In Situ GWAS in Two Range‐Edge Populations of European Beech (Fagus sylvatica L.) Identify Genetic Factors of Adaptation to Climate Change Authors: Lauren Clément, Andrea Modica, Isabelle Lesur Kupin, Christophe Plomion, Catherine Bodénès, Grégoire Le Provost, Sylvie Oddou‐Muratorio, Katrin Heer, Marie Mirouze, Marie‐Christine Carpentier, Bartosz Ulaszewski, William Marande, Jean‐Marc Aury, Simone Duprat, Corinne da Silva, Benjamin Noel, Corinne Cruaud, Karine Labadie, Joëlle Amselem, Johann Confais, Hadi Quesneville, Jean‐André Magdalou, Olivier Panaud, Ivan Scotti Published: 2026 ABSTRACT ABSTRACTIn forest tree populations, the timing of budbreak (TBB) depends on several factors, both environmental and genetic. The genomic architecture underlying this trait is still not fully characterized. So far, common garden experiments have highlighted a few genomic regions with little heritability, while the whole spectrum of TBB variation observed in the wild still remains unexplored. We performed an in situ genome‐wide association study (GWAS) to investigate the genetic and environmental bases of phenotypic variation in budbreak in two wild range‐edge populations of the temperate continental tree, European beech ( Fagus sylvatica L.) surveyed in two consecutive years. We found distinct significant loci between populations and among years, suggesting that local adaptation has occurred and that genotype‐by‐environment interactions contribute to the observed variation. The phenotypic variance explained by the significant loci is between 0.2% and 54% for field data and 0% to 59% for modelled data, which suggests that while some traits are polygenic, some others are influenced by a few loci with large effects in the population, which demonstrates the relevance of conducting in situ GWAS in multiple populations. Our findings provide novel insights into the complex genetic architecture of TBB in forest trees. The genetic diversity of budbreak control provides opportunities for selection that could be used through assisted migration to help prepare beech forests to face climate change.Keywords: beech ( Fagus sylvatica L.), budbreak, GWAS, local adaptation, QTL Introduction Trees follow a recurring annual developmental cycle that includes budbreak, flowering, fruiting and dormancy. These phenological stages must be precisely synchronized with the environment to ensure that growth and reproduction occur under optimal conditions (Singh et al. 2017). Phenological adaptation promotes species distribution and plays a key role in determining an individual's fitness (Chuine 2010). By allowing trees to track local climate, Timing of Bud Break (TBB) trait contributes to the length of the growing season, fecundity and water use efficiency (Battey 2000; Rowland et al. 2020). Trees must initiate budbreak early enough to maximize the growing season while avoiding late‐frost damage. To achieve this, they rely on environmental cues such as winter chilling (chilling requirement), spring warming (heat requirement) and day length (photoperiod). Multiple regulatory genes are involved in the signalling cascade from these cues to growth (Singh et al. 2017), but additional biological processes beyond the signalling cascade are likely involved, suggesting that TBB is a highly polygenic trait. European beech ( Fagus sylvatica L.) is a deciduous tree of the Fagaceae family and one of the dominant tree species in European temperate, mid‐altitude forests. Its natural distribution extends latitudinally from southern Norway to southern Italy and longitudinally from western Spain to Romania (Giesecke et al. 2007). European beech is increasingly vulnerable to the challenges posed by climate change. Dendrochronological surveys found that drought significantly reduced tree growth, with edge populations showing greater variability in growth, suggesting heightened sensitivity to water stress (Stjepanović et al. 2018). Schuldt et al. (2020) summarized the devastating effects of the extreme 2018 drought on European forest ecosystems, reporting that millions of trees across at least 2450 km2 were impacted. Following this event, over two‐thirds of European beech trees exhibited visible drought damage (Pfenninger et al. 2021). According to climate projections, drought events will become increasingly frequent and severe under future climate scenarios (Calvin et al. 2023), leading to a growth decline of European beech by up to 50%, depending on region and climate scenario (Martinez Del Castillo et al. 2022). Moreover, a north‐eastward shift in the species' range, driven by mortality in southern marginal populations, is forecasted, potentially leading to a reduction in the overall distribution area of up to 61% by 2100 (Saltré et al. 2015). With such severe changes, it becomes increasingly important for trees to avoid periods of drought and take full advantage of water availability in spring. In response, many forest species are shifting toward earlier spring phenology, with documented advances of up to 20 days (Körner and Basler 2010; Piao et al. 2019; Vitasse et al. 2009). Modelling studies have shown that TBB in beech may retain adaptive responses to warmer climates thanks to the genetic diversity present even within small local populations (Kramer et al. 2010), although the response to increased spring temperatures can be constrained by insufficient winter chilling (Fu et al. 2015). Range‐wide common‐garden experiments have shown that TBB can differ among populations, indicating that the trait is genetically controlled, heritable (Frewen et al. 2000; Howe et al. 2003; Hurme et al. 1997; Jermstad et al. 2001) and adaptive both at the continental (Robson et al. 2013) and at the local scale, if populations encompass steep environmental gradients (Gauzere, Teuf, et al. 2020). Despite these advances, the genetic architecture of TBB remains poorly resolved. Genomic approaches such as QTL mapping and GWAS allow the identification of loci associated with phenotypic variation without pedigree information and provide precise localization of genomic regions controlling traits (Bernard et al. 2020). However, three key aspects of this architecture remain unresolved. First, TBB is widely assumed to follow a polygenic architecture in trees across multiple taxa (Apuli et al. 2021; Bernard et al. 2020; Bükücü et al. 2020; Casasoli et al. 2004; Evans et al. 2014; Gabay et al. 2018; Jermstad et al. 2001; Kapoor et al. 2023; Lalagüe et al. 2014; Lesur et al. 2015; McKown et al. 2014, 2018; Meger et al. 2024; Riehl et al. 2023; Scotti‐Saintagne et al. 2004; Tylewicz et al. 2018; Watson et al. 2024). Yet, it remains unclear whether its variation is driven exclusively by many loci of small effect or also involves large‐effect loci. In Fagaceae, only a limited number of SNPs have been identified so far, each explaining a small proportion of the variance (0.4%–5.1%) (Lazic et al. 2024; Müller et al. 2017), including variants in CALLOSE SYNTHASE 1 and a CONSTANS‐like gene, leaving open the possibility that larger‐effect loci remain undetected. In contrast, a major QTL explaining 45.2% of the phenotypic variance has been reported in Prunus mume (Hsiang et al. 2024), suggesting that larger‐effect loci may exist in some populations. Second, it remains unclear to what extent the genetic architecture of TBB is shared across populations or instead reflects a population‐specific architecture. Although signals of local adaptation have been reported (McKown et al. 2018; Meger et al. 2021, 2024), the degree to which underlying loci are consistent or population‐specific remains unresolved. Third, the extent to which QTL detection depends on the reference genome, potentially leading to assembly‐dependent QTL detection, remains largely unexplored. In plants, it is well established that the choice of reference genome can strongly influence SNP detection and subsequent GWAS outcomes (Della Coletta et al. 2021; Gage et al. 2019). By identifying the genetic and phenotypic mechanisms that allow populations to cope with climatic stress, forest management and conservation efforts can be better tailored to support adaptation and long‐term sustainability under future climate scenarios. Across a species' range, populations respond to environmental gradients by changes in their vegetative and reproductive phenology, combining ‘instantaneous’ plastic responses with long‐term genetic differentiation (Jump et al. 2006) that in turn contribute to controlling population size, distribution and connectivity. The polygenic nature of the traits responding to selection makes it likely that local adaptation in each population relies on unique combinations of allele frequencies at underlying loci. In the context of adaptation to climate change, rear‐edge populations, located at the low‐latitude or low‐altitude limits of a species' distribution range could be suitable for the identification of adaptive genetic factors (Hampe and Petit 2005) that allow them to thrive in habitats that are suboptimal relative to the core of the distribution. In European beech, multiple studies have identified signatures of selection (Csilléry et al. 2014; Gauzere, Teuf, et al. 2020; Modica et al. 2024; Vitasse et al. 2009) and drought tolerance (Aranda et al. 2015) in rear‐edge populations. Patterns of variation in phenology at a local scale are well documented (Davi et al. 2011; Gauzere, Teuf, et al. 2020; Jean et al. 2023; Oddou‐Muratorio et al. 2018; Westergren et al. 2025): e.g., individuals from different stands within a population along a mountain slope can exhibit a TBB span as broad as that of populations along a spatially wider latitudinal gradient (Bernhardsson 2012; Gauzere et al. 2017; Gauzere, Klein, et al. 2020; Valen 1975). Hence, low altitude, rear‐edge populations may harbour alleles that are advantageous in drier conditions; identifying them may provide useful guidelines for management strategies such as assisted migration. This justifies the need for studies of genetic control of traits in individual populations. Such an approach can complement more commonly used common‐garden studies, which leverage a large amount of the extant genetic variation from multiple provenances, but can tell little about the actual genetic diversity available for adaptation in individual populations. In this study, we focused on two French beech forests located at the ecological margins of the species' distribution: La Massane forest, situated in the eastern Pyrenees in south‐western France, recognized as a UNESCO World Heritage Site; and the mixed forest of Mont Ventoux, located in south‐eastern France and a unique example of a highly diversified ecosystem, spanning from Mediterranean to Alpine ecological zones (also famous as a stage of Tour de France). We analysed the genomic architecture of TBB and, while confirming that this phenological trait is polygenic, we provide hints that its genetic control differs between populations; that it depends upon environmental conditions—thus suggesting genotype‐by‐environment interactions; and, additionally, that using genomic resources derived from different lineages leads to detecting different genomic architectures—suggesting that genome‐level structural variation affects the genetic control of traits. These findings confirm that genomic studies in situ can unravel a broad spectrum of genetic factors controlling trait variation that may be involved in local adaptation and become useful for assisted migration programs. Materials and Methods Study Species The distribution of beech extends latitudinally from southern Norway to southern Italy and longitudinally from western Spain to Romania (Giesecke et al. 2007). While it typically grows at altitudes between 1100 and 1900 m, it can also be found at much lower elevations, down to 110 m above sea level in some areas (Bontemps et al. 2013; Dulamsuren et al. 2017). This species provides key ecosystem services and supports high levels of biodiversity. In mainland France alone, forest trees are estimated to store 1.26 billion tonnes of carbon. Each year, French forest biomass absorbs approximately 60 million tonnes of CO2—equivalent to around 13% of the country's annual greenhouse gas emissions (IGN 2018). In addition to its ecological value, European beech is widely used in various industries, including furniture‐making, musical instruments and flooring (San‐Miguel‐Ayanz et al. 2016). This wide range of uses makes it an economically important species for forestry. Study Sites The study was conducted on two populations originating from two glacial refugia at the dry edge of the species' distribution range (Magri 2008; Magri et al. 2006): La Massane Forest Reserve and a stand on the northern slope of Mont Ventoux, both in Southern France. The former belongs to the Western phylogeographic cluster, whilst the latter sits in a contact between the Western and Central lineages (Postolache et al. 2021). La Massane Forest Reserve (42°28′41″ N, 3°01′26″ E) is a 336 ha protected area consisting mainly of European beech trees. It is located within the 11,396 ha beech forest of the Albères Massif in south‐west France. It was selected as an advanced research site for studying climate