A set of 27 tissue samples from Solanum lycopersicon cv. ciliegia plants, comprising all tomato organs at different developmental stages, was processed with a commercial kit. Purified total RNAs had a mean value of 1.98 (SD = 0.09) for 260/280 nm ratios and showed, after denaturing electrophoresis, sharp and intense 18S and 25S ribosomal RNA bands with a practical absence of smears. The level of genomic DNA (gDNA) contamination in each RNA preparation was estimated by real-time PCR through the amplification of an alpha-tubulin gene sequence (tables 1 and 2). Only RNA samples from mature roots and immature green fruit gave a contamination signal, but with threshold cycle (Ct) values higher than 35. The cDNAs obtained from contaminated RNA samples were controlled during the corresponding RT-PCRs by means of reverse transcriptase-minus amplification reactions (RT-minus controls).

A total of 11 genes were selected as candidates for normalization of gene expression measures during tomato development studies (table 1). These include 7 classical (GAPDH, EFα1, TBP, RPL8, APT, DNAJ and TUA) and 4 novel (TIP41, SAND, CAC and SGN-U346908 – from now on referred to as "Expressed") housekeeping genes. In order to control for gDNA contaminations in the cDNA samples, PCR primers were designed on different exons (table 2) or spanning an exon-exon junction (forward primer for GAPDH and TIP41 genes), mainly guided by information about Arabidopsis genes. The performance of the amplification primers was tested by real-time PCR in two ways. First, aliquots from the 27 cDNA samples were pooled and used as template in amplification reactions with each primer-pair. A single band with the expected size (table 2) was obtained in each case without signs of primer-dimers formation (figure 1, odd lanes), as suggested by the previous melting curve analyses. Second, amplification primers were tested using gDNA as template (figure 1, even lanes). Seven primer-pairs yielded amplicons longer than those obtained with a cDNA template (table 2), whereas primers for GAPDH, TBP, RPL8, and SAND genes were unable to amplify genomic sequences. This result implies that intron position prediction in tomato genes was successful. As summarized in table 2, six amplicons obtained from gDNA have a melting temperature sufficiently different from those of corresponding cDNA amplicons to allow detection of gDNA interferences in a homogeneous assay. In the case of EFα1 primers, real-time RT-PCR should be followed by standard agarose gel electrophoresis. Absolute Tm values in table 2 should be considered with caution because they depend on the ionic strength of the actual reaction mix and the precision/accuracy of the real-time PCR platform.

Finally, in order to optimize PCR conditions, different primer concentrations were tested by real-time RT-PCR with the cDNA pool as template. Table 2 shows primer concentrations that provided the lowest Ct and thus the highest amplification efficiency.

Real-time RT-PCR was conducted on the 27 cDNA samples with the 11 primer-pairs. RT-minus controls were incorporated for mature roots and immature green fruit samples and only with the seven primer-pairs that yielded an amplification signal using gDNA as template. The 11 candidate control genes displayed a relatively wide range of expression level with mean Ct values between 21.1 (GAPDH) and 30.9 (EFα1). The RT-minus controls for mature roots and immature green fruit reached the fluorescence threshold only with APT, CAC and "Expressed" primers, but an extra treatment with DNase was not required because the Ct values of the mentioned RT-minus reactions were at least 10 cycles higher than those in the corresponding RT-PCRs, exceeding the minimum of 5 cycles recommended by Nolan et al. 3. Amplification specificity was confirmed by melting analysis or, in the case of EFα1 primers, by agarose gel electrophoresis.

In order to identify the most stably expressed genes during tomato development, the entire Ct dataset was analyzed using three different statistical approaches that have been incorporated in free specific VBA applets. The "pairwise comparison strategy" 26, implemented in the geNorm software, evaluates the variation of relative quantity (RQ) ratios for each gene-pair along the sample series. The "model-based approach for estimation of expression variation" 27, supported by the NormFinder software, estimates intra- and intergroup variation, and thus subdivision of the sample set in at least two coherent groups is required for the correct application of this approach. In this sense, we initially established the following sample-groups: roots (n = 5), leaves (n = 7; including cotyledons), inflorescences (n = 9) and fruits (n = 6). Since a minimum of 8 samples/group is recommended 27, expression data from different organs were also combined into "vegetative" (roots and leaves; n = 12) and "reproductive" (flowers and fruits; n = 15) sample-groups. The third statistical approach determines the expression stability for each control gene as the coefficient of variation (CV) of the relative expression levels after normalization. This evaluation strategy has been incorporated in the qBase program although limited to the analysis of 5 candidate genes 28.

