To perform our study, we have selected samples from the East Timor (Figure 1 and Table 1). No endangered or protected species were involved in this study. Seeds of Eucalyptus urophylla were collected between in 1973 and preserved in the CIRAD Genetic Forestry Laboratory. At the time of the sampling, East Timor was a Portuguese territory and sampling authorization was obtained from the military governor (SUPDOC). As GPS technology was not available at the time of the sampling, coordinates were estimated using a 1/50,000 topographic map. The seeds had been stored in a cold room at 4˚C and 30% humidity. In 2003, some seeds from this collection were planted. The plantlets were maintained in humid tropical nursery at the following conditions: 28˚C day temperature, 25˚C night temperature, 70% humidity, pH4 substrate. Leaves were collected on 4 months old plantlets, dried in silica gel and stored at room temperature until DNA extraction. To perform our study, we have selected samples from the East of Timor Island (Figure 1 and Table 1). We concentrate on Timor Island because the E. urophylla stands growing there are bigger and likely more stable over time than in other islands [20] . DNA was extracted from 43 individuals according to a modified protocol adapted from Gawel & Jarret [34] and Saghai- Maroof [35] . Each individual is coming from a different mother tree. All the 43 genotypes were used for the microsatellite analysis while 18 individuals were used for gene sequencing and gene polymorphism studies.

A total of 17 microsatellite loci were used for this study. They were all previously described: EMB18 [36] ; EMB22, EMB27, EMB30, EMB32, EMB33, EMB37, EMB38, EMB42, EMB44, EMB47, EMB52, EMB56, EMB63 [37] ; FMRA1, FMRA4, FMRA5 [38] . They were selected according to several criteria: specificity of the amplification, high polymorphism level, belonging to different linkage groups. The genotyping was per-

Num: location number reported on Figure 1 map; Microsatellites: number of individuals used for the microsatellite analysis; Gene: number of individuals used for the gene sequence analysis; Lat, Long: latitude, longitude; Alt: altitude (expressed in meters). Latitude and longitude were estimated with a 1/50,000 topographic map.

formed on the 43 individual set and 3 DNA controls with known allele size were added to each 96 well plates. The PCR amplification was performed by multiplexing 2 primers. Five microliters of QIAGEN multiplex mix, 0.08 µM of both forward primer with 5’-tail-end M13 (CACGACGTTGTAAAACGAC), 0.10 µM of both reverse primer, 0.10 µM IRDye fluorescent-labelled M13-primer (700 or 800 nm) and 5.0 ng of genomic template DNA. A touchdown cycling programme was used: 95˚C for 15 min, 67˚C for 1.5 min, 72˚C for 1 min, followed by eight cycles of 94˚C for 30 s, 65˚C for 1.5 min with 2˚C decrease at each cycle, 72˚C for 1 min then 24 cycles at 94˚C for 30 s, 51˚C for 1.5 min, 72˚C for 1 min, and a final extension of 60˚C for 30 min. Amplified fragments were analysed at 700 and 800 nm by electrophoresis on an IR2-DNA analyzer (LI-COR 4200 sequencer) at the Montpellier Languedoc-Roussillon Genopole genotyping platform. Allele scoring was done with SAGA software (LI-COR).

