Amino acid substitutions in protein cores are significantly more destabilizing than substitutions on protein surfaces 29 30 . Therefore, we selected five buried residues encoding non-polar amino acids that could be mutated to polar residues with single nucleotide substitutions while maintaining a similar level of codon preference (Table 1). We used the DPX server 31 to identify buried residues of the LacZ protein based on its crystal structure (Protein Data Bank (PDB) code 1dp0). We then applied the I-Mutant2.0 algorithm 32 to confirm that the selected substitutions would be indeed destabilizing. Using site-directed mutagenesis, the five selected substitutions were introduced separately into plasmids containing lacZ under transcriptional control of the isopropyl β-D-1-thiogalactopyranoside (IPTG)-inducible lac promoter 33 . We then used a β-galactosidase assay 34 to experimentally confirm reductions in the catalytic activity of LacZ in all of the generated mutants (Table 1).

To determine whether the destabilized proteins tended to aggregate, we separated soluble proteins and proteins in inclusion bodies (see Materials and methods) and analyzed them by SDS-PAGE (Figure 1a). The three mutants with the lowest catalytic activity (F758S, I141N and G353D) were found in inclusion bodies (Table 1), the remaining two mutants (V567D and A880E) and WT proteins were found mainly in the soluble protein fraction. Next, by inspecting total cell extracts at different time points after IPTG induction, we confirmed that the total amount of protein synthesized in each mutant strain was similar to that in the WT. As shown in Figure 1b, similar amounts of LacZ are produced in the WT and either soluble (V567D) or insoluble (F758S) mutants. Quantitative analysis of the Coomasie stained bands also did not reveal any significant difference between the LacZ synthesis rates in WT and mutant strains (Figure 1c). Finally, because expression of misfolded proteins is expected to generate a heat shock response 35 36 , we used western blots to monitor the amount of the GroEL heat shock protein in induced and un-induced cells carrying WT and mutant lacZ (Figure 1d). In cells carrying WT lacZ, the concentration of GroEL increased when IPTG was added. However, in both the V567D and F758S mutants, the levels of GroEL in either induced or uninduced cells were equal or higher than that in induced WT cells.

Overall, the results described in this section demonstrate that: all engineered mutants have significantly reduced catalytic activities; soluble and insoluble mutants are expressed at the same level as WT; and the mutants induce a heat shock response, and in some cases aggregate in inclusion bodies.

The synthesis of WT or mutant β-galactosidase was initially induced by adding 10 μM IPTG. Using WT LacZ activity as a reference 37 , we estimated that about 30,000 molecules of β-galactosidase were present in each bacterial cell at this induction level. This approximately corresponds to half of the protein molecules expressed by a fully induced WT lacZ operon 34 . Cells expressing WT LacZ grew 13.5% slower on glycerol as the sole carbon source compared to uninduced cells (Figure 2a). If misfolded proteins indeed impose a significant extra cost on the bacterium, then similarly expressed mutant strains with destabilizing substitutions should lead to a more pronounced growth decrease compared to the one observed with WT LacZ. However, as shown in Figure 2a, the mutant strains grew as well as cells expresing WT LacZ, and, despite inclusion body formation, two of the mutants even grew significantly faster (see Discussion).

To further explore the potential toxicity of the destabilized proteins, we focused on two mutants (F758S and V567D). These mutants are examples of a completely aggregated and a soluble but destabilized LacZ protein, respectively. By varying the concentration of IPTG, we monitored the growth of cells with different levels of expressed LacZ proteins (Figure 2b). Importantly, no additional growth decrease was observed in the mutant strains compared to the WT at all IPTG induction levels. When no IPTG was added, resulting in a low expression level from the un-induced promoter, we also observed the same growth rate reduction in all constructs relative to cells carrying an empty pBR322 plasmid (Figure 2b).

We investigated the possibility that the toxicity of misfolded proteins was more pronounced on a relatively poor carbon source by measuring the growth of the E. coli V567D and F758S mutants and the WT on acetate. Although the overall growth rate on acetate was only about 60% of that on glycerol, we again did not observe any additional fitness (growth) decrease due to the destabilizing mutations (Figure 2c). This experiment confirmed that the observed results are not specific to a particular carbon source.

