DNA methylation in Arabidopsis has a genetic basis and shows evidence of local adaptation

1To better understand how genotype and environment interact to affect DNA methylation and transcription, we grew 150 Arabidopsis thaliana accessions from Sweden in two different environments, 10°C and 16°C, chosen because they lead to very different flowering behavior. Relying on existing genome sequence information, methylome- and transcriptome-sequencing data were generated (see ‘Materials and methods’).

2In plants, DNA methylation occurs on cytosines in the CG, CHG, and CHH contexts (where H is any nucleotide except for C), each of which is catalyzed by independent pathways. Consistent with previous results we found considerable variation between accessions regardless of context, even at the level of genome-wide averages (Figure 1A). Temperature, on the other hand, did not appear to affect genome-wide CG or CHG methylation, but had a significant effect on CHH methylation, levels of which were 14% higher at 16°C than at 10°C, on average (Figure 1A). To investigate the genetic basis of DNA methylation, we performed genome-wide association studies (GWAS) using different facets of average methylation as the phenotype. For global CG and CHG methylation, no associations reached genome-wide significance, while for CHH methylation a clear peak of association was observed on chromosome 4 (Figure 1—figure supplement 1). The association was even more significant when restricting attention to average CHH methylation of large transposons (Figure 1B), in agreement with the notion that this type of methylation mostly occurs in transposons in Arabidopsis.

3The association centered around a SNP at 10,459,127 on chromosome 4, 38 kb downstream from the locus AT4G19020, which encodes a homolog of the CHG methyltransferase chromo-methylase-3 that has recently been shown to catalyze both CHH and CHG methylation on transposons, and is thus an excellent candidate. A multi-locus mixed model that included the identified SNP (CMT2a) as a fixed effect revealed another SNP downstream of CMT2, at position 10,454,628 (CMT2b), 4.5 kb closer to CMT2 than CMT2a, and in complete linkage disequilibrium with it (i.e., the non-reference alleles at CMT2a and CMT2b are never seen together). Repeating the GWAS with both CMT2a and CMT2b as cofactors identified no further loci (Figure 1—figure supplement 2). Both non-reference alleles are common in southern Sweden, but are also found in the north (22.6% vs 9.5% and 30.6% vs 7.9% for CMT2a and CMT2b, respectively). Accessions with the non-reference CMT2a allele have on average more CHH methylation on transposons than those with the reference haplotype (p = 1.1e-03), while those with the non-reference CMT2b allele have lower levels of CHH methylation than the reference haplotype (p = 8.1e-03; Figure 1C). The associations were readily confirmed using an F2 population generated by crossing accessions with the CMT2a and CMT2b non-reference alleles (Figure 2). No significant differences in CMT2 mRNA levels were observed between the alleles in our data and limited Sanger sequencing of cDNA showed no evidence of splicing variants (although, as will be discussed below, we did detect a putative rare null allele). Several non-synonymous polymorphisms in the methyltransferase and BAH domains of CMT2 were detected (Supplementary files 1 and 2) but they do not explain the phenotype as well as the CMT2a and CMT2b SNPs.

4The effect of genetic variation on local CHH methylation was examined by calculating the methylation level in 200 bp sliding windows across the genome (100 bp overlap between windows) and running GWAS for the 200,000 differentially methylated regions (DMRs; see ‘Materials and methods’) that varied most between individuals. 36023 DMRs had at least one genome-wide significant association (Bonferroni-corrected p-value of 0.05; 7273 remain after correcting for multiple GWAS using an FDR of 0.05). 45% (15,031) of the DMRs had a significant cis-association within 100 kb, while the rest showed evidence of trans-regulation, including the dramatic effect of CMT2 on chromosome 4 which accounted for approximately 21% (7392) of all significant associations (Figure 3A).

5To quantify the regulation of DMRs, we partitioned the variance of CHH methylation into environmental (E), CMT2, CMT2 X E, cis, cis X E, trans, and trans X E using a mixed model (Figure 3B). The analysis confirmed substantial cis and trans associations, with the environment modulating the genetic effects rather than having a major direct effect. At least for the cis associations, a possible explanation is that SNPs tag polymorphic TE insertions, with the insertion allele being silenced in a temperature-sensitive manner.