The E. urophylla genome has not been sequenced yet and at the time of the experiments, the E. grandis W. Hill ex Maiden sequence was not released. In that context, primer design was based on gene sequences from other species found in databases. Primers were designed in order to get the whole sequence of the candidate genes (Table S1). Details of primer design for each gene are given in Supplementary Online Material. PCRs were performed in 10 mL reaction mixtures with 10 mM incubation buffer 1X CORE Kit Q-Biogene (Tampon CORE Kit Q-Biogene, MP Biomedical), 0.30 µM forward and reverse primers, 1.2 units of Taq DNA polymerase (Invitrogen) and 15 ng of template DNA. We used the following PCR program for the amplification: 94˚C for 4 min; then 35 cycles (40 cycles for CesAs) of 94˚C for 30 sec, an annealing step for 1 min at the primer’s optimized annealing temperature for 1 min and 72˚C for 1 min (2 min for C4Hs), followed by a final extension at 72˚C for 5 min. PCR products were loaded on agarose gels for electrophoresis and the quantity of PCR product was estimated by image analysis with the freeware Image J v1.45b (Wayne Rasband, NIH). One hundred ng of primer product were sent to High-Throughput Genomics Unit (HTGU, Department of Genome Sciences, Univ. of Washington, Seattle) for direct sequencing. For each gene, the overlapping sequences obtained after direct PCR sequencing were assembled and aligned with Codon Code Aligner software (Licor. 2002-2007 ver. 3.5.2), and manually edited. For heterozygote sequences due the presence of INDELs, the sequences of the two alleles were either deducted manually or with INDELLIGENT software [39] .

While Tripiana et al. [22] found no population structure on a E. urophylla sample encompassing our studied area, we verified the absence of any cryptic population structure on our 43 individual SSR data set by using Structure software (version 2.3.3) [40] [41] . We used sample group information [41] and an admixture model with correlated allele frequencies. 12 independent runs were performed with 100,000 MCMC repetitions and a 100,000 burn-in period for K = 1 to K = 4. The most likely number of clusters was determined using the log probability of data.

We used the Approximate Bayesian Computation (ABC) approach as implemented in DIYABC software [42] , [43] to determine the demographic scenario that may have shaped E. urophylla diversity pattern observed on SSR data. ABC uses coalescence tool to generate thousands of simulated data having a similar configuration then observed one in terms of individual number, marker type and number and relies on summary statistics to summarize and compare simulated and observed data. Simulations are launched from a set of user-defined demographic scenarios with parameters sampled from prior distributions. DIYABC software provides a full set of tools that we used to simulate SSR data and to perform scenario choice, parameter evaluation and several steps of model validation.

Based on Tripiana et al. [22] results, we considered a single panmictic population and investigated five simple scenarios accordingly: SNM Standard Neutral Model (constant population size), BOT: bottleneck, EXP: demographic expansion, EXPBOT: recent expansion and past bottleneck and BOTEXP: recent bottleneck and past expansion. Cycles of population bottlenecks and expansions are possible in E. urophylla due to the volcanic characteristics of Sunda Islands and motivate the exploration of scenario EXPBOT and BOTEXP. These five scenarios are detailed in Figure 2, along with parameter description and imposed constraints. Since we have only scarce information about E. urophylla tree, we gave the five scenarios equal prior probability and used non-informative parameter priors, i.e. sampled from uniform distribution (Table S2). For each of these five scenarios, we sampled 100,000 simulated data sets whose corresponding information (parameters and summary statistics) was stored in the so-called reference table used in all further analysis except model checking. We recorded mean allele number, mean genetic diversity [44] , mean allele size variance and mean Garza-Williamson’s M [45] as summary statistics. Scenario posterior probabilities were calculated using a local logistic regression procedure proceeding on the 10% closest simulated points and that, shortly saying, gives more weight to those closer to the observed data. We then performed a model checking step to verify if the scenario with highest probability fits properly the observed data. As advised by Cornuet et al. [43] who recommended not to use for model checking the same summary statistics as for model choice, we constructed a new reference table with same parameter priors than previously but with recording only mean allele number and mean allele size variance as summary statistics and used mean genetic diversity [44] and mean Garza-Williamson’s M [45] for model checking. We simulated then 1000 additional data sets using the parameter posteriors of the scenario with highest probability and compared their summary statistic with those of the observed data. The scenario with highest probability is considered as suitable if the summary statistics of the observed data are included in the 95% confidence interval of the summary statistic’s posterior predictive distribution. Parameter posterior distributions were estimated only for the scenario with highest probability using the 10% closest simulated points with a local regression procedure and a logit transformation of the parameters. Confidence in scenario choice was finally evaluated by calculating power and alpha risk associated with choosing the scenario with highest probability. Alpha risk was evaluated by generating 1000 additional simulated sets for each of the four rejected scenario using their respective prior and by assessing the proportion wrongly attributed to the scenario with highest probability based on the reference table. Power was calculated by generating 1000 additional simulated data sets for the scenario with highest probability and by calculating the proportion correctly assigned using the reference table. Confidence in parameter estimates was evaluated by comparing the known parameters of a set of 1000 newly simulated data set obtained for the most likely scenario to those estimated using the current reference table. As measure of parameter estimation error and bias, we calculated the average relative bias and the factor 2, i.e. the proportion of pseudo observed data set for which the point estimate is at least half and at most twice the true value [42] .

