Insight Into the Genetic Architecture of Local (Mal)Adaptation in Wild and Feral Brassica rapa Populations

ABSTRACT

As climate change drives rapid environmental shifts, crops face significant challenges due to their limited adaptive genetic diversity, threatening food security. Crop wild relatives and potentially feral populations can serve as valuable reservoirs of adaptive diversity for breeding. Exploring and conserving this diversity is therefore essential, as it may encompass a wide range of adaptations. In this study, we investigate the genetic basis of local climate adaptation and the risk of maladaptation under future conditions in spontaneous populations—naturally occurring in the wild—of Brassica rapa (field mustard). A total of 62 populations were sampled, revealing strong differentiation between feral (France, Slovenia) and wild (Algeria, Sicily) populations. A genome–environment association analysis identified genomic regions associated with climatic variables, with broader patterns suggesting a polygenic architecture. Feral populations broadened the climatic range explored, but their reduced genetic diversity limited our ability to capture the full spectrum of adaptive variation in wild populations, indicating that the genetic architecture of climate adaptation is only partly shared between the two groups. We also used genomic offset (GO) analyses to assess mismatches between current and predicted‐optimal genetic composition under 2050 climate projections. A common garden experiment indicated that GO captures some component of fitness variation, with feral populations providing additional predictive information beyond the environmental space occupied by wild populations. These findings highlight the complex genetic architecture of climate adaptation in B. rapa , as well as the potential and limitations of genomic tools for guiding conservation and crop improvement under climate change.

1. Introduction

Climate change is driving rapid environmental shifts, posing a major threat to many species, which must either adapt or migrate to persist (Intergovernmental Panel on Climate Change (IPCC) ²⁰²³). Among the numerous challenges associated with climate change, shifts in climatic conditions alone, such as increasing aridity, the intensification of extreme weather events, and greater climatic variability, can exert substantial selective pressures on plant populations (Hoffmann and Sgrò 2011; Bellard et al. ²⁰¹²). Given that adaptation through de novo mutations is unlikely to occur rapidly enough in most species or populations, the main mechanism for climate change adaptation relies on standing genetic variation (Barrett and Schluter 2008; Matuszewski et al. ²⁰¹⁵). This presents a critical challenge for agricultural crops, which often have low genetic variability to respond to changing environment due to the bottleneck of domestication followed by strong artificial selection (Olsen and Gross 2008; Flint‐Garcia et al. ²⁰²³; Gage et al. ²⁰¹⁷; Vaughan et al. ²⁰⁰⁷), which could limit the genetic diversity available to respond to changing environments, and thereby their adaptive potential. As this could threaten long‐term food security (Maxted and Kell 2009; Food and Agriculture Organization of the United Nations (FAO) ²⁰¹⁰), Crop Wild Relatives (CWRs) emerge as valuable reservoirs of adaptive genetic diversity (Maxted et al. 2006; Hajjar and Hodgkin ²⁰⁰⁷; Dempewolf et al. ²⁰¹⁷). As close relatives of domesticated crops, often belonging to the same species, these wild populations represent a crucial genetic resource to improve crop resilience in the face of environmental change (Maxted et al. 2010).

A useful concept for understanding the value of adaptive genetic diversity in CWRs is local adaptation. If populations have been present and stable in a given environment for a sufficiently long period of time, they are expected to have evolved genetic adaptations to local conditions that maximize their fitness in their specific environment (Kawecki and Ebert 2004). Consequently, conserving CWRs is not only essential for preserving the species themselves, but also for maintaining a critical reservoir of genetic diversity that could be instrumental for crop improvement. Ensuring that these wild relatives remain widespread and abundant in situ is particularly important, as they preserve a broad spectrum of adaptation to diverse environmental conditions, which can serve as genetic resources for future breeding efforts. However, many CWRs have restricted and often shrinking distributions, limiting their potential to contribute to crop adaptation under changing climatic conditions (Jarvis et al. 2008; Vincent et al. ²⁰¹³).

Feral populations, descendants of cultivated plants that have escaped cultivation, can be relatively common, often establishing themselves in environments where wild relatives are absent (Gressel 2005). When these feral populations persist over extended periods, they may adapt locally to their new environments (Bagavathiannan and Van Acker ²⁰⁰⁸; Pascher et al. ²⁰¹⁰; Qiu et al. ²⁰¹⁷; Mittell et al. ²⁰²⁰). In particular, as they have adapted to constrained environments outside the field, feral populations could provide a source of adaptive genetic diversity to strengthen the resilience of related domestic species. However, they typically exhibit low genetic variation, likely due to the historical bottlenecks associated with domestication, as well as additional bottlenecks related to their escape and establishment in the wild (Hernández et al. 2022). Whether feral populations can serve as valuable genetic resources remains an open question (Mabry et al. 2023). It likely depends on several factors: the demographic history of these populations, the species considered, the extent to which their adaptive genetic diversity has been depleted relative to wild populations, and whether they had time to readapt to their new environments. Nevertheless, given their closer genetic relationship to cultivated varieties, feral populations may act as an intermediary, facilitating gene flow between wild and domesticated populations. Therefore, it is crucial to assess and conserve both CWRs and feral populations by documenting their distribution, environmental range, and their adaptive genetic diversity to better understand their potential as resources for future crop adaptation, as well as the threats they face (Mabry et al. 2023).

In this study, we examine genetic local adaptation in the crop wild relative Brassica rapa , with populations collected in natural habitats, including wild forms and feral populations. This species is important for food production and is also one of the parental species of the allotetraploid Brassica napus napus (rapeseed), which is one of the most widely cultivated plants in the world (Raboanatahiry et al. 2021). B. rapa (Brassicaceae, 2n = 20) is an annual, primarily self‐incompatible species native to the Mediterranean basin (Song et al. 1988). It is thought to have been domesticated in Europe and Central Asia (Qi et al. 2017; Saban et al. ²⁰²³) and is now cultivated worldwide, comprising varieties with highly diverse morphotypes (Cai et al. 2021), used for their roots (turnips), leaves (Chinese cabbage), or seeds (yellow sarson). This morphological diversity, along with a well‐annotated genome (Wang et al. 2011), makes this species an excellent model to study local adaptation. Although the species is cultivated worldwide, the distribution of its wild populations appears to be relatively limited (McAlvay et al. ²⁰²¹). This study focuses on 45 wild and 17 feral populations, which have been first described and characterized in Gay et al. (2026).

The objectives of this study are (i) to investigate whether adaptive genetic variation arises in response to climatic variation in wild or feral populations, (ii) to evaluate whether the inclusion of feral populations influences our results, and to determine their relevance as genetic resources, (iii) to assess the risk of maladaptation under future climate change using a genomic offset approach. While we expect to find signals of local adaptation in wild populations, this may not be the case for feral populations, since agricultural practices may have alleviated the natural environment's selective pressures and they have potentially escaped cultivation only recently.

