The genetic code is degenerate; thus, protein evolution does not uniquely determine the coding sequence. One of the puzzles in evolutionary genetics is therefore to uncover evolutionary driving forces that result in specific codon choice. In many bacteria, the first 5–10 codons of protein‐coding genes are often codons that are less frequently used in the rest of the genome, an effect that has been argued to arise from selection for slowed early elongation to reduce ribosome traffic jams. However, genome analysis across many species has demonstrated that the region shows reduced mRNA folding consistent with pressure for efficient translation initiation. This raises the possibility that unusual codon usage is a side effect of selection for reduced mRNA structure. Here we discriminate between these two competing hypotheses, and show that in bacteria selection favours codons that reduce mRNA folding around the translation start, regardless of whether these codons are frequent or rare. Experiments confirm that primarily mRNA structure, and not codon usage, at the beginning of genes determines the translation rate.
Rare codons are enriched at gene start in many genomes. Genome analysis and experimental testing show that this enrichment evolved to keep the ribosome binding site free from stable mRNA structures, in order to facilitate efficient translation initiation.
The use of rare codons coincides with suppression of mRNA structures at the ribosome binding site across genomes.
There is preferential selection for synonymous codons that reduce GC‐content at the beginning of genes and a stronger pressure for rare codon usage in GC‐rich organisms.
Amino acids encoded by AU‐rich codons are preferred at gene start.
Experimental results show that mRNA structure at translation start strongly influences protein yield.
Which evolutionary constraints shape the genome of an organism? This question has been puzzling researchers for decades (Nirenberg et al, 1966; Sharp and Li, 1987; Trifonov, 1987; Herzel and Grosse, 1997; Karlin, 1998; Itzkovitz and Alon, 2007). Clearly, protein‐coding sequences and the segments regulating their expression are the main determinants of genomic sequences in bacteria. The genetic code is degenerate, allowing for diversity to encode one and the same amino acid. Except for tryptophan and methionine, amino acids are encoded by two to six different codons (Nirenberg et al, 1966). Therefore, it is important to understand which mechanisms shape codon choice.
On a genome‐wide scale some codons are preferred to others, termed codon usage bias (Grantham et al, 1980; Sharp and Li, 1987). Each genome shows a specific bias, although the origin of it remains unclear (Hershberg and Petrov, 2008). Driving forces for preferentially selecting specific codons (Plotkin and Kudla, 2011) could be the efficiency and accuracy of translation (Trifonov, 1987; Drummond and Wilke, 2008; Cannarozzi et al, 2010; Warnecke and Hurst, 2010; Tuller et al, 2010a; Chu et al, 2011; Gingold and Pilpel, 2011; Shah and Gilchrist, 2011; Chu and von der Haar, 2012), GC content (Lynn et al, 2002), environmental factors (Lynn et al, 2002; Singer and Hickey, 2003), or DNA folding (Herzel et al, 1999). The species‐specific codon bias mirrors the amount of the cognate transfer RNAs (tRNAs), i.e., highly used triplets are read by high‐abundance tRNAs and rarely used codons pair to low‐abundance tRNA (Ikemura, 1981; Dong et al, 1996). Triplets read by major tRNAs are translated faster than the codons read by minor tRNAs (Zhang and Ignatova, 2009). Highly expressed genes thus display a stronger bias towards usage of abundant codons reflecting tRNA availability (Sharp and Li, 1987; Kudla et al, 2009; Tuller et al, 2010b).
Within each gene, rare codons are far from being randomly distributed along the coding mRNA sequences; they tend to cluster and transiently attenuate ribosomal traffic, an effect suggested to synchronize translation with cotranslational folding of multidomain proteins (Komar, 2009; Zhang et al, 2009). Furthermore, rare codons may exert regulatory function under starvation (Elf et al, 2003; Zhang et al, 2010), but may also cause misfolding in highly expressed genes, and thus are avoided (Warnecke and Hurst, 2010).