The SNPs and INDELs were identified by comparison of the 18 genotypes. Polymorphism estimation and neutrality tests were performed with DNAsp software [46] . For all the analyses, the missing data and the INDELs were excluded in all the subsets. Nucleotide diversity on a base pair basis was estimated by Θ_W from the number of polymorphic segregating (S) sites [47] and by π [44] from the number of pairwise mismatches. These two site frequency spectrum related statistics are two estimators of 4Nμ, with N effective population size and μ the mutation rate per nucleotide and per generation. They are complementary since Θ_W is not sensitive to allelic frequencies and π is. To determine a possible non-equilibrium status of the studied genes, we measured Tajima’s D [48] . Tajima’s D quantifies the difference between π and Θ_w. A significant negative value of Tajima’s D can indicate an excess of rare variants as expected under positive selection. It can also reflect demographic event such as population expansion. A significant positive Tajima’s D value, at the opposite, indicates an excess of high frequency variants as expected under balancing selection or diversifying selection. It can also reflect population structure or bottleneck events. We calculated the rate of synonymous (dS) and non-synonymous (dN) substitution according to Nei’s [44] method. The ratio should provide insights into the long-term selective pressures on a full gene or a particular gene region and allows us to identify purifying selection and positive selection. As the later criteria for positive selection is overly stringent because positive selection often only acts on functional domain [49] we also performed a sliding window analysis with window length of 100 bp and step of 25 bp. We finally performed Hudson Kreitman and Aguade (HKA) test [50] and McDonald & Kreitman (MDK) test [51] with publicly available sequences of E. pillularis Smith for CesA1 (NCBI sequence AB591266.1), CesA3 (NCBI sequence AB591273.1) and CAD2 (NCBI sequence AB591254.1) genes as outgroup sequences. HKA and MDK neutrality tests are based on polymorphism to divergence comparison and are much less sensitive to demographic history than neutrality tests based on site frequency spectrum.

Demographic history strongly impact sequence polymorphism pattern and it is necessary to account for it while searching for selection imprint on candidate genes. To achieve that aim, we build the expected null distribution of several summary statistics for sequence data, conditional to the most likely demographic scenario inferred from the SSR data set, and used it to assess neutrality status of the candidate genes. We built this conditional distribution and detected outlier sequences using EGGLIB Python/C++ library and its ABC commands [52] . This library provides tools to easily perform coalescent simulations from inferred parameter posteriors and determine outlier status of candidate loci. Since DYIABC and EGGLIB do not use the same scenario parameters and that SSR and SNPs have different mutation rate and pattern, it is not possible to use the parameter posterior distributions inferred from SSR with DIYABC in our candidate gene analysis. We thus first calibrate scenario parameters for sequence data by running a new ABC analysis using EGGLIB, with candidate gene sequences as observed data and investigating only the most likely scenario inferred from SSR data. We simulated 10⁶ data sets, using uninformative priors and accounting for gene recombination. As summary statistics, we recorded average Θ_W and the relative frequency of 4 classes of minor allele frequency i.e. overall proportion of all polymorphic sites from all loci with minor allele frequency: below 0.125, between 0.125 and 0.25, between 0.25 and 0.375 and under 0.375. Parameter posterior distribution was established through simple rejection procedure [53] , keeping the 1% simulated data points closest to the observed data and applying a log transformation. To further proceed with posterior based coalescent simulations, we transformed the obtained continuous parameter posterior distributions into a single discrete multivariate distribution. We used the discrete multivariate posterior distribution of the parameters to proceed with two model checking procedures. As the studied candidate genes have different length, we simulated 10,000 times a 500 bp fragment from the posterior parameter distribution to be compared with the set of 500 bp long non overlapping fragments obtained through sliding window procedure on our candidate loci. We first applied a procedure similar to that used by DIYABC, by positioning the mean π, Θ_W and Tajima’s D calculated over the 44 observed 500 bp long non overlapping fragment in the corresponding posterior predictive distribution. We then compared the posterior predictive and observed distributions of π, Θ_W and Tajima’s D using Kolmogorov-Smirnov [54] test. We finally performed outlier detection by identifying the 500 bp observed fragments with Tajima’s D outside the 95% bilateral confidence interval of the posterior predictive distribution. Such outliers were considered as presenting evidence for selection signature.