The results of the three evaluation approaches are shown in table 3. It is noteworthy that definition of sample-groups had a notable effect on NormFinder output. Only two different NormFinder analyses were included in table 3 because we believe that sample grouping should not be arbitrary in an effort to adjust group sizes to increment statistical power, but rather that it reflects comparisons that researchers wish to make. It is remarkable that the NormFinder output with 4 sample-groups and CV ranking differ only in the relative position of the CAC and TIP41 genes. The results of the three statistical analyses exhibit several common features: i) CAC and TIP41 always rank as the most stably expressed housekeeping genes; ii), "Expressed", TBP and SAND also exhibit a remarkable stability of their expression levels and are always included among the 5 best performing reference genes; iii) GAPDH, EFα1 and TUA show unstable expression patterns and are always classified among the least reliable control genes.

Since the different statistical analyses applied to the expression data represent complementary strategies, we decided to combine results of the three evaluation approaches in a consensus rank, after averaging the two NormFinder outputs. For this purpose, genes were scored from 1 (most stable) to 11 (less stable) based on their relative position in each individual list. When two candidate genes are co-localized in a particular ranking (i.e., CV of the corresponding expression stability values ≤ 15%), both were scored with the average of the two consecutive positions. From the resulting consensus rank (table 3) it can be concluded that the best choice for normalization of expression measures in the entire developmental series are CAC and TIP41 genes, followed by "Expressed". Analysis of pairwise variation between two sequential normalization factors (NF) revealed that three genes are sufficient to calculate an accurate sample-specific NF as the geometric mean of their RQs. That is, the addition of a fourth control gene into the CAC/TIP41/Expressed combination does not significantly change NFs. The variation value for the pairs NF₃/NF₄ (V_3/4 = 0.118) is lower than the default cut-off value of 0.15 26. The mean M and CV values for the CAC/TIP41/Expressed genes in the complete developmental series are M = 0.537 and CV = 0.338. These values are inside the ranges M ≤ 1 and CV ≤ 0.5, which have been proposed by Hellemans et al. 28 as acceptable for heterogeneous sample panels, such as the space-temporal one surveyed in the present study. Unfortunately, reference values for assessing the relevance of NormFinder scoring have not been specified by the software's authors 27. In short, the CAC/TIP41/Expressed gene-triplet is recommended for accurate normalization of gene expression measures encompassing the complete development process in tomato.

To assess the validity of the procedure for the selection of control genes detailed above, the relative expression level of the ToFZY gene was estimated in five tomato tissue samples, using the control genes that we recommended for the normalization of gene expression measures in the whole developmental series. For this purpose, we used ToFZY specific primers described previously 31 and applied an efficiency-correction model for relative quantification 32. Our results (figure 2) were highly concordant with the transcriptional pattern of YUC1/YUC4 genes (the Arabidopsis homologous to tomato ToFZY gene) reported by Cheng et al. 33 based on histochemical analysis of GUS reporter lines and in situ hybridization.

The same evaluation procedure applied to data from the whole developmental series was tested on different sample combinations which, in our understanding would represent plausible experimental contexts. Cotyledons were always included in the leaf sample group because top-ranked genes were not affected by this combination. The unique exceptions to the analysis routine were the inference of consensus gene-rankings for individual organs (root, leaf, inflorescence and fruit), which were constructed without participation of NormFinder software because an estimate of the intergroup variation is not possible. Results shown in table 4 can be used as a guide for selection of suitable control genes that fulfil specific research needs with regard to the particular developmental series analysed. The complete consensus rankings are available as additional file 1.

The combinations of control genes recommended for the different sample subsets (table 4) are basically constructed with those that were ranked among the top five in the analysis of the whole developmental series (table 3), with the notable exception of TBP which is now downgraded in most consensus rankings (see additional file 1). It is clear that normalization of expression measures within organs have different requirements from comparisons between organs. On the one hand, two control genes are sufficient for accurate normalization in individual organs, as indicated by V_2/3 values lower than 0.15 (table 4). The recommended gene-pairs have mean stability values that are acceptable (M ≤ 0.5 and CV ≤ 0.25) for relatively homogeneous sample panels 28. In these cases, a third control gene is included (table 4, in brackets) for those wanting to use a minimum of 3 genes for calculating NFs, as suggested by Vandesompele et al. 26. In addition, the SAND gene is revealed as appropriate for normalization within organs, but less advisable for between-organs experiments. On the other hand, when the sample subsets were comprised of 2 or 3 different organs the evaluation procedure indicated that 3 control genes are necessary for a reliable normalization (see V_3/4 values in table 4). The expression levels of the 2 proposed triplets of control genes undergo oscillations comparable to those observed in the entire sample set, to judge by the values of the corresponding stability parameters (mean M and CV values in table 4). Moreover, control genes recommended for normalization in the complete developmental series (CAC/TIP41/Expressed) are also suitable for 3 different combinations of plant organs. The only exception is the subset integrated by inflorescence and fruit samples which can be suitably normalized with the following gene-triplet: CAC/SAND/RPL8. In this case, if RPL8 is substituted by TIP41, the next most stable gene in the inflorescence/fruit consensus ranking (additional file 1), the mean values for stability parameters would remain acceptable (M = 0.433 and CV = 0.341) and, at the same time, would allow the tool-kit of control genes for normalization during tomato development to be reduced to 4 components: CAC, TIP41, Expressed and SAND.