Genome‐wide analysis revealed that codon usage of the protein‐coding sequence differs between the beginning of the gene and the rest of the mRNA sequences (Eyre‐Walker and Bulmer, 1993; Tuller et al, 2010a). This suggests that there are different evolutionary pressures on codon usage at the beginning of genes than in the rest of the genome. An intriguing hypothesis is that these codons have been selected to reduce elongation speed at the beginning of genes (Tuller et al, 2010a). Such a ‘ramp’ in elongation speed could reduce the likelihood of ribosomal traffic jams along the mRNA, and may be advantageous as such traffic jams may result in premature termination or in protein bursts (Dobrzynski and Bruggeman, 2009). However, other studies have demonstrated that the same region is selected for reduced mRNA folding (Gu et al, 2010; Keller et al, 2012), required for efficient translation initiation (McCarthy and Bokelmann, 1988; de Smit and van Duin, 1990; Kudla et al, 2009; Kertesz et al, 2010).
This finding raises the intriguing possibility that the abnormal codon usage at the beginning of genes is not selected for to reduce elongation speed (‘ramp hypothesis’), but is a side effect of selection for reduced stability (‘structure hypothesis’). To address this fundamental question we analysed 414 bacterial genomes. We found that deviation of codon usage strongly correlates with the suppression of mRNA secondary structure at gene start. Furthermore, we found that rare codons are only selected if they are AU rich, whereas abundant codons are repressed if they are GC rich. Thus, our genome analysis shows that rare codons are not selected because they are rare, but to weaken the mRNA structure. The hypothesis is further corroborated by experimental measurements of the translational efficiency of two Escherichia coli genes with various synonymous starting sequences, underpinning the functional relevance.
Unusual codon usage and reduced folding around the translation start site
To analyse codon usage and mRNA folding energy around the translation start site, we collected a panel of 414 bacterial genomes from the BioCyc database (Keseler et al, 2011). We aligned protein‐coding genes with respect to the translation start, and removed the first genes of transcriptional units (TUs), to avoid spurious signals due to misannotated untranslated regions (UTRs). We then used Kullback–Leibler divergence (KLD) to calculate how strongly codon frequency deviates at each position from the overall codon usage (for details, see Materials and methods). For example, in E. coli the first eight codons show higher values of KLD than a null model (Figure 1A, and Supplementary Figure S1a and b), indicating an unusual codon usage in that region, which is consistent with previous findings (Tuller et al, 2010a).
In addition, we calculated folding energy of 39‐nucleotide (nt) long stretches around each position (Hofacker, 2009), which corresponds to the length of a ribosome binding site (Beyer et al, 1994). This allowed us to quantify typical mRNA secondary structure in a position‐specific manner. Typical folding profiles showed reduced mRNA stability and less‐paired nucleotides around the translation start site (Figure 1B, and Supplementary Figures S1c, d and S2 for E. coli), similar to previous observations (Kudla et al, 2009; Gu et al, 2010).
To compare the deviation of codon usage and reduced mRNA folding between different organisms, we defined two scores: average deviation in the codon usage within the first five codons (ΔCU) and change in folding energy at translation start when compared with typical mRNA folding elsewhere in coding regions (ΔG, see Materials and methods). Clearly, both scores varied strongly among the organisms. The correlation between these scores with genomic features confirmed that suppression of the mRNA structure occurs primarily in GC‐rich organisms (Figure 1C; Gu et al, 2010). This is consistent with a stronger pressure to reduce the secondary structure propensity in GC‐rich organisms, where mRNA sequences tend to form more stable secondary structures due to the GC content. More surprisingly, we also found unusual codon usage primarily in GC‐rich organisms (Figure 1D). Rare codons typically are more slowly elongated, irrespective of GC content (compare Supplementary Figure S3 for an analysis of the tAI scores, a commonly used measure for elongation speed based on tRNA gene copy number (dos Reis et al, 2004)). This suggests that attenuation of elongation occurs primarily in GC‐rich organisms. According to the ‘ramp hypothesis’, this would imply that different bacteria would be subject to different pressures to prevent ribosome jamming, and this pressure would be correlated with GC content. Alternatively, codon usage deviates as a side effect of another selective pressure. As the mRNA regions with unusual codon usage and suppressed mRNA structure overlap (Figures 1A and B, and Supplementary Figure S1), the structural constraints required for efficient initiation is one of the prime candidates to exert such a selective pressure. In line with this ‘structure hypothesis’, we find that genomes with suppressed structure show also deviations in codon usage, and the two scores correlate strongly across genomes (r=0.93; Figure 1E, and Supplementary Figures S4 and S5).