Each gene consensus sequence was built by alignment of the overlapping fragments. The consensus sequence was then compared with the most similar gene used for the primer design in order to predict the 5’ and 3’ UTR, as well as the exonic and intronic regions. Hundred percent of the E. urophylla CAD2 gene (EuCAD2) was sequenced with a total length of 5408 bp, containing about 2500 bp of promoter region (GQ387647). We found two EuC4H copies and hundred percent of both EuC4H1 and EuC4H2 genes were sequenced. The sequenced genomic DNA length was 5020 bp for EuC4H1 (JX270996) and 2850 bp for EuC4H2 (JX270997). The two sequences were aligned with the three C4Hs of Populus trichocarpa Torr & A. Gray ex Hook and the two C4Hs of Populus tremuloides Michx to identify the localization of the exons and introns which have standard GT/AG splicing sites. A dramatic difference between the two EuC4Hs primary structures was observed: EuC4H1 contains 3 exons (785 pb, 134 pb and 599 pb) and 2 introns (676 pb, 1294 pb), while EuC4H2 had 2 exons (866 pb, 745 pb) and 1 intron (105 pb). Actually, the sequence of the EuC4H2 exon2 was highly similar to the exon 2 and the exon 3 of EuC4H1. At the nucleotide level, the similarity between the whole genes EuC4H1 and EuC4H2 is only 25%. However, when we compare the coding sequences, there is 62% similarity. The length of the sequenced EuCesA1 (JX270998), EuCesA2 (JX70999) and EuCesA3 (JX271000) genes are, respectively, 6723 bp, 5529 bp and 5493 bp. EuCesA1 and EuCesA3 have 13 exons while EuCesA2 has 12 exons. When we compare the mRNA nucleotide sequence for each copy pair, we obtain 62% to 64% of similarity. As for the whole sequence of each gene, we get 41% to 46% of similarity.

Each gene was sequenced on 18 individuals (Table 1). Sequences are available from GenBank under accession nos. KF970937- KF971152. On the full sample, the number of sites without missing data and gaps varied from 1027 bp (EuC4H2) to 4615 (EuCAD2), with number of segregating sites S varying from 39 (EuC4H2) to 170 (EuCAD2). Taking all genes together, we had 19,798 sites without any gap or missing data, cumulating 661 segregating sites. The overall level of diversity was 0.00651 for π and 0.00744 for θw. However, at a gene level, the diversity varied on a twofold extends: for π, from 0.00479 (EuCesA1) to 0.00899 (EuC4H2) and for θw, from 0.0051 (EuCesA3) to 0.00916 (EuC4H2). For the six studied genes, ratio always stayed bellow one at the full length scale, varying on almost a ten fold extend, from 0.0167 for EuCesA3 to 0.1341 for EuCesA2. Considering the sliding window analysis with window length of 100 bp and step of 25 bp, 3 gene regions exhibited ratio over 1:1.786 between 1360 and 1509 bp in EuCesA2 (beginning of exon 5), 1.029 between 3363 and 3462 bp in EuCesA3 (exon 9) and 2.73 between 5343 and 5442 bp in EuCesA1 (exon12). Tajima’s D values measured at whole gene scale varied from −1.05718 (EuC4H1) to 0.31332 (EuCesA3) and none of them were significantly different from zero according to standard test procedure based on Wright Fisher equilibrium hypothesis. Summary statistics recorded on each gene of the full sample are presented in Table 2. Table 3 presents detailed value of π, θw and Tajima’s D for fragment defined in the 500 bp non overlapping fragment set. Note that three of these fragments have no polymorphism at all. Neither HKA nor MDK test results indicated significant deviation from neutral expectation for the evaluated genes (EuCesa1, EuCesa3 and EuCad2).