6The effect of temperature on CHH methylation could also be seen at the local level. We defined ‘temperature DMRs’ by looking for windows that differed significantly between temperatures. Comparing 16°C–10°C, each accession on average gained CHH methylation at ∼400 temperature DMRs and lost it at ∼200 temperature DMRs (false discovery rate = 0.05). CHH methylation is associated with transposable elements (TEs; ), and in agreement with this, 79% of the temperature DMRs where methylation was gained at 16°C were located within 500 bp of an annotated TE (with 60% directly overlapping one). These temperature DMRs were enriched in a small subset of TEs (835, or 2.7% of total, permutation based p-value = 0.05) that were more highly methylated than other transposons, but with lower methylation levels immediately adjacent (Figure 4A). Compared to TEs without temperature DMRs, these ‘variable’ TEs also tended to be euchromatic (Figure 4B), highly expressed (Figure 4C), and recently inserted (‘evolutionarily young’ TE insertions for which orthologs are not present in Arabidopsis lyrata comprised 75% of the variable TEs vs 68% of non-variable TEs). At the super-family level, members of the SINE, SINE-like, Helitrons and Mutator-like DNA TE superfamilies were over-represented among the variable transposons, and at the family level, 36 families were over-represented, including the AtREP, Vandal and HAT DNA transposons, as well as COPIA78/ONSEN and META1 retroelements (Table 1). Interestingly, COPIA78 has been shown to become active in response to heat stress apparently due to heat-shock promoter elements in its LTR regions.

7In order to gain further insight into the mechanisms responsible for variation in CHH methylation, we bisulfite-sequenced knockout lines of CMT2 (SAIL_906_G03) and DCL3 (dcl3-5, a component of the RdDM pathway), and identified 10,138 DCL3-dependant DMRs and 33,422 CMT2-dependent DMRs as described in section ‘DMR calling on DNA methylation mutants’ of the ‘Materials and methods’. As expected under the assumption that CMT2 is responsible for the massive GWAS peak on chromosome 4, the GWAS peak at this locus remains if we consider only the CMT2-dependent DMRs, but not for DCL3-dependent DMRs (Figure 5—figure supplement 1). Furthermore, while CHH methylation varied with temperature at both DCL3- and CMT2-dependent DMRs (Figure 5), DCL3-dependent DMRs were much more strongly associated with previously identified temperature DMRs (4703 out of 10,138 DCL3-dependent DMRs, or 46%, overlapped temperature-sensitive DMRs, whereas the corresponding numbers for CMT2-dependent DMRs were 2299 out of 33,422, or 7%; Fisher's exact p-value < 2.2e-16), suggesting that much of the temperature variation in CHH methylation is due to components of the RdDM pathway. This result is consistent with previous findings showing that RNA silencing is less active at lower temperatures.

8Interestingly, we observed one accession from northern Sweden, TAA-03, with almost undetectable levels of CHH methylation at CMT2-dependant DMRs (Figure 5). Further investigation suggested that it has a deletion or rearrangement in CMT2, as we were unable to map reads between positions 2813 and 4944 (intron 7 to exon 16, Figure 5—figure supplement 2). Sanger-sequencing indicates the insertion of a stretch of TC dinucleotide repeats of at least 330 bp. The same deletion appears to be present in three more accessions from northern Sweden (TAA-14, TAA-18, and Gro-3) a situation reminiscent of the homologous CMT1 gene, which seems to be non-functional in most Arabidopsis accessions. Although CMT2 null alleles have no obvious phenotype, the gene is highly conserved in plants (with the exception of maize;; ).

