Haplotype Blocks Are Associated With Rapid Local Adaptation to Environmental Shifts in Wild Barley Authors: Evgenii Potapenko, David Schwarts, Tali Mandel, Nimrod Ashkenazy, Dana Fuerst, Guy Atsmon, Sagi Snir, Abraham B Korol, Michael B Kantar, Avi Bar‐Massada, Sariel Hübner Published: 2026 ABSTRACT ABSTRACTGenomic mechanisms of local adaptation must be highly responsive in geographic regions where climate is changing rapidly. The Levant region is a critical biodiversity hotspot and the distribution edge for many species, including the wild ancestor of domesticated barley. This region is under an accelerated desertification process, thus enforcing a rapid genomic response to the projected environmental changes. To elucidate the genomic basis of rapid local adaptation, we studied wild barley populations using an ecological‐genetic sampling design that decouples environmental variation from demographic background. We collected and sequenced 300 wild barley individuals and evaluated the phenotypes of 3600 progeny plants over 3 years. Our genomic analyses revealed that local adaptation is associated with clusters of candidate genes forming haplotype blocks. These clusters are enriched with environment and stress responsive genes, including flowering time regulators, drought and heat responsive genes. We identified six candidate adaptive haplotype blocks which span 1–8 Mbp and are distributed across chromosomes 1H, 2H, 4H and 5H, each segregating as two major haplotypes. Additionally, we integrated over 2600 occurrence records into ecological and evolutionary modelling to assess the genomic vulnerability of populations to projected future climates. Our study identifies candidate genomic regions and environmental drivers of local adaptation in wild barley and highlights the advantage of haplotype blocks architecture in orchestrating an efficient response to rapid environmental change. We highlight the ecological factors most strongly associated with the observed evolutionary responses and provide insights and guidelines for biodiversity conservation and implementation of crop wild relatives in breeding.Keywords: abiotic stress, climate change, haplotype blocks, landscape genomics, local adaptation, population genomics, sampling design, wild barley Introduction Accelerated climate change and destruction of natural habitats, specifically in the past century, are threatening biodiversity, agriculture, and human health worldwide (Díaz et al. 2019; Hitz and Smith 2004). To keep pace with the projected accelerating environmental change, species must rapidly evolve through genetic adaptation. Adaptation to a changing environment is driven by natural selection, which can act gradually on numerous small contributing alleles (Fisher 1999) or on a few large‐effect loci (Barton and Keightley 2002). Apart from the strength and type of selection, the pace and direction of the subsequent evolutionary change is determined by other factors including linkage disequilibrium, migration, and genetic drift (Charlesworth et al. 2003). Local adaptation in a heterogeneous environment evolves through divergent selection favouring altered beneficial genetic variation at different habitats while overcoming contrasting forces of gene flow and genetic drift (Kawecki and Ebert 2004). Theory predicts that local adaptation can be accelerated by favouring the evolution of linkage between beneficial loci (Hill and Robertson 1966; Lenormand and Otto 2000; Yeaman and Whitlock 2011). Thus, adaptive response to environmental change can be accelerated when multiple beneficial alleles are inherited together as a non‐recombining haplotypic genomic architecture (Barrett and Schluter 2008; Hansen 2013). Other modes of local adaptation are attributed to specific events of chromosomal modifications like inversions and deletions (Jayakodi et al. 2020; Todesco et al. 2020), yet their mode of accumulation in response to environmental change remains elusive. Beyond understanding the mechanisms of adaptation, the accessibility of genomic data now enables the development of predictive models for population dynamics from an evolutionary perspective by estimating the adaptive potential to future climate (Exposito‐Alonso et al. 2019; Fitzpatrick and Keller 2014; Waldvogel et al. 2020). The impact of climate change is specifically alarming in biodiversity hotspots, and in regions that are rich with crop wild relatives that harbour beneficial alleles for breeding (Khoury et al. 2021). The eastern Mediterranean basin (Levant) is bridging Africa and Eurasia and considered one such global biodiversity hotspot and a home to many crop wild relatives (Lev‐Yadun et al. 2000; Zohary et al. 2012). Wild barley ( Hordeum spontaneum ), the progenitor of cultivated barley (Lev‐Yadun et al. 2000; Zohary et al. 2012), occurs in large stands along a wide range of habitats in the southern Levant, from the Mediterranean tree‐line zone at high elevation (1600 m above sea level) to desert climate near the Dead Sea (−400 m). As a major crop wild relative, it represents a precious genetic resource threatened by ongoing environmental change (Khoury et al. 2021), thus providing an ideal system to study local adaptation along wide environmental gradients. However, the geographical settings in this region induce a tight link between latitude and environmental clines (north‐Mediterranean, south‐desert). This correlation may confound the identification of adaptive genetic variation because population structure and selection are often detected as similar genomic outliers of differentiation (Hübner et al. 2009; Lasky et al. 2012). Moreover, the high rate of self‐fertilization in wild barley (Abdel‐Ghani et al. 2004) intensifies the effect of demographic differentiation between populations, over local adaptation through selection. Previous collections of wild barley in this region (Hübner et al. 2009; Nevo et al. 2005) were not designed to tackle these issues and thus are less suitable for identification of genome‐environment associations and development of predictive evolutionary models. Here, we describe the establishment of a new wild barley collection designed specifically to overcome the confounding correlation between demographic population structure and local adaptation. Whole genome sequence data were generated for the entire collection and used to identify genetic patterns involved in local adaptation. Understanding these mechanisms and identifying the underlying genes enabled the development of predictive models for risk assessment which can inform conservation and management strategies to mitigate the adverse effects of climate change. Material and Methods Sampling Design and Plant Material The studied area lies within longitude 34–36 and latitude 29.3–33.5. To obtain a wide, yet balanced representation of the environmental and genetic variation of wild barley populations, sampling sites were selected based on climatic data obtained from the Israeli Meteorological Services (IMS), available eco‐geographic raster layers, and available genetic information from the Barley1K collection (Hübner et al. 2009). The sampling strategy objective was to break the linkage between environmental and geographic distances by selecting locations that are characterized by similar environmental conditions (e.g., precipitation, temperature, soil type) but that are found at different geographical regions. To target such locations for sampling, the studied area was divided into six polygons. Three polygons were defined based on the assignment (> 85%) of populations from the Barley1K collection to the known ecotypes' borders in this region (Hübner et al. 2013) and three polygons between ecotype regions were defined as intermediate zones. Within each polygon, five sampling sites were selected based on the environmental conditions and soil in each habitat (Figure S1). The exact sampling locations were selected to maximize the representation of environmental variation and minimize anthropogenic disturbance. At each site, mature spikes were sampled from 10 plants that were separated by at least 10 m from each other, and geographical coordinates were recorded. Thus, 300 individual plants were sampled across 30 different sites. One seed from each plant was germinated and grown under controlled room conditions and was self‐crossed to reduce potential heterozygosity. This procedure (single‐seed‐descent and selfing) was repeated three times to obtain a fixed accession from each plant. Climate Data In all models for present and future conditions, we used the bioclimatic variables obtained from the WorldClim v2.1 database (Fick and Hijmans 2017). Non‐relevant variables for this region were excluded (e.g., precipitation in the warmest month), and multicollinearity was controlled using a maximum correlation coefficient of 0.8 between bioclimatic variables. The remaining seven bioclimatic variables were BIO1 (annual mean temperature), BIO2 (mean diurnal range), BIO3 (isothermality), BIO4 (temperature seasonality), BIO7 (temperature annual range), BIO12 (annual precipitation) and BIO15 (precipitation seasonality; Table S1). Future climatic variables were obtained following a conservative approach by averaging three Coupled Model Intercomparison Project Phase 6 (CMIP6) climate models (Eyring et al. 2016): MPI‐ESM1‐2‐HR, UKESM1‐0‐LL, and IPSL‐CM6A‐LR, for years 2061–2080 under mild (SSP1‐2.6) and severe (SSP5‐8.5) socioeconomic pathways. Data were downloaded from the WorldClim v2.1 database using the ‘geodata’ v0.6‐2 R package (Hijmans et al. 2024) and processed with the ‘raster’ package (Hijmans 2024a). Environmental variation across sampling sites was explored using a principal component analysis (PCA) of a 30 × 7 site by bioclimatic factors matrix containing the per‐site values of the seven retained bioclimatic variables. Variables were z‐score standardized prior to analysis, and PCA was applied to the standardized matrix using the prcomp function in R (R Core Team 2023). DNA Extraction, Library Preparation and Sequencing Genomic DNA was extracted from a leaf tissue of a single young seedling using the CTAB buffer (100 mM Tris HCl pH 8.0, 20 mM EDTA pH 8.0, 1.4 M NaCl, and 2% CTAB, 0.5% Sodium bisulfite), followed by Phenol/Chloroform purification protocol. Genomic DNA obtained from each sample was sheared to an average size of 350 bp using Covaris E220. Sequencing libraries were prepared using the NEBNext UltraTM II End Repair/dA‐Tailing Module (NEB—E7546) followed by the NEBNext UltraTM II Ligation Module (NEB—E7595). Samples were multiplexed with the Illumina‐TruSeq DNA UD Indexes (96 indexes, 96 samples) (Illumina—#20022370), and post‐ligation cleanup was done with 0.8× volume of homemade size‐selection Magnetic‐Beads (MB) washed twice with 80% ethanol. Libraries were amplified using KAPA HiFi HotStart ReadyMix (2×) (KAPABIOSYSTEMS—KK2600) and 10 μM of Illumina compatible primers. Finally, each library was quantified using the QubitTM dsDNA HS Assay Kit (Invitrogen—Q32854) and qualified using high‐sensitivity D1000 ScreenTape (Agilent) to a mean size of 550 bp. Pooling, titration, and sequencing on the NovaSeq 6000 system using the S4 flow cell with 2 × 150 bp cycles were done at the Roy J. Carver Biotechnology Center, University of Illinois at Urbana‐Champaign. Whole genome sequencing was targeted to at least 5× coverage for each accession, and additional sequencing was performed for samples that did not generate adequate coverage. Variant Calling Raw reads were evaluated for quality and GC content using FastQC software v0.11.5 (Andrews 2010) and adapters, and low‐quality data were trimmed using the fastp v0.20.0 software (Chen et al. 2018). Cleaned reads were aligned to the Morex reference genome v2 (Monat