Natural variation in cross-talk between glucosinolates and onset of flowering in Arabidopsis

Naturally variable regulatory networks control different biological processes including reproduction and defense. This variation within regulatory networks enables plants to optimize defense and reproduction in different environments. In this study we investigate the ability of two enzyme-encoding genes in the glucosinolate pathway, AOP2 and AOP3, to affect glucosinolate accumulation and flowering time. We have introduced the two highly similar enzymes into two different AOPnull accessions, Col-0 and Cph-0, and found that the genes differ in their ability to affect glucosinolate levels and flowering time across the accessions. This indicated that the different glucosinolates produced by AOP2 and AOP3 serve specific regulatory roles in controlling these phenotypes. While the changes in glucosinolate levels were similar in both accessions, the effect on flowering time was dependent on the genetic background pointing to natural variation in cross-talk between defense chemistry and onset of flowering. This variation likely reflects an adaptation to survival in different environments.


Introduction
To maximize plant fitness in challenging, ever-changing and unpredictable environments, organisms must coordinate growth and defense to respond to diverse combinations of biotic and abiotic factors. Interactions between biotic and abiotic factors force growth and defense to coevolve for local adaptation. Optimizing this necessary co-evolution requires the combination of various mechanistic solutions including the plasticity provided by regulatory networks to respond to environmental changes. Additional solutions are provided by the diversity inherent in genetic variation within a plant species, consequently providing different solutions depending on the environment (Pigliucci, 2003). Together, regulatory networks and genetic variation establish the potential for diverse solutions across a species to optimize growth and defense across highly varied environments (Burow et al., 2010;Paul-Victor et al., 2010;Woods et al., 2012).
To identify genes involved in cross-talk between growth and defense, we focused on Arabidopsis natural variation studies. Numerous studies have identified key genes controlling natural variation in the plastic timing of flowering time and thereby reproduction (Koornneef et al., 1991;Michaels and Amasino, 1999;Johanson et al., 2000;El-Din El-Assal et al., 2001;Salomé et al., 2011;Ward et al., 2012;Grillo et al., 2013;Méndez-Vigo et al., 2013).
Similarly, natural variation studies in the primary Arabidopsis defense compounds, the glucosinolates produced from tryptophan (indole glucosinolates) and methionine (aliphatic glucosinolates), have aimed at understanding the diversity in glucosinolate profiles (Kliebenstein et al., 2001a,b;Hirai et al., 2005;Keurentjes et al., 2006;Wentzell et al., 2007;Rowe et al., 2008;Jensen et al., 2014). Cross-talk between the networks controlling flowering and glucosinolate profiles seems to occur, as glucosinolate biosynthetic genes within the GS-AOP locus have been associated with not only glucosinolate biosynthesis, but also the control of onset of flowering (Figure 1), (Atwell et al., 2010;Kerwin et al., 2011). Hence, the genes in this locus may help illuminate the cross-talk between defense and flowering time in Arabidopsis.
Based on the catalytic properties of AOP2 and AOP3, we hypothesized that the previously observed difference in these two genes' potential regulatory effects may depend on their FIGURE 1 | AOP2 and AOP3 have been associated with natural variation in different phenotypes. The GS-AOP locus encoding the glucosinolate biosynthetic enzymes AOP2 and AOP3 has been associated with variation in glucosinolate profiles due to their enzymatic activities. The same genes have been linked to changes in glucosinolate levels and onset of flowering in different natural variation studies. enzymatic function and substrate availability, i.e., the levels of 3msp and 4msb. The production of these substrates is largely controlled by the allelic status of GS-ELONG encoding different methylthioalkylmalate synthases. In the absence of a functional AOP2 or AOP3, the presence of MAM1 (methylthioalkylmalate synthase 1) leads to accumulation of 4msb as the major SC aliphatic glucosinolate, while accumulation of 3msp is attributed to the presence of MAM2 (Figure 2), (de Quiros et al., 2000;Kroymann et al., 2001Kroymann et al., , 2003. GS-ELONG and GS-AOP show an epistatic interaction for glucosinolate accumulation (Kliebenstein et al., 2001b). This epistasis might be linked to the differences in AOP2/3 substrate availability depending on the allelic state at GS-ELONG, which would hint at specific feedback effects of the different products formed by AOP2 and AOP3.
In addition to their enzymatic and regulatory role in glucosinolate biosynthesis, both AOP2 and AOP3 have also been linked to flowering time control. Introduction of a functional AOP2 to the AOP null accession Col-0 shortened the circadian clock period and altered flowering time (Kerwin et al., 2011). In contrast, the AOP3 gene was found to have a hit in a genome wide association study to the control of the transcript level for a key flowering regulator, the MADS-box transcription factor FLC (Flowering Locus C), (Atwell et al., 2010). Yet the associations between the AOPs and flowering time remain to be tested in different accessions. However, AOP2 and AOP3 may mediate cross-talk between chemical defense i.e., glucosinolates and onset of flowering, which is critical for adaptation to environmental settings.
Hence, we tested if these two enzyme encoding genes, AOP2 and AOP3, have different effects on glucosinolate levels and flowering time in two different accessions, Col-0 and Cph-0, which express neither AOP2 nor AOP3. We introduced AOP2 and AOP3 into the two AOP null accessions that differ at the GS-ELONG locus and thus accumulate different AOP substrates. The use of multiple transgenic lines enable direct comparison of the effects of the three different AOP alleles (AOP2, AOP3, AOP null ) on glucosinolate accumulation and flowering time. Thus, this approach allowed us to systematically investigate the differences in the regulatory roles of all three AOP alleles, which is not possible in natural variation studies such as GWAS or in QTL mapping e.g., using RILs or F2 populations. The introduction of AOP2 and AOP3 led to different changes in the glucosinolate profiles in the two accessions. While the regulatory effects on glucosinolate levels were similar in both accessions, our study shows that AOP2 and AOP3 possess different abilities to change flowering time dependent on the genetic background.