Using Structure, we found no cryptic population structure. Increasing group number from 1 to 4 did not increase likelihood model. Average likelihood for K = 1 was −3727, average likelihood for K = 2 was −3764, average likelihood for K = 3 was −3822 and average likelihood for K = 4 was −4122 with low standard error between repetition of K (from 3.28 for K = 1 to 384.05 for K = 1). As an additional proof of no population structure, examination of individual repartition among group for K > 1 shows high admixture level for individual with minor group contribution.

The reference table was built with a total of 513,800 simulated data points, evenly distributed across scenarios: 102,500 for scenario SNM, 102,361 for scenario BOT, 102,951 for scenario EXP, 102,797 for scenario EXPBOT and 103,191 for scenario BOTEXP. Logistic approach indicated the highest posterior probability for scenario EXP with a mean value of 0.527 and confidence interval [0.502; 0.551]. BOTEXP and EXPBOT had lower probability, respectively 0.311 with confidence interval [0.287, 0.334] for BOTEXP and 0.134 with con-

l: length in base pair, excluding gap and missing data, S: number of segregating sites, π: pi, θ_w: θ defined by Watterson, Dtaj: Tajima’s D, dN/dS: ratio of non synonymous to synonymous polymorphism.

“Firstbase” and “lastbase” define the fragment boundaries; π: pi (ref) θ_w: θ defined by Watterson (ref), Dtaj: Tajima’s D; “Sim P value” indicated significant P values for Tajima’s D calculated using Tajima’s D expectation under sudden expansion demographic history and no selection, *: 0.025 > P and P > 0.975, **: 0.005 > P and P > 0.995.

fidence interval [0.119, 0.149] for EXPBOT. The two remaining scenarios have considerably lower probability, below 1% for scenario BOT and below 5% for scenario SNM. We kept only the scenario EXP for further analysis. Using the new reference table constructed for model checking and using only 2 summary statistics, scenario EXP still appears to be the most likely scenario. Model checking procedure further indicated that scenario EXP is compatible with observed data with a P(simulated < observed) of 0.265 for mean genetic diversity and of 0.403 for mean Garza-Williamson’s M. On the contrary, scenario BOTEXP, the second scenario in terms of posterior probability, was not compatible with observed data as P(simulated < observed) for mean genetic diversity was 0.976. Main characteristics of parameter posterior distributions for scenario EXP are presented in Table 4. For most parameters, the 95% confidence interval appeared to be broad, but examination of the posterior distributions indicated that they generally markedly differed from the priors (Figure S1). The posterior distribution of N2/Ne indicates a marked expansion (5% confidence interval [0.001; 0.110]), with a median value of 0.014. Confidence in parameter estimate is good (Table 4), with a mean bias below 1, except for N2 (mean bias 1.448) and Fact 2 above 0.7 (except again for N2, Fact 2 = 0.634). These values indicate that parameter estimation is reasonably good most parameters except N2. We finally assessed confidence in scenario choice by calculating associated power and alpha risk. The power to identify scenario EXP is good (81.5%) and its alpha risk is moderate (20.35%).

