Sugar beet plants are out-breeding, with a tendency to suppress self-compatibility and therefore are naturally highly variable at the population level. This makes it difficult to interpret genetic data unless the experimental populations are fixed to some extent. The level of genetic variability can be reduced by selecting lines that are generated through single seed decent. To achieve this, we first selected C600 lines of the genotype bb, lacking the early bolting gene B and therefore unable to bolt without prior vernalization. These lines are considered to carry the late bolting gene Lb-lb 11 which is known to be linked to the monogerm character 12 . Siblings from one of these C600 bb lines were grown to maturity, vernalized for 18 wks and scored for bolting and flowering time. Sibling plants which bolted and flowered within 1-2 days of each other and designated C600 MB1-7; C600 MB1-13 and C600 MB1-35 were inter-crossed to provide bulk seed for our experiment. This bulked seed lot, designated C600 MB1 SibA, was then used to raise plant material for the experiment described here. In our hands, the generation time of biennial plants can be reduced to 1 year by artificial vernalization, in a controlled environment chamber, without compromising seed quality and quantity. Thus it took just over 2.5 yrs (including seed maturation/drying) to generate material suitable for our experiment. A second bb genotype, Roberta, a proprietary commercial cultivar was included in our experiment. We could then evaluate and quantify differences and/or similarities between our experimental line and commercial varieties, which are normally generated as hybrids of 3 different genetic backgrounds, combined in 2-way crosses between F1 cytoplasmic male sterile plants and a pollinator line 13 .

It is generally accepted that vernalization alters meristem competence to flower 14 and, in sugar beet evidence exists to suggest that vernalization signals are perceived in the leaves 15 . Further, our previous studies (unpublished) have indicated that vernalization-dependent GA-induced developmental processes leading to reproductive growth appear to be localized to the apical shoot meristematic tissues. The role of GA in floral regulatory networks is well established 16 and has been demonstrated for bolting and flowering in sugar beet 8 17 . To perform transcriptome-scale analysis of associated changes in gene expression, we harvested between 30-50 plants per treatment and micro-dissected apical tissues (Figure 1B), under the stereo microscope. A total of 184 apices were used in this experiment, distributed amongst the treatments as indicated (Figure 1A). Dissections were carried out ensuring that as much of the vascular and leaf tissues as possible were removed, whilst taking care to retain the meristem (Figure 1B iv). Following total RNA extraction, in-house quality assurance was carried out by first ensuring that there was no genomic DNA contamination. This was achieved by conducting a no-RT endpoint PCR reaction targeted at the housekeeping BvEF1a gene, using primers: L1: GATTCCCACCAAGCCTATGG and R1: GATGACACCAACAGCGACAG, optimised at 150 nM and 60°C. Next, the integrity of the RNA samples was confirmed on a standard denaturing formaldehyde RNA gel and by BvEF1a RT-PCR. Minimum 30 μg aliquots of total RNA for each treatment were then sent for custom sequencing prior to which the samples were quality controlled further by the service provider.

Digital gene expression (DGE) profiling using next generation sequencing depends on a reference transcriptome, which was not available for sugar beet prior to this study. Pooled total RNA from all six samples was therefore used to generate a normalised cDNA library which was then sequenced using the Illumina HiSeq2000 platform (100 bp single read module - see Figure 2A). The application of short reads is now significantly improved by the increase in read length to 100 bp, such that it now provides high throughput and good value for money and is therefore commonly used for de novo transcriptome assemblies in non-model species 18 19 20 . Here, this generated a total of 6.6 Gb of sugar beet transcript sequences. De novo assembly using the software tool Velvet/Oases 21 and http://www.ebi.ac.uk/~zerbino/oases/, yielded a total of 225,385 unique transcripts which corresponded to 165,742 unique loci. The assembly software tools were chosen because of their previous application for the assembly of similar RNA -seq data sets 19 . In this first pass assembly, the N50 value for all loci was 1185 bp and for large transcript loci, 1573. A BLAST search of the assembly against itself revealed that there was no redundancy although we found that 250 of the 17,186 unique entries (as of 17 March 2011 update) in the public sugar beet EST database (EST-DB) hosted at http://compbio.dfci.harvard.edu/cgi-bin/tgi/gimain.pl?gudb=beet mapped to more than 1 of the large transcript loci. To verify our assembly further, we performed a second pass assembly using Minimus 22 , which gave an assembly with an N50 value of 1678 bp. In contrast to the Velvet/Oases assembly, a significant proportion of these "doubly assembled" loci mapped to multiple Arabidopsis peptide sequences, thus suggesting that the Minimus assembly comes at the potential cost of over-merging, a feature that has been previously reported for peas 23 . We therefore elected to use the Velvet/Oases assembly in order to retain maximum information for subsequent analysis. The Minimus assembly output is however freely available (Additional file 1). Since there is as yet no consensus on the single best algorithm for sequence assembly 24 25 , we have made our raw data (Accession ID