Generation of Expression Constructs
For generation of expression constructs, we extracted genomic DNA from leaf tissue using the CTAB method (Clarke, 2009). The genomic sequence of AOP2 was amplified from Col-0 gDNA expressing the AOP2 allele of B. oleracea BoGSL-ALK under the control of the CaMV 35S promoter (Li and Quiros, 2003) with the primers 5 ′ -ggcttaauATGGGTG CAGACACTCCTCAAC-3 ′ and 5 ′ -ggtttaauTTATGCTCCAGAG FIGURE 2 | Enzymatic functions of MAMs and AOPs in the aliphatic glucosinolate pathway. The chain length of aliphatic glucosinolates is controlled by GS-ELONG: expression of MAM2 in the absence of MAM1 leads to C3 glucosinolates, MAM1 is required for the production of C4 glucosinolates, and MAM3 is responsible for the production of LC glucosinolates with C8 as the predominant chain length. The C3 glucosinolate, 3-methylsulfinylpropyl glucosinolate (3msp), can be converted to 3-hydroxypropyl glucosinolate (3ohp) by AOP3 or to 2-propenyl glucosinolate (2-prop) by AOP2. The C4 glucosinolate, 4-methylsulfinylbutyl glucosinolate (4msb), is converted to 3-butenyl by AOP2 glucosinolate (3-but), which is further converted by GS-OH to 2(R/S)-hydroxy-3-butenyl glucosinolate. Here, we report that 4msb can also be converted by AOP3 to give 4-hydrozybutyl glucosinolate (4ohb), (see Figure 3).

Generation of Transgenic Plants
Agrobacterium tumefaciens (strain PGV38 c58) was transformed with either the 35S:AOP2 or the 35S:AOP3 expression construct for transformation of Col-0 and Cph-0 by floral-dip (Clough and Bent, 1998). Positive transformants carrying the 35S:AOP2 construct were selected on MS plates with 50 µM kanamycin. T1 seeds carrying the 35S:AOP3 construct were selected by repeated spraying with 300 µM Basta at the 4-leaf stage. In the T1 and subsequent generations, the presence of the transgenes was confirmed by PCR on genomic DNA. For AOP2, we obtained multiple insertion lines in both Col-0 and Cph-0, whereas for AOP3 multiple lines were obtained in Cph-0, but only one line in Col-0.

Plant Growth
For all experiments, seeds were sown in a randomized design and cold stratified at 4 • C for at least 2 days. The plants were grown in climate chambers at 80-120 µE/(m 2 * s) light intensity, 16 h light, 20 • C, and 70% relative humidity. Flowering time was measured as the period between cold stratification and the day, when the inflorescence reached 1 cm in height and normalized to the corresponding WT.