We rely on whole genome sequencing data of pooled samples representative of 62 wild and feral populations. We considered these populations spontaneous, as they occur naturally outside of cultivated fields, without human management and undergo natural selection. We first apply a genome‐environment association (GEA) approach to identify genetic variants whose allele frequencies covary with environmental variables, while accounting for population structure in order to minimize false positives due to shared ancestry. GEA has proven to be a robust method to detect loci under selection (Forester et al. 2018; De Mita et al. ²⁰¹³) and has been widely applied in studies of local adaptation. Previous work in Brassica rapa focused on germination (Tiret et al. 2025), flowering time (Jia et al. 2020; Shea et al. ²⁰¹⁸) or drought response (Franks et al. 2007) suggests that these traits could have been shaped by populations' local climate. Yet, this study is, to our knowledge, the first application of GEA in this species. Then, we assess whether the wild and feral populations studied show some risk of maladaptation in the future, or in other words, we address the following question: “do these populations show patterns of adaptation that might increase their vulnerability in the case of climate change?” To achieve this, we use the Genomic Offset (GO), which represents a potential “mismatch” between current allele frequencies and the optimal frequencies expected in a new environment, based on the GEA results, and provides a measure of the risk of maladaptation of those populations in the face of climate change. (Fitzpatrick and Keller 2015; Bay et al. ²⁰¹⁸; Rhoné et al. ²⁰²⁰; Jia et al. ²⁰²⁰; Gugger et al. ²⁰¹⁸).

2. Materials and Methods

2.1. Sampling and Sequencing

Sixty two spontaneous populations of Brassica rapa , that is, populations naturally occurring in the wild, outside from cultivation, were collected in Algeria ($n=28$ populations), Italy ($n=17$), France ($n=16$) and Slovenia ($n=1$) (Figure 1a). Previous analyses of population diversity and structure, including B. rapa landraces from the same countries, showed that spontaneous populations collected in France and Slovenia ($n=17$) are ferals, that is, populations living in the wild but descending from domesticated lines (Gay et al. 2026). Sampling protocols and species‐identification controls (flow cytometry, chromosome counting, chloroplast sequencing) are detailed in Falentin et al. (2024). Except for two populations (Slovenia and French Pyrenees), most populations were large (hundreds to thousands of individuals), in habitats ranging from semi‐natural environments to field margins and perennial plantations; collection details are in Appendix S1.

FIGURE 1.

Populations sampling and genetic structure. Colours indicate the country of origin, Algeria (n = 28, orange squares, wild), France (n = 16, blue triangles, feral), Italy (n = 17, bluish green circles, wild) and Slovenia (n = 1, reddish purple diamond, ferral). (a) Sampling locations for the 62 spontaneous populations of Brassica rapa , (b) Plot representing the first two components of the “Random Allele” PCA carried on the spontaneous set (n = 62). PC1 separate feral to wild population, while PC2 distinguish wild Algerian and Italian populations.

Thirty seeds per population, representing 15–30 mother plants collected in situ, were germinated and cultivated in a greenhouse to collect an equivalent quantity of leaf tissue that was pooled per population for DNA extraction. Then, 10 out of these 30 plants per population were transferred to the field and grown under pollen‐proof cages (two cages with five plants each, Tiret et al. ²⁰²⁵) at a single location (Le Rheu, France, 48°06′N 1°47′O) to multiply the seeds and avoid maternal effects for the common garden experiment (see section “Predicting the loss of fitness in a common garden with the GO”).

2.2. Genomic Data Processing

DNA libraries were constructed following Illumina protocols and sequenced using the Illumina HiSeq technology to achieve an average depth of 50×. The genomic data were processed following a standardized pipeline to ensure high‐quality SNP detection. First, raw sequencing reads were cleaned and trimmed to remove adapter sequences and low‐quality bases using fastp (v0.20.1; Chen et al. ²⁰¹⁸). Cleaned reads were then aligned to the B. rapa reference genome ( B. rapa C1.3 var. rapifera, Breton landrace; Maillet et al. (2025)) with bwa mem (v0.7.17; Li and Durbin ²⁰⁰⁹) run with default options. Read alignments with a mapping quality < 20 or duplicates were further removed using the view (option ‐q 20) and markdup programs from SAMtools (v1.14; Danecek et al. ²⁰²¹). Freebayes (v1.3.1; Garrison and Marth ²⁰¹²) was then used for variant calling with the options ‐K (aka pooled‐continuous), ‐C 1 and ‐F 0.01 as recommended when analysing PoolSeq data. In addition, we used the options ‐m 30 and ‐q 20 to set the minimum mapping quality for an alignment to be used to 30 and the minimum base quality for a base to be considered to 20. Finally, clumping was disabled (option ‐E −1) to have as many simple variants as possible and we retained only bi‐allelic SNPs that had at least five reads for the minor allele over all the samples (−G 5). To improve computational efficiency, variant calling was performed in parallel on a computer grid for 7550 non‐overlapping 50 kb windows covering the whole genome assembly. The resulting vcf files were concatenated and reordered using the concat and view programs of bcftools (v1.20; Danecek et al. ²⁰²¹), filtering low quality variant with option ‐i QUAL $>$ 20. Finally, the vcf was parsed and further filtered using the vcf2pooldata function of the R package poolfstat (v3.0; Gautier et al. ²⁰²²), discarding SNPs < 5 bp away from indels, with an overall minimum allele frequency (MAF) < 0.01, and with extreme overall coverage values (outside the 1st and 99th percentiles). The final dataset consisted of 14,457,097 SNPs, spanning the 10 chromosomes of B. rapa genome assembly and 670 unanchored scaffolds.

2.3. Population Structure and Genetic Diversity

Genetic structure was assessed with a “random allele” PCA (randomallele.pca, poolfstat), and quantified using hierarchical $F_{\mathrm{ST}}$, grouping populations by wild or feral status (computeFST, poolfstat (see details in Gautier et al. ²⁰²⁴)), decomposing overall differentiation ($F_{\mathrm{ST}}$) into within‐group ($F_{\mathrm{SG}}$) and between‐group ($F_{\mathrm{GT}}$) components.

For each population, nucleotide diversity ($\pi$) was estimated (npstat v.1; Ferretti et al. ²⁰¹³) in 10,000‐bp sliding windows (windows $<9,000$ bp discarded), averaged across retained windows for feral and wild groups, and compared between groups with a Wilcoxon test.

2.4. Genome‐Environment Association (GEA) Analysis

2.4.1. Climate Data

Daily climate data (temperature, soil moisture) were extracted from the ERA5‐land reanalysis dataset (Hersbach et al. 2020) (Copernicus Climate Change Service) at 1 km² resolution using KrigR (Kusch and Davy 2022). Unlike interpolation‐based datasets such as WorldClim and CHELSA, ERA5 blends observations with reforecast weather models, giving greater consistency and higher temporal resolution.