ERP000947

In order to focus subsequent analysis on those transcripts/loci most likely to correspond to protein coding genes, we also determined the number of large transcripts (> 0.5 kb) which, in our de novo assembly would have required at least 6 independent 100 bp reads. We identified a total of 53,175 large transcripts, in the size range 0.5 to 8.729 kb, corresponding to 15,493 loci and hence putative protein coding sequences. To further substantiate our hypothesis that these were coding sequences, we explored the overlap between our de novo assembled loci and "large transcript loci" with the sugar beet EST-DB (as above), largely identified by conventional Sanger sequencing of EST collections from different sugar beet tissues. BLAST sequence similarity searches 26 27 set to a stringency of 100 bp overlap and 98% sequence identity of all 17,186 sugar beet EST-DB entries against all our 15,493 large transcript loci identified 7,925 loci common to both (Figure 2B), that is 46% of all the unigene ESTs currently in the public sugar beet EST-DB. This overlap rose to 12,810 unigene EST sequences, equivalent to 75% of the sugar beet EST-DB entries, following a comprehensive BLAST analysis with all our 165,742 shoot apex transcriptome loci.

Taken together, this analysis demonstrates a significant overlap of our new transcriptome with the existing sugar beet EST database. Moreover, the 7,568 large transcript loci with no matches in the current sugar beet EST-DB (Figure 2B) represent potential candidates of previously unknown (novel) sugar beet genes although, the possibility that some of these sequences may be the result of mis-assembly cannot be overlooked. Finally, our finding that 4,376 unigene ESTs in the existing public sugar beet database have no corresponding matches in our transcriptome is not unexpected since the public EST database is assembled from a wider range of tissue sources and experimental conditions, in contrast to the single source of shoot apices used in the current study. We elected not to perform a full-scale amalgamation of our transcriptome with the public EST database and to instead provide our database as a uniquely defined resource for vernalized and GA-treated shoot apices. In this way, it is more readily available to both breeders and academics with a specific interest in gene discoveries associated with the induction of reproductive growth in sugar beet. To facilitate such exploitation of our dataset further, we have mapped all loci to the Arabidopsis proteome by BLASTX alignment (cut off 1 × 10^-10) with the 27,416 peptide sequences in the TAIR database (version TAIR10), and have made these data freely available (see Additional file 3).

Having generated a sugar beet shoot meristem reference transcriptome, our next goal was to perform quantitative comparisons of the transcriptional programmes in shoot apices with respect to vernalization and/or GA treatment. We also aimed to investigate the potential impact of genotypic differences. To achieve this, un-normalised cDNA libraries were generated from subsamples of the same RNA as originally used to generate the reference and extracted from apices of the genotypes C600 variously treated with vernalization and GA, and, non-vernalized Roberta, variously treated with GA. A total of six independent libraries consisting of C600/untreated (sample A); C600/GA treated (sample B); Roberta/untreated (sample C); Roberta/GA treated (sample D); C600/vernalized (sample E) and C600/vernalized and GA treated (sample F), as shown in Figure 1A, were sequenced using the Illumina HiSeq2000 50 bp single read module. For each library/sample, the total number of counts for each sequencing read were determined and the reads were mapped back to our newly established reference transcriptome using Bowtie software 28 . Mapping programs continue to evolve 29 30 hence, we are mindful of the fact that slightly different results may be obtained with alternative mappers. Nevertheless, in this study, between 13.786 million and 20.360 million uniquely mappable reads, representing between 71% - 77% of total reads, were obtained from the six libraries (Figure 3A) thus, providing good coverage for the differential expression profiling.

