the fact that some groups are small (for example, only two samples from Former Yugoslavia are available) makes it necessary to be careful to draw unrestricted conclusions about the indicated migration events. Population structure in our hazelnut panel was assessed with ADMIXTURE and DAPC, which provided complementary results. ADMIXTURE indicated many individuals carry ancestry from multiple Ancestral Clusters, evidencing gene flow, whereas DAPC identified fewer clusters, assigned most individuals to a single group and highlighted the distinctness of Eurasian varieties. This contrast reflects differences in method design. While ADMIXTURE estimates continuous ancestry proportions, DAPC maximizes among-group and minimizes within-group variance [75], making DAPC better for detecting discrete, island-like structure and ADMIXTURE more sensitive to admixture and continuous differentiation. Together, they support a broad split into three DAPC clusters, finer substructure into five Ancestral Clusters, and widespread admixture across our populations. The differentially selected loci in Eurasian samples provide novel candidate targets for genetic improvement We identified four loci under diversifying selection between Eurasian and the other samples and investigated the closest genes to each locus as well as overrepresented gene categories within half LD decay distance up- and downstream. The influence of candidate variant loci on these genes and their functional interpretation are merely hypothesis-generating and will need validation through experimental approaches to assess gene expression and protein levels. Here, we suggest these genes as a starting point for future investigations. The strongest candidate outlier locus on chromosome 6 (position 2,083,200) (Fig. 6; Table 4) is located in the coding sequence of a probable calcium uniporter (Cav06g02620). In Arabidopsis, Mitochondrial calcium uniporters (MCUs) are involved in the detection of environmental stress by modulating stromal Ca2+ in chloroplasts [76]. Upstream of the chromosome 1 outlier, a BTB/POZ protein family domain-containing protein was found (Table 4). This protein family is known for its involvement in plant growth, development [77] and defense regulation [78]. tRNA methyltransferases, one of which was found downstream of the outlier locus in chromosome 7, are crucial for nucleoside modifications on transfer RNA. Such modifications were found to be involved in stress tolerance and yield in Arabidopsis and crop species [79]. Finally, a probable CCAAT-binding factor was found upstream of the outlier variant on chromosome 11. These transcription factors have been shown to be involved in stress responses in Arabidopsis and crop plants [80] and for the regulation of flowering in plants [81, 82]. In the wider range of LD decay distance around the outlier candidates, we found ovule development and tubulin-assembly proteins to be among the overrepresented categories with the highest number of genes (Additional file 7). Tubulins form microtubules and were found to be able to induce remodeling of the plant microtubule cytoskeleton in response to stress [83], Many of the genes close to the SNPs under selection have been shown to play important roles in plant development and stress responses in A. thaliana, suggesting that diversifying selection of these genes might constitute an adaptation mechanism to local environmental conditions in the Eurasian samples. Experimental evidence will be required to support this hypothesis. We suggest these genes together with the whole set of genes in the LD decay range around the outlier loci (Additional file 6) as a starting point for future studies aimed at exploring the breeding and adaptation potential of the distinct Eurasian varieties. Future validation should include transcriptomic profiling to evaluate expression levels of these genes, as well as functional assays such as overexpression or knockdown experiments to test their roles in stress tolerance and phenological adaptation. Marker trait associations to accelerate genetic gain in hazelnut breeding Nut size and shape are among the most important traits considering yield and requirements of the industry. Unlike conventional phenotyping approaches that rely on manual measurements using digital calipers [84], the photographing and digital measuring (Additional file 1) employed in our study enhances reproducibility, as the original images can be reanalyzed or even used for new studies. The absence of provenance-based significant differences between nut traits (Additional file 8) suggests limited phenotypic differentiation between populations and is consistent with the low genetic structure observed. A GWAS of 151 both genotyped and phenotyped varieties allowed us to determine 10 SNPs to be significantly associated with important nut traits (Fig. 7; Table 5). The deflated genomic inflation factors (λ = 0.73–0.87) observed in FarmCPU models