Statistics
We used R version 3.0.1 (2013-05-16) for statistical analysis (Team, 2013). For the WT and insertion lines significance was tested using the lm and Anova function for the following linear model GLS = Experiment + Genotype + insertion line nested within Genotype + Experiment:Genotype with specific differences tested post-hoc using the pairwise t-test function with a Holm-adjustment for multiple testing. Summary statistics was calculated using the SummaryBy from the doBy package (Højsgaard and Halekoh, 2014).

Generation of the Col-0 × Cph-0 F2 Population
To create the Col-0 × Cph-0 F2 mapping population, Col-0 and Cph-0 WT plants were grown until the flowering stage and Cph-0 was used to pollinate Col-0. The F1 plants were genotyped for MAM1 and MAM2 to insure the presence of a copy of each allele, i.e., heterozygosity, before the F1 plants were selfed. The F2 population was investigated for flowering time by measuring the number of days from stratification till the primary inflorescence reached 1 cm (File S1 in Supplementary Table 1). For 171 plants, we obtained both genotype and phenotype results, which could be used in QTL mapping.

Genotyping by MassARRAY
For genotyping, genomic DNA was extracted using the Qiagen DNeasy 96 Plant Kit (Qiagen) according to manufacturer's protocol. 100 sites were chosen for Sequenom MassARRAY in an attempt to get full coverage of the genome. However, we did not have any previous knowledge on the Cph-0 accession and only 53 of the SNPs turned out being polymorphic between the parents. These SNPs were used to generate genetic maps for each mapping population using the Haldane function (File S2 in Supplementary Table 1).

QTL Mapping
Windows QTL Cartographer Version 2.5 was used for composite interval mapping determining significant thresholds for flowering time by doing 1000 permutations to estimate the 0.05 significance levels (Wang et al., 2012). The main-effect markers were validated and tested for Two-Way epistatic interactions using lm type II and ANOVA in R version 3.0.1 (2013-05-16) including the most significant marker for each QTL.

Results
AOP2 and AOP3 Alter Glucosinolate Profiles in Col-0 and Cph-0 Biosynthesis of the glucosinolate core structure gives rise to methylthioalkyl glucosinolates, which are further converted to methylsulfinylalkyl glucosinolates by GS-OX FMO1-5 Sønderby et al., 2010). Methylsulfinylalkyl glucosinolates accumulate in AOP null accessions, i.e., accessions that express neither a functional AOP2 nor AOP3 in leaves (Figure 2). To investigate the relative effects of AOP2 and AOP3, we introduced them to two AOP null accessions accumulating different SC methylsulfinylalkyl glucosinolates. We used the Col-0 accession accumulating mainly 4msb and a so far undescribed accession accumulating 3msp as its major SC glucosinolate, Copenhagen-0 (Cph-0).
In agreement with previous work using the Brassica oleracea or Arabidopsis accession Pi AOP2 (Li and Quiros, 2003;Wentzell et al., 2007;Neal et al., 2010), constitutive expression of AOP2 in the Col-0 background lead to the formation of 2-propenyl and 3-butenyl glucosinolate from 3msp to 4msb, respectively (Figures 2, 3B,C, for all individual glucosinolates see Figure  S1). The presence of a functional GS-OH within Col-0 further modified the 3-butenyl side chain to 2R-and 2S-2-hydroxy-3butenyl ( Figure 3C), . AOP3 expression in Col-0 led to the conversion of 3msp to 3ohp and interestingly, we also detected small amounts of 4-hydroxybutyl glucosinolate (4ohb), (Figures 3B,C). Thus, AOP3 appears to be able to convert 4msb to 4ohb in planta, an activity that could not be detected in vitro (Kliebenstein et al., 2001c). However, based on the ratios of substrates and products in Col-0, AOP3 seems to have a preference for 3msp.
In the Cph-0 background, 4msb is absent and instead, 3msp is the major SC glucosinolate. Consequently, introduction of the enzymes into the Cph-0 led to fewer different glucosinolate structures than in Col-0. As expected, AOP2 expression resulted in the formation of 2-propenyl glucosinolate from 3msp, whereas AOP3 formed 3ohp from the same substrate ( Figure 4B, for all individual glucosinolates see Figure S1).