Tomato (Solanum lycopersicon) cv. ciliegia plants were maintained under growth chamber conditions at 25 ± 2°C with standard potting compost in 9 cm diameter pots. The relative humidity was kept around 60% with a 12 h photoperiod (120 mol PAR m^-2 s^-1). Plants were moved to a glasshouse when the 5-leaves stage was reached.

Sampling of the developmental series was prolonged over a 5-month period and comprised a total of 27 samples. The primary root that emerges through the seed coat was harvested at 72, 78 and 96 h following water imbibition. The proximal and distal portions of the mature root were collected at the 7-leaves stage. Cotyledons were excised at 96 h after seed imbibition. Six leaf samples were harvested per individual at the 6-leaves stage and always came from the apical leaflet. A total of 9 inflorescence developmental stages were established on the basis of bud sizes (8 samples; from 1 to 8 mm) and flower opening, as proposed by Brukhin et al. 38. Seeds and pericarp were gently removed from fruits at 3 different developmental stages: immature green, breaker and red stages. After collection, samples were immediately frozen in liquid N₂ and stored at -80°C until RNA extraction.

Total RNA was purified from tissue samples using the Spectrum™ Plant Total RNA Kit and on-column DNase I digestion, following the manufacturer's recommendations (Sigma-Aldrich). The amount of starting sample was 50 mg and this required the pooling of tissues from 5–30 individuals depending on the size of the material recovered. RNA was quantified using absorbance at 260 nm, whereas its purity was assessed based on absorbance ratios at 260/280 nm. The integrity of purified RNA was confirmed by denaturing agarose gel electrophoresis and ethidium bromide staining. Genomic DNA was isolated from young leaves (100 mg) using the GenElute™ Plant Genomic DNA Miniprep Kit (Sigma-Aldrich) according to the manufacturer's instructions and checked by standard agarose electrophoresis.

We selected 11 potential reference genes (table 1) that belong to different functional classes to reduce the chance that the genes might be co-regulated, with the possible exception of RPL8 and EFα1 since both participated in the translation process. This group of genes comprised several classical housekeeping genes which are commonly used as internal control for expression studies 724353940, such as GAPDH (glyceraldehyde-3-phosphate dehydrogenase), EFα1 (elongation factor α1), TBP (TATA binding protein), RPL8 (ribosomal protein L8), APT (adenine phosphoribosyl transferase), DNAJ (DnaJ-like protein) and TUA (alpha-tubulin). Based on expressed sequence tag data, the GAPDH, DNAJ and TUA genes have been proposed as internal controls for expression analyses involving different tomato organs 30.

The set of candidate reference genes also included less conventional housekeeping genes which showed highly stable expression levels in analyses of microarray data-sets from Arabidopsis 23, such as TIP41 (TIP41-like protein), SAND (SAND family protein), AT5G46630 (clathrin adaptor complexes subunit) or AT4G33380 (expressed sequence). Some of these genes have also showed to be stably expressed in a variety of grapevine tissues 35. Potential homologs to the corresponding Arabidopsis housekeeping genes were identified in the tomato unigen collection of the SOL Genomics Network (Cornell University; http://www.sgn.cornell.edu) via amino acid sequence comparisons with the BLASTX application.

The amplification primers for real-time PCR were designed using PRIMER3 software 41. In order to control for genomic DNA contamination, amplification primers were targeted to different exons (table 2). Information about exon positions in tomato GAPDH and EFα1 genes was directly available from databases. For the remaining tomato housekeeping genes, exon/intron boundaries were predicted through alignments 42 involving amino acid or nucleotide sequences from Arabidopsis and tomato, and based on information about exon positions from Arabidopsis Genome Project. The performance of the designed primers (table 2) was tested by real-time PCR using either tomato cDNA or genomic DNA templates.

Real-time amplification reactions were performed using SYBR Green detection chemistry and run in triplicate on 96-wells plates with the iCycler iQ thermocycler (Bio-Rad). Reactions were prepared in a total volume of 20 μl containing: 4 μl of template, 2 μl of each amplification primer (optimized concentration in table 2), 10 μl of 2× FastStart SYBR Green Master (Roche Applied Science) and 2 μl of fluorescein as normalization dye. Blank controls were run in triplicate for each master mix. The cycling conditions were set as follows: initial denaturation step of 95°C for 10 min to activate the FastStart Taq DNA polymerase, followed by 45 cycles of denaturation at 95°C for 15 s, annealing at 60°C for 30 s and extension at 72°C for 30 s. The amplification process was followed by a melting curve analysis, ranging from 60°C to 90°C, with temperature increasing steps of 0.2°C every 10 s. Baseline and threshold cycles (Ct) were automatically determined using the Bio-Rad iQ Software 3.0.