change, due to its status as a low‐altitude Mediterranean beech forest subject to significant environmental disturbances (Garrigue 2018). The area studied for this work covers only 9.3 ha and ranges in altitude from 645 to 740 m. Within that area, 150 beech trees were randomly selected among dominant and co‐dominant trees. Between 1976 and 2017, the average temperature was 12.4°C, with an average minimum of 8.1°C and an average maximum of 15.7°C. The lowest temperature ever recorded was −13°C in January 1985, and the highest was 37.3°C in August 2003 (Garrigue et al. 2018). Since this last publication, temperature records have been repeatedly broken, with a current record of 40.7°C in July 2023. During the period of interest for spring phenology (1st January to 1st May), the temperatures were between −0.1°C and 17.4°C for 2022 and between −2.5°C and 20.8°C for 2023 (see Figure S1 and Table S1). Mont Ventoux (44°10′28′′N, 5°16′44′′E) sits in southern France, peaks at 1910 m and is one of the 52 glacial refugia of European beech (Magri 2008; Magri et al. 2006). The beech population covers approximately 4000 ha (approximate estimate from the map reported in Barbero et al. (1978)) and ranges in altitude from 750 to 1700 m above sea level (Lander et al. 2021; Oddou‐Muratorio et al. 2024). On the northern slope, a mixed forest of beech, Abies alba Mill. and Quercus spp. spans uninterrupted from 900 to 1500 m.a.s.l, across a steep gradient of environmental conditions from continental to Alpine (Oddou‐Muratorio et al. 2024). The studied tree stand ‘N1’ covers 1.3 ha at 1020 m.a.s.l. on the northern slope (Bontemps et al. 2013). Within this plot, 162 out of the 164 identified reproductive trees (Oddou‐Muratorio et al. 2024) were sampled. Temperatures have been recorded at the stand level each hour since 2007. During the period of interest for spring phenology (1st January to 1st May), the temperatures were between −3.7°C and 13.1°C for 2009 and between −9.4°C and 14.2°C for 2010 (see Figure S1 and Table S1). Genome Assembly and Annotation A de novo genome assembly of a 157 years‐old tree, referred to as MSSB, was obtained using a combination of short‐ and long‐read sequences, optical mapping and Hi‐C scaffolding. This tree is located in a climate refugium at the species' xeric range margin in south‐western France (Saint‐Symphorien, 44°25′41.138″N 0°29′23.125″W). All the details of the protocols used for the assembly, as well as for the annotation of genes and transposable elements are given in Data S1. Phenological Survey Phenological surveys were carried out in each forest according to the protocols described below; dates of observation of a given phenological stage were recorded as ‘day of year’ (DOY). The protocols applied at the two sites differed in multiple ways; therefore, the whole phenotyping protocol is described separately for each stand (see below and Data S1, Figure S2 and Tables S2 and S3). Phenotyping in La Massane forest was conducted in two consecutive years, twice a week between March 4th and April 29th 2022 and between March 13th and April 30th 2023 (Figure S3). The survey was performed with binoculars. The Biologische Bundesanstalt für Land‐ und Forstwirtschaft, Bundessortenamt und Chemische Industrie (BBCH) (Finn et al. 2007; Meier 2001) scale was employed to record phenological stages (Figure S4). Phenological monitoring was conducted by the same observer and with the same frequency in both years, ensuring consistency in data collection. There was an average gap of 2 and 2.5 days between each field record for 2022 and 2023, respectively. Though not all trees have been directly observed at the stage BBCH 09, the short intervals between observations (Table S4) made it possible to assign a date for that stage for all trees as follows: if a tree had not reached the BBCH stage on a given day, but had exceeded that stage 2 days later, the stage was attributed the day between the two visits (I. Chuine, pers. comm.). This way, TBB was recorded for all trees. In particular, we recorded two different observation series: the ‘Observation days’ series, that only comprises the scoring of BBCH observed directly in the field; and the ‘Imputed days’ series, that also includes dates assigned as described above, in addition to the direct field scores. On Mont Ventoux, the phenotyping campaign took place in 2009 and 2010 on 162 and 163 individuals, respectively. In 2009, observations occurred biweekly from 23rd March to 4th May and amounted to 15 phenological records; in 2010 the phenotyping took place once a week from 8th April to 10th May for a total of eight records, as described in Jean et al. (2023). While the mean gap between observations (2.33 and 1.4 days) was smaller than in La Massane, observations were sparser in the time span where BBCH stages changed faster (Table S4). Therefore, a date in the intercalar days was too arbitrary. The observers carried out the observations from the ground using binoculars and assigned phenological stages based on the scale of Davi et al. (2011) (Figure S4): they divided the canopy of each tree into an upper and lower part of equal size, and each observer independently assessed the developmental state of the buds from different angles from the lower and the upper canopy. To obtain comparable measurements between Mont Ventoux and La Massane, we used tables 2 and 3 from the Appendix of Jean et al. (2023) and we first converted these five scores to as many univocal BBCH scores. Modelling the BBCH ~ Date Relationship The dynamics of budbreak can be conceptualized as a sigmoidal growth curve that relates BBCH stages to Day of the Year (DOY), where the empirical appearance of each stage can vary from tree to tree and from year to year, depending on local conditions and year‐to‐year variations of meteorological conditions (i.e., the process can slow down or speed up due to colder or warmer periods respectively during budbreak (Fu et al. 2012)). To smooth out these variations, we modelled the BBCH stages as a logistic function of the DOY by fitting a curve based on the observed BBCH stage—DOY pairs, thus estimating the expected date for BBCH 09. These modelled TBB values were then used as phenotypes for GWAS. In addition to providing modelled TBB values, the model allowed us to compute model slope a at the inflexion point to capture variability in the overall individual budbreak dynamics (Crispim et al. 2015; Duan et al. 2021). Sensitivity of the model to data quantity and quality was assessed (Appendix S2 and Table S5). Based on the BBCH scores recorded in the field and the modelled analysis (Appendix S2, Table S6), a total of 11 phenology related traits were then computed for La Massane forest and 7 for Mont Ventoux forest (Table S6): TBB: date at stage BBCH 09 is reported for the first time; measured in Julian days (‘Day Of Year’, DOY); TBB: date at stage BBCH 09 is reported for the first time; measured in Julian days (‘Day Of Year’, DOY); Degree days (DD) were computed as the cumulative sum of temperatures from 1st January until TBB, taking into account only days on which the average temperature was above 5°C, which is reported as the optimal threshold for Fagus sylvatica (Dantec et al. 2014), for each forest, independently; measured in Celsius degrees (°C) (Savolainen et al. 2007). Sensitivity analyses using alternative thresholds (3°C and 7°C) have been done (Figure S5); Degree days (DD) were computed as the cumulative sum of temperatures from 1st January until TBB, taking into account only days on which the average temperature was above 5°C, which is reported as the optimal threshold for Fagus sylvatica (Dantec et al. 2014), for each forest, independently; measured in Celsius degrees (°C) (Savolainen et al. 2007). Sensitivity analyses using alternative thresholds (3°C and 7°C) have been done (Figure S5); Earliness of TBB: for each year, each tree was assigned to one of two classes (early/late) with a classification performed using K‐means clustering (k = 2). This trait was computed twice: based on the distribution of TBB values and based on the distribution of DD values and for each year of survey; binary, a‐dimensional. These were computed with or without switching trees (see ‘Switching’ below); Earliness of TBB: for each year, each tree was assigned to one of two classes (early/late) with a classification performed using K‐means clustering (k = 2). This trait was computed twice: based on the distribution of TBB values and based on the distribution of DD values and for each year of survey; binary, a‐dimensional. These were computed with or without switching trees (see ‘Switching’ below); Switching: switching was defined as individuals reversing their earliness category (early vs. late) between 2022 and 2023 based on date (or DD) at stage BBCH09—derived field data (TBB and DD) (see Section 3). Differences in earliness class (early or late) between years (2023–2022) were identified for each tree and for trees for which earliness status changed, it was labelled as ‘switching’. The proportion of switchers was computed after excluding missing values. This phenotype has been assessed only on La Massane data because the phenotyping procedure was carried out using a different timeframe between the 2 years within Mont Ventoux forest; Switching: switching was defined as individuals reversing their earliness category (early vs. late) between 2022 and 2023 based on date (or DD) at stage BBCH09—derived field data (TBB and DD) (see Section 3). Differences in earliness class (early or late) between years (2023–2022) were identified for each tree and for trees for which earliness status changed, it was labelled as ‘switching’. The proportion of switchers was computed after excluding missing values. This phenotype has been assessed only on La Massane data because the phenotyping procedure was carried out using a different timeframe between the 2 years within Mont Ventoux forest; Earliness of TBB without switching trees: we removed trees that displayed a switching phenotype between 2022 and 2023 and re‐classified the remaining trees using the K‐means clustering for early/late as described above, using the distribution of the TBB and DD values; Earliness of TBB without switching trees: we removed trees that displayed a switching phenotype between 2022 and 2023 and re‐classified the remaining trees using the K‐means clustering for early/late as described above, using the distribution of the TBB and DD values; Modelled TBB: based on the regression of BBCH stages onto their respective DOY using a logistic curve, we computed modelled TBB values; measured in Julian days (‘Day Of Year’, DOY). See above for details of the interpolation method; Modelled TBB: based on the regression of BBCH stages onto their respective DOY using a logistic curve, we computed modelled TBB values; measured in Julian days (‘Day Of Year’, DOY). See above for details of the interpolation method; Modelled DD: DD value computed using modelled TBB; measured in Celsius degrees (°C). Modelled DD is calculated on the nearest integer value of the decimal integer value of modelled TBB, rounded down. See above for details of the interpolation method; Modelled DD: DD value computed using modelled TBB; measured in Celsius degrees (°C). Modelled DD is calculated on the nearest integer value of the decimal integer value of modelled TBB, rounded down. See above for details of the interpolation method; a: slope of the logistic model at the inflexion point. See above for details of the interpolation method. a: slope of the logistic model at the inflexion point. See above for details of the interpolation method. DNA Extraction, Sequencing and Filtering Steps DNA extractions were performed using the Qiagen DNeasy kit following the manufacturer's protocol, starting from leaf material that was instant‐frozen in liquid nitrogen immediately upon collection. DNA quality and quantity were assessed by fluorometric assay prior to shipment to the sequencing laboratory. Genome sequencing data for the 150 beech trees for La Massane and 162 for Mont Ventoux were generated using the Illumina platform (2 × 150 bp), achieving an average coverage depth of 15X. Raw‐sequence quality filtering, read mapping and SNP calling were carried out using the BeechMarkET pipeline as described in (Mishra et al. 2022) and available at https://github.com/ulaszewski/beechMarkET, based on the two available reference genome assemblies, Bhaga (Mishra et al. 2022) and MSSB (see Data S1 and methods for genome assembly). For the purposes of GWAS, further SNP filtering was carried out using vcftools (v0.1.16) and BCFtools (v1.10.2) (Danecek et al. 2021; Li 2011). Only biallelic SNPs were kept. Then we removed sites with more than 10% of missing genotypes. The remaining missing data were imputed using BEAGLE (v5.4) with default parameters (see Figure 1, block 1). Flowchart of the analysis including filtering the VCF files, GWAS pipeline and training/testing analysis. Colour‐code: Blue for VCF; orange for steps involving randomisation; red for deterministic filtering steps; violet for GWAS analyses; azure for post‐GWAS processing; water green for predictions; and light green for results. GWAS