9It has recently been suggested that natural variation in CMT2 is associated with adaptation to climate, although the alleles identified in that study do not overlap with the ones identified here. Given the sensitivity of CHH methylation to growth temperature observed here, we next investigated the correlation between DNA methylation and the climate of origin. While CHH methylation was moderately correlated with photosynthetically active radiation (PAR) in spring (Pearson's r = 0.38), and CHG showed correlation with aridity (r = 0.35) and the number of frost-free days (Pearson's r = 0.30), by far the strongest signal was a strong positive correlation between CG methylation and latitude (Pearson's r = 0.70), as well as with a number of environmental variables that co-vary with latitude in our sample, such as minimum temperature and daytime length (Table 2, Figure 6A). As a result of the strong latitudinal correlation, accessions originating from northern Sweden (minimum temperature below −10°C) had on average 11% higher global CG methylation compared to those from the south (Figure 6A). The correlation appears to be driven by gene body methylation (GBM): as the correlation for CG methylation on transposons was much weaker (Figure 6A, Figure 6—figure supplement 1). Because the methylation variation observed for genes with average CG methylation below 5% appeared mostly to be noise (Figure 6—figure supplement 2, see also the ‘Materials and methods’ section), we classified genes into ‘unmethylated’ and ‘having GBM’ using this as a cutoff (5%). We also eliminated genes showing a transposon-like pattern of methylation in which not only CG, but also the CHH and CHG contexts are highly methylated. In what follows, we use GBM to refer only to gene body CG methylation for this filtered set. In order to better understand the observed variation in GBM, we examined CG methylation at the single nucleotide resolution within GBM containing genes. Although methylation was detectable (using a cut-off of 1%) at a similar number of sites in the north and south (1085292 vs 1079443 CG sites), those in the north showed a distinct skew towards higher methylation levels (Figure 6—figure supplement 3). Likewise when the difference between average methylation levels in the north and south was calculated individually for each of the cytosines, the majority of cytosines showed a small increase in the north compared to the south (Figure 6—figure supplement 4). From this we concluded that there is a general small increase in methylation of most CG dinucleotides in GBM genes, rather than large changes in a specific subset. GBM primarily occurs on long, evolutionary conserved genes that tend to be moderately-to-highly expressed, and is positively correlated with gene expression. Genes with higher GBM tended to be more highly expressed in our data as well, and—more interestingly—accessions with higher average GBM showed slightly higher average expression of methylated genes (although the correlation was weak, Figure 6—figure supplement 5). Given that northern accessions had higher GBM, this meant that genes with GBM were on average more highly expressed in northern than in southern accessions, while unmethylated genes showed little difference (Figure 6B). GBM has previously been shown to be anti-correlated with temperature-dependent gene expression. While no large-scale north-south expression differences were observed between 10°C and 16°C in our data, northern accessions showed considerably less variation in expression between the two temperatures for genes with GBM (Wilcoxon p-value = 1.2e-05), while no such difference was observed for genes without it (Figure 6—figure supplement 6).

10As for CHH DMRs, the genetic basis of GBM was examined using a variance-component approach (Figure 7A). The results were dramatically different: relative to CHH methylation, the trans effects for GBM are massive, and the environment appears to have no effect (in agreement with the observation that only CHH methylation levels vary with temperature, see Figure 1A). To identify the genes responsible, we also performed GWAS for each gene with GBM (Figure 7B). A total of 3241 significant associations were found for 2315 genes. 43% of these genes had a significant cis-association (within 100 kb of the gene of interest), which could represent local variants affecting methylation directly, or indirectly by affecting gene expression. No evidence for major trans-acting loci like CMT2 was found, but 69% of all significant associations were in trans. A comparison of the direction of the effect of GBM-associated SNPs in cis and trans revealed a striking pattern (Figure 7C). While the non-reference alleles of cis-SNPs were 1.18 times more likely to be associated with decreased rather than increased GBM (p = 2.01e-04), the non-reference alleles of trans-SNPs were 3.48 times more likely to be associated with increased GBM (p = 2.2e-16), and the non-reference alleles at the 15 trans-SNPs that were associated with GBM at five or more genes were always positively correlated (Figure 7C). Furthermore, while cis-SNPs showed a wide distribution of allele frequencies similar to random SNPs, trans-SNPs showed a much more limited distribution of frequencies (Figure 8A) and were also much more strongly correlated with latitude (Figure 8B,C). The correlation between GBM and latitude thus appears mostly to be due to trans-acting SNPs.

11The 15 highly associated trans-SNPs were largely limited to northern Sweden, and were in strong linkage disequilibrium with each other (Figure 8—figure supplement 1). A. thaliana from northern Sweden show signs of multiple strong selective sweeps and harbors many polymorphisms that appear to be involved in local adaptation (specifically to minimum temperature; ). The 15 SNPs were more than ninefold over-represented in the previously identified sweep regions (empirical p-value = 5.1e-03) and over fivefold over-represented within 2 kb of SNPs in the 1% tail of those associated with minimum temperature (empirical p-value = 3.1e-04), (Table 3). The ancestral state could be determined for 10 of the 15 SNPs, and in 8 of these cases, the non-reference allele was derived, further supporting sweeps in northern Sweden.

12That the difference in GBM between north and south is likely to reflect local adaptation is also clear from its relative magnitude. The north vs south divide explains a much higher fraction of the additive genetic variance for GBM (Qst = 0.772; see ‘Materials and methods’) than of the SNP variance (Fst = 0.187). This strongly suggest that the phenotypic differentiation is driven by selection rather than genetic drift.

