Gene Architectures that Minimize Cost of Gene Expression

Abstract

Gene expression burdens cells by consuming resources and energy. While numerous studies have investigated regulation of expression level, little is known about gene design elements that govern expression costs. Here, we ask how cells minimize production costs while maintaining a given protein expression level and whether there are gene architectures that optimize this process. We measured fitness of ∼14,000 E. coli strains, each expressing a reporter gene with a unique 5' architecture. By comparing cost-effective and ineffective architectures, we found that cost per protein molecule could be minimized by lowering transcription levels, regulating translation speeds, and utilizing amino acids that are cheap to synthesize and that are less hydrophobic. We then examined natural E. coli genes and found that highly expressed genes have evolved more forcefully to minimize costs associated with their expression. Our study thus elucidates gene design elements that improve the economy of protein expression in natural and heterologous systems.

Keywords: expression cost; gene expression; genome evolution; optimal gene architecture; synthetic biology; synthetic library; systems biology; translation efficiency.

Figure 1. 5′ Gene Architectures Affect Cost of Gene Expression at a Given Expression Level

(A) We utilized a synthetic library of ~14,000 E. coli strains, each expressing a GFP construct with a unique 5′ architecture that includes a promoter, ribosome binding site (RBS), and an 11-amino-acid-fused peptide. There were two different promoter types, four RBSs, and 137 amino acid fusions that were each synonymously re-coded to 13 different versions (see Goodman et al., 2013 for full details). (B) FitSeq methodology to measure relative fitness of strains in a pooled synthetic library. First, the library was grown six independent times for ~84 generations, and samples were taken at generations 0, ~28, ~56, and ~84. Then, unique 5′ gene architectures were simultaneously amplified and sent for deep sequencing, which allowed to follow the frequency of each variant in the population over the course of the experiment. Finally, a relative fitness score was assigned for each variant based on its frequency dynamics. (C) GFP expression level (as measured by Goodman et al., 2013; x axis) versus fitness effect (based on results of repetition C; y axis) of each variant in the library (Pearson correlation, r = −0.79, p < 10⁻²⁰⁰). Fitness effect comes from the burden of expressing unneeded proteins on cellular growth and is calculated by analyzing the frequency dynamics of each variant (see Experimental Procedures). We defined fitness residual as the difference between a variant’s observed and expected fitness. The expected fitness is calculated from the regression line between GFP expression and fitness (black line). Some variants consistently demonstrated positive (blue dots, n = 975) or negative (red dots, n = 815) fitness residual sign. Other variants showed extremely low fitness residual, and we termed those variants as “underachievers” (purple dots, n = 80). The group size of positive, negative, and underachiever variants are significantly much higher than expected by chance (Supplemental Information). These results suggest that certain 5′ gene architectures can increase or reduce the cost of gene expression. See also Figure S1A. Inset: positive (blue violin plot) and negative (red violin plot) fitness residual variants come from the same distribution of GFP expression level (Wilcoxon rank-sum, p = 0.46). Black line represents the median value. Thus, the effect of GFP levels on fitness was successfully factored out, thus allowing us to elucidate other molecular mechanisms that tune expression cost at given expression levels. (D) Fitness and fitness residuals demonstrate different distributions. While most variants showed negative fitness values, fitness residual is more similar to a normal distribution, though with a negative tail.

Figure 2. Higher Ratio of GFP Protein/mRNA Minimizes Cost of Gene Expression

(A) Although coming from the same distribution of GFP levels, positive variants (blue violin plot) demonstrate lower mRNA levels of the GFP gene compared to negative variants (red violin plot) (effect size = 58.26%, Wilcoxon rank-sum, p = 1.6 × 10⁻⁹). Consistently, underachiever variants (purple violin plot) show higher mRNA levels compared to negative variants (effect size = 68.04%, Wilcoxon rank-sum, p = 9.6 × 10⁻⁸). Black line represents the median value. (B) Positive variants show higher translation initiation rates compared to negative variants (effect size = 61.9%, Wilcoxon rank-sum, p = 3.7 × 10⁻¹⁸). (C) Positive variants demonstrate higher translation efficiencies (protein/mRNA) compared to negative variants (effect size = 55.67%, Wilcoxon rank-sum, p = 3.4 × 10⁻⁵). Consistently, underachiever variants (purple violin plot) further show lower translation efficiencies compared to negative variants (effect size = 63.06%, Wilcoxon rank-sum, p = 1.1 × 10⁻⁴). Statistically significant differences (p < 0.05) are marked with an asterisk. See also Figures S1B and S2A.

Figure 3. Slow Translation Speed at Early Elongation, Achieved by Diverse Molecular Means, Reduces Expression Cost

(A, C, and D) Positive variants show lower values of codon decoding speed (A), stronger mRNA structures (C), and lower speeds due to higher anti-Shine Dalgarno affinities (D) compared to negative variants (effect size = 59.55%, 65.03%, and 63.82%, Wilcoxon rank-sum, p = 3 × 10⁻¹², 5.4 × 10⁻²⁸, and 6.3 × 10⁻²⁴, respectively). Statistically significant differences (p < 0.05) are marked with an asterisk. See also Figure S1B. (B) Mean folding energy of mRNA secondary structure according to window’s start position for positive (blue curve) and negative (red curve) variants; error bars represent SEM. Dashed lines mark different positions along the variable region upstream to the GFP. Black vertical line marks the beginning of window with the largest observed difference, which is found at nucleotide positions +4 of the ORF, just after the first AUG codon. The distributions at this window position are seen in (C). See also Figure S2B.