We next generated a matrix containing the tag counts for each locus in each of the six samples and to reduce bias due to the slight variation in sequencing depths between samples (Figure 3A), for each sample, the tag count as a ratio of the total number of mappable reads was multiplied by 10 million, thus normalizing the values to tag counts per 10 million mappable reads. To perform the subsequent global comparative analysis, only the loci with a tag count ≥ 10 in at least 1 sample were retained. This resulted in a 7-column matrix (Locus ID plus tag counts for six samples) with 23,460 rows being equivalent to the total number of loci with a tag count ≥ 10 in at least 1 sample (Additional file 4). Analysis of this matrix by hierarchical clustering based on Pearson correlation coefficients showed high correlation (0.91 - 0.99) between all 6 samples, consistent with their shared tissue origin (Figure 3B). Within this, the biggest separation was observed for the two Roberta samples, suggesting that genotype had a greater overall impact on the global transcriptome than either GA or vernalization (Figure 3B). To further explore the impact of genotype, vernalization and GA treatment on shoot apex transcriptional programmes, we next analysed our 6 digital gene expression profiles by principal component analysis (PCA). In agreement with hierarchical clustering, a PCA plot based on principal components 1 and 2 resulted in a separation based on genotype (Figure 3C). However, a PCA plot based on principal components 2 and 3 showed a clear separation on the 3rd component (Y-axis) between the non-vernalized and vernalized C600 samples A, B and E, F respectively (Figure 3D). Moreover, principal component 5 further separated the non-vernalized samples (irrespective of genotype), according to GA treatment (samples A, C and B, D) while principal component 6 clearly separated the vernalized C600 samples (E, F) according to GA treatment (Figure 3E). Taken together therefore, transcriptome-scale analysis of the digital gene expression profiles suggested that transcriptional responses to treatment can be revealed from the data generated here. Our data also revealed a major impact of genotype on shoot apex transcriptional programmes. This is important because vernalization and GA treatment induce reproductive growth 8 . Currently, it is generally accepted that response to these inductive treatments is affected by genotype although it is not clear how. The data reported in this study provide a platform for future experimentation to reveal the molecular basis of genetic components that influence bolting and flowering.

Having identified genotype as the main factor determining transcriptional variation in our 6 datasets, we next set out to determine a gene set that showed high confidence expression differences between C600 and Roberta genotypes. To achieve this, we took advantage of the fact that our analysis included 4 × C600 and 2 × Roberta samples, which allowed us to determine statistical confidence scores for any potential expression differences driven by the 2 genotypes. In other words, we were able to exclude the additional variability due to treatment, thus enabling the assignment of higher statistical significance to those genotype-driven expression differences that are not also affected by GA treatment or vernalization.

As shown in Figure 4A, most loci fall on or close to the horizontal intersecting the Y-axis at zero when plotting the average expression scores in the C600 versus Roberta genotype. However, 4,880 loci showed differential expression at a p-value < 0.01 which corresponds to 21% of the 23,460 loci analysed. When the same analysis was repeated for the large transcript loci, 1,966 were differentially expressed at p-value < 0.01, corresponding to 15% of the 13,107 large transcript loci with an expression tag count ≥ 10 in at least 1 sample (Figure 4B). A bias towards C600 was observed in the differentially expressed loci which we suspect may have been a reflection of the higher number of C600 samples in our analysis. Collectively, the analyses performed here clearly demonstrate that the transcriptome datasets generated for the current study enable the global identification of genotype-specific expression differences. All differentially expressed gene loci presented in Figure 4 are listed in Additional file 5.