We derived 19 bioclimatic variables averaged over 2010–2020, reflecting recent conditions prior to seed collection—a window balancing a stable climatic baseline against temporal proximity to sampling (variable list in Table S1). All 19 variables were included to maximize climatic information without imposing prior assumptions about which factors matter most.

We paid particular attention to several variables of likely biological relevance:

• –BIO08 (Mean Temperature of Wettest Quarter) and BIO16 (Soil moisture of Wettest Quarter): seeds need relatively high temperature and humidity to germinate, making these key autumn drivers; temperature also matters for spring development and flowering.
• –BIO06 (Min Temperature of Coldest Month) and BIO11 (Mean Temperature of Coldest Quarter): winter exposure to low temperatures (typically 4°C) is required to induce flowering through vernalization; winters that are too warm may prevent vernalization, while excessive cold may compromise survival.
• –BIO05 (Max Temperature of Warmest Month) and BIO14/BIO18 (Soil moisture of Driest Month/Warmest Quarter): summer heat and drought tolerance are critical for survival, particularly under climate change.

2.4.2. Model

Genome‐environment associations were identified using BayPass (Gautier 2015), using the “standard covariate model” to evaluate linear associations between SNPs and climatic variables. Associations were quantified with Bayes Factors (BF), with $\mathit{BF}\ge 20$ (Jeffreys' rule) indicating strong support; a randomly generated control variable served as a qualitative check for false associations. Adaptation to a multi‐factor environment is expected to result from polygenic selection involving many small‐effect loci; while such signals are hard to detect, we expected to identify at least loci of strong and moderate effect, and possibly some smaller ones.

To delineate genomic regions enriched with associated SNPs, we applied the local score approach (Fariello et al. 2017) (R compute.local.score function; BayPass manual v3.1 p34, option xi = 1, min.maf = 0.2, pval.local.score.thres = 0.05). This leverages linkage disequilibrium to detect regions with significant enrichment of candidate markers, combining evidence across adjacent SNPs and limiting the impact of multiple testing, since false positives are unlikely to cluster at specific locations.

Candidate‐region function was examined via the reference genome's gene annotation (GFF, B. rapa C1.3 var. rapifera) and Arabidopsis thaliana orthology. To assess whether results reflect true adaptive signals, we tested whether highly significant SNPs ($\mathit{BF}\ge 30$) were disproportionately enriched for missense versus synonymous variants, compared with the equivalent ratio among sites with no association (negative BF). Functional effects were predicted using EnsemblVEP (v113.0; McLaren et al. ²⁰¹⁶) with the reference GFF.

Both wild and feral populations were included to maximize sample size and environmental range, and to assess whether patterns of adaptation or maladaptation emerge in feral populations and how strong genetic structure influences GEA results. Analyses were repeated on three subsets: all populations combined (“spontaneous set”, n = 62), wild populations only (“wild set”, n = 45), and feral populations only (“feral set”, n = 17)—addressing whether feral populations are locally adapted and follow the same patterns as wild ones, whether structure masks adaptive signals, and whether the two groups should be analysed jointly or separately.

To assess detection power for the feral set (only 17 populations), we simulated neutral and adaptive loci (effect sizes 0.05–0.2) using the simulate.baypass function (v3.1; Gautier ²⁰¹⁵), with environmental and omega matrices from the feral and wild sets, then analysed the simulated genotypes with BayPass to estimate true‐ and false‐positive rates.

2.5. Testing Local Adaptation

If the GEA detects genetic variation due to local adaptation, genetic differentiation at candidate loci should correlate strongly with environmental distance. We tested this with an “isolation by environment” test (IBE; Wang and Bradburd ²⁰¹⁴), by analogy with isolation by distance (Rousset 1997): partial Mantel tests assessed the correlation between pairwise $F_{\mathrm{ST}}$ at candidate loci and environmental distance, controlling for geographic distance. Environmental distance was the Euclidean distance along the first climatic PC (PC1, 58% of variation; Figure S1).

$F_{\mathrm{ST}}$ between populations were computed with the compute.pairwiseFST function in the R package poolfstat using all SNPs versus subsets of candidate SNPs revealed as associated with PC1 by the GEA (hereafter called “adaptive sets”). To gain insight into the nature of local adaptation, we constructed multiple adaptive sets based on varying thresholds for the evidence of association, using the Bayes factor (BF). These thresholds were set to $\mathit{BF}=10$, $15$, $20$, $30$ and $40$, representing a gradient from moderate to decisive evidence of association. Precision of partial Mantel R was evaluated via jackknife standard errors (excluding one population at a time), and whether R varied systematically with BF threshold was tested via a population bootstrap: in each of 1000 replicates, populations were resampled with replacement, R re‐estimated at all five thresholds, and a second‐order polynomial ($R\sim \mathit{BF}+B{F}^2$) fitted to generate a bootstrap distribution for the quadratic coefficient.

Using candidate SNPs from the GEA on the full spontaneous set, we repeated partial Mantel tests separately for wild and feral populations to assess how much environmental variance within each group was captured by loci detected across the whole 62‐population set—testing whether feral and wild populations share a similar genetic architecture of local adaptation.

These IBE analyses served three purposes: assessing how well GEA‐detected SNPs predict environmental differences, exploring how genetic structure influences results across population subsets, and evaluating—via different candidate thresholds—whether the underlying genetic architecture is oligogenic or polygenic.

2.6. Genomic Offset Analysis

2.6.1. Future Climate Projections

Future climate projections were obtained from the Copernicus Programme's Climate Change Service (C3S) using 10 CMIP5 models (GFDL‐ESM2M, HadGEM2‐CC, HadGEM2‐ES, bcc‐csm1‐1‐m, ACCESS1‐0, CSIRO‐Mk3.6.0, IPSL‐CM5A‐LR, IPSL‐CM5B‐LR, IPSL‐CM5A‐MR, NorESM1‐M) under the intermediate Representative Concentration Pathway (RCP 4.5) for the 2050 horizon (average for 2040–2060). Bioclimatic variables were extracted once again, this time incorporating both precipitation and soil moisture (See Table S2).

Finally, future climate projections of populations were mapped as supplementary individuals onto the multivariate PCA space built using the 2010–2020 climate data of the populations (Figure S1).

2.6.2. Genomic Offset

Genomic Offset was estimated using the “geometric GO” method (Gain et al. 2023) with the R function compute.genetic.offset available from BayPass software (v3.0) package utilities (Gautier 2015; Camus et al. ²⁰²⁴), which estimates the Euclidean distance between current allele frequencies and their expected values under future environmental conditions: $\mathit{gGO}=1/\mathit{\text{nSNPs}}\times {\left(e-e^{*}\right)}^{\prime }{B}^{\prime }B\left(e-e^{*}\right)$, with $e$ the climate of the local environment, $e^{*}$ the projected environment, $B$ the matrix of regression coefficient. To obtain accurate regression coefficients, we re‐ran the GEA under the covmcmc model (default settings), with all covariables considered jointly.

