OPT: codon optimize gene sequences for E. coli protein overexpression

Abstract

The ability to overexpress proteins is valuable for biotechnology, but not all sequences are compatible with high yield. We previously analyzed the sequence features and mRNA folding stability of a large data set of 6,384 distinct gene constructs, and developed a model for protein yield. Our OPT.williams.edu server (1) predicts the probability an input sequence will produce protein at a high level when overexpressed in E. coli, and (2) returns optimized synonymous sequences designed to boost protein expression. Here we also present experimental evidence of the high yields of our OPT constructs for eight commercially produced proteins.

Graphical Abstract

Introduction

Protein production is essential for many biotechnology applications, ranging from drug production and discovery to structural biology. However, many native sequences express at low levels or not at all when cloned into expression systems. Engineering higher expression levels holds the promise of reducing production costs and enabling the study of low-abundance proteins.

One way to improve expression is by leveraging the redundancy of the genetic code, replacing the codons used to encode the protein with synonymous codons that are more optimal for the expression host. This approach not only modifies the codons used but also alters the base composition of the sequence, both of which have been shown to influence protein expression [1].

Synonymous codon usage patterns have been investigated extensively [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17]. Codon effects [3, 18, 19, 20, 21, 22, 23], folding free energy [19, 24], mRNA secondary structure, and (G+C) content have been suggested as candidate factors [25, 26]. Different optimization methods have been proposed for protein expression in Escherichia coli, but most of them rely on the codon usage of the host used for expression [23, 24] or experimental expression assays that use a genetic reporter, such as a fluorescent protein [27, 28, 29]. The underlying principles nonetheless remain uncertain, both because most studies have not analyzed large-scale experimental datasets and because the related inferences have not been tested systematically.

The emerging availability of large datasets of expression data for hundreds or thousands of gene sequences [18, 19, 10] now allow the use of bioinformatics methods to begin to illuminate the causes of variability of gene expression levels. In a previous study [1], we data-mined the NorthEast Structural Genomics (NESG) expression dataset [18, 30] to identify elements of mRNA that affect its translation. This study was based on experimental expression in E. coli of 6348 recombinant genes from a wide variety of source organisms, and expression was monitored on an SDS-PAGE gel. The NESG data set has the dual advantage of monitoring heterologous protein expression while avoiding possible artifacts that can be induced by the use of a genetic reporter. This allows a larger codon sampling space, which is useful when applied to producing heterologous proteins for biotechnology applications.

Our resulting modeling [1] confirmed that the factors most influential in protein expression are free energy, base composition, codon usage, length, and the amino acid repetition rate. We have shown [31] that all four bases have strikingly different average free energies (G < C < U < A), so base composition terms are strongly correlated with the local free energy without requiring computations in larger folding windows. Other modeling of synonymous-codons libraries of 154 or 1459 Fluorescent Proteins [19, 10] came to the similar conclusion that folding free energy near the start of the coding sequence (using slightly different windows) is the dominant factor in explaining the correlations with gene expression. Traditional codon metrics [33, 34] are not well correlated with gene expression, but our codon-influence metric is correlated with expression and mRNA stability [1].

Our initial analysis led to the development of two codon metrics (31C and 6AA) that produced high yields of protein in all 22 tested genes [1]. In this paper, we present refined optimization methods (AG29G and AG44G). We further demonstrate that our codon-optimization algorithms based on these principles produce high-level protein overexpression in E. coli from 15 of 15 optimized gene sequences for eight unrelated proteins from diverse phylogenetic sources.

Our OPT webserver allows the research community to evaluate whether a gene is likely to have high expression when over-expressed in E. coli. The OPT server also provides AG29G and AG44G designs using synonymous codons predicted to generate high protein yields, see Figure 1.

Figure 1:

Screen shot of our OPT webserver. The probabilities of high expression are predicted for the input sequence, after AG head optimization, and for full AG29G and AG44G optimizations.