Having demonstrated that the transcriptome data generated for the current study allowed for the identification of genotype-specific differences in transcriptional programmes, we next explored whether the comparison of untreated shoot apices with GA-treated and vernalized samples would allow us to define responses to these two treatments at the transcriptome-scale. Comparisons were conducted only in the C600 genotype because it had the complete set of treatments. To this end, we performed BLAST searches against our reference transcriptome using cDNA sequences from putative sugar beet candidate regulators of bolting and flowering originally cloned in-house and/or identified by in silico BLAST searches of the sugar beet EST sequences in public databases (Sugar Beet Gene Index at http://compbio.dfci.harvard.edu/cgi-bin/tgi/gimain.pl?gudb=beet and GenBank) using known sequences from the Arabidopsis model. Of particular interest were GA metabolism genes, and apical shoot meristem identity genes expected respectively to participate in stem elongation (bolting) and to indicate the transition to flowering 10 16 . The expression profiles of matching gene loci with at least 100 bp overlap and ≥ 98% sequence identity were selected and then interrogated against our DGE dataset. This revealed that a greater change in expression profiles occurred in response to vernalization than to applied GA. Amongst the GA metabolic genes, we found that BvGA20ox1 (GenBank: DQ864510.1), which mapped to our Locus 24372, was up-regulated by vernalization, with a ~2-fold increase in expression in vernalized C600 apices (Figure 5A). Amongst the meristem identity genes, our data revealed the up-regulation of a MADS domain protein, which mapped to Locus 6819 with strong homology to FRUITFUL-like sequences (e.g. the sugar beet GeneBank: BQ584677 - BLASTN p value = 4.6 × 10^-107 and spinach GenBank: ACE75945.2 - BLASTX p value = 1 × 10^-102). Locus 6819 maps to the Arabidopsis APETALA 1 gene (TAIR: AT1G69120.1, BLASTX p value = 2 × 10^-70) although it does not show significant similarity to the sugar beet cDNA recently deposited at GenBank, labelled AP1 (GenBank: HQ454504.1). MADs domain transcription factors like APETALA1 are, in dicotyledonous plants, generally associated with floral organ development and therefore downstream of the floral transition 31 .

To interrogate vernalization responses further, we used our DGE profile matrix to rank transcripts according to expression levels in vernalized apices. We selected only those transcripts that were expressed in both non-vernalized and vernalized samples and discovered that a RAV1 (Related to ABA-insensitive 3/

v

We next used our DGE matrix to identify all gene loci whose expression profiles across our C600 samples either positively or negatively correlated with BvRAV1-like (see Figure 5C and 5D, respectively). For a p-value > 0.95, 732 genes showed positive correlation, and 34 genes showed negative correlation (all gene sets are listed in Additional file 7), thus demonstrating that this approach can readily identify gene sets with highly correlated expression profiles. Given that BvRAV1-like encodes a putative repressive transcription factor, with a negative regulatory role on flowering 33 , we suspected that amongst the correlated genes may be candidates with recognised functions in reproductive growth, including some which may be direct targets of BvRAV1. To test this, we analysed the smaller negatively correlated 34 gene set by first mapping each of the 34 transcripts to the Arabidopsis proteome, to identify homologous gene loci and selected those with E-value scores of 1 × 10 ^-10 or less. A total of 20 transcripts satisfied this criterion as indicated in Table 1 and were further analysed by proxy, using the best matched Arabidopsis gene locus IDs to conduct Gene Ontology (GO), Plant Ontology (PO) and network analysis using AraNet 36 , to enable inference of putative gene function and regulatory networks. A total of 18 out of 21 (including BvRAV1) of the Arabidopsis gene homologues were present in the AraNet database and as indicated in Table 2 the associated GO terms, by collective analysis of the gene set, were enriched for plant terms including some that we considered to be consistent with reproductive growth (e.g. as for AT4G2600, LOB, RAV1) and plant development (e.g. as for AT5G14450-lipid metabolism; AT1G68830-starch metabolism; LTP4; AT1G16070-transcriptional regulation; AT3G58690-protein modification) as well as hormone responses (MYB30; AT2G45640-abscisic acid response). This was supported further by GO analysis of the 20 query genes negatively correlated with the BvRAV1-like homologue, in combination with the top 200 associated new candidate genes revealed by AraNet. In this case, the 39 PO enriched terms included at least 13 developmental processes that could be directly associated with reproductive growth (see Table 3). Next, we conducted more specific GO analysis of individual query loci, based on inferences from direct assays, mutant phenotype, genetic interaction, physical interaction, expression patterns and traceable author statement. This enabled the assignment of putative gene functions based on network neighbours. Consequently, (as indicated in Additional file 8) we found that loci such as AT3G58690, AT4G25720, for which there was previously no functional information (from the gene set analysis - Table 2), may also be involved in processes associated with reproductive growth including regulation of meristem organisation, primary shoot apical meristem specification and flower development; while AT1G56580 is associated with GA biosynthesis, signalling and cell growth. Full details of enriched GO terms by inference from network neighbours for each of the 18 genes are given in Additional file 8 which also shows the 6 (out of 3,063) generally enriched (p < 0.05) GO terms (out of 3,063) and the 57 (out of 5,048) enriched (p < 0.05) InterPro Domains for the 20 query genes and their associated new candidates from AraNet.