Because climatic variables are often highly correlated, regressions used principal components (PCs) of climate data rather than the raw variables, avoiding redundancy. We retained five PCs, capturing 98% of climatic variability (see Figure S1), and included a randomly generated fake variable to evaluate the relative importance of each climatic covariable, as in (Gain et al. 2023).

GO was calculated for all populations and mapped across the study area. To account for inter‐model variability, we used the median GO value across the 10 climate models.

2.6.3. Predicting the Loss of Fitness in a Common Garden With the GO

To test whether GO estimates translate into fitness loss, we used phenotypic data from a common garden at Le Rheu (France, 48°06′N 1°47′O), where the 62 populations were sown in mid‐September 2022 in an alpha lattice design (three repetitions per population). At maturity, up to five plants were harvested per plot, pods threshed in bulk, and seed number and thousand‐seed weight (TSW) recorded; average seeds per plant was total seeds divided by plants harvested, and TSW × seed number served as a fitness proxy (reproductive biomass).

Under local adaptation and stabilizing selection, GO is expected to be proportional to the logarithm of fitness (Gain et al. 2023), so we expected a negative correlation between GO and our fitness proxy—while noting that a common garden is not the same as future conditions. To test the added value of genomic information, we compared this GO‐fitness correlation with the correlation between fitness and simple environmental distance to the common garden site, that is, ${\left(e-e^{*}\right)}^{\prime }\left(e-e^{*}\right)$ from the GO formula above, without the regression coefficients.

3. Results

3.1. Population Diversity and Structure

Read counts within populations reveal that, of the 14.9 million SNPs analysed, approximately 9.4 million polymorphic markers are exclusive to wild populations (Algeria and Italy), 4.3 million are shared between wild and feral populations and 1.3 million are exclusively found in feral populations (France and Slovenia). Hence, most of the molecular diversity observed in wild populations consists of private alleles, and although a large proportion of the polymorphism is shared between wild and feral populations, the wild populations still harbour 2.5 times more SNPs than the feral ones. This reduced diversity in markers is further supported by a lower estimated nucleotide diversity ($\pi$) in feral populations (0.004647059) compared to wild ones (0.009155556) (Wilcoxon test p‐value = 0.0000000002629). Feral populations were found to be comparable in terms of molecular diversity to cultivated populations from different countries (Gay et al. 2026).

The first two dimensions of the random allele PCA (Figure 1b) identify three distinct clusters of populations: PC1 separates feral populations (France and Slovenia) from wild populations, while PC2 differentiates Algerian from Italian wild populations. In addition, the variability observed among wild populations is considerably larger than among feral populations. Accordingly, the differentiation between groups explain a large part of the global differentiation. The overall $F_{\mathrm{ST}}$ was 0.328 $\pm 0.000786$, the $F_{\mathrm{SG}}$ (i.e., the overall differentiation within groups of wild and feral populations) was 0.122 $\pm 0.000228$ (more specifically $F_{\mathrm{ST}}$ = 0.116 $\pm 0.000254$ among wild populations and 0.137 $\pm 0.000464$ among feral populations), and finally the $F_{\mathrm{GT}}$ (differentiation between groups of wild and feral populations) was 0.24 $\pm 0.000848$.

3.2. Genetic Architecture of Local Adaptation to Climate

3.2.1. Genome‐Environment Associations

For each set of populations (spontaneous, wild and feral), genome scans were performed to detect associations between SNPs and individual climatic variables, with Manhattan plots representing BFs and local scores along the genome.

The spontaneous populations showed a substantial number of significant associations with the 19 bioclimatic variables, with many SNPs show decisive evidence (nSNPs = 3805, $\mathit{BF}\ge 20$). The application of the local score approach resulted in a final set of 29 highly significant candidate genomic regions present in most variables, with BIO01 and BIO04 having the most regions with four. Figure 2 shows the results for BIO14 (soil moisture of the driest month), which contained some of the strongest candidate regions, including one on chromosome 1. Most detected regions overlapped or lay close to a gene (Table S3), and Arabidopsis thaliana orthologs were retrieved if found (Table S4). The functionally annotated candidates included three SWEET‐family sugar transporters (SWEET6, SWEET7 and SWEET14), a Photosystem I subunit (PSAL), a putative disease‐resistance gene of the NB‐LRR class, and two F‐box proteins, a class broadly involved in growth, development and hormone‐mediated signalling.

FIGURE 2.

Local score values (Lindley process) along the genome for the association between allele frequencies and BIO14 (precipitations of the driest month). The horizontal bars indicate the significance thresholds, which vary slightly between chromosomes due to differences in length and LD (see Fariello et al. (2017) for details).

The wild set (45 wild populations) identified 4491 SNPs that exceed the Bayes Factor (BF) threshold of 20. Using the local score approach, we detected 57 significant candidate regions (Table S5). Among these 57 regions, only 8 regions had been detected previously with the whole set of spontaneous populations. Almost all variables contain regions with the exception of BIO04, with BIO03 and BIO05 containing as many as 11. Most detected genes had no identifiable ortholog in A. thaliana . Among those that did (Table S6), four were also recovered in the spontaneous set (coding for PDE327, PES1, SMC1 and the UPF0725‐family protein). The most notable candidates in terms of function were two heat‐shock proteins (HSP17.4A and HSP17.6C) and a gene involved in chloroplast development (PTAC7).

Figure 3 illustrates the pattern commonly observed in genomic regions detected with the wild set, but not previously detected with the whole spontaneous set: the cline between allele frequencies and environmental variables observed in wild populations is not observed in feral populations, because the reference allele (from the cultivated landrace B. rapa C1.3 var. rapifera reference genome assembly) is fixed. Even though the alternative allele could have been advantageous in feral populations, where annual precipitations are more abundant, it was lost. Most markers detected in the wild set are fixed or near fixed in feral populations, with a median minor allele frequency of 0.003384363 (Figure S2). This could be seen as a sign of maladaptation, caused by domestication and/or feralization bottlenecks. Overall, the analysis of the spontaneous set seems to miss many of the adaptive variants that segregate in wild populations.

FIGURE 3.

Reference allele frequencies for a top candidate SNP identified in the wild dataset (n = 45, Algerian and Italian) but not in the full spontaneous dataset (n = 62), shown across all populations. Each point represents one population, coloured by country of origin: Algeria (n = 28, orange squares), France (n = 16, blue triangles), Italy (n = 17, bluish‐green circles), and Slovenia (n = 1, reddish‐purple diamond). Among the wild populations (Algeria and Italy), the reference allele frequency tends to decrease with annual precipitation (BIO12), suggesting that the alternative allele is favoured in more humid environments. In contrast, the reference allele is fixed in the feral populations (France and Slovenia) despite their humid conditions—a mismatch that may indicate potential maladaptation.

Results for the feral‐only analysis are not shown, as some SNPs and local scores were also associated with the randomly generated fake variable, casting doubt on the results. This does not imply that feral populations are not adapted to climate; rather, the false positives most likely stem from the limited number of populations and the restricted environmental range sampled. Accordingly, the simulated feral set showed a marked drop in true positive rate compared to the simulated wild set (0.2 versus 0.42), despite similar false discovery rates.