et al. 2019) using BWA‐MEM2 v2.2.1 (Vasimuddin et al. 2019; Table S2). Duplicated reads were marked using the gatk v4.1.9 and bam files were indexed with samtools v1.9 (Danecek et al. 2021). Variant discovery analysis was conducted with HaplotypeCaller as implemented in gatk4 (Poplin et al. 2018). In gatk4, gvcf files were produced for each sample using the HaplotypeCaller tool and combined to a single data set with GenomicsDBImport and GenotypeGVCFs. The obtained variants were filtered following the recommended hard filtering approach in gatk4 and based on the distribution of density plots generated for each statistic (Figure S2). During filtering, only single nucleotide polymorphism (SNP) data were kept and filtered for QualByDepth < 20, MappingQuality < 50, FisherStrand > 70, ReadDepth < 1000 and > 2000, QUAL < 140, StrandOddRatio > 2, and MappingQualityRankSum < −2.5. In addition, heterozygous sites were set as missing data, minimum minor allele frequency was set to 5%, maximum missing data were 10%, and the minimum number of reads to call an allele was 2. All data processing and variant calling were conducted on the Amazon Web Services (AWS) system by splitting the pipeline into two phases. The first phase included the quality checks, trimming, alignment, and bam files processing and generation of gvcf files; all these steps were performed for each sample independently; thus, a total of 300 instances were deployed with 96 cores per instance. Once complete, all data were passed into an elastic file system. The second phase included the development of a genomic database and genotyping across all samples. To reduce the load, the gvcf files were divided into 1000 segments per chromosome, producing a total of 7000 segments which were processed on 5000 spot instances to reduce costs. Final vcf files were concatenated and filtered locally. For more details and code, see https://github.com/hubner‐lab/. Phenotyping Phenotyping was conducted in a common garden setup over 3 years in an insect‐proof net‐house located in the Hula Valley, Israel (33°09′08.4″ N 35°37′15.8″ E). Before each growing season, seeds were stored in paper bags and oven‐dried (30°C) for 14 days. After drying, seeds were dipped in 4% sodium hypochlorite solution for fungal prevention and were sown in mid‐October in germination papers, watered, and put in dark‐cold (4°C) conditions for 10 days for vernalization and dormancy break. After this treatment, germinating seedlings were transferred into 5‐L pots with commercial soil mix (‘Ram 158’, Tuff Merom Golan Ltd., Merom Golan, Israel) and were grown under full irrigation for the rest of the growing season. The common garden experiment was conducted in a randomized block design with 4 replicates for each genotype (Figure S3; Table S3). The phenotypes recorded included flowering time (days from sowing to heading), plant height (tallest tiller length), flag‐leaf length, spike length, number of tillers, number of grains per spike, grain weight and total above‐ground biomass at maturity. Flowering time, plant height and spike length were recorded in all three seasons (2019, 2020, 2021); number of tillers in two seasons (2019, 2020), grain weight in two seasons (2019, 2021); and flag‐leaf length, number of grains per spike, and total above‐ground biomass at maturity were recorded only in 2019 (Table S3). Phenotype data were recorded daily during the growing seasons following the Zadoks scale definitions for development and phenology (Zadoks et al. 1974) using the FieldBook application v 5.0.7 (Rife and Poland 2014). The dates at which each plant attained a developmental stage (following Zadoks scale) were recorded and subsequently converted into the number of days between each stage. To account for variations in developmental progress among genotypes throughout the growing season, time‐based measurements were converted to growing degree days (GDD) with a base temperature of 0°C. To achieve this, temperature data were sourced from the official meteorological service, specifically from the Kfar Blum station (no. 8473). Species Distribution Modelling Species distribution models (SDMs) relate species occurrence to environmental predictors to produce a continuous map of habitat suitability, here defined as the model‐predicted probability that wild barley is present at a given location. We fitted the SDM using the random‐forest algorithm (Breiman 2001) implemented in the R package randomForest v 4.7‐1.1 (Liaw and Wiener 2002) using the default settings (500 trees, two variables tried at each split). Variable importance was quantified as the mean decrease in node purity, reflecting how effectively each variable separates classes within the random forest model (Breiman 2001). We restricted our analyses to wild barley samples which occurred in sites with more than 100 mm of rainfall per year to avoid sporadic records, resulting in 2641 presence points. Accordingly, we selected at random a sample of 2641 background points to reflect the distribution of environmental conditions in the entire region. We divided the data into training and testing sets based on an 80%–20% split, ensuring that each set had a nearly equal number of presence and background points. We extracted the values of seven environmental variables (distance to nearest river, mean diurnal temperature range, precipitation seasonality, precipitation of the wettest quarter, elevation above sea level, cation exchange capacity of the soil, and soil pH index measured in water solution) from their respective raster layers, for all presence and background points and training and testing sets. Using the values extracted from the training set, we fitted a random‐forest model on the training data, with the dependent variable being the presence or absence of wild barley in the site (a binary variable). We then predicted the habitat suitability for wild barley across the entire study area by running the random‐forest model on the full set of raster layers of environmental variables. To evaluate model performance, we re‐ran the model using the set‐aside testing dataset and calculated its predictive accuracy according to the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. To predict future distribution of wild barley, we re‐ran the random‐forest model using future conditions under two climatic scenarios (RCP45 and RCP85) from CMIP6 (Eyring et al. 2016) for 2070. We generated predictions of habitat suitability for wild barley across the entire study area under each combination of year and scenario. Population Structure Analyses Population stratification analyses were conducted using pruned variants to avoid the effect of linkage disequilibrium (LD) on population structure inference. The SNP dataset was LD‐pruned with PLINK v1.90 software (Purcell et al. 2007) with a minimum LD cut‐off of r 2 = 0.1, window size 100 kbp, and step size 20 kbp. Population structure was assessed using the sNMF function from LEA v3.18.0 package (Frichot and François 2015) with the number of ancestry populations (K) estimated based on principal component analysis (PCA) and structure plots for K equals 3, 4, and 5 (Figure 1b, Figure S4a–d). Exploring genomic regions associated with local adaptation. (a) A map depicting the sampling design and sites. The geographic area was split into six polygons in accordance with ecotype borders, and intermediate zones. Within each polygon, five sampling locations were targeted based on ecological conditions. The legend denotes the five environmental selection categories: Highest CV (precipitation stability over years), Soil type (light rendzina), Min Prec. (lowest annual precipitation), Hot (highest temperatures) and Cold (lowest temperatures). (b) Principal component analysis based on genomic data obtained for all 300 accessions with pie charts representing assignment to the three genetic clusters identified by structure analysis. (c) Locations and gene flow between sampling sites. Colour gradient represents annual precipitation variation (blue: High; red: Low), pie charts show the assignment to clusters, and pie chart size indicates Tajima's D values (smaller pies represent lower values). Black lines between populations denote gene flow. (d) Circular plot of candidate adaptive haplotype blocks (HBs). From outer circle inwards: Gene abundance per 1 Mbp windows, positions of the six selected adaptive HBs (red vertical lines), significant associations between SNPs and climatic factors based on EMMAX, and LFMM2. SNPs associated with temperature factors are marked in red, and precipitation in blue. Dashed lines highlight SNPs within overlapping HBs. (e) Linkage disequilibrium heatmap of a representative HB (chr1H:421,725,937‐424,368,069), illustrating the extent of genetic linkage across this genomic region. Top panel: Focused Manhattan plot highlighting association signals for isothermality (BIO3). Middle panel: Local LD heatmap for the same region in samples carrying the adaptive HB. Bottom panel: LD heatmap of the corresponding region in samples lacking the adaptive HB. In both heatmaps the colour key indicates the pairwise LD measure D′. To evaluate the correlation between genetic and environmental variation while controlling for geographic distance, we performed a partial Mantel test using the vegan v2.6‐8 R package (Oksanen 2009). Genetic distances were computed as Jaccard dissimilarity based on the Q‐matrix obtained from sNMF analysis. Geographic distances between sampling sites were calculated as Vincenty distances using their longitude and latitude coordinates. Environmental distances were computed as Euclidean distances from the seven selected bioclimatic variables. This approach allowed us to test the association between genetic and environmental distance matrices while accounting for the effect of spatial structure. To quantify the contribution of individual climate variables and spatial autocorrelation to genome‐wide variation, we followed the analytical framework of Lasky et al. (2012) for 30 unique sampling sites. The genomic response matrix Y site (30 × 111) was obtained by averaging the first 111 principal components (explaining ≥ 80% of individual‐level genomic variance) within each site. The climate predictor X clim (30 × 7) comprised the seven bioclimatic variables retained after multicollinearity filtering (BIO1, BIO2, BIO3, BIO4, BIO7, BIO12, BIO15). Geographic predictors (X PCNM) were represented by principal coordinates of neighbour matrices (PCNM; Borcard and Legendre 2002) computed on the Vincenty‐ellipsoid inter‐site distance matrix (truncation threshold = 63.5 km; distances above threshold set to 4×). Spatial distances were computed with the ‘distance’ function in terra v1.7‐78 R package (WGS‐84 geodesic; Hijmans 2024b). Among the 17 positive‐eigenvalue PCNM vectors, 6 were retained by forward selection using the ‘ordiR2step’ function of vegan R package (P in = 0.05, R 2‐scope stopping criterion; Blanchet et al. 2008). Variance partitioning was performed with the ‘varpart’ function of vegan package (Oksanen 2009), yielding climate‐unique, geography‐unique, shared and unexplained adjusted R 2 fractions. The unique fractions were tested by partial redundancy analysis (pRDA) with 999 permutations. Three pRDA models were then fitted: (A) Y site ~ X PCNM, (B) Y site ~ X clim, and (C) Y site ~ X clim + Condition(X PCNM). The per‐variable P x metric (Lasky et al. 2012) was computed as: P x = Σ k (|r xk | × λ k )/r 2, where r xk is the Pearson correlation of predictor x with model‐fitted site scores on canonical axis k, λ k is the eigenvalue of axis k, and r 2 is the total variance available. P x therefore expresses the explanatory contribution of each individual environmental variable, which refers to the