Nucleotide sequences of highly expressed genes are significantly constrained by selection for amino acid codons corresponding to abundant tRNAs 38 39 40 . A recent experimental analysis by Kudla et al. 41 suggests that non-optimal codons can directly influence E. coli growth (fitness). Using 154 variants of GFP with multiple random synonymous substitutions, these authors found a significant positive correlation between codon optimality and bacterial growth rate. An important role played by the nucleotide-level selection in evolution of E. coli proteins is also supported by a high correlation between the rates of non-synonymous (Ka) and synonymous (Ks) substitutions (Figure 3b; Spearman's rank correlation r = 0.66, P-value < 10^-10). In addition, the partial correlation between Ka and mRNA expression, controlling for Ks, is small (r = -0.14, P = 7 × 10^-9), whereas the partial correlation between Ks and expression, controlling for Ka, is significantly higher (r = -0.38, P < 10^-10).

Although selection for optimal codons at the nucleotide level should significantly affect the rates of both synonymous and non-synonymous substitutions 40 , there are additional constraints specifically acting on non-synonymous sites 42 43 . Many of these additional constraints affect the propensity of proteins to misfold and aggregate. For example, it has been reported that highly expressed E. coli proteins are more soluble than proteins with lower expression 44 45 46 . It is likely that the observed increase in solubility is necessary to avoid protein aggregation and non-functional binding 47 mediated by non-specific hydrophobic interactions. Using the genome-wide protein solubility data for E. coli proteins obtained by Niwa et al. 48 , we indeed observed a significant correlation between solubility and expression (Figure 3c; Spearman's r = 0.27, P < 10^-10). Importantly, the observed selection for solubility does not explain the correlation between the protein evolutionary rate and expression (Figure 3a; r = -0.45, P < 10^-10); the partial correlation between Ka and expression, controlling either for solubility or for the fraction of charged residues, is still significant (r = -0.42 and -0.41, respectively; P < 10^-10).

The positive correlation between solubility and expression is in agreement with an increase in the fraction of charged residues (Figure 3d; r = 0.28, P < 10^-10) and a simultaneous decrease in the fraction of hydrophobic residues (r = -0.16, P < 10^-10) in highly expressed E. coli proteins. We observed similar results by analyzing E. coli protein duplicates (paralogs) with different expression levels. By directly comparing duplicates expressed at different levels, many confounding factors, such as differences in folding topology or protein secondary structure, are removed. The analysis of 370 E. coli paralogs (see Materials and methods) demonstrated a decrease in the fraction of hydrophobic residues (paired Wilcoxon signed rank test, P = 7 × 10^-4) and a simultaneous increase in the fraction of charged residues (P = 7 × 10^-6) in the duplicates with higher expression levels.

The analysis of 602 E. coli protein structures currently available in the PDB (see Materials and methods) confirmed a significant increase in the fraction of solvent-exposed charged residues in highly expressed proteins (r = 0.18, P = 6 × 10^-6). While such an increase may lead to higher protein stabilities 49 , a proposed consequence of selection for translational robustness 22 , we did not detect strong correlations between mRNA expression and other structural features usually associated with increased protein stability 18 22 . For example, we did not observe a significant increase in the fraction of buried hydrophobic residues (r = 0.06, P = 0.13) 50 51 52 or an increase in the average number of contacts per residue (contact density) in highly expressed E. coli proteins (r = 0.02, P = 0.96). Neither did we find a decrease in the fraction of residues in loops or unstructured protein regions (r = 0.07, P = 0.06) 53 . Our analysis of experimentally determined E. coli protein stabilities assembled in the ProTherm database 54 also failed to reveal any significant correlation between protein stability, measured either by protein melting temperature (r = -0.14, P = 0.46) or folding free energy (ΔG, r = -0.08, P = 0.70), and mRNA expression level (Figure 4a,b). We also did not detect significant changes in the contact order, a structural measure strongly associated with folding speed 55 56 , in highly expressed bacterial proteins (r = -0.01, P = 0.8).

Overall, the computational analysis described above suggests that, at least based on the currently available datasets, an increase in folding speed and/or protein stability for highly expressed bacterial proteins is unlikely to play a major role in constraining the protein molecular clock in E. coli.