In the spontaneous set, highly significant SNPs ($\mathit{BF}\ge 30$) show a moderate signal of enrichment in missense variants compared to synonymous ones, although not significant (fisher test Odd Ratio = 1.33, p‐value = 0.067). In the wild set, highly significant SNPs were significantly enriched for missense variants (Odd Ratio = 3.184197, p‐value = 0.000000000000196).

3.2.2. Patterns of Isolation by Environment as Evidence for Local Adaptation

Using the adaptive SNPs identified in the GEA with the spontaneous set (62 populations), we found a significant pattern of isolation‐by‐environment (IBE), with a strong correlation between environmental distance and genetic differentiation ($F_{\mathrm{ST}}$) (Mantel R = 0.8832, for $\mathit{BF}\ge 20$ as strong support for association) (Figure 4). In contrast, while a correlation was observed at the genome‐wide level (Mantel R = 0.5985), it was only due to the differences between the two groups forming two clusters. Indeed, when decomposed, the Mantel R were low, in wild −0.08 and 0.2137 in feral. Although a correlation is expected given that adaptive SNPs were detected as significantly associated with climatic variables by the GEA model, the observed relationship is strong, which gives additional confidence to the detected variants. This result supports the presence of local adaptation to climate in our populations.

FIGURE 4.

Patterns of Isolation by Environment: For each populations pair, $F_{\mathrm{ST}}/\left(1-F_{\mathrm{ST}}\right)$ is plotted as a function of the euclidean environmental distance on climate PC1. Red dots represent genetic distances computed from the adaptive SNP set $F\ge 20$, while black dots represent genetic distances on the whole genome. Due to genetic structure, two clusters emerged here, the upper group corresponding to distances between wild and feral populations.

Interestingly, the strongest correlations were observed for the adaptive sets composed of SNPs exhibiting $\mathit{BF}\ge 15$ (n = 13,549 SNPs, Partial Mantel R = 0.8824, Table 1), although the different sets are comparable with overlapping standard errors. Still, R appeared to rise then fall as a function of the BF threshold. This pattern was confirmed by the resampling‐based curvature test: the quadratic coefficient was negative in $99.9\%$ of bootstrap replicates for the spontaneous set—indicating a concave shape ‐, with an optimum adaptive sets estimate at a treshold of $21\mathit{BF}\pm 4$ suggesting that climatic adaptation involves a large number of loci. This is notable given that we identified several loci with strong effects using the GEA combined with local scores, indicating that, although a few key genes may contribute significantly, adaptation to climate likely follows a polygenic model, involving both large and small effect variants.

TABLE 1.

Partial mantel correlation coefficients (R ± jackknife SE) for isolation‐by‐environment test, for different sets of adaptive SNP identified from both GEA: The spontaneous set ($n=62$) and the wild set ($n=45$).

Sets	$\mathit{BF}\ge 10$	$\mathit{BF}\ge 15$	$\mathit{BF}\ge 20$	$\mathit{BF}\ge 30$	$\mathit{BF}\ge 40$
Spontaneous	$0.8713\pm 0.0689$	$0.8832\pm 0.0589$	$0.8812\pm 0.0355$	$0.8824\pm 0.0455$	$0.8245\pm 0.0932$
Wild	$0.8718\pm 0.0899$	$0.8967\pm 0.0669$	$0.9037\pm 0.0594$	$0.8921\pm 0.0715$	$0.7719\pm 0.1661$

Note: In columns, the adaptive SNP sets were created using markers with varying levels of support for association with the environment, ranging from moderate (Bayes Factor $\ge 10$) to decisive ($\mathit{BF}\ge 40$), to explore how polygenic adaptation can be.

Given the strong population genetic structure within the spontaneous set, with two distinct clusters (ferals versus wilds), we further investigated the IBE pattern by analysing populations subsets separately. Partial Mantel tests (corrected for geographic distance) performed within and between groups (feral and wild populations) show a consistently high correlation between environmental distances and adaptive genetic distances across all populations subsets, except for wild populations where it drops markedly (r = 0.40 for $\mathit{BF}\ge 20$, Table 2). This suggests that the SNPs identified by the GEA on the spontaneous set primarily explain environmental differentiation in feral populations but fail to fully capture the adaptive divergence in wild populations (see Figures S3 and S4 for a representation of IBD and IBE within groups).

TABLE 2.

Partial mantel correlation coefficients (R ± jackknife SE) from isolation‐by‐environment tests in the spontaneous analysis ($n=62$), using markers with a Bayes Factor $\ge 20$ (strong support for association).

All populations	Wild populations	Feral populations	Inter group
$0.8812\pm 0.0355$	$0.3834\pm 0.1007$	$0.9031\pm 0.1595$	$0.8808\pm 0.0456$

Note: Tests were run separately across population subsets: All 62 populations, 45 wild populations only, 17 feral populations only, and between feral and wild groups (inter‐group). Wild populations show a markedly and significantly weaker IBE signal than the other subsets, suggesting that markers identified in the full spontaneous set capture only part of the genetic variation underlying local adaptation to climate within wild populations alone. However, the feral subset shows similar result with all populations on the spontaneous set.

A likely explanation is that feral populations exhibit lower molecular variation and this variation is mostly nested within that of wild populations. As a result, the adaptive markers detected by the GEA on the spontaneous set may disproportionately reflect adaptation in feral populations, and most of the adaptive variants specific to wild populations were not identified in the spontaneous set.

In agreement with this, the IBE pattern obtained from the adaptive variants identified by the GEA on the wild set is highly significant. Additionally, the strongest correlation between environmental and adaptive genetic distances is observed with 4000 adaptive SNPs (i.e., $\mathit{BF}\ge 20$, Table 1), and the quadratic coefficient consistently negative, with an estimated optimum around $20.87\mathit{BF}\pm 2$.

3.3. Genomic Offset

3.3.1. Genomic Offset Predicted in the Future

A Genomic offset (GO) approach was used to assess the potential maladaptation of B. rapa populations to future climate, by estimating a distance between current allele frequencies and those predicted as optimal under future environmental conditions. GO was calculated for each population based on 10 climate projection models for 2050, and the median GO value was used to mitigate inter‐model variability. The extrapolated GO values across Western Europe and North Africa reveal spatial patterns of potential vulnerability (Figure 5).

FIGURE 5.

Median of the 10 Genomic Offset estimates for the 10 different climate models, represented across the study area for the horizon 2040–2060. Red (respectively blue) areas represent the highest (respectively the lowest) differences between current and expected future frequencies along the genome.