genomic variation captured by the pRDA model attributable to that predictor, allowing the predictors to be ranked by explanatory weight. P x was computed from models B and C for the primary seven bioclimatic set. Average nucleotide diversity (π) and Tajima's D were calculated per population (sampling site) across whole chromosomes with VCFtools v0.1.16 (Danecek et al. 2011) and then averaged genome wide. Genome Environment Association Analyses Genome‐environment association (GEA) analyses were performed using three different approaches: the RDA implemented in the vegan v2.6‐8 R package (Oksanen 2009), a mixed linear model implemented in the program EMMAX (Kang et al. 2010), and a latent factor mixed model (LFMM2) implemented in the LEA v3.18.0 package (Frichot and François 2015). Prior to the analysis, the missing variant calls were imputed based on the admixture coefficients (K = 3) calculated with the ‘impute’ function in the LEA package (Frichot and François 2015). To improve the signal in GEA analysis, we calculated identity by state metric for each pair of individuals with PLINK v1.90 software (Purcell et al. 2007) and filtered one sample from each pair which had identity by state higher than 0.99. After removing individuals, additional SNP filtering was applied based on a minor allele frequency of 5% and a maximum of 10% missing data (Figure S5). In EMMAX and RDA analyses, population structure was corrected using the first 3 principal components derived from a PCA as covariates, which were chosen based on the proportion of variance explained (Figure 1b, Figure S4d). Additionally, in EMMAX analysis, a Balding‐Nichols kinship matrix was used to correct relatedness as a random effect (Kang et al. 2010). In LFMM2, we corrected for population structure using 15 latent factors chosen with cross‐entropy criterion to minimize residual genomic inflation (Frichot and François 2015; Figure S4e). For RDA, candidate SNPs were identified from the distribution of SNP loadings on the first constrained axes (Capblancq et al. 2018), with p‐values corrected via the genomic inflation factor. We evaluated the significance of the full model, each constrained axis, and each environmental term using permutation tests (999 permutations) with the anova.cca function in the vegan package. Across all GEA analyses, significance was determined after Bonferroni correction at α = 1.8 × 10−9. Following Bonferroni correction, RDA still returned a considerably large number of significant SNPs (Figures S6–S8), suggesting an inflated false‐positive rate. Therefore, to ensure high confidence, we only considered loci that were supported by significant SNPs identified in both EMMAX and LFMM2. Genome‐Wide Association Study Genome‐wide association studies (GWAS) were performed after imputing SNP calls with the ‘impute’ function in the LEA v3.18.0 package (Frichot and François 2015). To guarantee that imputation does not introduce bias, the analysis was conducted for several traits with and without imputation, indicating consistency. To correct for population structure, the first three principal components were used, and genetic relatedness was corrected with a Balding‐Nichols kinship matrix (Kang et al. 2010). Phenotypic observations were summarized per accession using random‐effects models fitted with the lme4 R package v1‐1‐35.5 (Bates et al. 2015). The random‐effects structure was specified per trait according to seasonal data availability (Table S3). For multi‐season traits: yᵢⱼₖ = μ + gᵢ + sⱼ + bⱼₖ + εᵢⱼₖ, where yᵢⱼₖ is the standardized phenotypic value of genotype i in season j and block k, μ is the grand mean, and gᵢ ~ N(0, σ 2 g ), sⱼ ~ N(0, σ 2 s ), bₖ ~ N(0, σ 2 b ), εᵢⱼₖ ~ N(0, σ 2 e ) are the random effects for genotype, season, block and residual, respectively. The same blocks design was used in all three seasons and for single‐season traits the season term sⱼ was omitted. Phenotypes were standardized before fitting and the Best Linear Unbiased Predictions (BLUPs) of the genotype effect, ĝᵢ, were extracted and used as the GWAS response variable. Broad‐sense single‐plot heritability was estimated from the fitted variance components: H 2 = σ 2 g / (σ 2 g + σ 2 s + σ 2 b + σ 2 e ) for multi‐season traits; H 2 = σ 2 g / (σ 2 g + σ 2 b + σ 2 e ) for single‐season traits (Falconer and Mackay 1996). To correct for multiple testing, we applied a permissive exploratory significance threshold of p < 1 × 10−4 to each GWAS scan (Figure S9). Due to the large number of SNPs, standard corrections (e.g., Bonferroni) are overly conservative and eliminate all signals. The threshold used in the GWAS enabled us to reduce much of the false positive calls, however, candidate regions and genes were not determined based on this threshold alone rather by combining the results from GWAS and GEA analyses. To evaluate the performance of the analyses we calculated the genomic inflation factors (λ GC) and quantile‐quantile (QQ) plots for each GWAS scan (Figure S10). Identifying Haplotype Blocks and Candidate Genes The GEA analyses highlighted genomic regions with dense clusters of associated SNPs (Table S4). To delineate the boundaries of each region, we used the ‐‐blocks command in PLINK v1.90 (Purcell et al. 2007). Within each of the GEA‐highlighted regions, the SNP with the lowest p‐value was used as the representative anchor. Samples were split into two groups according to the allele carried at that anchor SNP. Prior to the analysis, we filtered variants with a minor allele frequency < 0.2 to mitigate the influence of rare alleles on contingency analyses and thinned the data to one SNP per 1 kbp. The region haplotypic block boundaries were defined using the following parameters: ‐‐blocks‐max‐kb 10,000 to limit the maximum range to 10 Mbp, ‐‐blocks‐strong‐lowci 0.51 to restrict the analysis to SNPs in strong LD, and ‐‐blocks‐inform‐frac 0.60 to set a minimum of 60% variants within a candidate block are informative, excluding regions of sparse LD where block boundaries cannot be reliably estimated. To visualize the patterns of LD across the identified genomic regions, we employed LDBlockShow v1.41 tool for each region in each group (Dong et al. 2020; Figures S11 and S12). The SNPs significantly associated with phenotyping traits were subsequently overlapped with the defined candidate haplotype blocks (HBs). To assess whether significantly associated SNPs are overrepresented in HBs, we performed a one‐tailed Fisher's exact test using ‘fisher. test’ function in R (Figure S9; Table S5). A 2 × 2 contingency table was constructed, comparing the number of significant SNPs within the HB to those outside the HB, as well as the non‐significant SNPs within and outside the HB. The test was conducted with the alternative hypothesis set to ‘greater’ to evaluate the overrepresentation of significant SNPs in HBs. The p‐values were computed using Monte Carlo simulation for enhanced accuracy (Table S5). The visualization of GWAS and GEA results were obtained using the CMplot v4.5.1 R package (Yin et al. 2021) and the topr v2.0.2 R package (Juliusdottir 2023). To estimate the proportion of variance explained (PVE) by significantly associated SNPs within the identified HBs for each bioclimatic factor and phenotypic trait, linear models were constructed. For each HB, a separate model was built with each bioclimatic factor and phenotypic trait (BLUPs) as the response variable and the set of significantly associated SNPs from the HB as explanatory variables. The proportion of variance explained was determined by calculating the adjusted r 2 value for each model. To assess whether candidate HBs fall within low‐recombination compartments of the genome, we used the recombination rate landscape of the Morex V2 pseudomolecules from the barley 50 K iSelect array re‐aligned to V2 coordinates (Bayer et al. 2017; James Hutton Institute 2020 release; Figure S13). Pericentromeric intervals were defined at two thresholds: T 1 = 0.1 cM/Mbp (strict) and T 2 = 0.5 cM/Mbp (permissive). To further test whether candidate HBs overlap with a polymorphic inversion, we performed a genome‐wide local PCA screening using the lostruct v0.0.0.9000 R package (Li and Ralph 2019). Local PCAs were computed in non‐overlapping 1000‐SNP windows across all chromosomes using the imputed SNP dataset. For each window, a PCA retaining K = 2 principal components was computed. Between‐window distances were calculated as the Frobenius‐based principal subspace distance (npc = 2). Multidimensional scaling (MDS) was applied with the ‘cmdscale’ function in R to the pairwise distance matrix retaining two MDS dimensions for a total of 32,555 windows (range: 3593–5476 per chromosome). Because MDS1 was mainly affected by the pericentromere—chromosomal arm LD contrast characteristic in barley, we applied the outlier criterion to both MDS1 and MDS2, to minimize the chance of missing inversion signals on secondary axes (Figure S14). Genomic windows with |MDS1| or |MDS2| > 4 × median absolute deviation (MAD) from the per‐chromosome median were classified as structural outliers. To identify signatures of recent positive selection among samples sharing candidate adaptive HBs, we used RAiSD v2.9 (Alachiotis and Pavlidis 2018). RAiSD detects regions under selection based on a composite μ‐statistic that integrates local patterns of polymorphism, site frequency spectrum, and linkage disequilibrium. Signatures of recent positive selection within and around the HBs were screened in intervals of ±3 Mb of flanking sequence for individuals carrying the minor haplotype and major haplotype (Table S4). RAiSD analysis was conducted independently for each subset with default parameters (Figure S15). For visualization, the μ‐statistic was scaled to match the p‐values from the GEA analysis (Figure 2e). Genomic footprints of local adaptation under climate change. Species distribution model (SDM) predicted for present (a) and future (b) climate, where the colour gradient on the map depicts the suitability score. In both panels, pie charts mark the sampling sites and depict the proportion of assignment of each population to the three genetic clusters. (c) Relative contribution (mean decrease in node purity) of the environmental predictors in the SDM, including pH (PHIHOX), distance from river (ds.river), cation exchange capacity of the soil (CECSOL), elevation (dem.ascii), mean temperature range (BIO2), precipitation seasonality (BIO15), precipitation of wettest quarter (BIO16). (d) Circular plot of candidate adaptive genomic regions. From outer circle inwards: Gene abundance, locations of adaptive HBs (green), significant associations in the GEA analyses (red dots), associations with flag length (blue dots); dashed lines indicate overlaps between GEA and GWAS for flag length. (e) Detailed view of a candidate adaptive HB on chromosome 4 showing significant associations with BIO3 as detected by EMMAX (red) and LFMM2 (blue), along with the selective sweep signals (black; scaled in the same range for visibility). At the top panel, the entire chromosome is plotted with the marked region in red rectangle. The middle panel is a zoom‐in to the HB region where horizontal red dashed line corresponds to Bonferroni correction and the candidate adaptive genes are indicated at the bottom panel. Genetic Offset Modelling To model the genomic offset among populations, we used a gradient forest analysis as implemented in the gradientForest package (Ellis et al. 2012). Two models were used to evaluate the vulnerability of populations to climate change. A ‘neutral’ model was trained using 100,000 randomly sampled SNPs across the entire genome and the ‘adaptive’ model was trained using candidate environment‐associated SNPs following a multistep process. For