Method We developed the analysis workflow to study the genetic architecture of each trait using all combinations of data from the two populations, the two phenotyping years and the two genome references. Therefore, for each trait, we carried out eight GWAS tests (with the exception of switching, which was only defined for La Massane). For each of these combinations, the data were processed in two alternative ways: (i) (Figure 1, block 2) the Beagle‐imputed VCF was submitted to further filtering steps (MAF > 0.1, test for Hardy–Weinberg Equilibrium with p‐value > 10−3) to produce the Filtered Main VCF. The number of SNPs after all the filtering steps are recorded in the Table S7. For each population, the VCF was LD‐pruned using PLINK v1.9 (Chang et al. 2015) (window size 50 Kb; window step size: 10; r 2: 0.1) to reduce marker redundancy prior to ADMIXTURE analysis and then submitted to an ADMIXTURE (v1.3.0) (Alexander et al. 2009) analysis to identify potential genetic subgroups, testing a range of K from 1 to 10. The Filtered Main VCF was submitted to the actual GWAS using the FarmCPU model (Fixed‐and‐random‐model Circulating Probability Unification (Liu et al. 2016) as implemented in the rMVP package (v1.4.0) (Yin et al. 2021), Appendix S3). The GWAS were run using the function MVP() with a maximum of 1000,000,000 iterations of the FEM/REM cycle, 1000 permutations (maxLoop parameter), EMMA as variance decomposition method (Kang et al. 2008) and a static bin and a kinship matrix K estimated by MVP.Data(). All GWAS analyses were performed using a kinship matrix. Kinship was estimated following the VanRaden method (VanRaden 2008). The distribution of relatedness among all pairs of individuals was examined for each population using the resulting kinship matrix. Supplementary figures illustrating these distributions are provided (Figure S6). The analyses were carried without genetic subgroups, as no significant structure was found in the ADMIXTURE analyses (i.e., K = 1 for both populations). The significant signals have been identified using the Bonferroni threshold (0.05/k, where k is the number of SNPs used for each GWAS). The individual polygenic score (PGS; a.k.a. ‘breeding value’ (Lynch and Walsh 1998; The International Schizophrenia Consortium 2009)) was calculated as the weighted sum of genotypic assays multiplied by the estimated allelic effects: where G ij is the genotype of individual i for SNP j and β j is the estimated allelic effect. The observed phenotype was then regressed against the polygenic score as follows: The pseudo‐heritability was computed as the proportion of phenotypic variance explained by the model. It should be noted that, because the prediction is carried out on the same data as the estimation of model parameters (i.e., allelic effects), this estimate of pseudo‐heritability is not a reliable indicator of predictive power and is reported here only as a summary statistic describing the GWAS results. (ii) (Figure 1, block 3) To correctly assess the predictive power of the GWAS, we proceeded as follows: each group of quantitative traits with the same sample size was bootstrapped eight times (i.e., eight independent subsamples were randomly drawn from the global dataset). For each bootstrap, a subset of ⅔ of the phenotyped individuals were set aside as the ‘training’ data set, and the remaining ⅓ were set aside as the ‘testing’ data set. The training data sets were submitted to the same filtering steps and GWAS algorithm as in the point (i) above. For each training replicate, we then identified all GWAS signals that also appeared in at least one additional replicate; if the loci were in LD, we picked the one with the highest allelic effect. We finally used these signals (called ‘consensus’ signals) for the computation of PGS and pseudo‐heritability of each testing replicate, using the same approach as in point (i) above. Notice that, because we use smaller samples than in point (i), the power estimated is likely a lower bound for the ‘true’ predictive power of the full model described in (i). We therefore estimate predictive power in a conservative way. Annotation of Significant SNPs The closest genes to the significant SNPs have been identified using bedtools closest and bedtools intersect (v2.31.1) (Quinlan and Hall 2010). Annotation of the significant SNPs on the genes was performed using the SnpEff tool (v5.2) (Cingolani et al. 2012). Identification of Significant Regions (‘QTLs’) We observed some clustering of significant SNPs for several traits (see Section 3). In 32% of the FarmCPU runs of pseudo‐QTN were included as cofactors to control for the effect of previously identified SNP in the following iterations (see Table S8a–d). In instances where no pseudo‐QTN were identified, the algorithm may have detected loci in linkage disequilibrium with each other—that had similar or identical genotypes—as independent loci with similar, or identical, allelic effects. When FarmCPU failed to initiate, a standard GLM was used instead. However, it is well established that with such models, highly linked SNPs also tend to emerge, as the GLM cannot effectively discriminate among SNPs in strong linkage disequilibrium (Kaler et al. 2020). The inability of the algorithm to identify cofactors may be caused by our limited sample size, causing the confidence intervals of allelic effects to include zero—thus removing them from the list of potential cofactors. Therefore, a series of closely linked SNPs may appear as artefactually tightly linked QTL. For this reason, we considered it more parsimonious to associate clusters of significant SNPs in linkage disequilibrium with a single Quantitative Trait Locus (QTL), as in Lazic et al. (2024). Consequently, two QTLs were declared identical between two different analyses (assemblies, sites, or years) if the significant SNPs supporting them are in linkage disequilibrium (as defined above). Linkage Disequilibrium (LD) analysis was conducted using plink (v1.9) with the following parameters: ‐‐ld‐window‐kb 10,000, −‐ld‐window 99,999 and ‐‐ld‐window‐r2 0.2. The threshold used for clustering was r 2 > 0.75. A high LD threshold (r 2 > 0.75) was used within analyses to ensure that SNPs clustered into the same QTL are strongly correlated and likely represent the same underlying causal locus. In order to compare the QTLs detected between the two populations, we cross‐validated the detection of significant SNPs as follows: we subsetted the VCF file of each population × assembly with the significant SNPs identified in the other population on the same assembly. Then we used the allelic effects, estimated for a given trait in one population, using a given reference genome, to predict the values of the same trait in the other population, using the same reference genome. To do this, we used individual‐SNP allelic effects. We expected that, if the genetic bases detected in the two populations differ, the loci detected in one population would systematically lead to a poorer estimation in the other population than in the one in which the loci were detected (Appendix S4). Testing for GxE Interactions We identified distinct suites of QTLs driving the same trait in each combination of Year × Population (see Section 3). Whilst identifying different signals for the same trait across years hints that the same genotypes may respond differently to inter‐year climatic variation (i.e., different environments), especially in polygenic traits (Lynch and Walsh 1998), it does not directly demonstrate a direct interaction between genotype and environment (GxE). We adopted a multilocus linear mixed model (MLMM) inspired by Segura et al. (2012) in the form Phenotype ~ QTL + Year + (QTL × Year) + K + ε to understand the effect of GxE on a trait's architecture in order to test whether fixed interaction term (QTL × Year) explains a significant degree of the residual phenotypic variance. For our model we used as fixed terms genotype (i.e., QTLs) and year of the phenotype (i.e., Year); and as random effect the kinship matrix (i.e., K). The function relmatLmer() from the R package lme4qtl (Ziyatdinov et al. 2018) fitted two mixed linear models to each combination of trait × population × assembly (except those concerning earliness and switching), one with and one without the GxE term (Appendix S5). Model fits were compared through ANOVA using a likelihood ratio test. We chose the best fitting model on the basis of the AIC (Napier et al. 2023). We expect a better fit for the models that include GxE to explain residual variance, considering the extent of the inter‐annual climatic oscillations at each site. Cross‐Mapping Signals From Different Genome Assemblies To compare the signals obtained between the two genome assemblies, we used a modified version of the ‘Rosetta Stone’ method developed by Modica et al. (2024). In brief, for each significant MSSB SNP, we extracted a 1000‐bp neighbourhood region (500 bp up‐, 500 bp downstream) from the sequence of the MSSB assembly and BLASTed it onto the Bhaga assembly. Then, the positions in the Bhaga genome (the ‘Subject’ sequence in BLAST output) corresponding to the most‐likely alignment were recorded, and the position corresponding to the target SNP was identified as the Bhaga‐mapped position of the MSSB SNP. Graphics and Statistical Analysis Graphical outputs were obtained using the packages ggplot2 (Wickham 2016) and tidyverse (Wickham et al. 2019). All graphics and statistical analyses were performed with RStudio (v4.4.0). For the statistical analyses, a significance threshold of p < 0.05 was applied. The following tests were performed: normality was assessed using the Shapiro–Wilk test; normality was assessed using the Shapiro–Wilk test; differences between early and late budbreakers were tested with Student's t‐test; differences between early and late budbreakers were tested with Student's t‐test; weather variation was evaluated using an F‐test for intra‐year variance between years, and a paired t‐test to compare the means between years; weather variation was evaluated using an F‐test for intra‐year variance between years, and a paired t‐test to compare the means between years; to test whether switching exceeded random expectations, permutation tests (10,000 iterations) were conducted by randomly shuffling phenotypes between years; to test whether switching exceeded random expectations, permutation tests (10,000 iterations) were conducted by randomly shuffling phenotypes between years; directional asymmetry for switching was assessed using an exact binomial test (H0: p = 0.5). directional asymmetry for switching was assessed using an exact binomial test (H0: p = 0.5). Results MSSB Reference Genome Assembly and Comparison Between Assemblies A new, high‐quality chromosome‐scale genome assembly of Fagus sylvatica MSSB was generated. Following assembly and manual curation based on Hi‐C contact maps, we successfully reconstructed the twelve chromosomes of the Fagus sylvatica genome. The final assembly, totaling 487 Mb, comprised 158 scaffolds, with the 12 largest scaffolds corresponding to the chromosomes (N90 = 24.59 Mb). The total assembly size was slightly smaller than that of the previously published Fagus sylvatica Bhaga genome, also used in this study (Mishra et al. 2022). However, the Bhaga assembly showed a higher level of duplications (Figure 2A,B) compared with the MSSB assembly, which likely explains the observed difference in genome size. This was also supported by BUSCO analysis (Manni et al. 2021), which revealed a greater proportion of duplicated genes in Bhaga (Table 1). Finally, a comparison of the two assemblies demonstrated a strong conservation of the overall chromosomal structure across the twelve chromosomes (Figure 3). KAT plots for (A) Bhaga and (B) MSSB genome assemblies. TABLE 1Quality metrics and comparison of Bhaga and MSSB genome assemblies.Assembly Fagus sylvatica Bhaga Fagus sylvatica MSSBCumulative size535 Mb459 Mb# scaffolds1212Max size73.9 Mb60.6 MbN50 (L50)46,564,880 (5)38,910,952 (5)N90 (L90)32,559,338 (11)25,954,076 (11)# Ns0.6 Mb (0.12%)1.6 Mb (0.35%)# contigs6495218Max Size1.0 Mb14.1 MbN50 (L50)138,851 (1028)5,304,236 (29)N90 (L90)34,034 (4129)1,284,088 (95)Merqury Completeness77.03%78.11%Merqury (Qv)29.0 (1/800 bp)34.7 (1/2951 bp)BUSCO (C; D) N = 2326 C:97.7% [S:88.2%, D:9.5%] F:0.7%, M:1.6% C:98.5% [S:94.2%, D:4.3%] F:0.4%, M:1.1% Quality metrics and comparison of Bhaga and MSSB genome assemblies. C:97.7% [S:88.2%, D:9.5%] F:0.7%, M:1.6% C:98.5% [S:94.2%, D:4.3%] F:0.4%, M:1.1% Dot plot comparison between MSSB (x axis) and Bhaga (y axis) genome assemblies. Differences in Weather Between Years and Between Sites To assess whether interannual temperature dynamics differed between the surveyed years, we compared both the mean daily temperatures and their variability over the January–May period at each site. At La Massane, daily temperature variance between January and May differed significantly between 2022 and 2023 (F‐test, p = 0.007), indicating greater variability across years, whereas daily mean temperatures remained statistically similar (paired t‐test, p = 0.205). Similarly, at Mont Ventoux, temperature variance differed