The populations sampled in Slovenia and French Brittany exhibit the highest GO values, suggesting that they may be particularly at risk under future climate conditions. Moderately high GO values are also observed along the Po and the Rhone valleys, as well as in central Spain, whereas most coastal regions show lower values, likely due to the buffering effect of large water bodies. However, these projections should be interpreted with caution in the wild native range (Algeria and Italy), as the environmental variation in wild populations was less well explained by the GEA analysis on the “spontaneous set” (as seen in Table 2, partial Mantel R = 0.46 for wild populations), on which the GO predictions rely. Moreover, extrapolating beyond the sampled populations may introduce uncertainty, as we do not know if Brassica rapa is present in areas where no population was sampled. Previously published niche models suggest that the species occurs in the wild in most of Europe (figure 4B in McAlvay et al. (2021)), but even so, extrapolating to the whole distribution range assumes that the relationship between allele frequencies and environment we estimated in our sample is valid for the whole distribution range, which may not be true.

The first PC of the PCA on climatic data (Figure S1), which explained 52% of the climatic variability, contributed the most to the GO prediction (Figure S5). This PC summarizes the effect of many variables and mostly reflects a gradient from warm and dry climate to cool and wet regions.

3.3.2. Predicting a Fitness Proxy in a Common Garden Experiment

To evaluate whether GO estimates correlate with fitness reductions in a novel environment, we compared, for each population, the GO prediction in the experimental site in Le Rheu (France, 48°06′N 1°47′O) with the average reproductive biomass (number of seeds × seed weight) measured in this site. We observed a significant negative correlation (Figure 6a, R ² = 0.36, p‐value = 0.00000033), indicating that populations with higher GO values tended to produce a lower reproductive biomass in the common garden, as expected if GO estimates accurately predicted population's fitness in a new environment.

FIGURE 6.

Fitness proxy (averaged reproductive biomass, number of seeds × seed weight) measured in a common garden, as a function of (a) Genomic Offset values based on the spontaneous set, and (b) Environmental distance. The blue lines represent the regressions on the 62 populations. As the higher reproductive biomass observed in feral populations might result from artificial selection on crops before feralisation, we also performed the regression on the 45 wild populations only, as shown by the brown lines.

To assess whether genomic information actually improves the predictions compared to purely environmental metrics, we replaced the GO in our linear model by the environmental distance alone (${\left(e-e^{*}\right)}^{\prime }\left(e-e^{*}\right)$, as estimated in the GO framework in Gain et al. (2023)). We found that GO provided a better prediction of fitness decline than environmental distance alone (Figure 6b; R ² = 0.22), which supports the added value of genomic data in assessing vulnerability to climate change.

We also reanalyzed the GO‐fitness relationship using only wild populations from the spontaneous analysis, to account for potential biases introduced by feral populations. Indeed, intentional selection for high seed production during the domestication process may have driven differences between wild and feral populations that are not related to local adaptation. While the relationship remained, it was weaker (R ² = 0.11, p‐value = 0.028). This is not unexpected given that the GO estimation was based on SNPs identified using the spontaneous set (62 populations), which, as we showed above, does not fully capture adaptive variation in wild populations. Interestingly, environmental distance alone failed to predict our fitness proxy in this case (R ² = 0).

Finally, this relationship was investigated based on GO values obtained from the GEA analysis on the “wild set” (45 populations, Figure S6). No correlation was observed between these GO values and the fitness proxy in the common garden data (R ² = 0). Yet, this was expected given that the environment of this common garden represents a strong extrapolation, far out of the environments considered in the training dataset.

4. Discussion

This study investigated genomic data from spontaneous populations (wild and feral) of Brassica rapa from four different countries in Europe and North Africa. Our objectives were to uncover the genetic basis of local adaptation to climate in these populations and to assess their potential maladaptation under future climatic conditions. Particular attention was paid to the inclusion of feral populations to evaluate their potential as genetic resources.

4.1. Signatures of Adaptation to Climate Variation in Spontaneous Populations of Brassica rapa

The GEA of the whole spontaneous set revealed a large number of markers associated with environmental variables, which is consistent with the expectations of polygenic adaptation (Rockman 2012; Barghi et al. ²⁰¹⁹), whereby many variants of small effects contribute to local adaptation. This result confirms that the populations studied are locally adapted to their climate, in agreement with previous results in the genus Brassica (Johnson et al. 2023; Nelson et al. ²⁰¹⁴). The top SNPs identified in the GEA showed a slight (although not decisively significant) enrichment in missense variants relative to synonymous variants, lending some support to their functional relevance. Interestingly, many associated markers and regions were located in upstream or intergenic regions. This is not unexpected, as such regions can influence gene expression, offering a potentially rapid route to environmental adaptation (Walther et al. 2007; Bunnik et al. ²⁰¹³). Still, the local score approach pinpointed several genomic regions with a highly significant excess of associated SNPs, which could reveal a few loci with a strong effect. Among candidate genes, the three SWEET family sugar transporters stand out, and it could be consistent with selection along a water‐availability gradient, since SWEET transporters and the sugars they mobilize contribute to osmotic adjustment and tolerance to osmotic stress (Saddhe et al. 2021). However, functional annotations for candidate genes rely on orthology to a distantly related species and should therefore be interpreted with caution, pending functional validation.

Finally, the isolation‐by‐environment (IBE) test (Figure 4) provides further support for local adaptation in our populations, with an optimum recovered for an intermediate number of top SNPs. However, a closer examination of this signal revealed that the adaptive markers identified from the spontaneous set poorly explained the environmental distances observed in the wild population, and so greatly reduced the power to detect adaptive diversity in the wild populations. When studying populations with important genetic structure, it therefore seems to be insightful to repeat genome × environment association analyses for different subsets, to best characterize the global and specific adaptive variants, particularly in studies ranging over a large geographic range.

4.2. Signatures of Adaptation in Wild Populations

Wild populations showed clear signatures of local adaptation, particularly in the GEA focusing on these populations (i.e., the “wild set”). Here, a greater number of variants and candidate regions were detected compared to the spontaneous set. This is consistent with a strongly polygenic basis of adaptation, as expected for natural populations exposed to long term, stable selective pressures. In this case, the enrichment of missense variants among the top candidate SNPs was stronger and highly significant, providing further evidence that changes in these coding regions have functional consequences that matter for climatic adaptation. Among candidates genes, the heat‐shock proteins and the chloroplast development gene are noteworthy, pointing towards thermal and photosynthetic adaptation, which could be relevant in the southern Mediterranean populations.

Reassuringly, several of the variables we identified a priori as biologically relevant (e.g., BIO05, BIO06, BIO08, BIO11, BIO14, BIO18) were indeed recovered among the top associated markers, notably those related to summer heat and drought tolerance and to winter vernalization requirements. However, their relative contribution differed between datasets: in the spontaneous set, associations were mainly driven by average temperature and seasonality, whereas in the wild set they were more strongly linked to extreme temperature and low‐precipitation conditions, suggesting that the climatic pressures underlying local adaptation may differ between the two sets.