Methods

Codon Influence Model

The NorthEast Structural Genomics (NESG) data set with more than 20,000 recombinant genes from a wide variety of source organisms were scored for expression and solubility [18, 30]. From this set, we selected 6348 genes with less than 60% pairwise sequence identity [1]. Of these, we then used the 1973 with the highest expression (E5) and 1754 with undetectable expression (E0) to develop a multiple logistic regression model. The model’s predicted probability of high expression [1] follows the logistic function

$$p(\theta )=\frac{e^{\theta }}{1+e^{\theta }},$$ (1)

and depends on multiple sequence features:

$$\begin{array}{l}\theta =2.0+0.054G_{UH}-1.5I\\ +6.6a_H-6.3a_H^2+0.8u_{3H}-1.8g_H^2\\ +0.078s_{7-16}+0.063s_{17-32}+0.86s_{\text{all}}\\ -16r-1.8d_{AUA}-0.0012L-520/L,\end{array}$$ (2)

where $G_{UH}$ is the folding free energy of the pET21 vector plus first 48 nucleotides of the coding sequence, indicator variable $I$ equals 1 when $G_{UH}<-39\text{kcal}/\text{mol}$ and $\%G{C}_{2-6}>0.6$ and equals 0 otherwise, $a_H$ and $g_H$ are the fractions of adenines and guanines in codons 2–6, $u_{3H}$ is the fraction of uracils at the third codon position, $s$ is our mean codon-influence parameter, $r$ is the amino acid repetition rate, $d_{AUA}$ equals 1 if the sequence contains the di-codon AUAAUA, and $L$ is the length [1].

Codon optimization approaches

Our AG29G and AG44G gene optimization approaches are based on Eq. (1). At the beginning of the coding sequence, the head, having weak RNA secondary structure is key to increase the ribosome initiation rate. We have found that the folding free energy of A,U,C,G bases differ dramatically [31], with adenines favoring single-strandedness and guanines favoring double-strandedness. Therefore, in codons 2–6 we make ${\text{AG}}_{\text{head}}$ codon substitutions to increase the A base content and reduce G content. These substitutions produce the best head sequences.

For the remainder of the gene, the tail, codon usage are the most influential model terms [1]. We use 29 or 44 Good codons, which include the top scoring codon for each amino acid and 9 or 24 others with the smallest offset in the codon influence metric [1].

We note that the ${\text{AG}}_{\text{head}}$, 29G, and 44G codon substitution sets are quite distinct from the optimal CAI or tAI codons [33, 34] (Table 1 and Fig. S6). Fig. S7 shows an alignment of native vs. optimized nucleotide sequences for one target gene, human proinsulin.

Table 1:

In the design of synonymous sequences, our AG, 29G, and 44G methods reduce the native degeneracy of the genetic code (the number indicated in column RND) to eliminate the worst codons in the head and tail regions. The methods sample from the subsets of codons listed.

AA	RND	tAI	CAI	AG	29G	44G
Ala	4	GCA	GCU	GCA	GCG	GCG,GCT,GCA
Arg	6	CGU	CGU	GCT	CGC	CGC,CGT
Asn	2	AAC	AAC	AAU	AAC	2
Asp	2	GAC	GAC	GAU	GAU	2
Cys	2	UGC	UGC	UGU	2	2
Gln	2	CAG	CAG	CAA	CAA	2
Glu	2	GAA	GAA	GAA	GAA	2
Gly	4	GGC	GGU	GCA	4	4
His	2	CAC	CAC	CAU	CAU	2
Ile	3	AUC	AUC	AUU	AUU	AUU,AUC
Leu	6	CUG	CUG	UUA	CUC,CUG	CUC,CUG
Lys	2	AAA	AAA	AAA	2	2
Met	1	AUG	AUG	AUG	AUG	AUG
Phe	2	UUC	UUC	UUU	2	2
Pro	4	CCG	CCG	CCA	CCC,CCG	CCC,CCG
Ser	6	UCC	UCU	AGU	UCA,AGC	UCA,AGU,AGU
Thr	4	ACG	ACC	ACA	ACC	ACC,ACA,ACG
Trp	1	UGG	UGG	UGG	UGG	UGG
Tyr	2	UAC	UAC	UAU	UAU	2
Val	4	GUA	GUU	GUA	GUA	GUA,GTT,GTG