significantly between 2009 and 2010 (F‐test, p = 0.019), but no significant difference was detected in daily mean temperatures (paired t‐test, p = 0.997). Across both study sites, interannual comparisons revealed differences in variability but not in mean values. Diversity of Budbreak Phenotyping of budbreak showed that it is variable both within each population (between years) and between populations (Figure 4 and Table S9). TBB was between 94 DOY and 117 DOY in 2022 and between 101 DOY and 119 DOY in 2023, in La Massane. For Mont Ventoux, TBB was between 104 DOY and 124 DOY in 2009 and between DOY 119 and 130 in 2010. Distribution of TBB (left) and DD (right) for both forests, both years. Massane is shown at the top, Ventoux at the bottom. The colour code corresponds to the year: Green for 2022, orange for 2023, purple for 2009 and pink 2010. Figure 4 shows in addition that TBB distribution in both populations and both years is not unimodal. This led us to classify trees in both sites as ‘early’ and ‘late’ based on the K‐means analyses. In La Massane, the thresholds were 103 DOY in 2022 and 110 DOY in 2023, while these values were 109 DOY for 2009 and 121 DOY for 2010 in Mont Ventoux. As for DD, in La Massane, TBB occurred between 506°C and 770°C for 2022 and 646°C and 875°C for 2023, while in Mont Ventoux, these values ranged between 252°C and 431°C for 2009 and between 306°C and 378°C for 2010 (Figure 4 and Table S9). Based on K‐means analyses, in La Massane, the thresholds were 602°C in 2022 and 749°C in 2023, while these values were 326°C in 2009 and 327°C in 2010 in Mont Ventoux. Based on the K‐means clustering analysis, individuals were consistently separated into two distinct groups for both TBB and DD across all sites and years. The differences between the ‘early’ and ‘late’ are highly significant (t‐tests, p‐value < 5e‐12 to p‐value < 5e‐65). These results indicate that the clustering robustly distinguishes early versus late phenotypes in all years and sites considered (Results S1). At La Massane, there was a switch in TBB ranking between the years. Some trees were classified as early in 1 year and late in the other, or vice versa. We refer to these individuals as switching trees. Among the 132 trees phenotyped in La Massane over two consecutive years, 20.5% (27/132) exhibited a switch in TBB ranking. This proportion was significantly higher than expected under random variation (permutation tests, p < 0.0001). Switching was also strongly directionally biased, with 85.2% of individuals (23/27) transitioning from early to late budbreak (binomial test, p = 0.00031) (Figure S7). The same analysis has been done on DD (Figure S8). Logistic Models of BBCH Dates The modelled TBB values were significantly higher than field TBB, especially for Mont Ventoux (p‐value of the paired samples based on the Wilcoxon test, alternative ‘less’—La Massane: 0.007 (2022), 0.01 (2023); Mont Ventoux: 2 · 10−16 for both years). The modelled dates were later than the field dates in both years and type of time series. On La Massane, the mean offsets were + 0.41 DOY for the ‘Observation days’ series and + 0.76 for the ‘Imputed days’ series (Results S2); the mean R 2 of the curves ranged between 0.63 and 0.73. The different methods to downgrade the time series increased the offsets (between −2.5 and 0.66 DOY depending on year, type of time series and method of degradation) while slightly improving the R 2 (0.69–0.86) due to the fewer data points. In a graphical assessment of the individual combinations of the parameters of the logistic models, the four combinations of Populations × Year discriminated along the first PC (67% of the variance) even though they exhibited extensive overlaps; yet they were indistinguishable on the second PC, suggesting that the different phenotyping strategies had a modest impact in the shape of the curves (Figure S9 and Table S10). The offset was larger for Mont Ventoux than for La Massane (+5.38 DOY in 2009 and +2.71 DOY in 2010) and the R 2 of the curves spanned between 0.76 (2009) and 0.88 (2010). Downgrading the time series increased the offset up to 8.0 and 4.12 DOY in 2009 and 2010, respectively, with small changes in the mean R 2. GWAS A total of 62 and 72 significant SNPs were found across all trait and assembly combinations, for La Massane and Mont Ventoux, respectively (see Figure 5 and Table S8a–d and Figures S10–S67). The effect size of each SNP per trait ranged from 4.3% to 42.3% (Figure S68). Taking clustering into account (see methods), these correspond to 17 and 58 QTLs respectively, explaining between 0.2% and 54% for field data and 0% to 59% for modelled data of observed variance (Table S8a–d). Among them, 4 QTLs represent a region, varying in size between 3.6 kb and 853 kb; the remaining 71 QTLs are supported by single SNPs. Manhattan plots and QQ plots for TBB (BBCH 09 dates in Julian days), considering Bhaga assembly. Massane is shown at the top, Ventoux at the bottom. The colour code corresponds to the year: Green for 2022, orange for 2023, purple for 2009 and pink 2010. Overall, 12 out of the 67 GWAS analyses showed genomic inflation factors ranged between 0.9 and 1.1, indicating well‐calibrated statistical tests in a subset of analyses. Additional analyses were conducted to investigate patterns underlying deviations of the genomic inflation factor (λ) from the expected value of 1. Across all GWAS, λ values ranged from 0.511 to 1.366, with a mean of 1.078 and a median of 1.184. λ values in La Massane were significantly greater than 1 (Wilcoxon test, p = 1.48 × 10−7). In contrast, no significant deviation from 1 was detected for Mont Ventoux. No substantial differences were observed between genome assemblies, with similar median λ values for Bhaga (1.193) and MSSB (1.167) (see Table S11). Population structure analyses indicated a single genetic cluster in both populations (K = 1). In 24 out of 67 cases, a GLM was used instead of FarmCPU because no pQTN were identified. Estimation of Predictive Power La Massane There were no consensus SNPs for the field traits in 2022; the few consensus loci found for the modelled traits weakly correlated with the observed phenotypes. We retrieved only a few consensus loci for the field and modelled traits of 2023. The pseudo‐heritability obtained from all these loci was relatively small (0.05–0.074). In addition, the two assemblies performed similarly (Results S3 and Table S12a–j). Mont Ventoux We were able to find significant associations for all the combinations of Trait × Year × Assembly except for modelled_DD/2009/Bhaga. We found consensus loci only for 13 traits × assembly combinations out of 20, mostly for two out of eight replicates (Table S12k–t). The analyses on the 2009 phenotypes almost always yielded no pQTN and bore large genomic inflation factors (0.61–1.8). The traits which were successfully validated were field_TBB/2009/Bhaga (for which we found consensus loci in four out of eight replicates; H 2: 0.9%–16.5%) and a/2009/Bhaga (six replicates; H 2: 0%–13.6%). We found significant associations for all traits from 2010. Often, more than half of the replicates for a trait measured in 2010 had pQTN; the genomic inflation spanned from 0.47 to 2.15, skewed to low values and thus suggesting that there were several cases where some loci with relevant effect had gone undetected. The traits with highest pseudo‐heritability were field_TBB/2010/Bhaga (six out of eight replicates; H 2: 0%–17.6%), field_TBB/2010/MSSB (six out of eight replicates; H 2: 5.8%–11.9%), field_DD/2010/Bhaga (six out of eight replicates; H 2: 0.02%–11.0%) and modelled_TBB/2010/Bhaga (two replicates; H 2: 0%–15.2%). Comparison of QTL Between Populations, Years and Genome Assemblies To assess the extent to which environmental conditions, genetic background and trait modelling affected QTL detection, we compared the distribution of QTLs between years, populations, genome references and modelled vs. field data for each trait (Figure 6). A summary result for distances between all the QTLs, sorted by comparison, is in Table S13. Venn diagram showing the overlap of QTLs between sites and assemblies. Massane is shown on the left, Ventoux on the right, with the number of QTLs indicated in each section. In La Massane we found no QTL for 2022 and six for 2023; for Mont Ventoux, we found 33 QTLs for 2009 and 25 QTLs for 2010. No QTL was shared either between years within each population and nor between populations. In the same way, we compared the sets of QTLs found between the two populations. Figure 6 shows that there is no QTL in common between the two studied populations. The closest QTLs were separated by 490 kb and showed low linkage disequilibrium (r 2 < 0.1), well below the threshold used to define shared QTLs across analyses (r 2 > 0.75), indicating that they likely represent independent signals. This pattern suggests that the genetic architecture underlying TBB may depend on both environmental conditions and genetic background. Figure 7 shows a synthetic view of the map of all the QTLs identified in this study. QTL mapping. Colours indicate the assembly: Green for MSSB, blue for Bhaga and orange for common. Each grey bar represents one of the twelve chromosomes. Massane is shown at the bottom, Ventoux at the top. Each dot corresponds to a QTL. Out of 75 QTLs, only 20 (26%) were found to be common between assemblies when the comparison threshold was relaxed to include all the SNPs in linkage disequilibrium. The phenotypes concerned are: modelled TBB 2009, modelled DD 2009, DD 2023, DD‐based switching, TBB 2010, DD 2010 (all phenotypes except TBB‐based switching and a for all years). These results suggest that the use of more than one reference genome for GWAS may optimize the identification of QTLs, especially when conspecific populations with different biogeographic origins occupy a wide range and form hybrid zones (e.g., Mont Ventoux, see Section 4). We obtained 38 QTLs for field data and 37 for modelled data, of which two were common (modelled and field DD 2023). The QTLs explained between 0.2% and 54% of phenotypic variance for field data and between 0% and 59% for modelled data. We failed to validate seven of these traits: e.g., Modelled_TBB/2009/Bhaga carried the largest pseudo‐heritability among the traits, but no significant association was found in the replicates. We assessed whether the individual SNPs significantly associated with phenological traits in one population were polymorphic in the other as well. Among the significant SNPs identified at Mont Ventoux, 91% (30/33) (Bhaga) and 69% (29/42) (MSSB), respectively, were also polymorphic on La Massane; while 50% (13/26) (Bhaga) and 40% (12/36) (MSSB) of the significant SNPs found at La Massane were polymorphic on Mont Ventoux. All QTLs found at Mont Ventoux had at least one SNP that was also polymorphic in La Massane; conversely, only three QTLs found at La Massane (QTL 50, 51 and 66) had at least one SNP that was also polymorphic at Mont Ventoux (Table S14a,b). In a graphic evaluation, all the significant loci displayed allele frequency differences between populations within the 95% confidence interval (Figure S8a–d; Figure S69; range of differences in frequencies in Table S15). These patterns suggest that the two populations are not strongly differentiated at these loci, the frequencies of which oscillate and support a framework in which local adaptive patterns were driven by subtle differences in allele frequencies. When we used allelic effects estimated in one population to predict phenotypes in the other, loci driving a given trait in a specific population–assembly combination rarely produced a polygenic score (PGS) that significantly covaried with phenotypes in the other population (only 3 out of 34 trait–population–assembly combinations), and even then showed low predictive power (R 2 ranging from 2.6% to 4.5%) (Table S14a,b). These genetic regions appear to contribute differently in the two populations to the observed phenotypes in spite of not being overly differentiated, possibly due to a distinct response to short (plasticity) or a long term (evolutionary change) local selective regimes. Because we analysed traits measured across different years in the same individuals, the differences in detected QTLs indicate that the genetic architecture of each trait may shift with environmental conditions. In an ANOVA‐based likelihood ratio test comparing the two MLMM fitted to each trait × population × assembly combination, the model including the GxE interaction term consistently outperformed the model without it, across all combinations (p‐values: 0.0458 to 2.2−16, Table S16a–d). This strongly supports that genotype‐by‐environment interactions influenced the detection of year‐specific architectures. Additionally, we characterized two putative plasticity‐related phenotypes at La Massane (i.e., switching) and were able to identify 11 QTLs associated with them. This suggests that the response to environmental cues may be under genetic control. Functional Annotation Of the 134 significant SNPs in total, 96 (71%) are located in intergenic regions. Of the remaining intragenic variants, 23 are located in introns, eight