13Identifying the causal variants is challenging, a gene-ontology analysis of genes within 100 kb (the average size of the sweep regions, ), of the 15 trans-SNPs found enrichment of loci associated with mRNA transcription (GO0009299, p-value = 2.62e-03). Genes involved in epigenetic processes are not captured well by standard gene-ontology, but we found that genes from the plant chromatin database (www.chromdb.org/) were significantly overrepresented in these regions as well (permutation p-value = 0.012; Table 4).

14We also looked for genes whose expression variation is consistent with a causal role. We identified 68 genes within 100 kB of one of the 15-trans SNPs whose expression is highly correlated (Wilcoxon test p < 0.001) with the adjacent SNP after correction for population structure (Table 5). No significant enrichment of Gene Ontology terms was observed among these genes, and manual inspection identified no proteins directly involved in DNA methylation. Instead, a number of proteins involved in the regulation of gene expression and/or chromatin accessibility were present (Table 5). This may suggest that the increased gene-body methylation observed in the north is not directly due to increased DNA methylation, but may be caused by increases in gene expression driven either by differences in transcription factors networks or chromatin compaction. Identification of the causal variants behind this phenomenon should provide insight into how plants adapt to their local environment.

15In conclusion, genes with GBM are generally up-regulated and more heavily methylated in northern accessions, and the change appears to be due to trans-acting polymorphisms that have been subject to directional selection. The candidate regions show an overrepresentation of genes involved in transcriptional processes. We also found that CHH methylation of TEs is temperature sensitive, and identified a major trans-acting controller, CMT2. Taken together, these observations suggest that local adaptation in A. thaliana involves genome-wide changes in fundamental mechanisms of gene regulation, perhaps as a form of temperature compensation.

16A diverse set of 150 Swedish accessions were sown on soil and stratified for 3 days at 4°C in the dark. They were then transferred to environmentally controlled growth chambers set at 10°C or 16°C under long day conditions (04:00–20:00) and individual seedlings were transplanted to single pots after 1 week. When plants attained the 9-true-leaf stage of development, whole rosettes were collected between 15:00 and 16:00 hr and flash frozen in liquid nitrogen.

17In addition, a cross between the T550 and Brösarp-11-135 accessions was created and F2 seed obtained. 113 individual F2 lines were grown in the same manner as the accessions.

18For each accession, three plants were pooled and total RNA was extracted by TRIzol (Invitrogen, Carlsbad, California, 15596-018), DNase treated and mRNA purified with oligo dT Dynabeads (Life Technologies, Carlsbad, California). RNA was then fragmented using Ambion Fragmentation buffer (Life Technologies) and first and second strand cDNA synthesis was carried out using Invitrogen kit 18064-071. The ends of sheared fragments were repaired using Epicentre (Madison, Wisconsion) kit ER81050. After A-tailing using exo-Klenow fragment (New England Biolabs, Ipswich, Massachusetts, NEB M0212L), barcoded adaptors were ligated with Epicentre Fast-Link DNA Ligation Kit (Epicentre LK6201H). Adaptor-ligated DNA was resolved on 1.5% low melt agarose gels for 1 hr at 100 V. DNA in the range of 200–250 bp was excised from the gel and purified with the Zymoclean Gel DNA recovery kit (Zymo Research). The libraries were amplified by PCR for 15 cycles with Illumina PCR primers 1.1 and 1.2 with Phusion polymerase (NEB F-530L).

19Single-end 32 bp sequencing was performed at the University of Southern California Epigenome Center on an Illumina (San Diego, California) GAIIx instrument using fourfold multiplexing.

20For each accession two individual plants were pooled and total DNA was extracted using CTAB and phenol-chloroform. Approximately two micrograms of genomic DNA was used for MethylC-seq library construction and sequencing (92 bp paired-end) by BGI.

21Illumina sequencing data from 180 published Swedish genomes were combined with sequencing data from another 79 (1001genomes.org), which had been processed using the same pipeline to yield polymorphism data for a total of 259 accessions (including those used for MethylC-seq and RNA-seq here). Linkage disequilibrium calculated using the R package LDHeatmap (version 0.9.1; ).