each of the seven bioclimatic variables, we retained the top 0.01% of SNPs ranked by p‐value across both EMMAX and LFMM2 (circa 2000 SNPs per scan). We then pooled these top SNPs by method and took the intersection of the two sets. Consequently, a SNP was retained if it appeared in the top 0.01% of at least one EMMAX scan and at least one LFMM2 scan. This approach ensured every candidate was supported by both statistical frameworks, even if the two methods flagged the SNP under different, inter‐correlated climatic variables. This procedure yielded 14,564 unique candidate SNPs, which constituted the ‘adaptive’ SNP set used to train the gradient forest model. The maximum number of splits was calculated by the formula log2(0.368*n/2), where n is the number of sampling sites. The correlation threshold was set to 0.5 and the number of bootstrapped trees to 1000. In each model, allelic composition across SNPs was set as the response variable, the seven major bioclimatic variables were included as predictors in addition to 8 principal coordinates of neighbour matrices (PCNM) components calculated with vegan v2.6‐8 (Oksanen 2009) to correct the effect of spatial heterogeneity and structure. PCNMs were computed from sampling coordinate Euclidean distances, retaining approximately half of the positive eigenvectors to model positive spatial autocorrelation without overfitting (Borcard and Legendre 2002). To model the effect of current and future environmental conditions the present and future bioclimatic variables were used, respectively. To identify the extent of stress along space using genetic models, we performed a PCA transformation on the predictions from each model (‘adaptive’ and ‘neutral’). Subsequently, we applied Procrustes rotation to align the datasets in a comparable orientation. The residuals were then extracted to estimate the differences between the ‘neutral’ and ‘adaptive’ models, thus high similarity indicates lower stress. To estimate the adaptive risk in populations, we employed a genetic offset approach. Initially, projected future climate data were used to predict the genomic composition that would be adaptive under future climatic conditions. Genetic offset was then calculated as the Euclidean distance between the current adaptive genomic composition and the predicted future adaptive genomic composition. Mantel test implemented in vegan v2.6‐8 (Oksanen 2009) was used to test for correlation between suitability scores obtained from the SDM and the genetic offset predictions for mild and severe socioeconomic scenarios using 999 permutations. Results Haplotype Blocks Are Associated With Rapid Local Adaptation We sampled 300 individual plants from 30 different sites across Israel following an ecological‐genetic sampling design to overcome the correlation between adaptive divergence and population genetic structure (Figure S1). To achieve this, we targeted populations that occur under similar environmental conditions (e.g., precipitation, temperature, soil type) but that are located at different geographical regions (Figure 1a). To guarantee that the populations sampled at similar environmental conditions indeed represent different genetic background, we used available genetic information from a previous collection to draw the borders of each genetically distinct cluster (Barley1K; Hübner et al. 2009). Based on this analysis, six polygons were defined, three of which represent distinct genetic clusters (ecotypes), and three represent intermediate zones between ecotypes. Within each polygon, five sampling sites with low effect of human disturbance (e.g., natural reserves) were selected, and 10 representative accessions were collected in each site. The entire collection was whole genome sequenced, and 33 million variants were obtained after stringent filtering (Figure S2; Table S6). To evaluate population structure and potential admixture, the full dataset of 33 M SNPs called across 300 accessions was pruned to reduce bias imposed by extensive linkage disequilibrium (LD > 0.1), leaving a total of 879,897 SNPs for PCA and structure analyses (Figure 1b, Figures S4 and S5). Wild barley populations were largely assigned to clusters following the expected split to known ecotypes with a gradual transition between them along the defined intermediate zones (Hübner et al. 2013; Figure 1b,c). In accordance with our hypothesis, the new sampling strategy effectively mitigated the confounding effect of isolation by distance, allowing for a more precise evaluation of environmental drivers of genetic variation. To assess the extent of correlation between genetic and environmental variation, we performed a partial Mantel test by accounting for geographic distances. The analysis was performed for the new collection and compared to a previous collection of wild barley from the studied region (Barley 1 K). A significant association between environmental variation and genetic differentiation was detected in the newly established collection (r = 0.21, p = 0.001, n perm = 999; Figure S16b). In contrast, this relationship was not statistically significant in the Barley1K where the signal was masked by the geographic distances (r = 0.066, p = 0.065, n perm = 999; Figure S16a). To further explore how our sampling design represents spatial and environmental variation and enables us to distinguish between them, we selected seven bioclimatic variables (Table S1) from the 19 variables in the WorldClim2 database (Fick and Hijmans 2017) after excluding highly correlated variables and those with low ecological relevance to this region (e.g., BIO14—Precipitation in driest month; Figure S17). The selected variables are annual mean temperature (BIO1), mean diurnal range (BIO2), isothermality (BIO3), temperature seasonality (BIO4), temperature annual range (BIO7), annual precipitation (BIO12), and precipitation seasonality (BIO15). A principal component analysis (PCA) of these seven variables across the 30 sampling sites showed that the first two principal components capture 80.6% of the total variance. PC1 (47.8%) separated high precipitation northern sites from arid southern sites, and PC2 (32.8%) reflected a gradient of temperature seasonality and isothermality (Figure S18). To disentangle the genomic variance uniquely attributable to climate from the variance confounded with spatial structure, we applied a variance partitioning framework (Lasky et al. 2012). The selected seven climatic variables and six spatial predictors (6 PCNMs) together explained 31.8% of the genomic variance at the site level. The fraction unique to climate was 9.5% (p = 0.001, n perm = 999) and the fraction unique to geography was 9.3% (p = 0.001, n perm = 999). The shared fraction, representing genomic variance jointly attributable to climate and geography, was 12.9%, leaving a residual unexplained fraction of 68.2%. Using the P x metric, which quantifies the explanatory contribution of each individual variable, the most influential bioclimatic predictor in the unconditional climate RDA (R 2 adj = 18.8%) was BIO12 (annual precipitation; P x = 16.37%). After removing spatial autocorrelation in the conditional RDA model (R 2 adj = 9.5%), BIO1 (annual mean temperature) had the strongest contribution (P x = 9.86%; Figure S19). The environmental PCA and the recovery of independent and significant fractions of genomic variation attributable to climate and to geography show that environmental and spatial signals are separable in the new collection. Building on this climatic framework, we next performed genome‐environment association (GEA) analyses to identify genomic regions underlying local adaptation (Figures S6 and S7). To improve the association signal, we calculated the identity‐by‐state (IBS) metric across all accessions and filtered out samples with high genetic similarity (IBS > 0.99), leaving 218 accessions for the GEA analyses. The SNP dataset was filtered accordingly to capture the changes in minor allele frequency and missing data, leaving a total of 27,239,486 SNPs for the GEA analyses (Figure S5). We used three complementary GEA approaches to identify candidate adaptive loci: EMMAX, LFMM2, and RDA. After Bonferroni correction (p < 1.8 × 10−9), the RDA analyses yielded a high proportion of significant signals compared to other analyses (Figure S8). The excess of RDA signals is likely attributed to the high marker density in our dataset (~27 M SNPs). We therefore treated RDA as a complementary exploratory analysis and restricted the candidate genomic regions identification to SNPs jointly supported by EMMAX and LFMM2 (Figure 1d, Figures S6 and S7). Altogether, 894 SNPs passed the Bonferroni threshold in both EMMAX and LFMM2 and were classified as high‐confidence candidate environment‐associated SNPs, where the majority clustered within six genomic regions on chromosomes 1H, 2H, 4H and 5H. We focused on these six regions for linkage disequilibrium (LD) analysis and identification of haplotypes. To define the candidate adaptive haplotype genomic boundaries, we selected one representative SNP with the lowest p‐value from each of the six regions and calculated LD around it using all samples carrying the same allele (Figure 1e). This approach enabled us to identify the haplotypic architecture among individuals that are adapted to similar environmental conditions but originated from different populations. Based on these analyses, we delineated the boundaries of the corresponding haplotype blocks (HBs) at each of the six genomic loci (Figures S11 and S12; Table S4). Because the large pericentromeric regions of barley can passively generate extended LD, we investigated whether the six candidate HBs overlap with compartments characterized by low recombination or structural variations such as polymorphic inversions. Recombination rates were estimated using a published genetic map from the barley 50 K iSelect array. Using a strict threshold (T 1 = 0.1 cM/Mbp), all six haplotype blocks lie exclusively in the distal long arm compartments, located 151 to 540 Mbp from the centromere midpoint, and exhibit mean local recombination rates between 0.173 and 1.089 cM/Mbp (Figure S13). Under a more permissive threshold (T 2 = 0.5 cM/Mbp), two blocks (chr1H:421,725,937–424,368,069 and chr5H:466,549,570–468,907,420) fall at the borders of the extended zone of low recombination and thus require cautious interpretation, whereas the remaining four HBs reside in the high recombination compartment. Most of the defined HBs were located in gene‐dense regions and no bias was identified compared to other similarly gene‐rich regions (Figure 1d). To explicitly test if the structural integrity of these HBs could stem from polymorphic inversions rather than recent selection, we performed a genome‐wide local PCA over 32,555 non‐overlapping windows across all seven chromosomes. Outlier windows were concentrated at mid‐chromosome positions, which aligns with the established pericentromeric architecture of barley (Figure S14). No outlier windows overlapped with any of the six candidate HBs on either axis. Combined with the spatial recombination analysis, the lack of an inversion signature indicates that the extended haplotype structure observed at the candidate loci is unlikely to be a passive consequence of either pericentromeric recombination suppression or structural variation. Moreover, the proportion of variance explained (PVE) in each bioclimatic variable by significantly associated SNPs across HBs was calculated using linear models indicating that the SNPs explain a considerable portion of the variation among temperature‐related variables (PVEBIO2 = 53%, PVEBIO3 = 68%, PVEBIO4 = 38%, PVEBIO7 = 25%). Thus, the identified clustered genetic architecture within HBs can potentially