The population genetic structure, particularly marked between feral and wild populations, strongly influenced our analyses of the spontaneous set. Wild and feral populations appear to share a number of adaptive variants (Table 2), but a substantial fraction of adaptive variants were private to wild populations (e.g., Figure 3). This is not unexpected, since a GEA approach primarily detects signals shared across most populations, meaning that adaptive paths specific to a group or to isolated populations may go undetected; correcting for population structure could further have masked variants restricted to wild populations, insofar as this structure is itself correlated with adaptive genetic variation. This pattern of private adaptive variants likely also reflects a biological difference between the two sets: the substantially higher genetic diversity of wild populations relative to feral ones (Matuszewski et al. 2015) constitutes a larger reservoir of standing variation from which local adaptation can draw, and which is largely unavailable to feral populations. Taken together, the comparison of the GEA conducted on the two sets of populations indicates that wild and feral populations differ in the genetic architecture underlying their adaptation to climate.

4.3. Feral Populations as Genetic Resources

The patterns of IBE based on the GEA with the spontaneous set (Figure 4) also supported the existence of climate‐driven local adaptation in feral populations (Table 1), with a strong correlation between the genetic distance matrix at adaptive SNPs and the environmental distance matrix. This finding is particularly noteworthy given the uncertain history of these feral populations. They may have undergone natural adaptation to local conditions, particularly if they have persisted outside of human management for extended periods, such as feral rye (Burger and Ellstrand 2014), feral rice (Lin et al. 2021), and feral cabbages (Mittell et al. 2020). It is also possible that the currently established feral populations are those that successfully adapted after multiple introductions and are particularly resilient. Finally, another likely scenario is that these feral populations derive from locally adapted landraces, from which they have inherited their adaptation to their current environments. Further analyses, such as tests for adaptation in closely related landraces, are needed to choose between these two scenarios of re‐adaptation as ferals versus pre‐existing adaptation, keeping in mind that they are not mutually exclusive and that admixture events (feral‐domesticated or even feral‐wild) are likely (FitzJohn et al. ²⁰⁰⁷).

Conversely, the relatively low molecular diversity of feral populations (Matuszewski et al. 2015) could constrain their ability to adapt to local conditions. Our study indeed suggests some level of maladaptation, with the loss of alleles that would otherwise have been advantageous (see Figure 3). Even though feral populations tend to have lower population sizes compared to wild populations (Table S1), strong genetic drift is unlikely to explain this observed maladaptation, as we would not expect to observe such a consistent pattern of diversity loss at a specific locus across populations. It is also possible that feral populations have adapted through alternative pathways not present in wild populations, or that the lost alleles were not beneficial in their specific environmental conditions, independently of climate. Alternatively, the domestication history of feral populations may have favoured rapid changes through modifications in cis‐regulatory regions (Ballinger et al. 2023; Zou et al. ²⁰⁰⁹; Bogan and Place ²⁰¹⁹). This would lead to regulatory changes playing a comparatively greater role than mutations in coding regions, as we observed on the spontaneous set.

From the perspective of studying feral populations, including the more numerous wild populations increased the statistical power to detect adaptation. Indeed, the association analysis we conducted on the feral set showed some false positives associated with a fake variable, and the pattern of IBE was not as significant as in the spontaneous set. Our simulations confirmed that this low power is most likely due to the small sample size (17 populations). Including wild populations with the spontaneous set, while accounting for population structure using the omega matrix, significantly increased our power to detect adaptive variants for feral populations. One potential drawback of this approach is that SNPs private to the feral populations and associated with a given climatic variable are likely to remain undetected. However, such SNPs are expected to be rare, as most of the genetic diversity of feral populations is also present in wild populations.

4.4. Genomic Offset and Common Garden Experiment

We identified populations presenting an important genomic offset in their future environment (2050) in Slovenia and French Brittany, while others appeared to be at a lower risk. As GO values are relative, there is no absolute threshold for significance. Interpretation therefore depends on the context and comparative values (Gain et al. 2023). Additionally, projections of GO values on a map (Figure 5) should be treated with caution, as they assume a continuous distribution of the species and extrapolate the genetic architecture detected using a limited training dataset to this entire distribution. Nevertheless, this information may still be insightful, as such a map represents the expected climate change, weighted according to the relative importance of the different climatic variables in shaping the genomic diversity of the training set. However, this assumes that the genomic variants involved in local adaptation were adequately identified, which we showed was not the case for wild populations in the spontaneous set GEA. This limitation of the GEA affects the reliability of the GO estimates. Although we also performed GO analyses on the wild set, we were unable to validate these predictions in a common garden experiment. Indeed, the fitness data we used were measured in a common garden in Le Rheu, in French Brittany, a site that lies far outside the climatic range of wild B. rapa populations. We know that predictions outside the range of the training set are unreliable (Lind and Lotterhos 2025) and, as expected, we found no correlation between fitness values in Le Rheu and GO predictions based on the wild set. A proper test to validate these GO predictions based on the wild set would require fitness measures in a common garden within the geographic range of wild B. rapa , that is, in Sicily or Algeria.

In contrast, and despite the limitations mentioned above, the relationship between GO predictions deduced from the spontaneous set and our fitness proxy in the common garden was informative. After excluding feral populations to avoid potential biases related to seed traits, the fitness proxy is still significantly correlated with the GO, and this relationship is consistently stronger than with the simple environmental distance. This result is particularly noteworthy, as the IBE pattern indicated that the candidate loci identified explained only 46% of the environmental variation in wild populations (see IBE Table 2), and the common garden lies well outside the environmental range of these populations. In this context, including feral populations, despite the fact that feral and wild populations do not fully share the same adaptive variation, allows projection into environmental conditions not represented in the wild populations.

4.5. What Does GO Predict?

Gain et al. (2023) proposed a mathematical model concluding that GO measures should be negatively correlated with the logarithm of fitness, a statement supported by a simulation study and a common garden experiment. This paper therefore suggests that GO measures can not only predict a distance to optimal allele frequencies, but are also related to a loss of fitness. While Genomic Offset analyses have become increasingly common in the literature, potential conceptual issues with what GO actually predicts have recently been underlined (Lotterhos 2024). In particular, the authors questioned which “fitness offset” the GO actually predicts, pointing out the conceptual difference between predicting fitness of different genotypes in a same common garden and predicting fitness of a same genotype in several potential future conditions (“Home vs. Away” and “Local vs. Foreign” in Kawecki and Ebert ²⁰⁰⁴). Despite this useful clarification, some ambiguity about what genomic offset is predicting remains. Indeed, the difference in genetic composition between current allele frequencies and the optimal frequencies expected in the new climate (Capblancq et al. 2020) actually corresponds to a difference in genetic composition in the new environment (i.e., assuming very little to no evolution) and how it will affect fitness relative to the optimal genotype in this environment (Gain et al. 2023). Thus, we emphasize, along with (Capblancq et al. 2025), that GO cannot be used to predict how the fitness of a population will evolve between environments, but that it rather quantifies the relative fitness difference between this population and a perfectly locally adapted population in the new environment.