The cDNA samples for real-time RT-PCR experiments were synthesized from 1 μg of total RNA and random nonamer primers, using the First-Strand Synthesis System of Sigma-Aldrich. The cDNAs were diluted to a final volume of 200 μl. A mixture of the 27 diluted cDNA samples was used for selecting the optimal concentration of each PCR primer pair (table 2), in a 0.2–0.6 μM range and based on the generation of lowest Ct values. The PCR efficiency was determined for each primer pair in its optimal concentration (table 2) with the DART-PCR workbook 43, which uses fluorescence data captured during the exponential phase of each amplification reaction. The amplicons obtained with each primer pair from the cDNA mixture and from a random subset of individual cDNA samples were checked by electrophoresis on 2% agarose gels and ethidium bromide staining.

The possibility of genomic DNA contamination in the RT-PCR assays was controlled in two ways, and through the ability of amplification primers to generate different amplicons from genomic DNA than from cDNA. First, each primer pair was tested by real-time PCR using tomato genomic DNA as template (1 ng). The melting temperature and the size of the amplicons obtained in these reactions were annotated (table 2) and considered further in the analyses of the RT-PCR results. Second, for each of the 27 RNA samples, a quantity equivalent to the cDNA used in the amplification reactions (i.e. 20 ng of total RNA) was amplified by real-time PCR using primers targeted to alpha-tubulin sequences (table 2). These primers provided a great power for detecting genomic DNA contaminations. RNA samples giving alfa-tubulin amplification were further controlled by means of RT-minus amplification reactions.

The suitability of candidate control genes was evaluated by applying three different statistical approaches to expression data (i.e. Ct values). These strategies provide complementary measures of gene expression stability among cDNA samples. In the first approach, Ct values were converted into relative quantities (RQs) using the sample with the lowest Ct as calibrator and taking into account the amplification efficiencies calculated for each primer-pair (table 2), and then imported into geNorm v.3.5 software [26; http://medgen.ugent.be/~jvdesomp/genorm/]. This program estimates an expression stability value (M) for each gene, defined as the average pairwise variation of a particular gene with all other control genes in a given panel of cDNA samples. Genes with the lowest M values have the most stable expression. Housekeeping genes are ranked by geNorm through the elimination of the worst-scoring candidate control gene (this is, the one with the highest M value) and recalculating of new M values for the remaining genes. After this procedure is completed, two candidate genes are always top ranked because expression ratios are required for gene-stability measurements. The geNorm program also allows the minimal number of control genes required for calculating an accurate normalization factor (NF) to be determined, as the geometric mean of their RQs, but restricted to the gene ranking previously defined by the same program. This statistical procedure was adapted to any list of ordered genes in a homemade Excel worksheet. For this aim, first a NF is calculated for each sample with the two top ranked genes. Then, the most stable remaining gene is stepwise included and the NF is recalculated in each step. Finally, a pairwise variation of two sequential NF (V_n/n+1) was estimated as the standard deviation of the logarithmically transformed NF_n/NF_n+1ratios, reflecting the effect of including an additional gene. A pairwise variation of 0.15 is accepted as cut-off 26, below which the inclusion of an additional control gene is not required for reliable normalization.

In the second evaluation approach, Ct values were log-transformed and imported into the NormFinder software [27; http://www.mdl.dk/publicationsnormfinder.htm]. This strategy is based on a mathematical model of gene expression that enables estimation of the intra- and intergroup variation, which are then combined into a stability value. Candidate control genes with the minimal intra- and intergroup variation will have the lowest stability value and will be top ranked. For adequate application of the NormFinder program, the sample set should be subdivided into at least two coherent groups, each one ideally integrated for a minimum of 8 samples.

In the third evaluation approach, the coefficient of variation of normalized relative expression levels was calculated for candidate genes throughout each developmental series tested. This statistical approach, proposed by Hellemans et al. 28, has been implemented in the qBase software http://medgen.ugent.be/qbase/ but only 5 candidate genes can be simultaneously evaluated. In order to overcome this limitation, we incorporated in an Excel worksheet the formulas 11, 13 and 15 described in 28 for calculating normalized relative quantities (NRQs). First, mean Ct values were transformed into RQs using the specific amplification efficiency of each primer-pair and the sample with the lowest Ct as calibrator (formula 11). Then, a sample-specific NF was calculated as the geometric mean of the RQs estimate for all candidate genes (formula 13). Finally, NRQs were calculated as the ratio of the RQ estimated for a gene/sample pair and the corresponding sample NF (formula 15).