E. coli K12 strain GP4 (W3102, XA 21Z, lacI^q) was used in all experiments. lacZ was expressed from the IPTG-inducible lac promoter in plasmid PIV18 33 ; PIV18 is a pBR322 derivative that carries a mutation in the Shine Dalgarno sequence of the lacZ transcript that increases translation efficiency. Site directed mutagenesis was carried out using Stratagene's QuikChange Lightning kit (Stratagene, Cedar Creek, TX, USA). pBR322 was used as the empty plasmid control.

For each construct, a sweep of colonies was grown overnight on Luria-Bertani (LB) liquid media supplemented with 100 μg/ml ampicillin. Overnight cultures were diluted by a 1:100 factor and grown on M9 minimal media suplemented with 0.5% casaminoacids, 0.25 μg/ml thiamine, 100 μg/ml ampicillin and either 0.4% glycerol or acetate as carbon sources. We transfered 300 μl of cells with an OD600 of 0.5 to flasks containing 5.5 ml of prewarmed media suplemented with the appropiate amount of IPTG. Two hours after induction, OD600 was measured every 45 minutes. Growth rate was determined as the regression line slope of time and the logarithm of OD600.

The equivalent of 200 μl of cells at an OD600 of 0.7 was collected by centrifugation and lysed using Novagen's BugBuster (primary amine-free) Protein Extraction Reagent (Novagen, Merck, Darmstadt, Germany). Soluble proteins were retrieved after centrifugation of the lysed cells and aggregated proteins were then harvested following instructions for inclusion body purification described in the BugBuster reagent manual. Both fractions were saved in a 50 μl volume including 10 μl 4× SDS loading buffer, boiled, and electrophoresed on a 10% SDS polyacrylamide gel. Gels were stained with Coomassie blue and scanned for analysis. For the analysis of total protein, cells were lysed in BugBuster reagent containing rLysozyme and boiled after addition of 4× SDS loading buffer. Bands were quantified using the ImageJ program 70 .

Protein samples separated by SDS-PAGE as described above were blotted overnight onto a nitrocellulose membrane and incubated with Anti-GroEL antibody produced in rabbit 1:10 000 (Sigma Aldrich, St Louis, MO, USA). Blots were blocked with 5% non-fat dry milk, incubated with 1:3,000 anti-rabbit horseradish peroxidase conjugate antibody and visualized with Amersham's ECL Plus Western Blotting Reagent (GE Healthcare, Munich, Germany).

In the analysis we used 602 E. coli protein structures currently available in the PDB 71 . To prevent sampling biases, we filtered available PDB entries so that no two protein structures used in the calculations had sequence identity higher than 90%; similar results were obtained without filtering. We defined buried residues as those with a solvent accessible area smaller than 16% 72 73 . Solvent accessibility was calculated by the DSSP 74 program. The fraction of protein residues in loops was also calculated using DSSP. Two non-adjacent protein residues were considered to be in contact if any two of their non-hydrogen atoms were closer than 4.5 Å 75 . The protein contact density was defined as the average number of non-adjacent contacts per residue. Contact order was calculated as (L × N)^-1 × ΣΔS_ij, where N is the total number of contacts, L is the total number of residues in the protein and ΔS_ij, which is summed over all contacts, is the number of amino acids separating contacting residues 56 . In vitro solubility data for E. coli proteins was obtained directly from the study of Niwa et al. 48 .

Orthologous open reading frames and protein sequences from E. coli and Salmonella enterica were used to calculate Ks and Ka values. The E. coli-Salmonella orthologs were determined as bi-directional best hits using protein BLAST 76 . Ka and Ks values were calculated using the maximum likelihood method implemented in the PAML package 77 . The mRNA expression data reported by Lu et al. 78 were used to calculate the correlations. For the analysis of duplicated genes, we defined duplicates as pairs of E. coli proteins having more than 40% sequence identity that could be aligned for at least 80% of their total length using BLAST. In the analysis of duplicates, we used expression data from 466 experiments in the Many Microbes Microarrays Database 79 . We selected for the analysis only the pairs for which one paralog had higher expression values in more than 80% of the reported experiments.