reflect stringent correction for population structure, favoring specificity over sensitivity. While this conservative adjustment may have reduced our power to detect minor associations, it strengthens confidence in the loci that surpassed the significance threshold. Our phenotypic data was collected during the single growing season of 2022, and the varieties originated from three distinct environments (Türkiye, Italy, and Spain) providing geographic and climatic variability within the dataset. However, potential genotype-by-environment (G × E) interactions influencing GWAS results cannot be excluded. Measured nut traits might have been affected by environmental conditions, potentially masking true genetic associations. At the same time, false-positive associations cannot be excluded in case of biased influence of environmental variables on certain genotypes. The stringent association thresholds we applied, evidenced by the deflated genomic inflation factors, reduced the likelihood of spurious correlations. The moderate effects of the associated SNPs, highlighted by the PVE values (6–23%) (Table 5), is in line with the polygenic characteristic of nut morphological traits, where multiple loci of small to moderate effect collectively shape the phenotype. At the same time, the relatively high value observed for the SNP in chromosome 7 associated with nut perimeter suggests the presence of a locus with a particular relevance. Besides that, the PIC values observed for all significant SNPs (0.14–0.36) confirm that these markers are informative (Table 5), as this suggests that the associations reported here are supported by markers with adequate allele frequency distribution, avoiding biases that could arise from very rare or very common variants. Whereas prior work established the feasibility of GBS-derived SNPs for DNA typing in hazelnut [24], our study extends this to population structure, selection scans, and genome-wide association with nut morphology. Few large-scale population studies have investigated the global genetic diversity of C. avellana, and to our knowledge, only one study to date determined Quantitative Trait Nucleotides (QTNs) via GWAS in C. avellana, focusing on a moderate number of only Turkish varieties [85]. In this study, a SNP on chromosome 6 at position 28,735,711 was identified to be associated with nut size. Our study identified three SNPs associated with nut size located on chromosomes 1, 4, and 11. This discrepancy likely reflects the allelic differences in the populations studied, and the larger diversity represented by the global collection of hazelnuts. The closest genes for half of the trait-associated variants showed a high (> 90%) query cover with an Arabidopsis functionally annotated gene (Table 6). Among them, SPIKE1 (SPK1), found downstream of the nut maximum caliber-associated variant on chromosome 3 (position 7,456,383), is responsible for normal cell shape (e.g. epidermis cells) and tissue development, suggesting it may influence nut size and shape in hazelnut. A gamma secretase subunit (Presenilin-1) with unclear function [86] was found downstream of the nut minimum caliber-associated SNP on chromosome 9 (position 12,275,332). Since proper endomembrane trafficking, one of Presenilin’s presumable activities [87, 88], is vital during seed development, it cannot be excluded that the gene can influence nut caliber. A second significantly associated variant on chromosome 9 (position 9,387,630) was found to lie in the presumable promoter region of an N-acetyl-l-glutamate kinase, involved in Arginine biosynthesis [89] playing a key role in plant growth and development. A Serine Type Protease (SCPL5), suggested to be involved in nitrogen remobilization [90], was found downstream a variant associated with nut perimeter on chromosome 9 (position 831,705). Recent evidence indicates that serine proteases may have specific roles during seed development [91]. A heat shock protein (HSP21) involved in maintenance of chloroplast and seedling development during heat stress [92, 93] was found downstream of the nut size-associated variant on chromosome 4 (position 34,896,462). In addition to the closest gene to each variant, the genes in LD distance around the associated loci were investigated via GO analysis (Additional file 11), offering insights into the biological processes and molecular functions that might be influencing the nut traits. Among the 32 overrepresented categories, 4 are associated to methyltransferase activity, including quercetin 3-O-methyltransferase (GO:0030755), luteolin O-methyltransferase (GO:0030744), myricetin 3’-O-methyltransferase (GO:0033799), and caffeate O-methyltransferase (GO:0047763) (Additional file 11). Recent studies showed that the overexpression of O-methyltransferase 1 in Solanum lycopersicum enhanced seed germination [94], establishing a link between this type of enzyme and seed development. Furthermore, terms like