AU‐rich codons are selected at the beginning of coding regions
Mechanistically, the choice of specific codons could suppress secondary structure by reducing the GC content at the beginning of genes. Indeed, the GC content in E. coli is strongly reduced within the first four to six codons (Figure 2A and Supplementary Figure S6a, dashed line). Furthermore, we observed a strong decrease in the GC content in the third base (GC3 content; Figure 2A and Supplementary Figure S6a, solid line). As synonymous codons differ mainly in their third base, this suggests a preferential selection of codons with A or U at the third position (AU3) to reduce folding propensity. A null model with randomly shuffled synonymous codons (SSC) within each gene (Figure 2A and Supplementary Figure S6a, dotted line) also shows a slight reduction in the GC3 content. This implies that amino acid selection is also biased towards higher AU3 content, albeit the major influence on GC3 is due to the preferential choice of AU3 codons. We also observed a general decrease in the GC content at the first and second nucleotide position (GC1 and GC2, respectively) within the first four to six codons (Figure 2B and Supplementary Figure S6b, solid lines). A similar trend is also detectable with the null models with conserved amino acid sequences (SSC), but almost vanishes if all codons are randomly shuffled (shuffled codons—SC). Taken together, these results suggest that amino acids encoded by AU‐rich codons are chosen preferentially at the gene start (Figure 2B and Supplementary Figure S6b).
Properties of rare and abundant codons
The reduction of GC content at the beginning of genes might be a consequence of selection of slowly translated and consequently of rare codons. To investigate this, we analysed the properties of rare and abundant codons across many bacterial species. For each genome, we defined subsets of the 15 codons with lowest and highest abundance, and termed these the sets of rare and abundant codons, respectively. With this definition, the set of rare codons typically contained 5% of all codons of an organism, whereas on average 50% belonged to the set of abundant codons (Figure 3A). Intriguingly, we observed that the GC content strongly determines the overall codon distribution. In bacteria with a GC content of 0.5, the abundant codons were about fivefold more abundant in the genome than the rare codons. These differences were more pronounced in bacteria with more extreme GC content, with up to 100‐fold difference in frequency (Figure 3B). This dependence is symmetric, i.e., GC‐rich and AT‐rich genomes show the same disparity.
Furthermore, the GC content of an organism is also reflected in the GC content of rare and abundant codons in a particular genome (Figures 3C and D). In AT‐rich organisms, such as Thermoanaerobacter tengcongensis, rare codons are GC rich and abundant codons are AU rich. Similarly, in GC‐rich organisms, such as Aeromonas hydrophila, AU‐rich codons are enriched among the rare codons and abundant codons show high GC content. In contrast, bacteria with intermediate GC content show no particular selection for GC‐rich or GC‐poor codons. Thus, in organisms with GC content of around 0.5, rare codons are not per se biased towards a particular GC content, and a selective pressure for rare codons is unlikely to strongly influence the GC content. Most importantly, in the context of selection of alternative codons the GC3 content of codons in the sets of rare and abundant codons is primarily determined by the GC content of the organism (Figure 3E). In GC‐rich organism, virtually all abundant codons show a G or C at the third position, whereas virtually all rare codons end with an A or a U.
Similarly, when we use tAI scores to classify codons according to their translation elongation speed, we find that in GC‐rich organisms slow codons are AU3 rich, whereas fast codons are GC3 rich (Supplementary Figure S7).
Rare codons that reduce GC content are preferentially selected in E. coli
In E. coli, rare codons are enriched at the beginning of genes and abundant codons show a slight decrease (Figure 4 and Supplementary Figure S8). We next asked whether rare codons are selected per se or because of their GC content. We divided the sets of rare and abundant codons in two subsets: those with G or C at the third position (GC3) and those with an A or U (AU3). If there is a selective pressure to increase the frequency of rare codons downstream of the translation start to slow down early elongation (‘ramp hypothesis’), we expected an increase of rare codons irrespective of their GC3 content. In contrast, if codons are chosen in order to weaken the mRNA structure we expected an asymmetry between GC3 and AU3 codons. Within the set of rare codons enriched at the start of genes in E. coli, we detected a clear asymmetry between AU3 and GC3 codons in the set of rare codons (see Figure 4A and Supplementary Figure S8a). Rare AU3 codons are enriched at the beginning of genes, whereas rare GC3 codons are not effected in their frequency, i.e., they remain rare.