We eliminate any homopolymer runs from our AG29G sequences in order to avoid difficulties with synthesis and with unintended frameshifting. For example, if the base following an AAA codon is an A, we replace that codon with AAG. We also typically resample the tail 100 times and picked the sequence with GC content nearest 50% since some gene synthesis companies had difficulty when the G+C content was too high or low.

Experimental methods

Optimized gene sequences were synthesized and cloned into the pET21-NESG expression vector (https://dnasu.org/DNASU/GetVectorDetail.do?vectorid=317) by Twist Bioscience (South San Francisco, CA). For most in vivo protein overexpression experiments, cells were pre-grown overnight at 37 °C in LB with 100 μg/ml ampicillin and also 25 μg/ml kanamycin for the BL21(DE3) cells harboring the pMGK plasmid. The proinsulin expression experiments were done in the E. coli SHuffle strain [37]. A second round of pre-growth was then conducted in chemically defined MJ9 medium with 0.4% (w/v) glucose and the same antibiotic concentrations. Those cells were diluted in a fresh aliquot of the same medium to an OD_600nm of 0.1 and grown at 37 °C to an OD_600nm of 0.5–0.7 prior to induction with 1 mM IPTG and shifting the culture to 17 °C for overnight growth with 250 rpm shaking. The induced cultures were collected by centrifugation, resuspended in lysis buffer (300 mM NaCl, 10 mM 2-mercaptoethanol, 50 mM monobasic sodium phosphate, pH 8.0), and sonicated prior to either directly loading them on the gel for the whole cell samples or to centrifuging them for 15 minutes at 4 °C at the maximum speed in an Eppendorf centrifuge and taking the supernatant for the soluble extracts. The genotype of the SHuffle strain is F´ lac, pro, lacIq /Δ(ara-leu)7697 araD139 fhuA2 lacZ::T7 gene1 Δ(phoA)PvuII phoR ahpC^∗ galE (or U) galK λatt::pNEB3-r1-cDsbC (Spec^R, lacI^q) ΔtrxB rpsL150(Str^R) Δgor Δ(malF)3.

Results

Our OPT server (1) reports the probability of high expression for input sequences and (2) supplies users with AG_head, AG44G, and AG29G optimized sequences. The interface is shown in Fig. 1.

Computational Validation

In Ref. [1], we find that the head region (codons 1–16) is of enhanced significance, and is dominated by base composition and folding free energy terms. In the tail, codons are most influential. Our Ref. [1] codon metric, which was parameterized from a large data set, is only slightly correlated (see Fig. S3 and Extended Data Figure 4 of Ref. [1]) with popular metrics like CAI and tAI, which are determined from a smaller set of genes and with some theory input. To explore the characteristics of our model inferences, we performed the following computational analyses:

First, Fig. S1 shows that the predicted probability of high expression as trained on the E0 and E5 categories only, is also correlated in the intermediate E1 to E4 categories of the entire 6348 NESG dataset. We find that the model predictions are ordered consistent with the observed intermediate expression categories, see Figure S1.

Second, in Table 2, we use Eq. (1) to calculate the median and mean predicted probabilities of high expression for different codon-optimization strategies. Substituting the optimal CAI or tAI codons or randomly sampling all codons only slightly increases the probability compared to the native Wild Type (WT) gene sequences. In contrast, performing AG optimization of codons 1–16 produces an ∼10% increase in the probability of high expression, while full AG29G optimization produces a much greater increase, yielding median and mean probabilities of 98% and 99% respectively. Thus, the calculations in Table 2 predict that AG29G optimizations are more likely to yield high protein expression than several standard codon-optimization strategies.