cause missense variation, one causes a variant in the 5′ UTR, two cause a variant in the 3′ UTR and two cause variation in the splicing position (see Table 2). TABLE 2Functional annotation of significant SNPs, performed using the SnpEff tool.La MassaneMont VentouxNumberPercentageNumberPercentageTotal6210072100Intergenic variant49794765Intragenic variantIntron variant6101724Missense variant34575′ UTR variant12003′ UTR variant1211Splice region variant2323 Functional annotation of significant SNPs, performed using the SnpEff tool. Three candidates have been found among all these genes: a FRIGIDA‐like, CONSTANS‐like 6 and PHOTOPERIOD‐INDEPENDENT EARLY FLOWERING 1 genes. All the closest genes, and the genes within the QTLs, are in Tables S17 and S18. Discussion We identified a total of 75 QTLs associated with TBB, with 17 detected in La Massane and 58 in Mont Ventoux. No overlap was observed between the two forests, nor between the 2 years within each forest. Across the two genome assemblies, 26% of the QTLs were shared. Among these, several candidate genes were highlighted, including FRIGIDA‐like, CONSTANS‐LIKE 6 and PHOTOPERIOD‐INDEPENDENT EARLY FLOWERING 1 (PIE1). Advantages of Performing GWAS In Situ This study differs from most previous works in trees in two key aspects: (i) it was conducted directly in their natural environments (i.e., in situ), and (ii) each population was analysed independently rather than being merged into a pooled dataset. While common‐garden experiments remain highly valuable, especially for long‐lived species such as trees, they also present some limitations. In particular, when used to carry out GWAS, common garden data are likely to return only signals that are shared among most of the provenances; on the contrary, carrying out a GWAS in situ taps into the naturally occurring genetic diversity of populations and the local genetic architecture of traits, providing a timely opportunity to capture ongoing evolutionary processes in natural environments. Analysing each population independently further allows the detection of population‐specific genetic architectures, while maintaining the breadth of variation present within the species (Brachi et al. 2011). This approach therefore offers a powerful framework to study adaptation directly in situ, complementing the insights gained from controlled common garden experiments. Next‐generation sequencing methods have made the study of phenotypic diversity of wild populations more affordable. Several studies have performed GWAS in situ on multiple populations by targeting each site independently. Hiraoka et al. (2018) identified several QTLs for growth, wood properties and male fecundity in Japanese cedar ( Cryptomeria japonica D. Don) using independent analyses of each population and found that these loci were located close to each other or overlapped across populations. Similarly, Müller et al. (2019) applied this approach in Eucalyptus and detected QTLs for growth, with a majority being population‐specific. Ghosh Dasgupta et al. (2021) also used population‐independent GWAS for wood‐related traits in Eucalyptus and identified 45 QTLs common to two populations. As reviewed by Brachi et al. (2011), analysing populations separately is crucial for adaptive traits like TBB, as pooling populations can mask true associations that are confounded with population history. Polygenic Architecture of TBB In European beech, genomic studies have so far investigated the genetic drivers of TBB in a common garden context by analysing data from multiple populations sampled over different broad spatial scales. At a continental scale, Lazic et al. (2024) identified three QTLs associated with TBB across 100 populations, and Meger et al. (2021) detected eight QTLs for spring phenology from a panel of 380 candidate genes. Similarly, Postolache et al. (2021) identified 27 QTLs for phenology based on 139 candidate genes. The works of Müller et al. (2015, 2017) involved local German populations and detected seven phenology QTLs from a panel of 40 candidate genes. In other tree species, TBB also has a complex genetic architecture. For instance, in Populus trichocarpa , GWAS identified hundreds of SNPs linked to phenology and later studies pinpointed genes involved in thermal sensing, meristem regulation and symplastic mobility (Evans et al. 2014; McKown et al. 2014, 2018; Tylewicz et al. 2018). In Juglans regia , phytochromes and cytoskeleton genes are implicated (Bernard et al. 2020; Bükücü et al. 2020). In Fagaceae, QTLs for TBB are scattered across all chromosomes in Quercus robur and Castanea sativa and transcriptomic studies revealed involvement of genes regulating DNA replication, cell cycle, phytohormone responses and sucrose metabolism (Casasoli et al. 2004; Lesur et al. 2015; Scotti‐Saintagne et al. 2004). A CONSTANS‐like gene has been repeatedly associated with TBB in oak species (Alberto et al. 2013; Lind‐Riehl et al. 2014; Müller et al. 2017). Studies in other species identified up to 61 QTLs of variable sizes, involved in the control of TBB variation, suggesting that the trait has evolved under a polygenic architecture across several taxa (Apuli et al. 2021; Bernard et al. 2020; Bükücü et al. 2020; Evans et al. 2014; Gabay et al. 2018; Hsiang et al. 2024; Jermstad et al. 2001; Kapoor et al. 2023; Lalagüe et al. 2014; McKown et al. 2014, 2018; Riehl et al. 2023; Scotti‐Saintagne et al. 2004; Watson et al. 2024). Although these studies provide important insights into the genetic architecture of TBB in trees, most of the identified QTLs explain only a small fraction of phenotypic variance—for example, those reported by Lazic et al. (2024) accounted for only 1.4%–5.1%. In other species, concerning TBB, the identified QTLs explain between 1.2% (Coastal Douglas‐fir [ Pseudotsuga menziesii (Mirb.) Franco var. menziesii]) and 44.3% (Persian walnut ( Juglans regia L.)) (Bernard et al. 2020; Bükücü et al. 2020; Gabay et al. 2018; Jermstad et al. 2001; McKown et al. 2014; Meger et al. 2021; Müller et al. 2015, 2017). By contrast, our analysis identified QTLs explaining up to 59% of the variance, despite relying on a smaller sample size. Hsiang et al. (2024) also identified a single major QTL for TBB in Prunus mume , which explained 45.2% of the phenotypic variance using a newly haplotype‐phased genome. Overall, these findings highlight that, although most QTLs for TBB explain only a limited fraction of phenotypic variance, larger‐effect loci can occasionally be uncovered, particularly when leveraging improved genomic resources. Population‐Specific Architecture of TBB Interestingly, the QTLs identified by Lazic et al. (2024) were located 242 kb (chromosome 1), 11 Mb (chromosome 7) and 370 kb (chromosome 8) away from those detected in our study, respectively, suggesting that distinct genomic regions regulate TBB depending on the populations analysed. We found that distinct genetic factors underlie TBB in La Massane and Mont Ventoux forests and during the two pairs of survey years, trees at Mont Ventoux budded later than those at La Massane. This observation reinforces the need to perform GWAS independently in different populations, rather than combining them into a single analysis, to better capture the architecture of the trait that seems to be specific to each population. These differences in the inferred architecture between the two sites may stem from phylogeographic history. The populations investigated here may trace back to two glacial refugia located at the south‐eastern margin of the species' range (Magri 2008; Magri et al. 2006). Although both belong to the ‘Atlantic’ cluster identified by Postolache et al. (2021), La Massane lies at the cluster's core, while Mont Ventoux is situated at its edge, in a region where other populations display mixed ancestry. Such contrasting contexts likely contributed to the distinct genomic architectures of TBB revealed in the two forests. This divergence may reflect two non‐exclusive processes: (i) differences in the initial gene pool leading to variation in the genomic regions controlling the trait across populations (Rellstab et al. 2017) and/or (ii) local adaptation to distinct environmental conditions (Kawecki and Ebert 2004). Another possible explanation for the contrasting results lies in the differences in heat requirements between the two forests. The delay in Ventoux likely results from the tradeoff between a longer growing season and late frost avoidance (Gömöry and Paule 2011); at higher elevations, the evolutionary pressure for an early TBB is buffered by the plastic response to colder temperatures that delays bud break, thus reducing the risk of frost damage (Alberto et al. 2013; Gauzere, Teuf, et al. 2020; Kreyling et al. 2014; Vitasse et al. 2009). On average, trees at La Massane required 692°C (DD) to reach budbreak across the 2 years, whereas those at Mont Ventoux needed only 310°C—that is, La Massane trees required more than twice the accumulated heat. Similar patterns have also been reported in the literature. Wang et al. (2022) reported heat requirement differences of up to 241.8°C among various European beech provenances. Such differences in heat requirements between populations are consistent with broader geographic patterns reported in the literature. Regional‐scale studies (Chmura and Rozkowski 2004; Gömöry and Paule 2011) have reported clinal variation in phenological traits along latitudinal and longitudinal gradients, pointing to a geographic structuring of TBB. In European beech, Davi et al. (2011) found a significant correlation between TBB and elevation, which reflects an inverse relationship between the genetic component of TBB and altitude (Gauzere, Klein, et al. 2020). By contrast, common garden experiments in oaks revealed a direct relationship between TBB and elevation (Vitasse et al. 2009), and a similar positive correlation with elevation was observed in southern beeches (Nothofagus spp.) (Duboscq‐Carra et al. 2020). Our findings highlight the importance of studying multiple local tree populations independently in situ, rather than relying solely on large‐scale provenance trials. This approach provides a powerful framework to uncover the genetic bases of local adaptation in TBB, by capturing the diversity of signals present in natural populations without the constraints imposed by highly controlled experimental designs. Genotype × Environment Interactions For polygenic traits such as spring phenology, multilocus genotypes can exhibit variable phenotypic responses to environmental variation (Lynch and Walsh 1998). As a result, testing genotype–phenotype associations in the same individuals across contrasting years (i.e., environments) may uncover different sets of loci underlying the same trait, reflecting genotype‐by‐environment (G × E) interactions. While TBB is under genetic control, environment has a strong influence in driving the trait both directly and through GxE interaction (North and Kovaleski 2024; Singh et al. 2017; Wang and Ding 2023; Yang et al. 2021) thus often contributing to missing heritability (Brachi et al. 2011). Combining common garden experiments with genomic approaches has consistently revealed a polygenic architecture of spring phenology, where a limited number of SNPs are stable across years, alongside a larger set of year‐specific variants (McKown et al. 2014; Müller et al. 2015; Bükücü et al. 2020; Apuli et al. 2021; pers. comm., C. Geßler, D. Lazic, N. A. Müller), a pattern also observed in other plant models (Frachon et al. 2017, 2018). In some studies, only year‐specific associations were detected (Meger et al. 2021; Lazic et al. 2024). Such limited replication may arise from population structure, insufficient power to resolve highly polygenic architectures, or interannual climatic variability (Westergren et al. 2025). A key limitation of in situ GWAS is that environmental conditions are only partially characterized, making it difficult to identify the drivers of phenotypic variation and accurately estimate genetic effects (Santure and Garant 2018). In contrast, replicated common gardens allow partitioning additive (G and E) and non‐additive (G × E) components (Cappa et al. 2025; Papper and Ackerly 2021; Serrie et al. 2025). Here, we adopted an intermediate approach by leveraging interannual variation at the same site as a proxy for contrasting environmental conditions. At Mont Ventoux, the first 5 months of 2009 were warmer (3.849°C) than in 2010 (2.862°C), and QTLs detected in each year did not overlap, with the closest signals located 266 kb apart. At La Massane, although mean temperatures were similar between years, a cold spell in early April 2022 (1st–4th) likely delayed TBB, and no significant associations were detected. This may reflect either a highly polygenic architecture involving many small‐effect loci or context‐dependent genetic effects (Clauw et al. 2025). Overall, the lack of overlap across years suggests that distinct genomic regions may contribute under different climatic conditions, which is consistent with—though not sufficient to demonstrate—G × E interactions. Indeed, observing different signals across years does not constitute direct evidence for G × E, as limited sample size may reduce the power to detect consistent QTLs. If phenotypes measured across years correspond to the same underlying trait, they are expected to remain genetically correlated despite differences in inferred architectures (Falconer 1952; Yamada 1962; Meyer 2009). Using MLMM and likelihood‐ratio‐based ANOVA, we nonetheless found consistent support for G × E effects, as models including a G × E term provided the best fit across all trait × assembly combinations. The significance for the MLMM including the GxE term scaled with the number of QTLs associated with a trait: the more QTLs (parameters) were included in the model, the higher the statistical support for significant GxE, hinting that the detection power also constrains the assessing GxE interactions, which have more statistical support whenever tested on more complex architectures (Table S16a–d). However, given the limited number of years analysed, these results should be interpreted cautiously. In natural conditions, multiple biotic and abiotic factors interact—often non‐additively—making it difficult to fully capture environmental heterogeneity and necessitating a simplified modelling framework compared to those used in plant breeding (Anderson and Song 2020; Van Eeuwijk et al. 2016). These findings align with theoretical expectations for highly polygenic traits such as phenology, where G × E interactions can maintain genetic variation through balancing selection. In such systems, multilocus genotypes with high heterozygosity may buffer phenotypic variation across environments (Lewontin 1964; Gillespie and Turelli 1989). This dual role of genetic variation—modulating environmental responses while promoting local adaptation—can lead to genomic architectures composed of a few large‐effect loci and many small‐effect loci (Yeaman 2022). From a GWAS perspective, this translates into environment‐specific contributions of genomic regions. At La Massane, some trees exhibited phenological shifts of up to 15 days between years, likely reflecting the marked temperature differences between 2022 and 2023, combined with variation in individual sensitivity to these climatic cues. We detected six QTLs underlying TBB‐based switching (i.e., plasticity) in our data. Modelling studies have demonstrated that European beech displays strong phenological plasticity in response to interannual climatic variation (Duputié et al. 2015). Our analysis provides insights into the molecular basis of this plasticity as an heritable and genetically controlled‐trait (Cooper et al. 2019; Gauzere, Teuf, et al. 2020; Martínez‐Sancho et al. 2025). The genomic control of the facets of plasticity of spring phenology would be better elucidated through longer multi‐year TBB records. Longer time series would allow assessing variations in the switching phenomenon across environments; furthermore, it would be possible to model changes in the reaction norms and in the GxE profiles between years and analyse the slope as a synthetic phenotype (Kadam et al. 2017; Laitinen and Nikoloski 2019; Mangin et al. 2017). Different Results Between Different Reference Genome Assemblies In our analysis, the results varied depending on the genome assembly used for mapping, even though the population and phenotypes were the same. Forest trees are well known for their high genetic diversity, and populations located at the margins of the species' distribution often contain a substantial fraction of this diversity (Hampe and Petit 2005; Petit and Hampe 2006). Consequently, genetic differences between two individuals can be considerable, underlining the importance of selecting a reference genome that best represents the studied population. In plants, it is already well established that the choice of reference genome strongly influences SNP detection and subsequent GWAS outcomes (Della Coletta et al. 2021; Gage et al. 2019). Depending on the plant species, only a fraction of genes (33%–92%) belong to the core genome (i.e., shared genomic regions between all individuals of a species; (Bayer et al. 2020)). A substantial portion of the genome is absent in some individuals and the SNPs in these regions cannot be mapped, leading to inconsistent GWAS results when using different assemblies. Unpublished pangenome data assembled from 70 individuals from all across Europe indicate that variation in genomic content is more substantial between individuals rather than populations, with the result that genome size of beech varies between 410 and 569 Mb, with an average of 469 Mb (pers. comm., D. Lazic). Together with our assessment of duplicated regions, this partially explains the qualitative differences in QTL identification between assemblies. Using two reference assemblies made it easier to detect significant genotype–phenotype associations. Some signals appeared in both references or were located only a few base pairs apart, allowing us to group them into a single QTL. However, most signals were found in only one of the two assemblies, meaning they would have been missed if we had used a single assembly. Finally, structural variants also play a role in shaping GWAS outcomes. They can help explain why the proportion of variance explained by significant SNPs is sometimes low (Akakpo et al. 2020). Candidate Genes We identified a FRIGIDA‐like gene associated with late TBB and high DD requirement at Mont Ventoux (modelled.TBB/2010/Bhaga, modelled.DD/2010/Bhaga). In Arabidopsis thaliana , allelic variation at the FRIGIDA locus can account for up to 70% of the variation in flowering time (Shindo et al. 2005). Beyond flowering, FRIGIDA has also been shown to contribute to local adaptation in a pleiotropic manner, influencing growth, water use and phenology depending on its expression level (Lovell et al. 2013). In addition, FRIGIDA‐like genes have been implicated in dormancy regulation in Japanese apricot ( Prunus mume ) (Li et al. 2021). These findings suggest that FRIGIDA‐like genes could similarly contribute to dormancy and TBB regulation in beech. We also found CONSTANS‐LIKE 6 (COL6) to be associated with switching status in La Massane. In A. thaliana , CONSTANS (CO) and CONSTANS‐LIKE (COL) genes interact with Flowering locus T (FT) to integrate photoperiodic signals and regulate the circadian clock (Khanna et al. 2009; Zhang et al. 2023). More specifically, AtCOL6 has been shown to delay flowering through its interactions with AtCOL16 and AtCOL7 (Zhang et al. 2023). This gene has already been found to control TBB in European beech (Müller et al. 2017) and sessile oak (Alberto et al. 2013) at a local or regional level and in aspen provenances in controlled conditions (Riehl et al. 2023). Another candidate highlighted in our results is PHOTOPERIOD‐INDEPENDENT EARLY FLOWERING 1 (PIE1); individuals from La Massane population carrying the alternative allele at several sites in the genomic neighbourhood of this gene had a higher likelihood of exhibiting a switch in their TBB in response to temperature (DD‐based switching/MSSB). In A. thaliana , this gene activates FLC (Flowering Locus C), with loss‐of‐function mutations leading to early flowering and abnormal flower development (Noh and Amasino 2003). Although PIE1 has not yet been studied in beech, it is well established that in tree species, the molecular regulation of TBB is tightly connected to flowering pathways, particularly at the signalling level (Zhao et al. 2025). Finally, the majority of the identified QTLs (89% in both populations) map into non‐coding regions or introns. Since regulatory regions have been found to be involved in modulating GxE interaction (Gage et al. 2017) and that all the trait architectures we inferred were shaped by GxE, we hypothesise that, at the genomic level, spring phenology may be driven more by variants in the regulatory network of major genes playing a role in driving spring phenology as large as the variants at the major genes themselves than by different alleles in a gene coding sequence. In this context, we can expect these regions to harbour a larger degree of heterozygosity (Gillespie and Turelli 1989). Further transcriptomic and expression analyses are required to understand how the modulation of these candidate genes contributes to regulating TBB in European beech. Study Limitations A general issue with GWAS studies is the replicability of results across studies, which is hampered by the genomic context and the interaction between loci and sites (Holliday et al. 2012; NCI‐NHGRI Working Group on Replication in Association Studies et al. 2007; Santure and Garant 2018). Major‐effect loci have proven to be replicable between populations; instead, those that contribute less to the variation of the trait may be more difficult to detect in multiple contexts. These loci with smaller effect have been validated more successfully within rather than between populations (Schielzeth et al. 2018). In our experiments, the two populations displayed a different architecture for the same traits (replication between populations) and the loci identified in one population overall failed to predict the phenotypes in the other one. Bootstrapping the datasets often led to non‐overlapping results: we found consensus loci (i.e., loci appearing in more than one replicate) for 13 out of 20 analyses (Mont Ventoux) and seven out of 20 analyses (La Massane), with very few common loci (Table S12a–t). The PGS computed from these loci in the training subsets variably predicted the phenotypes in the testing subsets (R 2 from 0.2 to 17.6) (Results S3). The pseudo‐heritability estimates obtained using training/testing subsets are an approximate estimator (with larger sampling variance) and possibly a downward‐biased estimator of pseudo‐heritability that could be inferred from the main data set (Korte and Farlow 2013). Additionally, the variability of results across replicates suggests that some small‐effect loci that are harder to capture in small samples may have gone undetected in the main analysis, resulting in an excess of large‐effect loci (Table S19). In our case, almost all significant SNPs within the QTLs identified in one population failed to predict their phenotypes in the other one, even though most of them (40%–90%) were polymorphic in both. The exception was a QTL overlapping a gene coding for an LRR receptor‐like serine/threonine‐protein kinase, which predicted a small portion of trait variation in three combinations of trait × year × assembly (R 2: 2.6%–4.5%) (Table S14a,b). Sample size is a critical factor in GWAS, as statistical power strongly depends on the number of individuals included. While GWAS in humans often rely on very large cohorts, sometimes reaching hundreds of thousands of individuals (Visscher et al. 2012), studies in natural populations—particularly in forest trees—are typically conducted with more limited sample sizes (Alberto et al. 2013; McKown et al. 2014; Bernard et al. 2020; Bükücü et al. 2020). In this study, we analysed two independent, weakly structured populations, which reduces the risk of confounding due to population stratification. However, despite this advantage, the relatively small sample size (n = 150) likely limits the power to detect loci of small effect and may contribute to the variability observed across analyses. As a result, our findings should be interpreted with caution. Beyond sample size considerations, phenotyping methodology may also have introduced additional sources of variability. Although the scale developed by Davi et al. (2011) was converted to the BBCH code (Finn et al. 2007; Meier 2001), it is important to acknowledge that some bias related to the original scale may persist. The Davi et al. (2011) system divides the crown into two sections (upper and lower), whereas the BBCH code considers the crown as a whole. Additionally, the Davi et al. (2011) scale assigns a developmental stage to a tree as soon as a few buds reach it, making the assessment more subjective. In contrast, the BBCH scale relies on fixed thresholds based on the percentage of buds at a given stage. As a result, the Davi et al. (2011) scale likely introduces greater variability among trees classified within the same stage. The BBCH system, being more stringent due to its quantitative criteria, makes it more difficult to assign a tree to a particular stage. Moreover, the observers are not the same between the two forests. These differences in phenotyping approaches may have partially influenced the GWAS results obtained for the two forests. Another limitation of this study concerns the variability observed in genomic inflation factors (λ) across GWAS analyses. The use of the FarmCPU method, which iteratively fits fixed and random effects to control for confounding (Liu et al. 2016), may in some cases lead to overcorrection, resulting in λ values below 1, or may not fully capture all sources of confounding. In addition, the polygenic architecture of the studied traits can contribute to deviations from the null expectation, leading to moderate inflation of test statistics. Previous studies, such as The GIANT Consortium et al. (2011) and Singh et al. (2025), have shown that under a polygenic model, the median λ tends to