Network analysis revealed that 10 of the 18 genes in the AraNet database were highly connected within extended regulatory networks, the largest of which included 7 query genes, connected as shown in Figure 6. The unknown gene locus AT5G19300 was the central node, with links to loci broadly involved with translation, rRNA biogenesis and assembly, protein modification, signalling, hormone and cold responses. A gene involved in reproductive organogenesis encoding the WD40 domain protein SLOW WALKER1 (SWA1), which is known to mediate mitotic cell division during female gametogenesis 37 was also directly connected to this central AT5G19300 node. A branch of the network extending to AT4G25720 (Figure 6A) included well established floral transcription factors such as the retinoblastoma-associated protein FVE which regulates flowering time 38 ; and the PHD-type transcription factor MALE STERILITY1 (MS1), which regulates pollen and tapetum development 39 . A second extension of the network converged on AT4G26000 (Figure 6B), known to be associated with GO plant terms for gynoecium and shoot development (Table 2 & Additional file 8), while others converged on AT2G45640 largely associated with GO plant terms for cold and hormone regulated responses and including histone modification affecting traits such as juvenility, apical dominance and floral organ development (Table 2 & Additional file 8) and as supported by the presence of HISTONE DEACETYLASE1 (HD1) 40 in this branch of the network (Figure 6C). Of note, histone methylation has also recently been demonstrated to play a role in vernalization induced bolting in sugar beet 41 .

None of our candidate genes revealed highly dynamic differences in response to GA treatment. Therefore, we interrogated our dataset for all those gene loci that displayed similar expression levels across all three non-GA treated samples (C600; C600/vern and Roberta), and at the same time, at least a 2-fold increase in expression following the application of GA. These loci were therefore expected to represent those genes whose expression reflects a generalised response to GA treatment, irrespective of the genetic background. This analysis identified 19 gene loci with robust GA-induction under all three experimental conditions (see Figure 7), thus demonstrating the utility of our DGE dataset for the identification of gene sets of interest without the need to have any prior candidate gene information.

Results of BLASTX searches with these GA-induced genes are given in Table 4, showing that amongst the sugar beet transcripts with significant matches to known genes were included an acetoacetyl-CoA thiolase (Locus 12027), a GA-regulated gene (Locus 30091); and 3 genes encoding efflux-type pumps of the

A

b

c

m

a

t

e

All raw sequencing data have been submitted to the European sequencing archive at the European Bioinformatics Institute under accession number EBI:

ERP000947

As for many other agronomically important crop plants, the development of genomics resources for sugar beet has lagged far behind model organisms, with no public reference genome or expression array platforms reported to date. However, the compilation of a sugar beet gene index from a variety of EST collections such as those generated by 50 51 52 53 , has provided a valuable resource. It was therefore necessary to benchmark the RNA-Seq-derived transcriptome generated here against this existing sugar beet gene index. This analysis revealed several important points:

i) Approximately 75% of the 17,186 unigene ESTs in the sugar beet gene index had a counterpart in our RNA-Seq transcriptome. This demonstrated that a large proportion of genes previously defined using traditional sequencing technologies were recovered in our RNA-Seq based study.

ii) 25% of previously known sugar beet unigene ESTs were not recovered which may be indicative of our tissue type and/or conditions to which the plants were subjected. Further, RNA-Seq based studies of additional sugar beet tissues are likely to increase the proportion of ESTs recovered by RNA-Seq.

iii) We identified 7,468 large transcript loci with no counterparts in the sugar beet gene index collection. Since these loci contain transcripts assembled from at least 6 independent 100 bp reads, extending over a minimum of 500 bp of mRNA sequence, the majority of these loci can be expected to represent bona fide protein-encoding mRNA. Notwithstanding the possibility that some of these loci could be the result of mis-assemblies, our data may now have expanded the total known gene count of sugar beet by up to 40%. Such an enlarged gene count will not only facilitate future efforts to assemble a sugar beet reference at the genome level, but also represents an immediately exploitable resource for gene discovery by both basic scientists and the sugar beet breeding community. The original data sets may be accessed in the sequence read archive (SRA) database at EBI under study accession number EBI:

ERP000947

As a general resource to the wider sugar beet community, the reference transcriptome described here is limited by having been generated not from the whole plant but from a single tissue type and under specified treatments and conditions. As a resource, it is therefore most suited to the analysis of reproductive developmental transitions in the shoot apex and hence for applications in genetic crop improvement for bolting and flowering time control. It is widely recognised that Illumina HiSeq technology can generate a significant amount of reads that are difficult to incorporate in the final assembly with currently available bioinformatics tools (~23%-29% with our samples). This suggests that our assembly could improve further as new tools come on line.