facilitate a synchronized adaptive response to correlated environmental stressors. Climate Change Increases Geographic and Phenological Isolation Between Populations To characterize current habitat suitability for wild barley and estimate its future distribution under climate change, we implemented a random‐forest species distribution model (SDM). SDMs relate species occurrence to environmental predictors to produce a continuous estimate of habitat suitability, here defined as the probability of wild barley presence at a given location. The model was used to predict shifts in the current distribution of wild barley in the southern Levant under projected climate scenarios for 2070. To improve the model performance, the number of observations was expanded to include 2641 locations where wild barley populations were recorded in the last decade (BioGIS 2002). In the models, four static soil and geographic parameters (pH, cation exchange capacity, distance to river, elevation) were included together with three predicted bioclimatic factors (BIO2, BIO15, BIO16; Figure 2c). Collinearity between variables was controlled and erratic parameters (e.g., the normalized difference vegetation index, NDVI, and vegetation) that cannot be robustly predicted in the future were excluded. High prediction accuracy (AUC > 0.92) was obtained for the models under different scenarios, and the projected distribution of wild barley and suitability of habitats were well supported by available records of population densities (https://biogis.huji.ac.il/). Lower population densities are usually observed at the edge of the distribution range; however, these regions are also characterized by lower suitability which may reflect eco‐physiological constraints due to reduced water availability in desert habitats (Figure 2a). The highest habitat suitability under current conditions was obtained at lower elevations along the Mediterranean region and at high elevations in the desert, where precipitation is the main limiting factor (Figure 2a, Figure S20). Under the considered scenarios for future climate, a maximum reduction of 7% in suitability was predicted across habitats. Reduction in suitability was pronounced in the drier regions (annual precipitation < 300 mm) along the Jordan Valley in the east, and the Negev Desert in the south (Figure 2a,b). Moreover, the expected shift in precipitation was also reflected with a reduction in suitability along the southern coastal region (Figure 2a,b). The pattern of declining suitability manifested under mild and severe scenarios echoes the ongoing process of desertification across the Levant. Despite the predicted loss of habitats along this region, high suitability will be maintained along creeks of large drainage basins, and along mountain ridges where temperatures are expected to remain mild also under future conditions (Figure 2b). The better suitability along creeks and higher elevations is already noticed today in the desert and is expected to expand to other regions in the future. Desert populations are more isolated (average pairwise isolation by distance, IBD = 0.7) compared to Mediterranean populations (average pairwise IBD = 0.47; t = 3.54, p = 9.3 × 10−4; Figure S21), which is also reflected by a 35% reduction in nucleotide diversity in the desert. The lower diversity and reduced gene‐flow between desert populations are likely to further constrain the maintenance of minimal genetic variation necessary to respond and sustain future climate. This trend of desertification and loss of suitable habitats is likely to further increase isolation of populations, funnelling wild barley populations and probably other species to refugia habitats along creeks and higher elevations. To link the adaptive response to climate change with plant phenotypes, a common garden experiment was performed. A total of 3600 plants were grown and screened for key phenological and morphological traits over 3 years. A wide range of phenotypic features was recorded in accordance with the environmental variation along habitats (Tables S3 and S7). For example, flowering time extended over 74 days between the first flowering desert accessions (73 days from sowing) to the late flowering accessions from Mt. Hermon (147 days from sowing). A similar trend was also observed in morphological traits such as spike length, where spikes were four times longer in northern accessions (max = 22 cm) compared with desert ecotype (min = 5 cm). Among the measured traits, heritability was highest for flowering time (H 2 = 0.57) and lowest for number of tillers (H 2 = 0.012), implying that the adaptive response is attributed to synchronized flowering (escape strategy) while reduction in biomass is a plastic feature which is more responsive to changes in environmental conditions between years. To identify genomic regions and candidate genes that underlie the measured traits, we performed GWAS corrected for population structure and relatedness. To link the results with adaptation along an environmental gradient, the identified top signals from the GWAS (p < 10−4) were overlapped with the HBs identified in the GEA analyses. Interestingly, all six HBs overlapped with trait‐associated signals, of which four were supported by two or more phenotypic traits suggesting a pleiotropic regulation of adaptation and enhanced adaptive response projected by a clustered genetic architecture. Moreover, a selective‐sweep analysis using the composite μ‐statistic for individuals carrying the minor haplotype recovered a pronounced signal of selection around most of the HBs (Figure 2e, Figure S15). To assess whether overlaps between GWAS and GEA at the candidate adaptive HBs exceed random expectation, a Fisher's exact test was conducted supporting the concordance between GEA and GWAS signals at all six HBs (Figure 2d, Figure S9; Table S5). The PVE for each phenotype by HBs was highest for plant height (tiller length, 0.69), and lowest for the number of tillers (0.18). These results are consistent with the overall heritability of these traits. Following these results, the HBs were further explored for potential candidate adaptive genes that are associated with environmental variables and contribute to functional phenotypic response (Figure 2e; Table S8). All inspected genomic regions harboured clusters of tightly linked genes that are associated with response to abiotic stress. For example, a strong overlapping signal associated with plant biomass, spike length, flowering time, as well as isothermality (BIO3) and temperature seasonality (BIO4) was observed for the HB on chromosome 4H (chr4H: 601,511,403–607,147,956). This HB harbours a LEA gene which is associated with desiccation protection right next to the flowering time regulator CONSTANS, and the shoot and floral meristem regulator CLAVATA gene (Clark et al. 1995; Shim et al. 2017; Figure 2e). This region also harbours multiple heat shock chaperones involved in protein homeostasis under thermal stress (Faralli et al. 2015), and transcription factors such as ERF, ARF, and NAC, known to mediate hormonal signalling and transcriptional reprogramming in response to environmental stresses (Bouzroud et al. 2018; Puranik et al. 2012; Xie et al. 2019). Together, these genes form a tightly linked cluster within HBs and presumably contribute to coordinated responses to drought, temperature, and other associated abiotic stressors (Table S8). Another notable example is on chromosome 5H (chr5H:559,584,763–567,399,541), which exhibited significant associations with isothermality (BIO3), temperature seasonality (BIO4), and multiple phenotypic traits such as tiller length, flag length and grain weight. This HB harbours a cluster of functionally relevant genes, including the meristem development regulator CLAVATA, adjacent to BTB/MATH and glutathione transferase genes, conferring tolerance to abiotic stress (Lechner et al. 2011; Nianiou‐Obeidat et al. 2017). Additionally, this genomic region contains several stress‐responsive elements crucial for drought and heat tolerance including the early‐responsive to dehydration stress gene (ERD; Ziaf et al. 2011), and several ethylene associated genes (Xie et al. 2019). Genetic Offset Assessment Enables Prioritizing Areas for Conservation Climate change is expected to pose significant challenges to populations that fail to track the pace of environmental shifts. To characterize scenario‐based projections of population vulnerability for 2061–2080 under mild (SSP1‐2.6) and severe (SSP5‐8.5) socioeconomic pathways, we estimated the genetic offset of wild barley populations relative to their current allelic composition. For each scenario, three different climatic models (MPI‐ESM1‐2‐HR, UKESM1‐0‐LL, and IPSL‐CM6A‐LR) were averaged and used to improve the robustness of the predictions. A spatial model was constructed within a gradient‐forest framework using 14,564 SNPs identified among the top 0.01% signals in the GEA analyses for at least one bioclimatic variable, representing putative adaptive loci. The SNP dataset was used as the response matrix, and the 7 bioclimatic factors and 8 principal coordinates of neighbour matrices (PCNM) as predictors (‘adaptive model’). Based on these models we calculated the cumulative importance which quantifies the expected allelic composition across an environmental gradient. The main bioclimatic factor risking population persistence is annual precipitation (BIO12), with an increased impact on the allelic composition in regions with an average rainfall lower than 250 mm or above 750 mm, which marks the transition to the desert and northern high‐altitude habitats, respectively (Figure 3c). The next factor affecting the projected allelic composition is PCNM4, which corresponds to increased demographic isolation (isolation by distance) between populations. The third factor is isothermality (BIO3), which reflects the day‐to‐night oscillations in temperatures that have a strong effect on plant development (growth degree days); thus, increased diurnal temperature variability is predicted to significantly constrain adaptation to a changing climate (Figure 3a). To validate these results, a neutral model was constructed to reflect the proportion of variation explained merely by demographic processes and population structure. The model was calculated using a set of 100,000 SNPs randomly sampled across the genome (Figure 3b). Overall, the ‘neutral’ model explained less variation (mean R 2 = 0.27) than the ‘adaptive’ model (mean R 2 = 0.69). Predicting adaptive potential to climate change in wild barley populations. (a) Relative contribution of the bioclimatic (BIOs) and spatial (PCNM) factors to the ‘adaptive’ model. (b) Relative contribution of the same factors to the ‘neutral’ model. (c) Cumulative importance of isothermality (BIO3) and annual precipitation (BIO12) in the ‘adaptive’ (red) and ‘neutral’ (blue) gradient‐forest models. (d) Genetic offset analysis under mild scenario (SSP1‐2.6). (e) Genetic offset analysis under severe scenario (SSP5‐8.5). (f) Residuals of the Procrustes rotation for the predicted allelic compositions by ‘neutral’ and ‘adaptive’ models. The colour gradients from blue to red indicate regions under varying levels of selection pressure across wild barley populations. In all three panels (d–f), pie charts illustrate genetic clusters, with the size reflecting Tajima's D score (smaller pie corresponds to lower score), and lines between population pairs represent gene flow estimated via an isolation‐by‐distance approach. To highlight habitats where wild barley populations are under selective pressure, a spatial comparison was conducted using Procrustes superimposition between the allelic composition predictions of adaptive and