AOP2 and AOP3 have Differential Effects on Glucosinolate Levels in Col-0 and Cph-0
Several studies have suggested that both AOP2 and AOP3 influence the total level of glucosinolate accumulation as well as the specific structures being produced (Kliebenstein et al., 2001b;Wentzell et al., 2007;Rohr et al., 2009Rohr et al., , 2012Brachi et al., 2015). To directly compare the regulatory capacity of both genes, we quantified the glucosinolates in our different AOP2 and AOP3 lines. The introduction of AOP2 to Col-0 led to a several fold increase in SC glucosinolates (Figure 3A), which is in agreement with previously published studies (Wentzell et al., 2007;Burow et al., 2015). The high levels of SC glucosinolates correlate with high accumulation of glucosinolates down-stream of AOP2, i.e., 2-propenyl, 3-butenyl, and 2R/2S-2-hydroxy-3-butenyl (Figures 3B,C). Introduction of AOP2 to Col-0 does, however, not only change the levels of SC but also LC glucosinolates, which are not AOP2 substrates ( Figure 3A). In contrast to AOP2, the introduction of AOP3 into Col-0 did not change the total accumulation of SC or LC glucosinolates ( Figure 3A). Thus, in the Col-0 background, AOP2 and AOP3 might differ in their  Frontiers in Plant Science | www.frontiersin.org ability to regulate glucosinolate accumulation possibly depending on their difference in enzymatic properties. However, as only one AOP3 line was obtained in Col-0, we cannot completely rule out a regulatory effect of AOP3 on glucosinolate accumulation in this background. In this line, AOP3 expression caused a larger decrease in 3msp than AOP2 and 4msb levels were only significantly decreased in the AOP2 lines (Figures 3B,C). This difference may be critical in case the removal of substrates is important for the regulatory role through a feed-back mechanism. However, if the regulatory role of AOP2 depends on its enzymatic activity, not only the removal of substrates but also the accumulation of the specific glucosinolate products may mediate this function.
Next, we investigated potential regulatory effects of AOP2 and AOP3 in Cph-0 that accumulates 3msp as the major SC glucosinolate. Cph-0 expressing AOP2 accumulated high levels of SC glucosinolates (Figure 4A). Similar to the effect of AOP2 expression in Col-0, increased SC glucosinolate levels correlated with the amounts of the AOP2 product ( Figure 4B). In Cph-0, AOP2 only converted 3msp to 2-propenyl, suggesting that the effect of AOP2 on SC may not depend on the chain length of the available substrate. In contrast to Col-0, introduction of AOP2 to Cph-0 did not significantly affect LC glucosinolate accumulation. Thus, the regulatory effect of AOP2 on LC glucosinolates might differ between the two accessions. Similar to the Col-0 accession, AOP3 did not change the levels of total SC or LC glucosinolates in Cph-0 ( Figure 4A). Even though the AOP2 and AOP3 can both convert 3msp, AOP2 seems to more efficiently pull from the 3msp pool than AOP3 and possibly thereby increase the flux through the pathway in the Cph-0 accession. Taken together, AOP2 has a larger effect on the regulation of glucosinolate levels than AOP3 in both accessions, which might be explained by the difference in their catalytic activities.