Reverse genetic-based approaches for the identification of gene candidates associated with key agronomic traits in crop plants are useful and informative. However, they may be encumbered by the need for transgenic analyses in crop plants, some of which may be recalcitrant to transformation. Expression assays of target genes may also be limited by the ability to design suitable qRT-PCR primers. Here, we have demonstrated the utility of next generation sequencing-based DGE profiling, not only as a tool for screening known gene candidates, but also for selectively identifying genes that are directly regulated by specified cues. For example, a MADS box transcription factor was revealed to be strongly up-regulated by vernalization. This was unexpected, since in dicotyledonous plants, MADS box transcription factors examined so far are repressors of floral induction that are known to be down-regulated by vernalization, as exemplified by FLC 14 54 . Nevertheless, we cannot discount the possibility that MADS box proteins may have a different role in sugar beet, indeed, in the shoot apex transcriptome, we found that the sugar beet FLC homologue, (BvFL1 - 55 ) is not down-regulated by vernalization (data not shown). Further, the discovery that transcripts which encode an AP2/B3 BvRAV1-like protein are strongly up-regulated by vernalization, now establish a role for a member of this protein family in vernalization responses for the first time. In the vernalized sugar beet apex transcriptome, we identified at least 14 other transcripts encoding AP2/B3 domain proteins (data not shown), none of which were similarly up-regulated by vernalization. The distinctive behaviour of BvRAV1-like therefore suggests a different role in vernalization which may extend beyond the generalized role in cold stress responses as for AtRAV1 34 35 56 . Transcriptome-scale analysis also enabled inference of putative gene functions of co-regulated transcripts, providing further support of a role for BvRAV1 in vernalization-induced reproductive growth processes in sugar beet. Thus, although the type of network analysis based on inference from co-regulated genes as described for BvRAV1 may be regarded as speculative, it nevertheless provides a rapid and comprehensive approach to candidate gene discovery. Hence, our discovery of BvRAV1 as a likely regulator of vernalization provides an attractive new hypothesis, which certainly will require further experimental validation, but nevertheless already adds an interesting new dimension to the identification of novel breeding targets for the sugar beet crop.

Another powerful application of DGE data, in cases such as that described here where expression array platforms and reference genomes are not available, is the ability to selectively identify, at the transcriptome-scale, those genes that behave in a specified manner under given conditions. Thus, we discovered 19 sugar beet genes that are directly up-regulated at least 2-fold by GA treatment of shoot apices. Although our analysis was conducted in a manner expected to reveal genes whose expression patterns would represent a general response to GA treatment, we cannot exclude the possibility that a different set of genes might be revealed in experiments with alternative genotypes. Nevertheless, for the genes discovered here, amongst the 8 genes that we putatively annotated using BLASTX, four appear to be consistent with the expected physiological effect of applied GA. For example, the added GA is well in excess of normal physiological levels hence, we might speculate that an immediate response of the plant may be to actively remove the excess GA. Here, we found that two of the GA-induced genes are predicted to have efflux pump associated functions, as represented by the ABC transporter and MATE family protein genes 57 58 . The unknown Med27 domain protein gene is also plausible since GA participates in the hormonal orchestration of transcriptional regulation. It is therefore encouraging to find that one of the GA-responses in the sugar beet shoot apex appears to involve the up-regulation of a protein with strong similarity to a well characterised family of co-transcription factors. Detailed analysis of the remaining, as yet unknown genes will, in future, reveal new insight into the molecular basis of GA-dependent bolting and flowering mechanisms in sugar beet.

C600 MB1 sibA plants were raised from seeds generated in-house from a sibling cross whose original biennial parent was propagated by single seed decent from a line that was segregating for vernalization requirement. Genotype at the bolting locus (B-gene) was confirmed using a PCR marker as previously described 8 . The original C600 lines were a kind gift from Bob Lewellen, (Plant Geneticist, USDA ARS, Salinas, California - now retired), and are closely related to the previously described C600 line, PI 520748 59 .