Following results obtained from ABC analysis conducted on SSR data, we build the expected null distribution of several summary statistics for sequence data conditional to the EXP scenario and used it to assessed neutrality status of the candidate genes. We first evaluated the EXP scenario parameters using candidate gene data. Main characteristics of parameter prior and posterior distributions for scenario EXP are presented in Table 5 and they are graphically displayed in Figure S2. Posterior median value for the ratio of ancestral size to actual size is 0.395, with corresponding 95% being [0.098; 0.778]. Median posterior value for current population size (4μN₀) was 0.011 with a narrow confidence interval [0.007; 0.025]. For expansion time expressed in 4N₀ generations, median value was 0.273 with broad confidence interval ([0.026; 0.480]).

The parameter posterior distributions obtained for scenario EXP were further used to simulate ten thousand 500 bp long sequence fragments. Tajima’s D, Θ_W and π values were calculated on each simulated fragment and used to determine a 95% bilateral confident interval characterizing the expected values of these statistics under a neutral model. The two model checking procedures indicated that observed data are consistent with EXP demographic model and inferred parameters. Positioning Tajima’s D values calculated on the 500 bp non-overlapping observed fragment set allow identifying two fragments presenting outlying values as indicated in Table 3. These two fragments, located in EuCesA2 and EuCesA3 are positioned on Figure 3 along with gene exonic structure, and Tajima’s D sliding window analysis and SNPs associated with wood traits. The first fragment is positioned in EuCesA3, between base 4236 and 5132 and has a positive Tajima’s D (1.395, P value < 0.025). The second outstanding fragment is positioned in EuCesA2, between bases 1 to 922 and has also a positive Tajima’s D value (1.700, P < 0.025) Note that this EuCesA2 fragment contains 2 SNPs found to be associated with Acid Soluble Lignin by Denis et al. [17] .

IC 95% indicates 95% confidence interval; MeanBias indicates the average relative bias and Fact2 the average proportion of pseudo observed data set for which the point estimate is at least half and at most twice the true value; N2: ancestral effective population size; Ne: current effective population size; Te: time of expansion in generation number; Mean_μmic: mean mutation rate; Mean_P is average parameter of the Generalized Stepwise model geometric distribution.

THETA defines actual population size ´ mutation rate, ANCSIZE is the ratio of ancestral population size to actual population size and DUR indicates the duration of expansion. For each of these three parameters, prior distribution shape (U: uniform) is given with minimum (min) and maximum (max) values. Posterior mean, median and 95% confidence intervals (IC95) are also presented.

E. urophylla is among the rare eucalypts species not originating from Australia. It was formerly found to be not structured at the nuclear level [22] [31] and it is believed to have undergone several cycles of population expansion and bottleneck in relationship with the volcanic nature of Sunda Islands [31] . Our microsatellite data analysis based on a tree sample from seeds originating from a small region of Timor Island is consistent with the absence of genetic structure and with a sudden demographic expansion. If local extinctions due to volcanic eruption may have occurred, it didn’t affect markedly overall diversity pattern in the sampled region where the E. urophylla populations are known to be large and stable. This expansion is confirmed by the ABC analysis we performed on gene sequence data set alone (results not shown). Previous studies on E. urophylla didn’t investigate possible demographic expansion but only tested bottleneck [32] . One of the main consequences of population expansion on polymorphism pattern is an excess of rare variants and thus negatively skewed Tajima’s D values for sequence data. In such a situation, not accounting for population demography may lead to the detection of false positive hits for positive directional selection.

The two data sets we used (microsatellites and genes) are very different in terms of sample size and marker mutation rate and modality. However, it is interesting to compare their respective posterior distribution for ance-

stral to current population size ratio in ABC expansion scenario. Indeed it does not depend on mutation rate and is a measure of expansion intensity. Expansion appears much stronger in microsatellites data set (median value of ANCSIZE, the ratio of ancestral population size to actual population size, is 0.014% and 95% confidence interval [0.001; 0.110]) than in gene data set (median value 0.395% and 95% confidence interval [0.098; 0.778]). Two explanations can be proposed for this discrepancy. On the first hand, some uncertainties can be associated with parameter posterior distribution estimation from ABC procedure. For example, in the ABC analysis performed with the microsatellite data set, the mean bias for ancestral size N2 is 1.448, which is beyond the threshold of 1. Poor parameter posterior estimation is often the result of an observed data set insufficiently informative [42] . On the second hand, our gene data set is built only with wood formation genes that are likely impacted by natural selection. For example, balancing selection may generate intermediate frequency variants that may attenuate demographic expansion signal. Finally, the choice of the proper posterior demographic parameters is very important and critical because it strongly affects false discovery rate. In the particular case of this study, we choose to be conservative and to use the posteriors inferred from the candidates themselves to draw neutral expectations, with in mind the fact that they may not exactly correspond to selective neutrality.