We also noted a strong asymmetry between GC3 and AU3 abundant codons: abundant GC3 codons are strongly depleted at the beginning of genes, whereas abundant AU3 codons are slightly enriched (Figure 4B and Supplementary Figure S8b). Interestingly, also the null model with SSC shows a drop of GC3 and an increase of AU3 codons (grey lines in Figure 4B and Supplementary Figure S8b). This implies that amino acids preferentially encoded by AU3 codons are enriched and those encoded by GC3 codons are avoided, suggesting a strong evolutionary pressure on GC content that manifests itself in the amino acid sequence. When we used tAI scores to classify codons as slow and fast, we observed an enrichment of slow and a partial depletion of fast codons (Supplementary Figures S9 and S10). Again, in both groups the enrichment depends on the third base of these codons: the frequency of AU3 codons increases, whereas usage of GC3 codons drops. Taken together, our analysis shows that in E. coli rare codons are selected because of their GC content and not because they are rare and slowly translated, supporting the ‘structure hypothesis’.
Widespread selection for low GC content at gene start
We next asked whether the asymmetric enrichment of GC‐poor codons at the start of coding sequences is also detectable in other genomes. Bacteria largely differ in their GC content and, therefore, we first grouped them according to the GC3 content in their sets of rare or abundant codons. We observed that abundant codons are depleted only if the fraction of GC3 codons in that subset is above 50%, and rare codons are enriched when they have on average a AU3 content below 50% (Figure 5A and Supplementary Figure S11a). This finding further supports the ‘structure hypothesis’. Under the assumption that the ‘structure hypothesis’ dictates codon usage, we also expect that enrichment of rare codons depends on the GC content of the genome, because (i) mRNAs tend to fold more strongly in GC‐rich organisms and (ii) the set of rare codons is enriched of AU‐rich codons. Indeed, rare codons are only enriched and abundant codons depleted in genomes with a GC content >0.5 (Figure 5B and Supplementary Figure S11b). When we use tAI scores to classify codons with respect to their elongation speed, we find the same result: slow codons are only enriched at the beginning of genes if they are AU3 rich and fast codons are depleted if they are GC3 rich (Supplementary Figure S12).
In addition, the ‘structure hypothesis’ predicts an asymmetry in GC3 content such that GC3 content at the beginning of genes is always reduced, with stronger depletion for genomes with higher GC content. In the 414 analysed genomes, the GC3 content behaves strongly asymmetric. We observed a reduction in GC3 content proximal to the start codon in most genomes (Figure 5C and Supplementary Figure S13). Importantly, the reduction in the local GC3 content is stronger for genomes with a higher GC content, suggesting that a reduction in the GC content is the main driving force behind the unusual codon bias just downstream of the start codon.
Folding of mRNA at gene start strongly influences translation
We interpreted the suppression of mRNA folding around translation start as a mechanism that facilitates ribosome binding and translation initiation. Moreover, the ‘ramp hypothesis’ predicts that rare codons at the beginning of genes will have an impact on translation efficiency. To discriminate between these two hypotheses, we experimentally dissected the role of the codon usage and mRNA secondary structure on translation efficiency. We selected two genes from E. coli (pykA and ypdE) for which codon choice allows to vary folding energy and codon adaptation independently (Supplementary Figure S14). To investigate the effect of mRNA secondary structure, we chose synonymous codons such that the mRNA structure differs strongly in folding energy (minG and maxG for minimal and maximal folding energy, respectively), but the sequences have similar codon adaptation as the native sequence (wild‐type (wt)). Analogously, we studied the effect of codon usage by designing sequences that differ in the codon adaptation but have similar estimated folding energy (minCA and maxCA, for minimal and maximal codon adaptation, respectively). These sequences were fused upstream to a yellow fluorescent protein (Figure 6A), expressed in E. coli, and the translation efficiency was assessed by quantitative measurements of mRNA and YFP fluorescence levels. Alterations in codon usage or secondary structures had no influence on the mRNA abundance as confirmed by qRT–PCR (Supplementary Figure S15). In contrast, the protein yields differed largely (Figures 6B and C, and Supplementary Figure S16). Notably, protein expression between the constructs differed upto 20‐ and 60‐fold for ypdE and pykA, respectively. In line with the structure hypotheses, we found that the constructs with the minimal folding energy (i.e., the strongest secondary structure) showed by far the lowest expression, whereas constructs with the maximal folding energy yielded the highest protein expression (Figures 6B and C). Codon usage also influenced protein expression, but the effects were much weaker and rather inconsistent. For ypdE, the constructs with the minCA and maxCA had a reduced protein expression (Figure 6B). In contrast, pykA showed a slightly enhanced expression for minCA and a decreased expression for maxCA (Figure 6C). The induction of pykA resulted in very high protein levels, which may impair cellular physiology in general. We thus measured the protein levels also in the absence of the inducer β‐D‐thiogalactoside (IPTG). Although the strong influence of the folding energy on protein level remained, the construct with minCA showed no difference to the wt construct (Figure 6C). Taken together, these results clearly suggest that mRNA structure determines the translation yield, whereas the effects of codon usage are less obvious.