Effects on Flowering Time Differ between AOP2 and AOP3 and Depend on the Genetic Background
The glucosinolate biosynthetic genes AOP2 and AOP3 represent candidate genes for the integration of defense with reproduction as they are associated with the control of flowering time in both the laboratory and the field (Atwell et al., 2010;Kerwin et al., 2011Kerwin et al., , 2015. To test the ability of AOP2 and AOP3 to link glucosinolates and flowering time in different backgrounds, we measured flowering time in all of our lines (Figure 5). AOP2 has been identified as a QTL for altering circadian clock parameters and thereby flowering time (Kerwin et al., 2011). Accordingly, introduction of a functional AOP2 into Col-0 under 16 h light delayed flowering time by several days (Figure 5A). AOP3 has been associated with natural variation in flowering time and gene expression level of the MADS-box transcription factor FLC (Flowering Locus C), which is one of the major determinants of flowering (Shindo et al., 2005;Atwell et al., 2010). Analysis of the Col-0 AOP3 line showed no significant difference between Col-0 WT and Col-0 AOP3 lines ( Figure 5A). Thus, AOP2 but not AOP3 seems to influence onset of flowering in Col-0. Black Col-0 WT, n = 110, light gray Col-0 AOP2, n = 50, (2 independent insertion lines), and dark gray Col-0 AOP3, n = 32, (1 line). ANOVA with nesting and experiment interaction (min 2 repeats) shows that Col-0 AOP2 is significantly different from Col-0 WT and Col-0 AOP3, P < 0.001, whereas P = 0.45 for the Col-0 WT and Col-0 AOP3 comparison. (B) Flowering time relative to Cph-0 WT (41.6 days ±5.3). Cph-0 WT (black), n = 60, Cph-0 AOP2 (light gray), n = 73 (3 independent insertion lines), and Cph-0 AOP3 (dark gray), n = 60 (3 independent insertion lines). ANOVA with nesting of the different insertion lines and experiment interaction (two repeats) showed no significant difference between Cph-0 WT and the insertion lines; Cph-0 WT and Cph-0 AOP2 (P = 0.06) and Cph-0 WT and Cph-0 AOP3 (P = 0.22). Cph-0 AOP2 and Cph-0 AOP3 showed a significant difference (P < 0.01).
In contrast to its pronounced delay on flowering in Col-0, AOP2 had a suggestive ability to speed up flowering in Cph-0 using three independent transgenic lines (nested ANOVA; P = 0.06; Figure 5B). Similarly, AOP3 expression in the Cph-0 background did not significantly change flowering time compared to the WT ( Figure 5B). However, the AOP2 and AOP3 lines in Cph-0 showed a significant difference in flowering time from each other (P < 0.01), indicating that AOP2 and AOP3 have significant opposite effects on flowering time in Cph-0, which differs from the effects in Col-0.
Flowering Time Network Architecture is Critical for the Regulatory Effects of AOP2 and AOP3 While AOP2 and AOP3 had similar effects on glucosinolate accumulation across Col-0 and Cph-0, each gene differs in its ability to alter flowering time in the two accessions. One factor that may contribute to this difference is variation in the internal flowering time pathways in the two accessions. The Col-0 WT flowers earlier than Cph-0 WT indicating differences in the flowering networks. To identify the underlying loci that contribute to these differences and the variation in the AOP2/3 effect on flowering, we established a Col-0 × Cph-0 F2 population to map QTLs that control the major difference in flowering time between Col-0 and Cph-0. We genotyped 171 F2 plants for 100 SNPs using Sequenom MassARRAY and this resulted in identification of 53 polymorphic sites. For the Cph-0 accession with no prior sequence knowledge, we compared the SNP combination to the 1001 genome accession database (Cao et al., 2011;Schneeberger et al., 2011;Long et al., 2013) and did not find any annotated accession to have the same combination. The F2 population was also phenotyped for flowering time.
QTL mapping for flowering time revealed two loci (X29 and X188) as the major QTLs (Figure 6, Table 1). We found FT and FLC as the top candidate genes in these loci (candidate gene list from Grillo et al., 2013). Therefore, we analyzed the transcript levels of FT and FLC in the two accessions and found FT expression in Cph-0 to be 1-2% of the transcript level in Col-0, while FLC transcript levels were around 500 times higher in Cph-0 than in Col-0 ( Table 2). This is in agreement with the observed difference in flowering time between the two accessions, as FLC delays flowering time by repressing FT expression. This difference suggests that the ability to detect the influence of the AOP2 and AOP3 genes on flowering may depend on the allelic status at these known major effect flowering time genes in Arabidopsis. Further work is needed to assess how the AOP2 or AOP3 genes interplay with these known flowering time genes.