Despite its limitations, GO can still be a useful tool for predicting maladaptation. Combining GO with complementary approaches, such as species distribution models or estimates of genetic load, could improve our ability to identify vulnerable populations. Some studies have begun to integrate these frameworks (Jiang et al. 2025), but with little theoretical insight on how to combine them.

5. Conclusion

This study provides insight into the genomic basis of local adaptation in both wild and feral populations of Brassica rapa across Europe and North Africa. We show that adaptation in this species is highly polygenic, involving numerous small‐effect variants distributed throughout the whole genome, along with some loci of large effect.

Our results demonstrate that feral populations, despite their reduced molecular diversity compared to wild populations and their potentially complex demographic history (Gay et al. 2026), can exhibit clear signatures of local adaptation. This is consistent with the statement that neutral genetic diversity does not directly predict adaptive genetic diversity (Abson et al. 2026; Reed and Frankham ²⁰⁰³; Mittell et al. ²⁰¹⁵) and suggests that feral lineages may act as relevant reservoirs of adaptive alleles, especially in environments where the wild relative is absent. Nevertheless, the adaptive potential in feral populations may have been eroded by domestication and feralization events, increasing the risk of maladaptation under future climate change.

The different genetic architecture of climate adaptations between wild and feral groups posed both challenges and opportunities: while it reduced the power to detect adaptive variation in wild populations, the spontaneous analysis revealed clear signals of adaptation in feral populations. Incorporating feral populations also markedly expanded the range of environments available for maladaptation predictions. While the predictions remained imperfect, including feral populations allowed predictions that could not have been achieved by focusing exclusively on wild populations.

Our findings highlight the importance of considering feral populations alongside their wild relatives when studying local adaptation, especially when wild taxa have limited distributions.

Author Contributions

A.M.C., J.R. and L.G. designed the research and acquired the funding. C.F., G.D., M.T. and A.M.C. participated in the collection and the local description of the populations. C.F. was responsible for preparing the plants for DNA extractions, sampling leaf tissue and extracting DNA in pools. S.G. performed the analyses and wrote the manuscript with the help of L.G., S.B., M.G., J.R.

Funding

This work was supported by H2020 Prima funding (grant no. 1425), entitled “BrasExplor: Wide exploration of genetic diversity in Brassica species for sustainable crop production”.

Ethics Statement

This study complied with all applicable institutional and national regulations governing the collection of biological material. All samples were collected in accordance with relevant local regulations.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Bioclimatic variables used in this study. Variables follow the standard WorldClim/BIOCLIM definitions, with one modification: the precipitation based variables (BIO12, BIO19) were derived from soil moisture rather than precipitation. Quarters refer to consecutive 3 month periods.

Table S2: Bioclimatic and soil water variables extracted for the future climate projections. The 19 standard WorldClim/BIOCLIM variables (BIO1–BIO19) were complemented with five volumetric soil water variables. Unlike the present‐day dataset (Table S1), both precipitation and soil water were available here; this difference reflects the differing nature of the underlying datasets. Quarters refer to consecutive 3‐month periods.

Figure S1: PCA on the 24 climatic variables used in the genomic offset for one climatic model (NorESM1‐M). Each black dot represent the a population in its actual climate in the 2 first PC, while red dots represents future environment in the horizon 2050.

Table S3: Positions, genes and statistics for each regions detected with the local score approach: analysis on the spontaneous set. Lindley Score indicate the values for the local score, Peak BF (Bayes Factor) indicate the position of the marker with the highest support for association, and BF peak value the values in Bayes Factor for this marker. When the regions was not directly in a gene, nearest gene was indicated with the distance in base pair.

Table S4: Orthologs identified in Arabidopsis thaliana for the candidate genes in the spontaneous set, with their gene symbol and function. Functional descriptions follow TAIR annotations.

Table S5: Positions, genes and statistics for each regions detected with the local score: analysis on the wild set. Lindley Score indicate the values for the local score, Peak BF (Bayes Factor) indicate the position of the marker with the highest support for association, and BF peak value the values in Bayes Factor for this marker. When the regions was not directly in a gene, nearest gene was indicated with the distance in base pair.

Table S6: Orthologs identified in Arabidopsis thaliana for the candidate genes of the wild set, with their gene symbol and function. Functional descriptions follow TAIR annotations.

Figure S2: Minor Alelle frequency distribution of adaptive markers ($\mathit{BF}\ge 20$) detected in the wild set, represented for wild and feral populations. Frequencies were averaged per markers and per group (feral and wild). While most markers are present in intermediary frequencies in wild populations (median average MAF 0.1239), most of these markers are fixed or near fixed in feral populations (median average MAF 0.0033).

Figure S3: Patterns of Isolation‐by‐Environment and Isolation‐by‐Distance. For each population pair, $F_{\mathrm{ST}}/\left(1-F_{\mathrm{ST}}\right)$ is plotted as a function of the euclidean environmental distance on climate PC1 (IBE), or of the geographic distance (km, IBD). Black dots represent genetic distances on the whole genome, red for an adaptive set ($\mathit{BF}\ge 20$). The adaptive set was made from variant detected in the spontaneous analysis.

Figure S4: Patterns of Isolation‐by‐Environment and Isolation‐by‐Distance. For each population pair, $F_{\mathrm{ST}}/\left(1-F_{\mathrm{ST}}\right)$ is plotted as a function of the euclidean environmental distance on climate PC1 (IBE), or of the geographic distance (km, IBD). Black dots represent genetic distances on the whole genome, red for an adaptive set ($\mathit{BF}\ge 20$). The adaptive set was made from variant detected in the spontaneous analysis.

Figure S5: Variable importance of the climate‐data principal components, obtained when estimating the geometric genomic offset. The importance of predictor are assessed from the eigenvalues of the environmental effect‐size covariance matrix, following Gain et al. 2023. A random variable was included to assess the level of noise produced by the data.

Figure S6: Fitness proxy (averaged reproductive biomass, number of seeds × seed weight) measured in a common garden, as a function of Genomic Offset values based on the wild set (n = 45 wild populations). Correlation of determination is indicated in brown.

Appendix S1: Origins, locations, and general descriptions of populations included in “Insight into the genetic architecture of local (mal)adaptation in Wild and Feral Brassica rapa populations”.

Data S1: mec70596‐sup‐0003‐Supinfo3.tex.

Data Availability Statement

The short‐read resequencing data are available on NCBI's sequence read archive (SRA) under the project PRJNA1174687. The VCF file has been submitted to Zenodo and is available with the doi https://zenodo.org/records/18329477. Phenotypic data along with the scripts used for the analyses are available at https://github.com/SebastienGeneste/Brapa_mal_adaptation. Seeds of the accessions are available at the following Genetic Resource Centres: BrACySol for French accessions and Agricultural Institute of Slovenia for Slovenian accessions. For inquiries regarding Algerian, Italian, and Tunisian accessions, please refer to the contact specified in Table S3 in Tiret et al. (2025).