neutral models. Overall, extensive differences between models were observed for desert populations, indicating a stronger selective pressure that is gradually declining toward the coastal and northern regions (Figure 3f). Next, the genetic offset was calculated as the Euclidean distance between present and projected future adaptive allelic compositions to quantify the expected genomic vulnerability of populations to future climate. Geographically, higher genetic offset is predicted mainly in isolated desert populations under a mild scenario of climate change (SSP1‐2.6). However, under the severe scenario (SSP5‐8.5), risk is predicted to expand dramatically across the Levant, primarily in the desert (south) and along the Jordan Valley (east). The observed differences between scenarios emphasize the impact of policy and management on projected genetic offset. Significantly high genetic offset is also predicted in the southern coastal area and eastern Galilee in the north, where changes in temperature stability and precipitation may be too fast or excessive to enable local adaptation (Figure 3d,e). In these regions, maintenance of ecological corridors, transplanting, and other management practices are specifically urged and expected to have the strongest benefit. In several habitats, mainly along the northern coastal region of the Mediterranean Sea and at high elevations along mountain ridges, a low genetic offset was obtained (0.005–0.02), indicating an improvement of adaptation and suitability. These habitats are expected to obtain enough precipitation to provide refugia and enable the persistence of isolated populations (Figure 3d,e). Interestingly, in some cases considerable risk was also observed in populations with high overall genomic diversity, indicating that genetic variation alone is not enough to predict the adaptive potential (Figure 3d,e). Finally, the results of the genetic offset model and predictions of suitability obtained by the SDM were compared using a Mantel test, indicating a significant correspondence between ecological and evolutionary models (r M = 0.71; p = 0.001). Nevertheless, the genetic offset analysis was able to identify populations at high risk (e.g., southern coast and eastern Galilee) that were neglected by the SDM, recapitulating the importance of integrating genomic data into nature risk assessment models. Discussion Our study illuminates the adaptive potential and response of wild barley populations in the southern Levant to projected climate changes. The new ecological‐genetic sampling design enabled us to disentangle the confounding effects of environmental and geographic distances and highlight candidate genetic factors associated with adaptation. In previous studies of wild barley across the southern Levant and Fertile Crescent, strong population structure has hindered the detection of genomic adaptations along environmental clines (Chang et al. 2022; Hübner et al. 2009; Pankin et al. 2018; Russell et al. 2016). By sampling populations facing similar environmental stress but with distinct genetic backgrounds, we were able to reduce the effect of population structure and identify signals of adaptation at higher resolution (Figures 1d and 2d,e). However, despite the beneficial sampling design, detecting genomic associations with environmental clines remains challenging due to the high false positive rate of GEA analyses. To overcome these limitations, we have used a stringent approach with higher weight to a signal consistency over its intensity (Shaar‐Moshe et al. 2015). Thus, we focused on genomic regions supported by two different analyses with the cost of a presumable increase in the false negative rate. It is therefore likely that other genomic regions associated with environmental changes were missed in this study. It is also important to note that our study focused on wild barley populations in Israel, which represents only a fraction of the species distribution range along the Fertile Crescent. Inclusion of populations from other regions following a similar sampling design will likely highlight additional genomic factors and improve our understanding of adaptation to climate change in wild barley. Despite these limitations, Israel is characterized by exceptional ecological heterogeneity over short geographic distances, thus offering a unique sandbox for studying local adaptation. These settings allow for strong environmental contrasts with minimal confounding from large‐range isolation effects and population structure. In accordance with theoretical predictions (Hill and Robertson 1966; Lenormand and Otto 2000; Yeaman and Whitlock 2011) and recent empirical evidence (Battlay et al. 2023), our study highlights the advantage of clustered adaptive genetic variation in haplotype block (HB) architecture for enhancing orchestrated responses to environmental changes. The identified HBs were supported by the GEA and GWAS analyses and were found to be under selection (Figures 1 and 2). Notably, the modest SNP signal in our GWAS scans (λ GC = 0.83–0.96; Figures S9 and S10) is typical for quantitative traits in structured wild panels. Thus, GWAS serves only as independent evidence at the HB level, not for discovering novel candidate regions. Nevertheless, the enrichment of trait associated SNPs in GEA, combined with colocalized selective sweep signals, confirms these HBs hold phenotypic and adaptive relevance. Importantly, GEA and GWAS reflect statistical associations, not definitive causal adaptation. Similar patterns can stem also from nonadaptive processes like residual population structure, linkage disequilibrium hitchhiking, or shared demographic histories. While our ecological genetic sampling design minimizes geographic confounding in GEA signals (Figures S1 and S16), we cannot exclude these alternative processes for individual loci. Our recombination analysis placed the candidate HBs outside low‐recombination compartments. Moreover, we found no evidence of polymorphic inversions at the six candidate HBs using a local PCA screen, though very rare or near‐fixed variants could in principle escape detection. We therefore conservatively refer to these as candidate adaptive regions. Verifying genuine adaptive signals requires further research, including expression profiling under abiotic stress and functional validation of strong candidates. Interestingly, the HBs exhibit significant pleiotropic effects, wherein discrete genomic clusters coordinate the regulation of multiple phenotypes, most notably the synchronized relationship between plant height and flowering time (Figure S22). This genetic architecture is primarily shaped by bioclimatic variables and specifically mean diurnal range and water availability which act as the selective pressures defining regional adaptation. To survive these stressors, wild barley employs a high‐heritability “escape” strategy, particularly evident in the desert ecotype, which reaches flowering after 73 days to bypass terminal heat and drought stress. As climate change and desertification progress, the resulting reduction in habitat suitability suggests that barley's long‐term persistence will depend on the capacity of these conserved HBs to balance fixed adaptive traits with the physiological plasticity necessary to endure increasingly fluctuating environmental conditions. However, our single site greenhouse experiment over 3 years cannot resolve genotype and environment interactions because it measures traits under one standardized regime. These benign conditions may also mask selective pressures and GWAS signals for desert populations. Fully partitioning genetic and environmental variation requires further experiments across multiple contrasting field sites. Interestingly, genes within HBs were associated with different environmental stressors, where the key factors were response to drought, heat and regulation of flowering time, which are the main factors to adjust to climate changes (Chaudhry and Sidhu 2022). These regions showed a substantial linkage disequilibrium in individuals harbouring the same allele but from different geographic locations. This observation further supports the benefits of the sampling design but raises questions regarding the evolution of similar haplotypes in different regions. Considering the low outcrossing rate in wild barley (Abdel‐Ghani et al. 2004), it is likely that seeds are transported across locations, bringing beneficial non‐recombining HBs from other populations. However, the origin and evolution of these candidate adaptive HBs require further research. The HBs account for high proportions of variance explained (PVE) for critical adaptive traits, including tiller length and flowering time. Selective sweep analysis and reduced diversity provided additional support that these HBs are under natural selection, serving as genomic hubs that harbour tightly linked clusters of functionally diverse genes. It should be noted that filtering variants at a 5% minor allele frequency threshold is necessary with low coverage whole genome data but may truncate the rare alleles required for neutral theory statistics (Hemstrom et al. 2024). This filtering likely shifted Tajima's D upward, deflated π, and reduced RAiSD power to detect sweeps (Alachiotis and Pavlidis 2018). Consequently, we interpret π and Tajima's D as relative comparisons across populations rather than absolute neutral estimates. Similarly, we view the RAiSD sweep signals within the six HBs as conservative, rather than exhaustive. Despite these caveats, the HBs effectively bridge environmental sensing with developmental reprogramming. For instance, the HB on chromosome 4H integrates desiccation protection via Late Embryogenesis Abundant (LEA) genes with the flowering time regulator CONSTANS and the meristem development regulator CLAVATA, alongside thermal‐stress chaperones and transcription factors such as NAC and ERF. Similarly, the HB on chromosome 5H clusters central regulators like CLAVATA with early responsive to dehydration (ERD) elements and glutathione transferase genes to coordinate grain weight and biomass with temperature seasonality. Together, these linked genetic modules potentially enable the plant to orchestrate a complex, multi‐trait adaptive response that balances structural development with abiotic stress tolerance, providing the requisite plasticity to maintain fitness across fluctuating climatic gradients. Climate projections indicated that the sustainability of wild barley populations in the southern Levant is primarily threatened by reductions in precipitation and water availability (Figure 3a), which will drive populations toward creeks along drainage basins where water is more accessible, or higher elevations serving as refugia from unsuitable conditions (Figures 2b and 3d,e). This constrained dispersal pattern may increase population isolation and the risk of maladaptation, thus encouraging speciation and endemism (Harrison and Noss 2017). Previous studies suggest a predominant migration pattern from north to south (Hübner et al. 2012); given the lower adaptive potential to drought and heat stress in northern populations, the risk is expected to escalate with rising temperatures and reduction of rainfall. Nevertheless, our genetic offset analyses carry compounding uncertainties because they rely on environmental‐associated SNPs, gradient forest models, and future climate projections. These results therefore represent relative vulnerability rankings rather than exact forecasts of fitness loss. Because environmental gradients and population structure align in this region, offset signals may partly reflect demographic turnover rather than purely adaptive changes (Láruson et al. 2022; Rellstab et al. 2021). Consequently, conservation efforts should prioritize spatial rankings over absolute offset values, which require independent validation through reciprocal transplants or multiple environment trials. The projected adaptive response in wild barley is largely characterized by a shift to shorter life cycles and earlier flowering times, thus reflecting a strategy of phenological escape that has been observed in other species (Tyrmi et al. 2020). This shift is expected to reverberate throughout the ecosystem, affecting grazing fauna, grain‐feeding insects and mammals, and their predators. Therefore, future strategic planning of land use and infrastructure development in the Levant must account for the anticipated distribution shifts of wild barley and other herbaceous species to mitigate the risks of maladaptation and potential local extinctions. Author Contributions S.H. conceived the study, D.S. and G.A. established the collection, T.M., N.A. and D.F. conducted the sequencing, variant calling, and phenotyping, E.P., M.B.K. and A.B.