Genome-wide transcriptional analyses have until now focused almost exclusively on differential expression. Methods that have been developed to quantify absolute mRNA abundances have largely been ignored or proved impractical. However, the use of gDNA as a reference channel in traditional microarray experiments is increasingly common 4, and as this is equivalent to an equimolar concentration of all transcripts we have investigated using this as a normaliser that would allow us to generate a measure of genome wide mRNA abundances.

Initially we calculated relative mRNA abundances for M. tuberculosis growing aerobically in chemostat cultures as we had access to the RNA used for hybridization to the arrays, allowing us to experimentally validate our analysis. RNA and microarray procedures were carried out as described in the Materials and methods and 20. To obtain a measure of mRNA abundance, we first removed the local background fluorescence from each probe spot on the microarray. The fluorescence intensity from the RNA channel was then normalized to that of the gDNA channel, after which the technical and biological replicates were averaged.

We then found it necessary to correct for a probe length effect present in the data. In a traditional microarray (type 1) experiment, comparisons are made between two cDNA populations hybridized to one spot. Although in most cases it is necessary to control for factors such as the spatial dependent effects of hybridization, and normalizations such as loess are routinely implemented for this purpose 21, it is not necessary to control for individual spot variations as these would be negated when calculating fold expression ratios. In our case, where we are attempting to draw comparison between signals generated from different spots on the array, we are faced with additional factors that could introduce bias to our results - those that differ between spots on the array, for example, the probe GC content and length. Using the signal reported from a gDNA hybridization, we investigated these factors. We found that the length of the probes, which ranged in length from 60 to 1,000 bp, affected the hybridization whereby longer probes report higher signal intensities (Figure 1a). We suspect this may be because larger probes are able to bind more DNA and bind it more strongly than shorter probes. We corrected for this effect using a model of linear regression (Figure 1b; and see Materials and methods).

Finally, as the sum of all fluorescence intensities is equal to the sum of all labeled mRNA, the measures of mRNA abundance were converted from unintuitive ratios to a proportional value; parts per million (ppm). Table 1 shows the mRNA abundance values for the 50 most highly expressed genes of M. tuberculosis. Figure 1c shows the mRNA levels, in ppm, for each gene in the M. tuberculosis chromosome and their log distribution is shown in Figure 1d. It is clear that the mRNA abundances, even once log transformed, are not normally distributed, which reflects the observations of others 22.

In order to validate our estimates of abundance, we performed quantitative reverse transcriptase PCR (RTq-PCR) on a large selection of genes that we had predicted to span the mRNA abundance spectrum (n = 24). The RTq-PCR was carried out on a sample of the same RNA used for the microarray hybridizations. The measures of mRNA abundance as predicted from the microarray analysis show a good correlation (Spearman's rank = 0.86, p < 0.0001) with the absolute copy number as determined by RTq-PCR data (Figure 2).

Further validation of the method was provided by its high reproducibility when applied to data sets from independent laboratories using the same microarray designs. Correlations were determined for a variety of mRNA abundance data from both M. tuberculosis and the highly similar Mycobacterium bovis 23. Chemostat-grown M. tuberculosis showed a correlation of 0.8 (p < 0.0001) with the homologous genes in chemostat-grown M. bovis. The same was true of batch-cultured M. tuberculosis and M. bovis grown in different institutions. However, there was a lower correlation of 0.5 (p < 0.0001) between chemostat and batch cultured M. tuberculosis from different laboratories, suggesting that the method of culture significantly affects the transcriptome.

To explore the possibility that global measures of mRNA abundance are an important indicator of prokaryotic biology, we compared our mRNA abundance data with proteome and gene essentiality data. Demonstrating a correlation between mRNA and protein levels is difficult without the availability of genome-wide measures of protein abundance. We looked instead at the list of M. tuberculosis proteins identified to date from two-dimensional PAGE analysis and stored in an online database 24. As the mRNA abundances are not normally distributed, we determined the frequency of identified proteins in each quartile of the abundance distribution. Of the 283 unique proteins identified in M. tuberculosis cell lysates and supernatants, the majority (187 proteins, 66%) are expressed in the most abundant quartile (Figure 3a). Others have suggested that proteomic experiments have an intrinsic bias toward abundant proteins 1225 and this would support our hypothesis that the most abundant transcripts produce high levels of protein in bacteria. In addition, our analysis makes no allowance for differential rates of translation initiation or mRNA/protein degradation, so this finding reflects how tightly coupled the systems of transcription and translation are in prokaryotes 910.

As there is little correlation between reports of biological importance and gene induction 678 we instead looked at the correlation with mRNA abundance. We compared the genome-wide values of mRNA abundance with the genes identified as being essential in M. tuberculosis by genome-wide transposon mutant library (TraSH) screens 72627. For our RTq-PCR validated data from M. tuberculosis growing under aerobic chemostat conditions we found that there is a significant relationship between expression level and essentiality on a global scale (Chi-squared test for trend in proportions = 161.2, df = 1, p value < 0.0001; Figure 3b). This is in contrast to the lack of correlation with fold-induction upon infection 67 and illustrates the potential importance of determining mRNA abundances on a global scale. The correlation may reflect our previous finding that the more highly expressed a gene, the more protein is produced. Although there are obvious examples where proteins with essential functions, such as the cell division apparatus or many enzymes, need not be (or indeed cannot be) highly expressed 28, prokaryotic cells would waste considerable energy synthesizing large amounts of proteins that do not have essential functions.