22After demultiplexing, 36 bp RNA-Seq reads were trimmed from barcodes (4 nt) and mapped to the TAIR10 reference genome including known variation with the PALMapper aligner using a variant-aware alignment strategy. Two different sources of variants were used: (1) single nucleotide variants (SNV) and structural variants (SV) from genome sequencing (2.1) and (2) SNVs and SVs called in an initial alignment round of the RNA-Seq reads to the TAIR10 reference genome with PALMapper (relevant parameters: -M 4 -G 4 -E 6 -I 25000 -NI 1 -S). For both sources of variants we applied stringent filter criteria to reduce false calls: (1) genome variants had to appear in at least 40 strains with a minor allele count of at least 5 strains, (2) RNA-Seq variants had to be confirmed by at least 2 alignments within the same strain and had to have less than factor 2 many non-confirming alignments within the same strain. Variants from both sources were integrated into one file that was used for a second, variant-aware alignment round with PALMapper (relevant parameters: -M 2 -G 0 -E 2 -I 5000 -NI 0 -S). In variant-aware alignment mode, PALMapper builds an implicit representation of the reference genome that reflects all possible variant combinations that exist for a genomic region. The output is automatically projected to the TAIR10 coordinate system. To account for reads ambiguously mapping to multiple locations in the genome, we used a custom python script (Supplementary file 3) to remove all reads that showed at least one mapping additional to the best hit with the same edit distance. Additional hits were only counted as ambiguous, if they differed at least 3 nt in start and stop coordinates to the best hit. On average, 5.7 M reads were mapped per sample after removal of ambiguous reads. Low complexity libraries with less than 30% of mappable reads or samples with less than 800,000 mappable reads (6 in total) where excluded from further analysis.

23To correct for possible sample or data mix-ups, SNP were called from the RNA-seq alignments using a custom python script and compared to independently collected SNPs from the Arabidopsis 250K SNP array (Supplementary materials; ). Samples not matching the expected genotype were reassigned to the correct genotype where possible or otherwise excluded from further analysis.

24We quantified gene expression by counting the number of reads that were longer than 24 bp and that mapped to genes on all non-chloroplast and non-mitochondrial chromosomes. To obtain a stable quantification, we only used those reads that were uniquely mapped into the exonic regions of genes. Furthermore, we required that the reads did not map completely into regions where two genes overlap in order to avoid mixing quantifications of different genes. In the later text we will refer to this estimate as the raw expression estimate.

25We also quantified the gene expression when additionally accounting for SV, alternative splicing and repetitive sequences that can all bias gene expression quantification. This estimate will be referred to as sv-corrected expression. For this quantification we additionally filtered for reads that start in an insertion or deletions and their two neighboring bases, that mapped into regions that are not contained in all transcripts of a gene and reads which were mapped completely into regions which are repetitive based on a 50 bp window.

26After filtering (see ‘Filter for gene expression quantification’), there were 499 RNA-Seq libraries left for analysis. Next, we merged libraries per ecotype and environment, yielding 323 unique merged RNA-Seq libraries for a unique ecotype and environment (160 in 10°C, 163 in 16°C).

27We followed the low level normalization proposed by, jointly applied to the set of expression estimates across ecotypes and environmental backgrounds. First, we estimated effective library sizes as the median expression estimates across all genes. Based on this, we derived correction factors to adjust individual libraries for differences in size.

28Library-size adjusted raw counts were used to obtain standard read counts per million expression estimates for each gene.

29Reads were aligned as previously described to a modified pseudo-reference chromosome in which SNPs were inserted into the TAIR10 reference genome using NextGenMap (version 0.4.3; ) allowing up to 10% mismatch between the reads (-i 0.90) and the reference sequence and discarding reads that map equally well to more than one genomic location or have less than 45 nucleotides mapping without error to the reference sequence (-R 45). Average coverage was 12.6 X.

30To correct for sample or data mix-ups, the raw data was also aligned to the first chromosome of the Columbia-0 TAIR reference genome as described above and SNP calling performed using the BISsnp package. The polymorphism data were then compared to data from genome sequencing (1001genomes.org). Accessions that did not have the highest similarity to the expected genotype were excluded from further analysis.

31Methylation was estimated individually for each cytosine using a python script provided with the BSMAP software package. Conversion efficiency was estimated from the fraction of methylated cytosines in chloroplasts using the R software package (www.r-project.org, version 2.15.2). After eliminating one outlier, the samples had conversion efficiencies ranging from 99.25%–99.80% (mean = 99.59%). Genome wide average methylation levels were calculated separately for the CG, CHG and CHH contexts. The average variance between 11 biological replicates was 2.2%, 3.2% and 7.3% for CG, CHG and CHH methylation respectively, while for identical genotypes grown at different temperatures (111 pairs) CG, CHG and CHH methylation variance was 2.7%, 4.6% and 15.9% respectively. The variance in genome wide methylation levels for the 152 accessions grown at 10°C was respectively 7.6%, 9.2% and 13.2% for CG, CHG and CHH methylation, while for the 121 accessions grown at 16°C genome wide CG, CHG and CHH methylation varied 8.5%, 9.5% and 14.3% respectively.