Clearly, synonymous codons are not chosen randomly and here we present a new aspect that drives the selection of unusual codons. In many genomes, the frequency of synonymous codons at the beginning of genes differs from the overall codon usage in the genome. This has led to the ‘ramp hypothesis’, suggesting that rare codons at the beginning of genes were selected to slow down initial elongation, which may help to avoid traffic jams during translation (Tuller et al, 2010a). However, as the same region has been shown to be under selective pressure for mRNA structure, we investigated whether the unusual codon usage is mainly a side effect of structural constraints (‘structure hypothesis’). Such evolutionary pressure is likely to be strong as secondary structures at the ribosome binding region can reduce or even terminate translation initiation (de Smit and van Duin, 1990; Kudla et al, 2009). Both hypotheses make divergent predictions, most importantly regarding the use of AU‐rich and GC‐rich codons, allowing us to distinguish between these two hypotheses. Our bioinformatics analysis of many bacterial species with varying GC content showed that codons at the beginning of genes are selected that effectively decrease the propensity of mRNAs to form secondary structure around the ribosome binding site.
We also experimentally show that changing the folding energy while keeping the same codon usage at the beginning of native E. coli genes markedly affects translation efficiency. In contrast, alterations of the codon usage at constant folding energy had mild but inconsistent effects on protein yield. This suggests that translation efficiency is strongly modulated by the folding energy at translation start.
The need to disfavour strong mRNA secondary structure formation around the translation initiation site depends on GC content of the genome: if the GC content is higher, the mRNAs tend to form more stable structures and, therefore, the pressure to avoid these is stronger. As suppression of mRNA folding is primarily determined by decreasing the GC content, we observe an enrichment of AU‐rich codons at the beginning of genes. This, however, also explains why rare codons are used more often in this region: with increasing genomic GC content the frequency of AU‐rich codons throughout the genome drops, i.e., they become rare. As a consequence, we observe the strong correlation between unusual codon usage and suppression of mRNA folding. However, rare codons at the gene start lead to slower elongation, as codon frequency correlates with tRNA abundance. Indeed, in many organisms elongation speed may be slower at the beginning of genes, such as that in yeast (Ingolia et al, 2009). Our results indicate that this ramp in speed is most likely a consequence of the need to suppress mRNA structure and the resulting use of rare codons. Furthermore, our analysis suggests that the evolutionary pressure to keep the ribosome binding regions free of structure is very strong, and may even impact protein sequence. For example, the N‐terminal regions of E. coli proteins show an enrichment of amino acids encoded by codons with A or U at the first or second nucleotide position.
Structural constraints for efficient initiation also extend to higher eukaryotic organisms: shorter genes show less structure at the ribosomal binding sites, which results in higher translation rates (Ding et al, 2012). The role of codon usage in defining mRNA structure might not only be restricted to the translation start but also codon choice along the coding sequences may be shaped by various requirements to favour or disfavour specific secondary structures. For example, mRNA stability, micro‐RNA‐binding or RNA‐binding proteins may require certain structures, which would impact codon choice. Thus, evolution has to solve a multi‐dimensional problem: although efficient elongation and error reduction require the usage of abundant codons, certain structural requirements for the mRNA may need infrequent codons (Komar, 2009; Zhang et al, 2009). Thus, codon usage is shaped by many possibly conflicting constraints, which we are just beginning to understand.
