Parallel analysis of transcript and metabolic profiles: a new approach in systems biology

Introduction

Information flow in biological systems follows the sequence DNA to RNA to protein, with gene function generally described by the latter. Most current genomic research is carried out at the DNA and RNA levels; however, it is increasingly clear that a wide range of post-translational factors have functional importance in the cell. The current focus on the nucleic acid level is in part due to the fact that mature technologies are available that allow the sequencing (The Arabidopsis Genome Initiative, 2000; International Human Genome Sequencing Consortium, 2001) and analysis of the expression levels of complete genomes (or large proportions thereof) at the transcript level (Lockhart et al., 1996), whereas strategies to allow similar comprehensive studies of the protein and metabolite complements of the cell are still at relatively early stages in their development. However, rapid progress has been made in proteomic (Shevchenko et al., 1996) and metabolomic (Fiehn et al., 2000; Roessner et al., 2001) analyses in recent years, and combined studies of the transciptome and a subset of the proteome have been carried out in Saccharomyces cerevisiae (Futcher et al., 1999; Gygi et al., 1999; Ideker et al., 2001a), facilitating the use of systems-biology approaches (Ideker et al., 2001b; Kitano 2002; Oltvai & Barabasi, 2002).

Here, we describe the parallel analysis of potato tuber systems using a recently established platform for metabolic profiling based on gas chromatography–mass spectrometry (GC–MS) analysis in combination with the parallel analysis of gene-expression data using classical array technology. In carrying out these experiments, we set out to answer two questions: first, what is the relative power of the two phenotyping technologies to discriminate biological systems that either differ in developmental state or show well-characterized changes in response to the expression of transgenes? Second, can the combined analysis of transcript and metabolite profiling data present a new and meaningful approach in the identification of candidate genes for changing the metabolic composition of a biological system?

Results and Disscusion

When we designed these experiments, we decided to use a crop-plant system, the potato tuber system, to answer the above questions. The reasons for this were twofold: first, the potato tuber has well-defined but nevertheless highly related developmental stages and allows the assessment of several well-characterized transgenic situations; second, our group is experienced in the analysis of the potato tuber on both the biochemical and molecular levels (Fernie et al., 2002). As a first step, the expression of 1,000 genes (represented by at least two expressed sequence tags (ESTs), representing many transcription factors and a wide variety of biosynthetic genes) was analysed across development and in two transgenic situations using a custom array. After analysing the data sets obtained, we selected 279 ESTs that gave highly reproducible results with respect to the consensus sequence from which they were derived. We next performed a principal-component analysis (PCA) to determine whether various developmental stages could be discriminated from one another and from the transcript profiles of potato tubers that ectopically express a more efficient pathway of sucrose mobilization (Roessner et al., 2001). As shown in Fig. 1A, some of the developmental stages can be readily distinguished from each other on the basis of their transcript profiles. Most notable in this respect is the clear differentiation of tubers harvested ten weeks after transfer to the greenhouse from the other developmental stages. The exact reason for this discrimination is unclear. Examination of the upregulated clones suggests a higher metabolic activity at this stage; however, this finding is not consistent with changes in the metabolite levels documented in identical samples (see below). Despite the fact that these results are surprising, they are consistent for all three replicates harvested at this time point.

We also obtained the surprising result that the transcript levels of the transgenic systems could not be discriminated either from each other or from the corresponding wild-type tuber samples. It is important to note here that a similar situation was seen after PCA of the entire transcript data set (2,200 ESTs). Although it is clear that the tuber samples harvested ten weeks after transfer to the greenhouse were markedly different from those harvested at other developmental stages, the fact remains that transcriptional variation during development is greater than that after a relatively severe genetic perturbation of primary metabolism (Roessner et al., 2001).

Metabolic profiling was then carried out on identical samples to determine the levels of the major metabolites of primary metabolism, including sugars, sugar alcohols, organic acids, amino acids, as well as the nutritionally important ascorbate and tocopherol. When a PCA was carried out on the data set obtained from the metabolic profiling studies (Fig. 1B), a different situation was observed from that seen on analysis of the transcript data. In this case, the transgenic events clustered completely independently (both from the wild-type control and from each other); furthermore, on the basis of their metabolite complement, samples taken after ten weeks of growth were similar to those taken at other time points. As with the results presented for transcript levels, we believe that these data have important ramifications for the potential risks associated with transgenic organisms and theories of substantial equivalence (Kuiper et al., 2001; Trewavas & Leaver, 1999).

As stated above, this study was started with two main objectives. First, we wanted to compare the discriminatory power of both profiling approaches as described above. Second, we were interested in using the combined evaluation of both analyses as a new approach in experimental systems biology (for a definition, see Sweetlove et al., 2003). As a first step, we decided to run all data points through pairwise correlation analysis, determining for each transcript whether it is correlated to any of the metabolites. Of the 26,616 pairs analysed, 363 positive and 208 negative correlations were seen, the total number of 571 correlations being well above the number of correlations that might be expected by chance (266 at p < 0.01). Several representative examples of transcript–metabolite correlations are shown in Fig. 2 and are discussed in detail below, and the entire list is available in the supplementary information online.