Roberta is our lab standard commercial sugar beet cultivar, which in common with most commercial genotypes, is a combination of 3 genetic backgrounds and has an obligate requirement for vernalization. Although this is a legacy cultivar it is representative of current European varieties and is kindly maintained and provided to us by Gunter Diener at KWS SAAT AG, Einbeck, Germany.

Plants were grown and vernalized as previously described 6 except that photoperiod was set to 8 h light and they were vernalized for 21 weeks. Non-vernalized plants were chronologically younger than vernalized plants but considered to be at the same developmental stage as determined by the number of fully expanded leaves at apex harvest. During the experiment, plants were kept in a controlled environment room at 22°C, under short day (SD) conditions (8 h light) in light intensities of ~285 μmol m^-2 s^-1 of photosynthetically active radiation (PAR).

GA treatment started at 5 weeks post vernalization and was continued for two weeks immediately prior to the harvest date. In this case, GA₄ was applied to plant shoot apices (10 μL of 200 μM stock) on alternate days and always at 7.5-8 h after lights came on, being the time since the environmental cue (ZEITGEBER (ZT); in this case light) that sets the circadian clock is expressed, entraining biological and hence expression rhythms in the plants. Figure 1B i) - iii) shows the appearance of plants immediately prior to harvest, which was carried out consistently at 4 hours after the lights came on (ZT4), by harvesting plant shoot apices (Figure 1B iii)) into 50-mL Falcon tubes, and keeping them cool on ice until the shoot apices were dissected out as indicated (Figure 1B iv), while taking care to remove all leaf and vascular tissues as visualised under the stereo dissection microscope. Dissected apices were transferred to RNALater and stored at 4°C until total RNA was extracted using the Qiagen Plant RNAeasy Kit. Apex tissues were disrupted using ceramic beads in the Bertin Technologies Precellys 24 Lysis & Homoginization machine (supplied by Stretton Scientific Ltd, UK), and the Qiagen RLT buffer was selected for RNA extraction. The RNA was DNAse treated with Ambion DNA-free™ DNase Treatment and quantified using the NanoDrop 2000/2000c (Thermo Fisher) machine.

To create a reference sequence, aliquots of total RNA (~5 μg) from each sample (Figure 1A) were pooled and used to create a custom normalised cDNA library which was sequenced by Illumina HiSeq2000 (100 bp single read module) and processed for assembly into contigs and loci using the software tools Velvet/Oases (v1.04 and v0.1.21 respectively). Sequencing and assembly were provided as a custom service by Eurofins MWG GmbH, Ebersberg, Germany. A schematic of the process is shown in Figure 2A. To evaluate the selected assembly tools further, the reads were re-assembled in-house, using Minimus 22 .

cDNA libraries generated from sub-samples of each test sample were sequenced (Illumina HiSeq2000, 50 bp single read module). Reads for each sequence tag were counted and mapped back to the reference using the software tool Bowtie 28 . cDNA library construction, sequencing DGE profiling and mapping back to the reference transcriptome as above were carried out as a custom service by Eurofins MWG GmbH. The outsourcing of these tasks, to highly experienced and reputable service providers, and in particular library construction, is recommended to reduce the risk of potential technical problems which may introduce bias in the final data sets.

Reference transcripts assembled from the 6.6 Gb of total reads from the pooled cDNA samples were validated by direct comparison with EST sequences in the sugar beet gene index data base at http://compbio.dfci.harvard.edu/cgi-bin/tgi/gimain.pl?gudb=beet using BLAST, set to a stringency of 100 bp overlap and ≥ 98% sequence identity. Further validation was obtained by comparison with peptide sequences in the Arabidopsis TAIR database (version TAIR10), using BLAST and an E-value cut off of 1 × 10^-10. These analyses were carried out in-house.

A matrix was generated with tag counts for each locus in each sample, and normalised to counts per 10 million reads [(reads/total reads per sample) × 10 million]. Only those loci with tag counts of 10 or above in at least one experimental sample were retained for further analysis. All correlation analyses, hierarchical clustering and principal component analysis were performed using Matlab http://www.mathworks.com. This was carried out in-house.