Materials and methods
Genome sequences and annotation was collected from EcoCyc database (Keseler et al, 2011) version 13.6 for E. coli, and for the other 414 bacterial genomes from the BioCyc database (Karp et al, 2005) version 13.6, Tier3. We excluded the first gene of each TU, non‐chromosomal, non‐protein‐coding sequences and splicable genes (genes with programmed frame shifts). Furthermore, we removed TUs containing genes of nucleotide length, which are not multiple of 3 or which contained components that were incompletely annotated. Genes present in multiple TUs were assigned to the largest TU. Taxonomy was annotated using the NCBI taxonomy IDs and the Perl module Bio‐LITE‐Taxonomy‐NCBI‐0.08.
We used two null models: (i) SC, where codons were randomly permuted within each gene, and (ii) SSC, which preserves the amino acid sequence by shuffling only synonymous codons within genes. Start and stop codons, codons with sequencing errors (only in the case of SSC), and overlapping sequences were not shuffled.
Calculation of folding energy
We used the Vienna RNA Package, version 1.8.5, available at http://www.tbi.univie.ac.at/%7Eivo/RNA/, to predict free energy of RNA sequences (Hofacker, 2009). Gibbs free energy was calculated within a sliding window of 39 nts that correspond to the approximate number of nucleotides covered by a ribosome (Beyer et al, 1994), and this value was assigned to the nucleotide position of the window centre. Using this method, we defined suppression of mRNA structure around the start codon (ΔG=G0–Gb1) as the difference between average folding energy G0 within a −5 to +5 nt window around gene start. The baseline folding energy Gb1 estimated as the average within a 50‐nt window from nucleotide position +150 to +199.
Codon frequencies and KLD
For each set of synonymous codons, we determined the genome‐wide frequency qi,j of each codon within this set, where i=1…20 indicates the amino acid and j=1…Si indexes the synonymous codon (where Si is number of synonymous codons). In addition, we defined the position‐dependent codon frequencies, pi,j(k), for each codon position k relative to the translation start site (start codon is at k=0). In both cases, pseudocount regularization with a pseudocount of 1 is applied. The position‐dependent KLD(k) that quantifies the deviation of the codon usage at each position k is then calculated as:
Because of finite size effects, the KLD(k) is biased to values larger than 0 even if pi,j and qi,j stem from the same distribution (Herzel and Grosse, 1997; Roulston, 1999). The bias due to this finite size sampling was estimated using the SSC null model.
Enrichment of extreme codons
We define abundant and rare codons as the 15 most abundant and 15 most rare codons in each genome, measured by their codon frequencies. The total frequencies of rare frare and abundant fabund codons are the sum over the codon frequencies in the corresponding set, e.g. where fi denotes the frequency of the ith codon and the same definition applies for abundant codons. For each position k, we defined the fold change fc(k) of the rare codon frequency as: and, correspondingly, for abundant codons. Enrichment of rare and abundant codons at the beginning of genes was determined by averaging the fold change from position 1 through 5 () and subsequent calculation of Z‐values: where the estimate of the mean and variance were calculated using n=100 instances of the SSC null model.
Enrichment of slow and abundant codons was assessed analogously.
Evaluation of synonymous sequences
For each cytoplasmic protein, we generated in silico all synonymous sequences that differ in the first six codons. Only the beginning of the genes, i.e., 5′‐UTR and 31 codons, including ATG, was taken into account. To select appropriate genes to address codon usage and folding energy independently, we judged each synonymous sequence according to the following measures. We calculated the mean folding energies around the translation start (i.e., folding energies of 39‐nt long stretches centred at nucleotide positions 0, ±3 and ±6 relative to the translation start). Codon adaptation was estimated by determining the geometric mean of the relative tRNA abundance corresponding to codons from position 1–6 (Dong et al, 1996; Zhang et al, 2009; Zhang and Ignatova, 2009). The folding energy and codon usage profiles for the selected genes ypdE and pykA are shown in Supplementary Figure S14.