Figure 4. Usage of Expensive-to-Synthetize, Lowly Available, and Hydrophobic Amino Acids Decreases Fitness Residual

(A) N terminus amino acid fusions of negative variants are more expensive to synthesize compared to positive variants (effect size = 72.74%, Wilcoxon rank-sum, p = 7.4 × 10⁻⁶²). Underachievers utilize even more expensive amino acids (effect size= 72.75%, Wilcoxon rank-sum, p = 1.7 × 10⁻¹¹). See also Figures S1B andS2C. (B) The frequency ratio of amino acids between positive and negative variants is negatively correlated with the energetic cost of amino acids (Pearson correlation, r = −0.54, p = 0.01). Each amino acid is marked according to its one-letter code. (C) The frequency ratio of amino acids between positive and negative variants is negatively correlated with the demand/supply ratio of amino acids (Pearson correlation, r = −0.82, p = 10⁻⁴). Demand comes from occupancy of ribosomes on each transcript (see Experimental Procedures), and supply is the cellular concentration of each amino acid (Bennett et al., 2009). (D) Amino acid availability and energetic cost are correlated (Pearson correlation, r = −0.72, p = 1.8 × 10⁻³). (E) N terminus amino acid fusions of negative variants are more hydrophobic than positive variants (effect size = 69.11%, Wilcoxon rank-sum, p = 3.2 × 10⁻⁴⁴). N terminus fusion of underachievers are even more hydrophobic (effect size = 81.67%, Wilcoxon rank-sum, p = 7.7 × 10⁻²¹). See also Figures S1B and S2C.

Figure 5. Each Feature Affects Fitness Residual Independently

Correlation plots of each feature pair show lack of correlation in most cases and only weak correlations in other cases. For feature pairs with Pearson correlation of r > 0.1, we compared the difference in one feature while controlling for the second and vice versa. See also Figure S3. Black lines are the regression curves between each feature pair. Number at upper-left corner is the Pearson correlation.

Figure 6. Variant with Same N Terminus Amino Acid Fusion Demonstrate a Range of Fitness Residuals

(A–E) Each dot represents one of the 137 N terminus fusions in the library. The x axis and the y axis represent the mean value of a feature for the variants with either below-average or above-average Δfitness-residual, respectively. The vertical and horizontal error bars represent standard errors for each of the axes. A statistical difference for deviance from the X = Y line was observed for all features, suggesting that even at a given amino acid sequence, these mechanisms affect fitness residual and can minimize expression costs (t test, p values: A, mRNA levels, 6.2 × 10⁻³; B, initiation rates, 7 × 10⁻⁹; C, codon decoding speeds, 4.3 × 10⁻²; D, mRNA folding, 3.5 × 10⁻¹⁶; and E, aSD velocity, 7.6 × 10⁻⁷). d is Cohen’s d that calculates the effect size.

Figure 7. A Model that Predicts Fitness Residual Accurately Reveals that Fitness Residual of Natural Bacterial Genes Is Correlated with Their Expression Level

(A) A linear regression model based on all eight features predicts fitness residual accurately in a cross-validation test (Pearson correlation, r = 0.53, p < 10⁻²⁰⁰). See also Figure S4. (B) The weighted coefficients of each feature in the regression model demonstrating the relative contribution of each feature to fitness residual (p value for regression coefficient of mRNA level = 3.5 × 10⁻¹¹, initiation rate = 2.5 × 10⁻¹², TE_{GFP protein/mRNA} = 2.7 × 10⁻⁹, codon decoding speed = 8.7 × 10⁻³, mRNA folding energy = 1.5 × 10⁻⁵⁰, aSD velocity = 8.7 × 10⁻³, hydrophobicity < 10⁻²⁰⁰, and amino acid synthesis cost = 5.4 × 10⁻⁸⁰). The sign of the contribution of each coefficient shows whether a feature is associated positively or negatively with fitness residuals. Error bars represent standard error of the coefficient estimation. (C) Predicted fitness residuals of E. coli genes according to the regression model are correlated with their expression levels (Pearson correlation, r = 0.25, p = 2 × 10⁻⁵³), suggesting that natural selection shapes 5′ gene architectures in order to minimize costs of gene expression. (D) Distribution of fitness residual scores for E. coli genes as predicted by regression model that was trained on either experimental or mock data. The experimentally based model predicts a significant, higher range of fitness residuals (p < 10⁻⁵), suggesting that the mechanisms that we elucidate with the synthetic library also apply on natural genes. (E) Predicted fitness residuals of B. subtilis genes according to the regression model are correlated with their expression levels (Pearson correlation, r = 0.33, p = 10⁻⁹³), suggesting that our model also applies for other bacteria species. (F) Same as (D), only for B. subtilis genes.