Table 2:

Comparison of predicted probability of high expression of Native E. coli and the optimization methods.

Optimization	Median	$\langle p\rangle \pm {\sigma }_p$
Native	0.63	0.56 ± 0.27
${\text{AG}}_{\text{head}}$	0.74	0.66 ± 0.25
AG44G	0.91	0.84 ± 0.17
AG29G	0.99	0.97 ± 0.09

Experimental Validation

To test the AG29G and AG44G codon-optimization methods, we applied them to eight commercially produced proteins ranging in size from 9 to 117 kDa that originated from a variety of phylogenetic sources: (1) Firefly Luciferase, which is commonly used as a reported for gene expression; (2) E. coli β-galactosidase (LacZ), which is also used as a reporter; (3) E. coli Maltose Binding Protein (MBP), which is used for protein solubilization and affinity purification; (4) a self-cleaving fusion of MBP linked to an engineered variant of Tobacco Etch Virus protease called SuperTEV [38], which is a widely used sequence-specific protease; (5) Thermus aquaticus Taq polymerase, which is used for DNA amplification; (6) a fusion of the E. coli pro-OmpA signal sequence as a secretion tag in front of Elizabethkingia meningoseptica N-glycosidase F (PNGaseF), which cleaves N-linked oligosaccharides from glycoproteins [39]; (7) an anti-GFP nanobody [40], which was chosen to demonstrate the applicability of the methods to a class of protein that is generally difficult to express in the cytosol in E. coli; and (8) human proinsulin, which is the precursor of the peptide hormone used to treat diabetes.

For four of the target proteins (luciferase, the anti-GFP nanobody, MBP, and MBP-SuperTEV), one synonymous AG29G sequence and one synonymous AG44G sequence were evaluated in parallel (Fig. 2). For human proinsulin, two different AG29G and AG44G sequences were generated using different random samplings (Fig. S4) of those codon sets and tested in parallel. For PNGaseF, Taq polymerase, and LacZ, a single AG29G sequence was tested. These recoded sequence all show low correlation with the Relative Synonymous Codon Usage (RSCU) pattern found in highly expressed endogenous E. coli genes (Fig. S3). The AG29G and AG44G sequences were chemically synthesized, cloned into a T7 expression plasmid, and protein expression was assayed using the same experimental methods previously employed to generate the reference dataset that produced the AG29G and AG44G codon-optimization methods. Figure 2 shows SDS-PAGE gels of whole-cell and soluble extracts following overnight induction of protein expression from those plasmids in E. coli. All AG29G and AG44G recodings of all eight genes uniformly yield high-level protein expression.

Figure 2:

AG29G and AG44G recoded genes uniformly yield high-level overexpression in E. coli for 8/8 commercial protein targets. The images show Coomasie-Blue-stained SDS-polyacrylamide gels of whole-cell (left) or soluble (right) extracts. Protein expression methods were equivalent to those used to produced the reference dataset that generated the AG29G method [1, 35, 36]. In brief, expression was induced overnight in MJ9-glucose medium at 17 °C using a T7-polymerase expression plasmid (pET21-NESG). The experiments employed E. coli strain BL21(DE3) for all protein targets but human proinsulin, which was expressed in the E. coli SHuffle strain that has been engineered to have an elevated cytosolic oxidation potential to promote disulfide bond formation [37]. For the experiments in BL21(DE3), in addition to the T7 expression plasmid, the cells also contained the pMGK plasmid, which harbors the argU gene that encodes tRNA^AGA under endogenous promoter control. See the Methods section for additional details.