32The Bioconductor package Repitools (version 0.6.0; ) was used to average methylation over genomic features of interest (e.g., all genes, all transposons over 4 kB or a subset of transposons of interest). Pairwise DMRs were called individually for each accession using the R software package methylKit (version 0.5.6; ) using a window size of 100 bp, an FDR rate of 0.05 and a minimum fold change of 0.3. Overlap of DMRs with (TAIR10) genomic features such as transposons and genes was calculated using the Bioconductor package ChIPpeakAnno (version 2.8.0; ). For each accession, methylation data was smoothed independently for each context using the Bioconductor package BSmooth (version 0.4.5; ) using the default settings. Average methylation was then calculated for (overlapping) 200 bp sliding windows centered every 100 bp across the genome. Further analysis was limited to the 200,000 windows showing the most variance among accessions.

33Linear mixed models that correct for confounding by the genetic background using a kinship matrix calculated from genetic data were used throughout. To examine the effect of genotype on local CHH methylation variation, DMRs were defined by filtering the 200 bp methylation windows to remove those containing missing data (no coverage) in one or more accessions, then selecting the 105 remaining windows with the greatest variance in DNA methylation. For GBM, genes were filtered to remove those that had more than 0.05 average CHG methylation or less than 0.05 average CG methylation across the accessions (Figure 6—figure supplement 2).

34To investigate the relative contributions of genetic and environmental effects to methylation differences we used LIMIX, which efficiently estimates variance components using linear mixed models.

35For each DMR, we considered a linear mixed model with a fixed effect for the environment and random effects for the contributions from cis and trans genetic variants and variants from the CMT2 locus. Indicating with N and E respectively the number of samples and environments (E = 2), the NxE multivariate phenotype Y can be written aswhere μ is a Ex1 vector of environment-specific mean values, andwhere MVN(0,C,R) denotes a matrix normal distribution with mean 0, column covariance matrix C and row covariance matrix R. Rcis and Rtrans indicate the genetic relatedness matrices based on cis and trans variants respectively, where all variants within 50 kb from the DMR were defined as cis-acting and all others as trans–acting. Similarly, RCMT2 denotes the genetic relatedness matrix based on genotypes at the CMT2 locus. The row covariance of the noise component IN corresponds to an N x N identity matrix.

36The covariance matrices CCMT2, Ccis, Ctrans and Σ describe phenotypic correlations across environments due to these contributions, and were estimated from the data using maximum likelihood. For each DMR, we considered up to 10 random restarts for the optimization and stopped as soon as convergence was achieved. DMRs for which no convergence was achieved were discarded from genome-wide summary statistics.

37Once the model parameters have been estimated, the variance explained by environment can be calculated from μ, while environment-persistent and environment-specific effects from a given random effect can be estimated by decomposing the corresponding trait covariance into a shared and an independent component.

38MethylC-seq data for the 113 F2 individuals was mapped as described in section ‘Read mapping’ to the Columbia-0 TAIR reference genome. SNP-calling was done directly from the methylC-seq data using the BIS-SNP package. From these SNPs local haplotype was inferred for sequential 500 Mb windows which were then used to create a haplotype map using the R package R/qtl. Mapping was done using Haley-Knot regression with a 4 centimorgan steps size. Genome wide significance was estimated by permutation testing (1000 permutations).

39Pairwise DMRs were called for T-DNA mutants vs the wild-type control using the R software package methylKit (version 0.5.6; ) using a window size of 100 bp, an FDR rate of 0.05 and a minimum fold change of 0.3. Overlap between these DMRs and ‘temperature DMRs” calculated for the accessions was calculated and significance testing (Fisher's exact test) was calculated using R software.

40Fst was computed using the Hudson estimate as suggested in. We note that our estimate of 0.187 is consistent with the recent estimate of (although the samples only overlap in part). For Qst, we first estimated northern, southern, and overall additive variance using the Hasemann-Elston regression, and a SNP-based identity-by-state matrix, then calculated Qst as, where is the weighted average of variance within north and south populations, and is the variance between populations, obtained by subtracting from the overall additive variance.