‐M. performed the spatial analyses, E.P. and S.H. performed the genomic analyses, E.P., A.B.‐M. and S.H. wrote the manuscript with input from all co‐authors. All authors read and approved the manuscript. Funding This work was supported by the Israel Ministry of Science and Technology grant 4802, United States‐Israel Binational Agricultural Research and Development Fund (BARD) grant IS‐5668‐24, and the Russell Berrie Foundation (SH). Ethics Statement No ethical considerations are relevant to this submission. All plants were sampled under the permission restrictions of the Israeli Natural Parks Authority in Israel. Conflicts of Interest The authors declare no conflicts of interest. Supporting information Figure S1: Ecological‐genetic sampling design. The studied area was divided into six polygons in accordance with genetic assignment and ecotype boundaries. Within each polygon, five environmental categories were defined: precipitation stability, soil, annual precipitation, highest temperatures, and lowest temperatures (left to right). Sampling sites representing each category in each polygon were selected for sampling. Figure S2: Distributions of single nucleotide polymorphic sites statistics and thresholds (blue vertical dashed line) for hard filtering. Statistics were obtained from the gatk4 haplotype caller pipeline and included read depth (DP), mapping quality (MQ), variant quality (QUAL), mapping quality rank sum test (MQRankSum), strand odds ratio (SOR), Fisher strand bias test (FS), and quality by depth (QD) for each SNP in the barley collection. Figure S3: Images from the net‐house experiments conducted over 3 years. Figure S4: Population structure and assignment of individuals to cluster for K = 3 (a), K = 4 (b), and K = 5 (c). (d) Scree plot showing the proportion of genomic variance explained by principal components from the population‐structure PCA. The inflection point at PC3 was used to select principal components as covariates in EMMAX and K = 3 for the sNMF ancestry model. (e) Cross‐entropy criterion from sNMF for selecting the number of latent factors in LFMM2. The minimum cross‐entropy supports K = 15, which was used to model neutral population structure in the LFMM2 genotype‐environment association analysis. Figure S5: SNP filtering and analysis pipeline flowchart. A brief overview of variant filtering steps and the derivation of SNP datasets used in each downstream analysis. Figure S6: Stacked Manhattan plots of genotype‐environment association analysis using LFMM2 (top) and EMMAX (bottom) algorithms. Plots are organized by environmental variables: (a) BIO1, (b) BIO2, (c) BIO3, (d) BIO4, (e) BIO7, (f) BIO12 and (g) BIO15. Red dots indicate SNPs that show significant associations, surpassing the Bonferroni‐corrected p‐value threshold. Figure S7: Quantile‐quantile (QQ) plots for GEA scans (EMMAX and LFMM2). Expected vs. observed –log10(p) values for all SNPs tested across 7 BioClim variables (a–g). The dashed line is the null diagonal; the grey ribbon is the 95% pointwise confidence envelope. The genomic inflation factor (λ GC) is shown for EMMAX (LFMM2 λ GC = 1 by construction). Dark grey dashed line marks the Bonferroni threshold; N_sig gives the count of significant SNPs (EMMAX/LFMM2). Coloured dashed lines mark the top 0.01% SNPs per method; filled shapes correspond to top SNPs in each method (triangles = EMMAX, squares = LFMM2). Figure S8: Manhattan plot of redundancy analysis (RDA) for selected seven bioclimatic factors across the seven barley chromosomes. Significantly associated SNPs are highlighted in red, with the colour gradient below indicating SNP counts across the chromosomes. Figure S9: Manhattan plots of genome‐wide association study (GWAS) for adaptive phenotypic traits, highlighting top signal SNPs that coincide with adaptive haplotype blocks (HBs) in green: (a) Flag length, (b) Flowering time, (c) Grain weight, (d) Number of tillers, (e) Number of grains, (f) Plant biomass, (g) Spike length, (h) Tiller length. Figure S10: Quantile‐quantile (QQ) plots for GWAS scans (EMMAX). Expected vs. Observed –log10(p) values for all SNPs tested across 8 traits (a–h). The dashed line is the null diagonal; the grey ribbon is the 95% pointwise confidence envelope. The genomic inflation factor (λGC) and the number of significant SNPs are indicated in each plot. Filled triangles mark SNPs passing the threshold –log10(p) > 4. Figure S11: LD heatmaps for each of the HBs identified by GEA analyses: (a) chr1H:421,725,937–424,368,069, (b) chr2H:633,919,124–635,760,614, (c) chr4H:601,511,403–607,147,956, (d) chr4H:619,797,481–620,925,486, (e) chr5H:466,549,570–468,907,420, (f) chr5H:559,584,763–567,399,541. Figure S12: LD heatmaps for each genomic region corresponding to the putatively adaptive haplotype blocks (HBs) in samples not carrying the respective haplotype: (a) chr1H:421,725,937–424,368,069, (b) chr2H:633,919,124–635,760,614, (c) chr4H:601,511,403–607,147,956, (d) chr4H:619,797,481–620,925,486, (e) chr5H:466,549,570–468,907,420, (f) chr5H:559,584,763–567,399,541. Figure S13: Projection of haplotype blocks location on the recombination landscape. Recombination rate (cM/Mbp; blue line) along the seven barley chromosomes (1H–7H). Dotted lines and shaded intervals mark the two pericentromeric thresholds: a strict threshold of T 1 = 0.1 cM/Mbp (blue line, dark‐grey shading) and a permissive threshold of T 2 = 0.5 cM/Mbp (red line, light‐grey shading). Red vertical rectangles denote the six candidate haplotype blocks (HBs). All six HBs lie outside the strict pericentromeric compartment; under the permissive threshold, two HBs (1H, 5H) intersect the extended low‐recombination zone. Figure S14: Lostruct MDS scan along barley chromosomes hosting candidate adaptive haplotype blocks. MDS1 (upper panel) and MDS2 (lower panel) values for 1000‐SNP non‐overlapping windows are plotted against genomic position (Mb) for each of the four chromosomes hosting the six candidate HBs on chromosomes chr1H, chr2H, chr4H and chr5H. Grey points indicate non‐outlier windows, and red points indicate outlier windows with |MDS| > 4× MAD from the per‐chromosome median (dashed lines). Blue vertical bands mark the genomic intervals of the candidate HBs. No outlier windows fall within any candidate HB on either MDS axis, providing no evidence for polymorphic inversion signatures in these regions. Figure S15: Signals of selective sweeps identified within adaptive haplotype blocks (HBs). Each block is shown along with its surrounding genomic region (+/− 3 Mb). The block boundaries are indicated with red dashed lines. Figure S16: The distributions of permutation in the partial Mantel test. The test was conducted for the Barley1K collection (a; Hübner et al. 2009) and the new barley collection (b). The x‐axis represents the correlation coefficient values, while the y‐axis indicates the frequency of permuted correlations. The observed correlation coefficient is marked by a red vertical line in each distribution, showing its position relative to the null distribution generated from permutations. Figure S17: A heatmap of correlation between climatic variables. In the heatmap, the correlation coefficients between all bioclimatic variables included in the analysis are displayed, except BIO14 (Precipitation of Driest Month), which was excluded due to its lack of variance within the study region. Figure S18: Environmental PCA biplot with genetic ancestry overlaid. Ordination of 30 sampling sites in the space of two principal components derived from seven bioclimatic variables (BIO1, BIO2, BIO3, BIO4, BIO7, BIO12, BIO15; WorldClim v2.1). PC1 (47.8%) and PC2 (32.8%) together explain 80.6% of bioclimatic variance among sites. Pie‐slice proportions reflect per‐site mean sNMF (K = 3) ancestry coefficients. Arrows indicate the direction and relative strength of each bioclimatic variable's contribution to the ordination axes (loadings scaled × 3.2 for visibility). Figure S19: Estimating the contribution of climate variables to genomic variance. For each of seven selected bioclimatic variables, the outer bar (gold, full width) shows percent of variance explained (P x) from an unconditional climate RDA (R 2 adj = 18.8%), and the inner bar (blue) shows P x after partitioning out spatial structure captured by forward‐selected PCNM eigenvectors (partial RDA, R 2 adj = 9.5%). The exposed tail of each outer bar represents the component of a variable's climate signal that is spatially confounded. Variables are ordered by their partial RDA P x. Percentages labelled on the inner bars. Note that the two P x metrics are computed relative to different denominators (total vs. residual‐after‐conditioning inertia) and are not additive. Figure S20: Geographic representation of the species distribution models. From left to right maps display distribution model under current conditions, future mild (year 2070, RCP4.5), and future severe (year 2070, RCP8.5). Pie charts mark the sampling locations, and their size represents the Tajima's D score. The colours in the pie charts correspond to the proportion of assignment to genetic cluster (K = 3) based on the sNMF algorithm. The lines between populations indicate gene flow estimated using the isolation‐by‐distance metrics. Maps are coloured based on the suitability index calculated for each population and extrapolated across the studied area. Figure S21: Boxplot comparing isolation by distance in different regions. Comparison of isolation measured as correlation coefficients between geographical and genetic distances between desert and Mediterranean populations. A higher correlation coefficient corresponds to higher isolation between samples. In the graph, dots represent pairwise comparisons between populations, boxes correspond to 1st–3rd quartiles and the horizontal line depicts the median. Figure S22: A heatmap displaying the correlations between all phenotypic trait measurements. Table S1: Bioclimatic factors and geographic coordinates for the 30 sampling sites. Table S2: Sequencing and alignment statistics for whole‐genome sequencing data of Hordeum spontaneum dataset. Summary of sequencing and mapping metrics for Hordeum spontaneum whole‐genome sequencing (WGS), including read counts, alignment statistics, coverage, and mean sequencing depth for each sample. Table S3: Phenotypic traits measurements along three seasons. Table S4: Summary table for identified haplotype blocks. For each block are indicated the total number of SNPs (nSNPs), the number of SNPs associated with climate factors, specified representative SNPs which were used for defining adaptive haplotype blocks. Table S5: Overlap between GWAS signals of adaptive phenotypic traits and putatively adaptive haplotype