ketoreductase activity (GO:0045703) and acetoacetyl-CoA reductase activity (GO:0018454) point to enzymatic functions involved in the reduction of ketones and acetoacetyl-CoA, which are crucial for plant growth [95]. The significant enrichment of flavonol biosynthetic processes (GO:0051555) could indicate that flavonoids might affect some aspects of nut traits. Indeed, flavonoids are a key component of the skins of C. avellana [96]. The genes around variants found to be significantly associated with important phenotypic traits, like nut size and shape, are suggested as candidates that might affect the phenotypic traits or whose expression might be influenced by genetic differences between varieties. Our hypotheses still require validation through gene expression analyses (e.g., transcriptomics, RT-qPCR) or gene-editing techniques (e.g., CRISPR-Cas9). Nevertheless, examining candidate genes located within the LD decay distance of associated SNPs provides a sound starting point for identifying targets to improve desirable traits by breeding. Conclusions We aimed to contribute to the understanding of hazelnut genetic diversity by analyzing germplasm across North America, Asia, and Europe using genome-wide SNP markers. Our objective was to establish a robust foundation for further research into genotype-environment interactions and evolutionary processes in hazelnuts. The insights into the uniqueness of Eurasian varieties can stimulate research about their genetic distinctiveness and how this potential can be used in breeding efforts. When combined with the accurate sequencing of a diverse and representative set of varieties, GWAS are a valuable tool to accelerate breeding programs. This is especially relevant given the lengthy timeline of approximately 17 years for conventional breeding methods to develop and release new cultivars [97]. The results provided insights into the genetic and phenotypic diversity of C. avellana as well as genome-trait interactions. These findings provide a basis for future applications, as the genomic regions associated with the nut traits can be further studied to be used in marker-assisted selection to improve nut characteristics. Furthermore, parental choice can be better guided by prioritizing Eurasian varieties to increase genetic diversity when needed, ensuring the existence of a diverse genetic pool in C. avellana. Therefore, these results could be valuable resources for geneticists and breeders in developing improved hazelnut varieties. While our study deepened the knowledge about the genetic diversity of hazelnut and its association with key traits, some limitations should be acknowledged. Although our collection spans a wide geographic range, the moderate sample size and SNP density may have limited our ability to detect rare alleles or loci with small effect sizes. Additionally, GBS captures only a reduced representation of the genome, which means structural variants and regulatory regions outside the sequenced loci may have been missed. Our nut morphology data was collected during the single growing season of 2022, and since there is a strong year-to-year variation in nut development traits in hazelnut, our single-season dataset may not capture the full extent of phenotypic variability. Therefore, false-negative or false-positive associations cannot be completely excluded. Future studies should take this into account, preferably by collecting multiple-year data to get more robust results, or by introducing environmental covariates into the model and estimate random effects by using adequate statistical tests, such as best linear unbiased prediction (BLUP) in mixed models to mitigate noise. Furthermore, while our analysis showed that all significantly associated SNPs have PIC values above the informativeness threshold, incorporating PIC-based filtering in future GWAS in hazelnut can further refine association signals by exclusively prioritizing highly informative markers. Finally, while we identified SNPs under selection as well as SNPs associated with nut traits, further validation, such as transcriptomic approaches or gene editing, will be necessary to confirm their biological roles. Future studies using whole-genome sequencing coupled with broader sampling as well as trials in different environmental conditions that span across multiple years will be essential to address these limitations and to better inform breeding strategies. Supplementary Information Additional file 1. Sequencing statistics and passport information of the samples used in this study. Additional file 1. Sequencing statistics and passport information of the samples used in this study. Additional file 2. NJ-tree rooted on the Tombul variety. Varieties names are shown, and they are colored according to their geographical provenance. Numbers at the nodes indicate bootstrap values (2000 bootstrap resampling). Branch lengths are proportional to genetic distance (scale bar = 0.01), with longer branches indicating greater divergence between samples. Additional file 2. NJ-tree rooted on the Tombul variety. Varieties names are shown, and they are colored according to their geographical provenance. Numbers at the nodes indicate bootstrap values (2000 bootstrap resampling). Branch lengths are proportional to genetic distance (scale bar = 0.01), with longer branches indicating greater divergence between samples. Additional file 3. Values of cross-validation error for different values of ancestral clusters of the ADMIXTURE analysis, showing that the optimal number of clusters is 5. Additional file 3. Values of cross-validation error for different values of ancestral clusters of the ADMIXTURE analysis, showing that the optimal number of clusters is 5. Additional file 4. (a) Residual Plot based on TreeMix analysis for models with 0–5 migration edges (m). Residuals indicate the difference between observed and fitted covariance. Colors represent the strength and direction of residuals, with positive residuals indicating an underestimation of the observed covariance, and negative values representing an overestimation. (b) Model likelihood and variance explained as a function of migration edges. Top: Mean log-likelihood (±SD) across 5 iterations, proportion of variance explained, and the 99.8% confidence threshold. Bottom: Δm values showing the improvement in model fit with increasing migration edges, used to determine the optimal number of migration events. Additional file 4. (a) Residual Plot based on TreeMix analysis for models with 0–5 migration edges (m). Residuals indicate the difference between observed and fitted covariance. Colors represent the strength and direction of residuals, with positive residuals indicating an underestimation of the observed covariance, and negative values representing an overestimation. (b) Model likelihood and variance explained as a function of migration edges. Top: Mean log-likelihood (±SD) across 5 iterations, proportion of variance explained, and the 99.8% confidence threshold. Bottom: Δm values showing the improvement in model fit with increasing migration edges, used to determine the optimal number of migration events. Additional file 5. The four potential candidate loci under diversifying selection (qvalue < 0.05). Additional file 5. The four potential candidate loci under diversifying selection (qvalue < 0.05). Additional file 6. List of genes found in half LD decay distance up- and downstream of the four potential candidate loci under diversifying selection. Additional file 6. List of genes found in half LD decay distance up- and downstream of the four potential candidate loci under diversifying selection. Additional file 7. GO enrichment analysis results of the candidate genes in LD-decay interval distance from the potential candidate loci under diversifying selection (p < 0.01). Additional file 7. GO enrichment analysis results of the candidate genes in LD-decay interval distance from the potential candidate loci under diversifying selection (p < 0.01). Additional file 8. Box plot representation of nut size (a), perimeter (b), maximum caliber (c), and minimum caliber (d) of the 151 hazelnut varieties used for the genome wide association study (GWAS). Additional file 8. Box plot representation of nut size (a), perimeter (b), maximum caliber (c), and minimum caliber (d) of the 151 hazelnut varieties used for the genome wide association study (GWAS). Additional file 9. Pairwise correlation chart analysis between all the analyzed nut traits. The diagonal panels display the distribution of each trait with fitted density curves (red lines). The lower panels show bivariate scatterplots of trait pairs with a linear regression line (red line), while the upper panels present Pearson correlation coefficients. Significant correlations are indicated with asterisks: *** means p < 0.001, while ** means p < 0.01. Additional file 9. Pairwise correlation chart analysis between all the analyzed nut traits. The diagonal panels display the distribution of each trait with fitted density curves (red lines). The lower panels show bivariate scatterplots of trait pairs with a linear regression line (red line), while the upper panels present Pearson correlation coefficients. Significant correlations are indicated with asterisks: *** means p < 0.001, while ** means p < 0.01. Additional file 10. List of genes found in half LD decay distance up- and downstream of the SNPs associated to the nut traits. Additional file 10. List of genes found in half LD decay distance up- and downstream of the SNPs associated to the nut traits. Additional file 11. GO enrichment analysis results of the candidate genes in LD-decay interval distance from the SNPs associated to the nut traits (p < 0.01). Additional file 11. GO enrichment analysis results of the candidate genes in LD-decay interval distance from the SNPs associated to the nut traits (p < 0.01).