Plasmids and strains
We created five constructs (wt, minCA, maxCA, minG and maxG) for each gene with (i) a 50‐nt long native 5′‐UTR in which additional ATG codons were removed to avoid alternative start sites, (ii) the start codon ATG, (iii) the stretch of variable codons, (iv) a sequence of 24 unaltered codons, followed by three codons coding for glycine (Figure 6A, and Supplementary Tables S1 and S2). All constructs were synthesized de novo (Eurofins MWG Operon) and subcloned into pETDuet‐1‐yfp vector (a kind gift from IM Axmann) upstream of the yfp sequence. All plasmids carry a T7 promoter inducible by IPTG and an ampicilin resistance for selection. Plasmids were expressed in E. coli BL21(D3) strain.
Overnight cultures were grown in Luria Bertani broth (LB) containing ampicilin (100 μg/ml) at 37 °C. For measurements of YFP expression, overnight cultures were diluted 1:50 in 20 ml fresh LB medium containing ampicilin. Cells were induced at OD600∼0.5(±0.1) with IPTG at a final concentration of 45 μM. One hour after induction, cell cultures were diluted 1:5 in fresh LB medium containing ampicilin and IPTG, grown for another hour, collected by centrifugation and resuspended in tethering buffer (5 mM K2HPO4, 5 mM KH2PO4, 0.1 mM EDTA, 1 μM l‐methionine, 0.1% (v/v) lactic acid (pH 7)). The protein expression was analysed by flow cytometry and mRNA was quantified by qRT–PCR.
Quantification of gene expression
Median expression levels of fluorescent proteins were quantified in a population of approximately 105 cells by flow cytometry on a FACSCalibur (BD Biosciences) equipped with an argon 488‐nm laser. A measurement was triggered by forward scatter (fsc) and sideward scatter (ssc) events. FACSCalibur data files were imported for analysis into MATLAB using fca_readfcs.m (developed by Laszlo Balkay, University of Debrecen, downloaded from Matlab Central File Exchange, File ID: 9608). Only non‐zero measurements were taken into account for fsc and ssc values between the 10th and 90th percentile. The median value of the autofluorescence background, measured for control cells transformed with an empty pETDuet‐1 vector, was subtracted from all values. Measured fluorescence distributions are shown in Supplementary Figure S16.
Total RNA was isolated using the InviTrap Spin Cell RNA Mini Kit (Stratec Molecular) following the protocol for Gram‐negative bacteria. The RNA quality was judged by the absorbance ratio A260nm/A280nm. Samples were then treated with DNAseI (Fermentas) for 1 h at 37 °C to remove remaining DNA traces and the enzyme was inactivated by adding 50 mM EDTA and incubated at 65 °C for 10 min.
cDNA was produced from total RNA using Reverse Transcription Kit (Fermentas) with random hexamer primer. For each sample, a control was performed without Revert Aid Transcriptase. mRNA was quantified by qRT–PCR with gene‐specific primers on a real‐time PCR cycler (7500 Fast Real‐Time PCR System, Applied Biosystems) using Fast SYBR Green Master Mix (Applied Biosystems). The level of yfp mRNA was normalized to the level of gapdh mRNA as an internal standard. Two clear outliers were removed from the analysis.
We thank Naama Barkai and Gong Zhang for stimulating discussions; Melanie Anding for her help with the experiments; Hanspeter Herzel, Debora Marks, Ralf Steuer, Johannes Meisig and Adam Wilkins for commenting on the manuscript. This work was supported by the European Commission (BACTOCOM), Deutsche Forschungsgemeinschaft (DFG, through SFB 618, project A3 and SPP 1395/InKoMBio) and BMBF (FORSYS) to NB and ITN NICHE to ZI.
Author contributions: NB conceived the study; KB and NB designed the computational analysis; ZI, KB and NB designed the experiments; KB carried out the wet lab experiments with PS and RR, and performed the computational experiments. KB and NB wrote the manuscript with input from ZI.
Conflict of Interest
The authors declare that they have no conflict of interest.
This is an open‐access article distributed under the terms of the Creative Commons Attribution License, which permits distribution, and reproduction in any medium, provided the original author and source are credited. This license does not permit commercial exploitation without specific permission.
- Copyright © 2013 EMBO and Macmillan Publishers Limited