We investigated possible natural selection imprint on wood formation candidate genes using several complementary approaches: a test based on site frequency spectrum (Tajima’s D) with ad-hoc correction for neutral expectations, non-synonymous comparison to synonymous polymorphism (ratios) and two selection tests based on polymorphism to divergence comparison (MDK and HKA tests). As the latter approach required an outgroup sequence, we were not able to implement the two corresponding tests on the full sequence data set.

None of the studied gene exhibited a clear selection imprint on its full length but we could observe interesting trends in global dN/dS between copies of the same gene family, witnessing possible contrasted constrain level. According to Ohno’s [56] gene family classical model, contrasted constrain level can be expected as at least one member of the family should maintain the original function of the gene, and the other copies possibly undergoing pseudogenization or subfunctionalization. The three members of the CesA gene family we studied are all involved in secondary cell wall formation and possess homologous counterparts across angiosperms [27] . While these three genes seem to be co-expressed in secondary tissues, E. grandis CesA3 copy was found to display the highest expression level in Ranik et al. [27] study. In this work, the three CesA gene copies present very contrasted ratio, with almost a 10-fold ratio between EuCesA3 (0.0167) and EuCesA2 (0.1341). While these ratios indicate purifying selection in both cases, they suggest more constraints on EuCesA3 than in EuCesA1 and EuCesA2. Similarly, EuC4H2 (ratio of 0.0217) appears to be more constrained than EuC4H1 (0.1207). The only C4H copy of A. thaliana is expressed in all tissues and responds to light, wounding and fungal infection [28] , indicating that it plays various functions in phenylpropanoid metabolism that open a door for subfunctionalization of gene copies. Indeed citrus sinensis C4H1, ortholog of EuC4H2 is wound inducible unlike C4H2, ortholog of EuC4H1 [30] .

Based on site frequency approaches and by focusing on the 500 bp length non-overlapping fragment set, we found two 500 bp regions that could bear the imprint of natural selection. These two regions both present an excess of sites in intermediate frequencies that can be interpreted as balancing selection. The first region is located between bases 4236 and 5232 of EuCesA3 and covers part of intron 11, exon 12, intron 12 and part of exon 13. This particular region does not itself contain a SNP investigated by Denis et al. [17] but a nearby SNP (EuCesA3-3854) was found to be associated with cellulose and extractive related traits [17] . However, as we do not have proper outgroup sequence for this EuCesA3 region, it is not possible to make any final conclusion about its selective status. The second region is located in EuCesA2 between 1 bp and 922 bp and encompasses the four first exons, the three first introns and part of the fourth exon. It contains two SNPs (EuCesA2-0762 and EuCesA2-853) found to be associated with acid soluble lignin by Denis et al. [17] , among them one being non-synonymous. We also do not have a outgroup sequence for this EuCesA2 particular region but Sexton et al. 2011 found an unexpectedly high concentration of trans-subgeneric conserved SNPs in the nearby promoter interpreted as a signal of balancing selection.

Besides our work, several studies reported evidence of balancing selection in CesA genes in evolutionary distant tree species: in Pinus pinaster (Aiton) [19] , Pinus radiata D. Don [18] , in Eucalyptus genus in general [55] and in E. urophylla [56] in particular. This suggests the wide adaptive importance of CesA genes across trees and raises the question of functional significance of this result. As pinpointed by some authors, wood formation genes may play an important role in plant defence [57] , [58] although the modalities are not clear.