blocks. Table S6: Summary of major steps, software, and key parameters used in the GATK4 pipeline for variant discovery and filtering Hordeum spontaneum dataset. This table includes read quality control, trimming, alignment, duplicate marking, variant calling, hard filtering thresholds, and computational setup for processing all samples. Table S7: Summary table of measured phenotypic traits measured across three seasons (2019–2021). The table lists the minimum and maximum scores, mean, median, standard deviation, and heritability of each trait. Table S8: Genes of putatively adaptive haplotype blocks. Data Availability Statement Raw sequence data are deposited in the European Nucleotide Archive under project PRJEB79623. All analysis code and scripts used in this study are available at https://github.com/hubner‐lab/HordeumSpontaneumEnvGrad. References Associated Data Supplementary Materials Figure S1: Ecological‐genetic sampling design. The studied area was divided into six polygons in accordance with genetic assignment and ecotype boundaries. Within each polygon, five environmental categories were defined: precipitation stability, soil, annual precipitation, highest temperatures, and lowest temperatures (left to right). Sampling sites representing each category in each polygon were selected for sampling. Figure S2: Distributions of single nucleotide polymorphic sites statistics and thresholds (blue vertical dashed line) for hard filtering. Statistics were obtained from the gatk4 haplotype caller pipeline and included read depth (DP), mapping quality (MQ), variant quality (QUAL), mapping quality rank sum test (MQRankSum), strand odds ratio (SOR), Fisher strand bias test (FS), and quality by depth (QD) for each SNP in the barley collection. Figure S3: Images from the net‐house experiments conducted over 3 years. Figure S4: Population structure and assignment of individuals to cluster for K = 3 (a), K = 4 (b), and K = 5 (c). (d) Scree plot showing the proportion of genomic variance explained by principal components from the population‐structure PCA. The inflection point at PC3 was used to select principal components as covariates in EMMAX and K = 3 for the sNMF ancestry model. (e) Cross‐entropy criterion from sNMF for selecting the number of latent factors in LFMM2. The minimum cross‐entropy supports K = 15, which was used to model neutral population structure in the LFMM2 genotype‐environment association analysis. Figure S5: SNP filtering and analysis pipeline flowchart. A brief overview of variant filtering steps and the derivation of SNP datasets used in each downstream analysis. Figure S6: Stacked Manhattan plots of genotype‐environment association analysis using LFMM2 (top) and EMMAX (bottom) algorithms. Plots are organized by environmental variables: (a) BIO1, (b) BIO2, (c) BIO3, (d) BIO4, (e) BIO7, (f) BIO12 and (g) BIO15. Red dots indicate SNPs that show significant associations, surpassing the Bonferroni‐corrected p‐value threshold. Figure S7: Quantile‐quantile (QQ) plots for GEA scans (EMMAX and LFMM2). Expected vs. observed –log10(p) values for all SNPs tested across 7 BioClim variables (a–g). The dashed line is the null diagonal; the grey ribbon is the 95% pointwise confidence envelope. The genomic inflation factor (λ GC) is shown for EMMAX (LFMM2 λ GC = 1 by construction). Dark grey dashed line marks the Bonferroni threshold; N_sig gives the count of significant SNPs (EMMAX/LFMM2). Coloured dashed lines mark the top 0.01% SNPs per method; filled shapes correspond to top SNPs in each method (triangles = EMMAX, squares = LFMM2). Figure S8: Manhattan plot of redundancy analysis (RDA) for selected seven bioclimatic factors across the seven barley chromosomes. Significantly associated SNPs are highlighted in red, with the colour gradient below indicating SNP counts across the chromosomes. Figure S9: Manhattan plots of genome‐wide association study (GWAS) for adaptive phenotypic traits, highlighting top signal SNPs that coincide with adaptive haplotype blocks (HBs) in green: (a) Flag length, (b) Flowering time, (c) Grain weight, (d) Number of tillers, (e) Number of grains, (f) Plant biomass, (g) Spike length, (h) Tiller length. Figure S10: Quantile‐quantile (QQ) plots for GWAS scans (EMMAX). Expected vs. Observed –log10(p) values for all SNPs tested across 8 traits (a–h). The dashed line is the null diagonal; the grey ribbon is the 95% pointwise confidence envelope. The genomic inflation factor (λGC) and the number of significant SNPs are indicated in each plot. Filled triangles mark SNPs passing the threshold –log10(p) > 4. Figure S11: LD heatmaps for each of the HBs identified by GEA analyses: (a) chr1H:421,725,937–424,368,069, (b) chr2H:633,919,124–635,760,614, (c) chr4H:601,511,403–607,147,956, (d) chr4H:619,797,481–620,925,486, (e) chr5H:466,549,570–468,907,420, (f) chr5H:559,584,763–567,399,541. Figure S12: LD heatmaps for each genomic region corresponding to the putatively adaptive haplotype blocks (HBs) in samples not carrying the respective haplotype: (a) chr1H:421,725,937–424,368,069, (b) chr2H:633,919,124–635,760,614, (c) chr4H:601,511,403–607,147,956, (d) chr4H:619,797,481–620,925,486, (e) chr5H:466,549,570–468,907,420, (f) chr5H:559,584,763–567,399,541. Figure S13: Projection of haplotype blocks location on the recombination landscape. Recombination rate (cM/Mbp; blue line) along the seven barley chromosomes (1H–7H). Dotted lines and shaded intervals mark the two pericentromeric thresholds: a strict threshold of T 1 = 0.1 cM/Mbp (blue line, dark‐grey shading) and a permissive threshold of T 2 = 0.5 cM/Mbp (red line, light‐grey shading). Red vertical rectangles denote the six candidate haplotype blocks (HBs). All six HBs lie outside the strict pericentromeric compartment; under the permissive threshold, two HBs (1H, 5H) intersect the extended low‐recombination zone. Figure S14: Lostruct MDS scan along barley chromosomes hosting candidate adaptive haplotype blocks. MDS1 (upper panel) and MDS2 (lower panel) values for 1000‐SNP non‐overlapping windows are plotted against genomic position (Mb) for each of the four chromosomes hosting the six candidate HBs on chromosomes chr1H, chr2H, chr4H and chr5H. Grey points indicate non‐outlier windows, and red points indicate outlier windows with |MDS| > 4× MAD from the per‐chromosome median (dashed lines). Blue vertical bands mark the genomic intervals of the candidate HBs. No outlier windows fall within any candidate HB on either MDS axis, providing no evidence for polymorphic inversion signatures in these regions. Figure S15: Signals of selective sweeps identified within adaptive haplotype blocks (HBs). Each block is shown along with its surrounding genomic region (+/− 3 Mb). The block boundaries are indicated with red dashed lines. Figure S16: The distributions of permutation in the partial Mantel test. The test was conducted for the Barley1K collection (a; Hübner et al. 2009) and the new barley collection (b). The x‐axis represents the correlation coefficient values, while the y‐axis indicates the frequency of permuted correlations. The observed correlation coefficient is marked by a red vertical line in each distribution, showing its position relative to the null distribution generated from permutations. Figure S17: A heatmap of correlation between climatic variables. In the heatmap, the correlation coefficients between all bioclimatic variables included in the analysis are displayed, except BIO14 (Precipitation of Driest Month), which was excluded due to its lack of variance within the study region. Figure S18: Environmental PCA biplot with genetic ancestry overlaid. Ordination of 30 sampling sites in the space of two principal components derived from seven bioclimatic variables (BIO1, BIO2, BIO3, BIO4, BIO7, BIO12, BIO15; WorldClim v2.1). PC1 (47.8%) and PC2 (32.8%) together explain 80.6% of bioclimatic variance among sites. Pie‐slice proportions reflect per‐site mean sNMF (K = 3) ancestry coefficients. Arrows indicate the direction and relative strength of each bioclimatic variable's contribution to the ordination axes (loadings scaled × 3.2 for visibility). Figure S19: Estimating the contribution of climate variables to genomic variance. For each of seven selected bioclimatic variables, the outer bar (gold, full width) shows percent of variance explained (P x) from an unconditional climate RDA (R 2 adj = 18.8%), and the inner bar (blue) shows P x after partitioning out spatial structure captured by forward‐selected PCNM eigenvectors (partial RDA, R 2 adj = 9.5%). The exposed tail of each outer bar represents the component of a variable's climate signal that is spatially confounded. Variables are ordered by their partial RDA P x. Percentages labelled on the inner bars. Note that the two P x metrics are computed relative to different denominators (total vs. residual‐after‐conditioning inertia) and are not additive. Figure S20: Geographic representation of the species distribution models. From left to right maps display distribution model under current conditions, future mild (year 2070, RCP4.5), and future severe (year 2070, RCP8.5). Pie charts mark the sampling locations, and their size represents the Tajima's D score. The colours in the pie charts correspond to the proportion of assignment to genetic cluster (K = 3) based on the sNMF algorithm. The lines between populations indicate gene flow estimated using the isolation‐by‐distance metrics. Maps are coloured based on the suitability index calculated for each population and extrapolated across the studied area. Figure S21: Boxplot comparing isolation by distance in different regions. Comparison of isolation measured as correlation coefficients between geographical and genetic distances between desert and Mediterranean populations. A higher correlation coefficient corresponds to higher isolation between samples. In the graph, dots represent pairwise comparisons between populations, boxes correspond to 1st–3rd quartiles and the horizontal line depicts the median. Figure S22: A heatmap displaying the correlations between all phenotypic trait measurements. Table S1: Bioclimatic factors and geographic coordinates for the 30 sampling sites. Table S2: Sequencing and alignment statistics for whole‐genome sequencing data of Hordeum spontaneum dataset. Summary of sequencing and mapping metrics for Hordeum spontaneum whole‐genome sequencing (WGS), including read counts, alignment statistics, coverage, and mean sequencing depth for each sample. Table S3: Phenotypic traits measurements along three seasons. Table S4: Summary table for identified haplotype blocks. For each block are indicated the total number of SNPs (nSNPs), the number of SNPs associated with climate factors, specified representative SNPs which were used for defining adaptive haplotype blocks. Table S5: Overlap between GWAS signals of adaptive phenotypic traits and putatively adaptive haplotype blocks. Table S6: Summary of major steps, software, and key parameters used in the GATK4 pipeline for variant discovery and filtering Hordeum spontaneum dataset. This table includes read quality control, trimming, alignment, duplicate marking, variant calling, hard filtering thresholds, and computational setup for processing all samples. Table S7: Summary table of measured phenotypic traits measured across three seasons (2019–2021). The table lists the minimum and maximum scores, mean, median, standard deviation, and heritability of each trait. Table S8: Genes of putatively adaptive haplotype blocks. Data Availability Statement Raw sequence data are deposited in the European Nucleotide Archive under project PRJEB79623. All analysis code and scripts used in this study are available at https://github.com/hubner‐lab/HordeumSpontaneumEnvGrad.