First, as with any new approach, it is important to determine whether the data obtained are in agreement with observations made using different experimental strategies. This is clearly the case, with a strong negative correlation between sucrose and sucrose transporter expression (Fig. 2A), and a strong positive correlation between 4-aminobutyric acid and glutamate decarboxylase isoform I (Fig. 2B). As both of these relationships have been reported previously (Vaughn et al., 2002; Facchini et al., 2000), these examples provide confirmation of the validity of our approach. Second, many further correlations seem to have a functional basis that can be explained retrospectively. The positive correlations of both tryptophan and tyrosine with the β2-chain of tryptophan synthase (Fig. 2C,D), and of ornithine carbamoyltransferase with serine and cysteine (Fig. 2E,F), are two such examples. Third, although several of the correlations, such as those described above, were predictable, most of the correlations obtained using this approach were novel and not directly related to the biochemical pathway in which the gene products participate. Although it is clear that many of these correlations are due to chance, the fact that several correlations occurred between transcripts and metabolites from the same or related pathways might strengthen the interpretation of the linkages. An example of this includes aminotransferase, which correlates with both fructose-6-phosphate and glucose-6-phosphate (Fig. 2G,H). Such comparisons may offer hints as to the functions of the genes involved (the elucidation of which is a subject that requires a huge amount of research effort). It is also interesting to note that several transcripts correlate with more than one metabolite, as highlighted above for the aminotransferase. Other examples of this include glutamate decarboxylase isoform I, which correlates with both spermidine (Fig. 2I) and tyrosine (Fig. 2J), and the same is also true for 4-aminobutyric acid and tryptophan, and various transcription factors. Finally, it is exciting to see that nutritionally important metabolites, such as ascorbate, tocopherol and lysine, were closely correlated to the expression levels of various genes or transcription factors: ascorbate was negatively correlated with a homologue of the clock gene CONSTANS (Fig. 2K); tocopherol was negatively correlated with succinyl-co-enzyme-A synthetase (Fig. 2L); lysine was positively regulated by the transcription factor WRKY6 (Fig. 2M), S-adenosyl-L-methionine synthetase (Fig. 2N) and ornithine carbamoyltransferase (Fig. 2O), and was negatively regulated by caffeoyl-co-enzyme A O-methyltransferase (Fig. 2P). These unexpected results only reveal correlated parameters and do not allow the identification of causality between the transcript and metabolite. Although many possible explanations for such metabolite linkages can be postulated, including unknown metabolic links between pathways (such as, for example, the sharing of a common cofactor; or, in the case of transcription factors, interactions with previously undescribed regulatory genes), further, more direct experimentation is required to elucidate the mechanisms that underlie the correlations presented here. Nevertheless, we believe these essentially unexpected correlations to be of great potential use for biotechnological applications in which the goal is the modification of metabolite compositions through genetic means. Although these data do not, in any way, prove causal or even mechanistic links between molecular entities, the approach of linking transcript and metabolite data through pairwise correlation analysis provides a powerful tool for the rapid identification of candidate genes.

In conclusion, the discriminatory powers of transcript and metabolite profiling approaches are different, with metabolite profiling allowing greater resolution of the different systems studied here. Whether this reflects the fact that changes at the transcript level are less pronounced as compared with changes at the metabolite level or merely highlights limitations of the profiling technologies used, remains an open question. It is possible that this result merely reflects low sensitivity of ESTs as probes and that profiling using full-length complementary DNAs (Seki et al., 2001), or oligonucleotide-specific probes would allow greater descrimination. However, whatever the reason, these results suggest that the discrimination of biological systems should be performed at more than one level, something that is of particular importance in the establishment of substantial equivalence between transgenic and conventional crops. The second main conclusion of our work is that although the number of strong metabolite–transcript correlations was relatively small, as would be expected from previous experimental work aimed at comparing transcript and protein levels (Futcher et al., 1999; Gygi et al., 1999; Ideker et al., 2001) and mathematical studies (ter Kuile & Westerhoff, 2001), we believe they allow the generation of clear, testable hypotheses. Of particular interest in our data set was the correlation of genes with the essential amino acid lysine and with the vitamins ascorbate and tocopherol, as these linkages define candidate genes for the manipulation of the content of these nutritionally important compounds in plants. To our knowledge, only the combination of the two platforms presented here is able to provide such novel hypotheses; therefore, the pairwise correlation between transcripts and metabolites is a potentially powerful new tool for hypothesis creation in systems biology.

Methods

Supplementary Material