Discussion
In the field, glucosinolate profiles and herbivory resistance are strongly dependent on GS-AOP and GS-ELONG (Bidart-Bouzat and Züst et al., 2012;Brachi et al., 2015;Kerwin et al., 2015). Hence, the structures and levels of the glucosinolates produced depend on the allelic status of these two loci and their interaction plays an important role for plant fitness. The introduction of AOP2 into the AOP null accessions Col-0 and Cph-0 causes accumulation of alkenyl glucosinolates together with increased levels of aliphatic glucosinolates (Figures 3, 4), (Li and Quiros, 2003;Wentzell et al., 2007;Burow et al., 2015). In contrast, AOP3 expression in the same backgrounds led to formation of hydroxyalkyl glucosinolates without associated changes in glucosinolate levels. Thus, the production of alkenyl but not hydroxyalkyl glucosinolates influences the feedback regulation of glucosinolate biosynthesis in Cph-0 and possibly also in Col-0. Both AOP2 and AOP3 convert the same substrates, which suggests that it is the product being produced and not the substrate that is the determining factor. Consequently, the increased flux from primary metabolism into specialized metabolism in the presence of AOP2 may depend on the products of the enzyme mediating positive feedback regulation of the pathway. Recent work has suggested that this may occur by altering the jasmonate signal transduction pathway in lines that contain a functional AOP2 gene (Burow et al., 2015).
AOP2 significantly increased levels of aliphatic glucosinolates in both Col-0 and Cph-0 that differ in their GS-ELONG allelic status. This suggests that variation in the presence of 3C or 4C substrate availability is not the major determinant for the control of SC glucosinolates by the interaction of AOP2 and GS-ELONG. Only in the Col-0 background, LC glucosinolate levels were also significantly increased in the AOP2 lines, which could suggest that this effect requires the presence of C4 glucosinolates or variation in the MAM3, which catalyzes the production of the LC glucosinolates. The fact that LC levels differ between the two WTs suggests that there is MAM3 variation or in the regulatory network controlling LC glucosinolate accumulation.
As previously found, we could show that AOP2 delays flowering time in Col-0 (Kerwin et al., 2011), but there was only a suggestive effect of AOP2 on flowering in the Cph-0 background ( Figure 5). Mapping flowering time variation between Col-0 and Cph-0 suggested that these two accessions have natural variation in FLC and FT suggesting that this difference in the AOP2 effect on flowering time may be due to interactions with the known flowering time pathways. Supporting this hypothesis,  AOP2 has been shown to alter the circadian clock pathway that affects flowering time via regulation of FT (Kerwin et al., 2011). Thus, the regulatory effect of AOP2 on flowering time in Col-0 through FT might be altered by the competing regulation by the 500 times higher FLC levels in Cph-0. Recently, AOP2 was moreover found to mediate positive feedback regulation between the aliphatic glucosinolate biosynthetic pathway and jasmonate signaling (Burow et al., 2015). More specifically, expression of AOP2 in Col-0 led to increased transcript levels of MYC2. Interestingly, MYC2 is not only a key regulator of jasmonatemediated plant responses (Lorenzo et al., 2004;Dombrecht et al., 2007) including glucosinolate biosynthesis, but also involved in circadian oscillation of jasmonate signaling through direct interaction with the clock component TIME FOR COFFEE (Shin et al., 2012). The ability of AOP2 to alter flowering time in Col-0 might thus be linked to its regulatory input into jasmonate signaling. In contrast to AOP2, no clear effect was observed for AOP3 in either accession. Yet, AOP3 has been linked to natural variation in FLC expression (Atwell et al., 2010) indicating that AOP3 provides input to the flowering time network through a different mechanism than AOP2. If AOP3 can play a regulatory role in flowering time at all, the effect is minor and possibly requires a specific allelic status at the flowering time genetic loci.
Our study illustrates natural variation in the cross-talk between glucosinolate accumulation and flowering time. The ability to fine-tune this cross-talk differs between the two enzyme-encoding genes, AOP2 and AOP3, even though they arose from a recent gene duplication event and share high sequence similarity (Kliebenstein et al., 2001c). To fully understand the role of the GS-AOP locus in fine-tuning the regulatory cross-talk between glucosinolate profiles, jasmonate signaling, and the onset of flowering time, future studies in a larger number of genetic backgrounds will be required. Most importantly, studies in accessions expressing a functional AOP2 or AOP3 will reveal potential differences in the architecture of the regulatory networks in these backgrounds. Nevertheless, the specific regulatory effects and the dependency on the genetic background possibly reflect the plant's need to coordinate defense and reproduction when faced with different combinations of biotic and abiotic challenges. Hence, Arabidopsis seems to have evolved different cross-talk mechanisms linking defense and flowering time phenotypes to adapt to different environments.

Author Contributions
LJ, DK, and MB designed the study and interpreted the data. LJ and MB conducted the plant work. LJ, MB, and DK did the statistical analyses. HJ designed and carried out expression analyses. LJ and MB wrote the paper. DK and BH commented on the manuscript.