The genome-wide distribution of mRNA levels from M. tuberculosis cultured aerobically in the chemostat is typical of the distributions we have found (Figure 1d). Despite being log transformed, the distribution shows a skew to the right, suggesting the presence of a highly expressed subpopulation. As we and others have shown, the coupling of transcription and translation in prokaryotes means that there is an enormous material investment in transcribing a gene at a high level. We therefore analyzed the most abundant mRNAs in some detail, focusing on the 95th percentile of genes, that is, the 5% most abundant transcripts, n = 198 (Additional data file 1). Of these 198 genes, 89 (45%) were reported as essential in the TraSH experiments, which is significantly higher than the 23% of all genes that are essential (Χ² p < 0.001). Of the 89 essential genes, 76 (38%) were essential in vitro, 11 (5%) in vivo and 2 (1%) are essential for survival in macrophages 72627.

The most highly expressed gene in vitro was ppiA (Rv0009), a probable peptidyl-prolyl cis-trans isomerase involved in protein folding, which makes up 13,634 ppm (which is equivalent to 1.3%) of the total mRNA population. As would be expected, many of the very abundant transcripts belong to the protein synthesis machinery, including thirty ribosomal proteins, six translation initiation factors and various components of RNA polymerase.

Several of the very abundant genes have previously been characterized as highly expressed and extensively documented as important virulence determinants of M. tuberculosis. In particular, some members of the esx gene family have been shown to be critical in infection, although dispensable in vitro. The paradigms for this family are esat6 (Rv3875) and cfp10 (Rv3874), whose products form a secreted complex that interacts with host cells 29. Furthermore, they are potent immunogens with potential roles as both subunit vaccines and diagnostic agents 3031. The esx family in M. tuberculosis contains 23 esat6-like genes, with 11 esat/cfp gene pairs 32. Including esat6 and cfp10, we identified 12 members (52%) of this family as being amongst the most highly transcribed of all genes. One such pair of genes, esxV (Rv3619c) and esxW (Rv3620c), are adjacent to, but not transcribed with, the SNM (secretion in mycobacteria) operon containing Rv3616c (espA), Rv3615c (snm9) and Rv3614c (snm10), which we also find are very highly expressed. Two groups have recently shown that the SNM system functions to export both ESAT-6 and CFP-10 3334.

We have also observed five highly expressed transcripts belonging to the PE/PPE family; a set of approximately 100 genes encoding proteins with proline and glutamate rich motifs that are found exclusively within the mycobacteria 35. Some members of this family are located adjacent to esx genes, suggesting a functional association, and it is now known that a PE/PPE pair form a stable dimer, reminiscent of ESAT-6/CFP-10 themselves 36. Of the five PE/PPE genes in our very abundant transcript list, two are located adjacent to highly expressed esx family gene pairs: PPE18 (Rv1196) with esxKL (Rv1197/Rv1198), and PE19 (Rv1791) with esxMN (Rv1792/Rv1793). The significance of linked high expression levels between esx and PE/PPE genes suggests a co-functionality critical to M. tuberculosis biology.

Thirty-three percent of the coding sequences in M. tuberculosis were classified as having no known function in a re-annotation of the genome 37. A similar proportion (60 of 198, 30%) of the transcripts we have classified as very abundant in M. tuberculosis are annotated as unknown, hypothetical or conserved hypothetical proteins. The organism consumes considerable energy in their expression so it is likely they have important functions, and indeed 12 of these unknown genes are essential. Using both BLASTP and functional predictions generated with the hidden Markov model profile tool SHARKhunt 38, we were unable to ascribe any further functions to these genes with the exception of Rv0097, which has close homology to a taurine dioxygenase of the Streptomyces, Ralstonia and Chromobacterium species. It has been reported 39 that the Rv0097 homologue in M. bovis is located in an operon essential for the synthesis of the virulence associated lipids phthiocerol dimycocerosate esters (PDIMs) and could function as an α-ketoglutarate-dependent dioxygenase, the super family to which taurine dioxygenases belong.

Much effort goes into looking for genes whose expression is modulated in different environments. Although it is likely that most, if not all, genes are regulated to some extent, using the analysis described here it is possible to search for genes whose expression does not change significantly, which we term 'invariant'. These may represent genes whose functions are so important that they cannot be switched off. To identify these invariant genes, we compared the mRNA abundance in a total of six data sets including various growth conditions, such as chemostat, batch culture, low oxygen and growth in macrophages. We focused on genes that were within the 85th percentile and found that 133 genes were consistently highly expressed across all of the conditions tested (Table 2).