increase with the number of associated loci. However, these studies relied on much larger sample sizes (e.g., ~3900 individuals in The GIANT Consortium et al. (2011) and up to more than 1.8 million individuals in Singh et al. (2025), compared to 150 in our study), which strongly influences the behaviour of λ. Furthermore, differences in sample size across traits and years in our dataset may also contribute to inconsistencies in λ estimates. Altogether, these factors highlight the need for cautious interpretation of association signals, particularly for analyses showing substantial deviations from λ = 1. A key challenge in this study lies in disentangling G × E interactions from year‐specific environmental effects. In natural forest conditions, climatic and environmental variables fluctuate substantially between years, such that each year effectively represents a distinct environmental context. For this reason, we treated each year as a separate environment, as this approach more closely reflects the conditions experienced by trees in situ (Yamada 1962). However, the marked climatic differences observed between the surveyed years likely emphasized variation in the genetic control of spring phenology rather than revealing stable, shared genomic patterns. Consequently, the inferred architectures may partly reflect the specific climatic conditions of each year (Saltz et al. 2018; Liu and El‐Kassaby 2019). For instance, no significant associations were detected for spring 2022 at La Massane. Another important result is the absence of shared QTLs between La Massane and Mont Ventoux, despite only moderate differences in allele frequencies at significant loci. Although our analysis is based on only two populations, this lack of overlap points to a strong population‐specific component in the genetic architecture of TBB (Rellstab et al. 2017). Several factors may account for this pattern. Subtle differences in allele frequencies and linkage disequilibrium can reduce the likelihood of detecting consistent associations across populations, particularly when they have evolved under contrasting selective regimes (e.g., La Massane at low elevation versus the Mont Ventoux beech forest spanning an ~1000 m altitudinal gradient). In addition, limited sample sizes decrease statistical power, especially for detecting loci of small effect, further decreasing the probability of overlap between analyses. Taken together, the absence of shared QTLs should not be interpreted simply as a lack of reproducibility, but rather as the combined effect of limited power, population‐specific genetic backgrounds and potentially distinct adaptive trajectories. These results nonetheless highlight the value of conducting GWAS within individual natural populations, as this approach enables the detection of population‐specific genetic architectures that may be obscured in multi‐population analyses. Rear‐Edge Populations as ‘Adaptation Sentinels’ Climate change is exerting strong negative impacts on European beech forests. The severe droughts of 2003 and 2018–2019 caused reduced growth, premature leaf senescence and widespread dieback, often followed by long‐lasting insect and fungal infestations (Rukh et al. 2023). Such extreme events are expected to become more frequent and intense, even under optimistic climate scenarios (Calvin et al. 2023). By 2090, beech growth is projected to decline by up to 50%, depending on the region and climate model (Martinez Del Castillo et al. 2022). This decline is expected to be particularly pronounced in rear‐edge populations, such as those at La Massane and Mont Ventoux (figure 1c in Saltré et al. (2015)). These changes will inevitably have far‐reaching ecological and economic consequences. Rear‐edge populations occupy suboptimal habitats within a species' range; they are often smaller, more fragmented, more prone to extinction (Gaston 2009; Hampe and Petit 2005; Lawton 1993; Sexton et al. 2009) and more differentiated than core populations (Eckert et al. 2008; Pironon et al. 2015). Nonetheless, rear‐edge populations have globally survived throughout the Quaternary, their fate likely driven by adaptation to ceaseless climate changes. In general, differentiation at the rear edge may result from drift (Kitamura et al. 2015) jointly with adaptation to local conditions (Aitken et al. 2008; Mathiasen and Premoli 2016; Mikola 1982; Savolainen et al. 2007; Siol et al. 2010; Theraroz et al. 2024). In particular, for temperate European tree species, the rear edge lies in corrugated areas (like the Mediterranean basin or Southern European mountain ranges) where environmental patchiness allows multiple phenotypic optima to coexist within the same landscape, thus increasing adaptive diversity (Alberto et al. 2013; Endler 1977; Kremer et al. 2012; Scotti et al. 2023; Slatkin 1973). Cavin and Jump (2017) showed that, for F. sylvatica , the individuals most sensitive to drought are located in the species' range core, whilst rear‐edge populations appeared to have adapted to harsher climatic conditions. This makes the putatively adapted alleles of these populations valuable for maintaining the species stable in the core of its range under future climate scenarios, as they may already be pre‐adapted to these drier conditions. In this context, the identification of QTLs associated with early TBB in beech provides new opportunities for forest management. Earlier TBB in regions with low frost risk may allow trees to extend their growing season and mitigate the impact of drought stress. More broadly, understanding the genetic basis of this trait, analysing several populations independently in situ, could guide the selection of individuals and provenances for assisted migration strategies, thereby accelerating adaptation in the face of rapid climate change. Author Contributions L.C. and A.M. performed the analyses under the supervision of O.P. and I.S. W.M., I.L.K., CP, C.B., G.L.P., S.O.‐M., J.‐M.A., S.D., C.D.S., B.N., C.C. and K.L. were involved in the genome assembly of MSSB beech accession. J.A., J.C. and H.Q. were involved in genome annotation. M.M. and M.C.P. assisted L.C. in GWAS analyses. K.H. and B.U. performed the field assessment of phenology at the Mont Ventoux site. J.‐A.M. assisted L.C. in Massane forest field experiments. Funding This work was supported by ecole universitaire de recherche tulip GS, ANR‐18‐EURE‐0019, Horizon 2020 Framework Programme 822221, the French National Research Agency (ANR) under grant ANR‐24‐CE02‐1230 (FAGRESCUE), the Agence Nationale de la Recherche under the France 2030 program as part of the ‘Forest Resilience’ research program (PEPR FORESTT), reference ANR‐25‐PEFO‐0009 (TOOLS‐FOR‐ADAP). Conflicts of Interest The authors declare no conflicts of interest. Supporting information Figure S1: Temperature trend during the period of interest (DOY 1–121, 1st January to 1st May). Table S1: Temperatures in the two forests for each year. Table S2: TE annotations comparison for MSSB genome. Table S3: Classification for TE annotation in MSSB genome. Figure S2: Example of MITEs nested: DTM_6 in TIR_15. Figure S3: Calendar of Ventoux and La Massane field work. Figure S4: BBCH stages of budbreak for European beech. Table S4: Patterns of observations between populations. Table S5: Mean and range of values for the logistic parameters. Table S6: Sample size by trait. Figure S5: Distribution of cumulative degree‐days calculated using different temperature thresholds (3°C in red, 5°C in green and 7°C in blue), in (A) La Massane and (B) Mont Ventoux. Table S7: Number of SNPs after filtering steps, for each forest and each assembly. Figure S6: Distribution of frequency of kinship values per pair of individuals in (A) La Massane and (B) Mont Ventoux. Table S8: Summary of the GWAS results. Table S9: Dates of budbreak and degree of days for the BBCH 09 stage, in both forests, both years. Figure S7: Analysis of the switching trees in La Massane, stable trait in green and switching trait in blue: (A) Proportion of switching trees in La Massane, (B) Switcher type. Figure S8: Analyse of the switcher trees in La Massane, stable trait in green and switching trait in blue: (A) Proportion of switching trees in La Massane, (B) Switcher type, (C) Difference of degree days for BBCH 09 stage, (D) Boxplot of difference of degree days for BBCH 09 stage. Figure S9: PCA of the individual combinations of logistic parameters. Table S10: Statistics of the validation of the logistic model. Figure S10: Manhattan plot and corresponding QQ plot for the trait ‘a in 2022’, with MSSB assembly. Figure S11: Manhattan plot and corresponding QQ plot for the trait ‘a in 2023’, with MSSB assembly. Figure S12: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2022’, with MSSB assembly. Figure S13: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2023’, with MSSB assembly. Figure S14: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2022’, with MSSB assembly. Figure S15: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2022, without switching trees’, with MSSB assembly. Figure S16: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2023’, with MSSB assembly. Figure S17: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2023, without switching trees’, with MSSB assembly. Figure S18: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2022’, with MSSB assembly. Figure S19: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2022, without switching trees’, with MSSB assembly. Figure S20: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2023’, with MSSB assembly. Figure S21: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2023, without switching trees’, with MSSB assembly. Figure S22: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2022’, with MSSB assembly. Figure S23: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2023’, with MSSB assembly. Figure S24: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day Of the Year in 2022’, with MSSB assembly. Figure S25: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day Of the Year in 2023’, with MSSB assembly. Figure S26: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022’, with MSSB assembly. Figure S27: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023’, with MSSB assembly. Figure S28: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Degree Days’, with MSSB assembly. Figure S29: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Day Of the Year’, with MSSB assembly. Figure S30: Manhattan plot and corresponding QQ plot for the trait ‘a in 2022’, with Bhaga assembly. Figure S31: Manhattan plot and corresponding QQ plot for the trait ‘a in 2023’, with Bhaga assembly. Figure S32: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2022’, with Bhaga assembly. Figure S33: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2023’, with Bhaga assembly. Figure S34: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022’, with Bhaga assembly. Figure S35: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022, without switching trees’, with Bhaga assembly. Figure S36: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023’, with Bhaga assembly. Figure S37: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023, without switching trees’, with Bhaga assembly. Figure S38: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2022’, with Bhaga assembly. Figure S39: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2022, without switching trees’, with Bhaga assembly. Figure S40: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2023’, with Bhaga assembly. Figure S41: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2023, without switching trees’, with Bhaga assembly. Figure S42: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2022’, with Bhaga assembly. Figure S43: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2023’, with Bhaga assembly. Figure S44: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2022’, with Bhaga assembly. Figure S45: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2023’, with Bhaga assembly. Figure S46: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022’, with Bhaga assembly. Figure S47: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023’, with Bhaga assembly. Figure S48: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Degree Days’, with Bhaga assembly. Figure S49: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Day of the Year’, with Bhaga assembly. Figure S50: Manhattan plot and corresponding QQ plot for the trait ‘a in 2009, for Mont Ventoux’, with MSSB assembly. Figure S51: Manhattan plot and corresponding QQ plot for the trait ‘a in 2010, for Mont Ventoux’, with MSSB assembly. Figure S52: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2009, for Mont Ventoux’, with MSSB assembly. Figure S53: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2010, for Mont Ventoux’, with MSSB assembly. Figure S54: Manhattan plot and corresponding QQ plot for the trait ‘Day of the Year in 2009, for Mont Ventoux’, with MSSB assembly. Figure S55: Manhattan plot and corresponding QQ plot for the trait ‘Day of the Year in 2010, for Mont Ventoux’, with MSSB assembly. Figure S56: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2009, for Mont Ventoux’, with MSSB assembly. Figure S57: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2010, for Mont Ventoux’, with MSSB assembly. Figure S58: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2009, for Mont Ventoux’, with MSSB assembly. Figure S59: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2010, for Mont Ventoux’, with MSSB assembly. Figure S60: Manhattan plot and corresponding QQ plot for the trait ‘a in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S61: Manhattan plot and corresponding QQ plot for the trait ‘a in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S62: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S63: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S64: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S65: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S66: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S67: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S68: Histogram of SNP effect size distribution. A total of 164 SNPs were included; some appear multiple times because they are associated with several phenotypes. Table S11: Summary of lambda values, depending on the comparisons. Table S12: Training/Testing. Table S13: Summary table of closest significant SNPs, depending the condition of comparison. Table S14: Cross predictions by Trait × Year × Assembly. Figure S69: Difference in the frequency of the ALT allele between the two populations. Table S15: Range of difference in allele frequencies at the significant loci between the two populations. Table S16:: Summary statistics of the likelihood‐ratio comparing the MLMM testing for GxE interaction on the trait architecture. Table S17: Summary table of closest genes to the QTLs. Table S18: Summary table of genes into QTLs region‐scale. Table S19: Intersections between signals retrieved in the main dataset and consensus loci across replicates. Acknowledgements La Massane's samples are part of a broader project, BEECHGENOMES (BEECHGENOMES—France Génomique), which aims to generate genomic data for several hundred beech trees across Europe. This study was supported by the French Laboratory of Excellence project ‘TULIP’ (ANR‐10‐LABX‐41) and the University of Perpignan, with the support of the Occitanie Regional Council's program « Key challenge BiodivOc ». Andrea Modica's thesis was supported by INRAE's ECODIV Division and by the Horizon2020 FORGENIUS research programme (contract No. 822221). We thank La Massane Nature Reserve for granting site access and supporting our fieldwork. We would like to thank Nicolas Francillonne and Mariene Wan from URGI for adding the consensus library of transposable elements from Fagus sylvatica L. to RepetDB. Contributor Information Olivier Panaud, Email: olivier.panaud@univ-perp.fr. Ivan Scotti, Email: ivan.scotti@inrae.fr. Data Availability Statement All scripts and codes used in this study are publicly available in the GitHub repository and archived on Zenodo: Clément, L. and MODICA, A. (2025). QTL_Massane_Ventoux (Version 1.0.4) Zenodo. 10.5281/zenodo.17569317. GitHub repository: https://github.com/lauren220200/QTL_Massane_Ventoux. Phenotypic and genotypic datasets supporting the results are available in the Data Gouv repository: Clement, Lauren, 2026, ‘Metadata for the QTLs paper on La Massane and Mont Ventoux analysis’, https://doi.org/10.82233/MM9NQA, Recherche Data Gouv, VERSION PROVISOIRE. Raw data will be made available prior to publication at SRAENA short sequence repository. The TE consensus sequences and annotations are available on RepetDB, https://urgi.versailles.inrae.fr/repetdb (Amselem et al. 2019). References Associated Data Supplementary Materials Figure S1: Temperature trend during the period of interest (DOY 1–121, 1st January to 1st May). Table S1: Temperatures in the two forests for each year. Table S2: TE annotations comparison for MSSB genome. Table S3: Classification for TE annotation in MSSB genome. Figure S2: Example of MITEs nested: DTM_6 in TIR_15. Figure S3: Calendar of Ventoux and La Massane field work. Figure S4: BBCH stages of budbreak for European beech. Table S4: Patterns of observations between populations. Table S5: Mean and range of values for the logistic parameters. Table S6: Sample size by trait. Figure S5: Distribution of cumulative degree‐days calculated using different temperature thresholds (3°C in red, 5°C in green and 7°C in blue), in (A) La Massane and (B) Mont Ventoux. Table S7: Number of SNPs after filtering steps, for each forest and each assembly. Figure S6: Distribution of frequency of kinship values per pair of individuals in (A) La Massane and (B) Mont Ventoux. Table S8: Summary of the GWAS results. Table S9: Dates of budbreak and degree of days for the BBCH 09 stage, in both forests, both years. Figure S7: Analysis of the switching trees in La Massane, stable trait in green and switching trait in blue: (A) Proportion of switching trees in La Massane, (B) Switcher type. Figure S8: Analyse of the switcher trees in La Massane, stable trait in green and switching trait in blue: (A) Proportion of switching trees in La Massane, (B) Switcher type, (C) Difference of degree days for BBCH 09 stage, (D) Boxplot of difference of degree days for BBCH 09 stage. Figure S9: PCA of the individual combinations of logistic parameters. Table S10: Statistics of the validation of the logistic model. Figure S10: Manhattan plot and corresponding QQ plot for the trait ‘a in 2022’, with MSSB assembly. Figure S11: Manhattan plot and corresponding QQ plot for the trait ‘a in 2023’, with MSSB assembly. Figure S12: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2022’, with MSSB assembly. Figure S13: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2023’, with MSSB assembly. Figure S14: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2022’, with MSSB assembly. Figure S15: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2022, without switching trees’, with MSSB assembly. Figure S16: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2023’, with MSSB assembly. Figure S17: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using degree days in 2023, without switching trees’, with MSSB assembly. Figure S18: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2022’, with MSSB assembly. Figure S19: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2022, without switching trees’, with MSSB assembly. Figure S20: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2023’, with MSSB assembly. Figure S21: Manhattan plot and corresponding QQ plot for the trait ‘Field earliness using Day of the Year in 2023, without switching trees’, with MSSB assembly. Figure S22: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2022’, with MSSB assembly. Figure S23: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2023’, with MSSB assembly. Figure S24: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day Of the Year in 2022’, with MSSB assembly. Figure S25: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day Of the Year in 2023’, with MSSB assembly. Figure S26: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022’, with MSSB assembly. Figure S27: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023’, with MSSB assembly. Figure S28: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Degree Days’, with MSSB assembly. Figure S29: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Day Of the Year’, with MSSB assembly. Figure S30: Manhattan plot and corresponding QQ plot for the trait ‘a in 2022’, with Bhaga assembly. Figure S31: Manhattan plot and corresponding QQ plot for the trait ‘a in 2023’, with Bhaga assembly. Figure S32: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2022’, with Bhaga assembly. Figure S33: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2023’, with Bhaga assembly. Figure S34: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022’, with Bhaga assembly. Figure S35: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022, without switching trees’, with Bhaga assembly. Figure S36: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023’, with Bhaga assembly. Figure S37: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023, without switching trees’, with Bhaga assembly. Figure S38: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2022’, with Bhaga assembly. Figure S39: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2022, without switching trees’, with Bhaga assembly. Figure S40: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2023’, with Bhaga assembly. Figure S41: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Day of the Year in 2023, without switching trees’, with Bhaga assembly. Figure S42: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2022’, with Bhaga assembly. Figure S43: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2023’, with Bhaga assembly. Figure S44: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2022’, with Bhaga assembly. Figure S45: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2023’, with Bhaga assembly. Figure S46: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2022’, with Bhaga assembly. Figure S47: Manhattan plot and corresponding QQ plot for the trait ‘Earliness using Degree Days in 2023’, with Bhaga assembly. Figure S48: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Degree Days’, with Bhaga assembly. Figure S49: Manhattan plot and corresponding QQ plot for the trait ‘Switching status using Day of the Year’, with Bhaga assembly. Figure S50: Manhattan plot and corresponding QQ plot for the trait ‘a in 2009, for Mont Ventoux’, with MSSB assembly. Figure S51: Manhattan plot and corresponding QQ plot for the trait ‘a in 2010, for Mont Ventoux’, with MSSB assembly. Figure S52: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2009, for Mont Ventoux’, with MSSB assembly. Figure S53: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2010, for Mont Ventoux’, with MSSB assembly. Figure S54: Manhattan plot and corresponding QQ plot for the trait ‘Day of the Year in 2009, for Mont Ventoux’, with MSSB assembly. Figure S55: Manhattan plot and corresponding QQ plot for the trait ‘Day of the Year in 2010, for Mont Ventoux’, with MSSB assembly. Figure S56: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2009, for Mont Ventoux’, with MSSB assembly. Figure S57: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2010, for Mont Ventoux’, with MSSB assembly. Figure S58: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2009, for Mont Ventoux’, with MSSB assembly. Figure S59: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2010, for Mont Ventoux’, with MSSB assembly. Figure S60: Manhattan plot and corresponding QQ plot for the trait ‘a in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S61: Manhattan plot and corresponding QQ plot for the trait ‘a in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S62: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S63: Manhattan plot and corresponding QQ plot for the trait ‘Degree Days in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S64: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S65: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Degree Days in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S66: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2009, for Mont Ventoux’, with Bhaga assembly. Figure S67: Manhattan plot and corresponding QQ plot for the trait ‘Modelled Day of the Year in 2010, for Mont Ventoux’, with Bhaga assembly. Figure S68: Histogram of SNP effect size distribution. A total of 164 SNPs were included; some appear multiple times because they are associated with several phenotypes. Table S11: Summary of lambda values, depending on the comparisons. Table S12: Training/Testing. Table S13: Summary table of closest significant SNPs, depending the condition of comparison. Table S14: Cross predictions by Trait × Year × Assembly. Figure S69: Difference in the frequency of the ALT allele between the two populations. Table S15: Range of difference in allele frequencies at the significant loci between the two populations. Table S16:: Summary statistics of the likelihood‐ratio comparing the MLMM testing for GxE interaction on the trait architecture. Table S17: Summary table of closest genes to the QTLs. Table S18: Summary table of genes into QTLs region‐scale. Table S19: Intersections between signals retrieved in the main dataset and consensus loci across replicates. Data Availability Statement All scripts and codes used in this study are publicly available in the GitHub repository and archived on Zenodo: Clément, L. and MODICA, A. (2025). QTL_Massane_Ventoux (Version 1.0.4) Zenodo. 10.5281/zenodo.17569317. GitHub repository: https://github.com/lauren220200/QTL_Massane_Ventoux. Phenotypic and genotypic datasets supporting the results are available in the Data Gouv repository: Clement, Lauren, 2026, ‘Metadata for the QTLs paper on La Massane and Mont Ventoux analysis’, https://doi.org/10.82233/MM9NQA, Recherche Data Gouv, VERSION PROVISOIRE. Raw data will be made available prior to publication at SRAENA short sequence repository. The TE consensus sequences and annotations are available on RepetDB, https://urgi.versailles.inrae.fr/repetdb (Amselem et al. 2019).