These experiments were all performed in the standard E. coli protein expression strain BL21(DE3) [43] except for the experiment on human proinsulin, which was performed in the E. coli SHuffle strain that has been engineered to have an elevated cytosolic oxidation potential [37]. Commercial production of human proinsulin in bacteria employs refolding of fusion proteins that form inclusion bodies in vivo [50, 51, 52, 53]. Fusion proteins are employed because unfused proinsulin is otherwise efficiently degraded when expressed in a strain with a normal cytosolic oxidation potential. Our AG29G and AG44G genes for proinsulin similarly result in no detectable protein expression in BL21(DE3) (data not shown) but produce extremely high level expression of the unfused protein in inclusion bodies in the E. coli SHuffle strain (Fig. 2). Induction of disulfide bond formation in this strain presumably promotes inclusion body formation, which protects proinsulin from proteolysis.

The terminal density of the cell cultures is somewhat reduced upon overnight protein induction (Fig. S5), which is generally observed when heterologous proteins are overexpressed due to metabolic limitations and interference with cellular biochemistry. Some synonymously recoded gene sequences exacerbate the toxicity associated with protein overexpression [1]. However, induction of AG29G genes does not produce such a toxicity effect given that they show a uniform influence on cell growth that is proportional to the observed level of protein overexpression (Fig. S5).

We also compared the performance of the native wild-type (WT) E. coli K12 lacZ gene to our synonymously recoded AG29G gene in coupled in vitro transcription/translation reactions programmed with purified PCR products containing identical transcriptional promoter and terminator sequences (Fig. 3). Our AG29G gene produces substantially more of this 116 kDa protein in an S30 cell extract [41, 42, 43, 44, 45] under standard conditions, and it also drives robust synthesis of the protein in the PURExpress system [46, 47, 48] containing purified ribosomes and translation factors. Our codon optimization method developed based on large-scale in vivo protein overexpression experiments also performs well for in vitro protein expression.

Figure 3:

In vitro translation of an AG29G-recoded E. coli lacZ gene yields higher protein expression than the native gene. The image shows a Coomasie-Blue-stained SDS-polyacrylamide gel of coupled in vitro T7-transcription/E.-coli-translation reactions using either an E. coli S30 extract [41, 42, 43, 44, 45] (Promega Corporation, Madison, WI) or the PURExpress system containing purified E. coli ribosomes and translation factors [46, 47, 48] (New England Biolabs, Ipswich, MA). The reactions, which were conducted according to the manufacturers recommended protocols, were programmed with a purified PCR product amplified using primers designed to include the T7 promoter and terminator sequences according to the instructions for PURExpress. For the S30 reactions, a 25 μl volume containing 375 ng of DNA was incubated for one hour. For the PURExpress reaction, a 15 μl volume containing 225 ng of DNA was incubated at 37 °C for 2 hours. The “–” DNA template lanes contained no PCR product.

Conclusions

Eq. (1) describes the key mRNA sequence features that modulate protein expression in E. coli based on an analysis of a large expression dataset. This model was used to develop the AG29G and AG44G methods of synonymous recoding of genes. These methods are implemented by the OPT web server. It predicts the probability that an input gene sequence will overexpress protein at a high level in E. coli, and it also generates synonymous sequences that elevate that probability by optimizing either the first 18 nucleotides or the full sequence using the AG29G and AG44G methods.

Optimizing the first 18 nucleotides without optimizing the remainder of the coding sequence can yield large gains in the probability of high expression (Table 2). The ability to change the nucleotide sequence in that region of a gene via the primers used for PCR cloning provides a lower-cost approach to codon optimization than total gene synthesis, which is generally required to optimize the full coding sequence. The observed protein expression level generally increases as A content is raised or G content is lowered in the first 18 nucleotides [1, 49]. Therefore, trivially implemented sequence variations in this regions of the coding sequence can effectively be used as a rheostat to control protein expression level [49]. In some cases, tuning expression to a lower level using this approach can increase the proportion of the protein expressed in the soluble fraction and even the absolute yield of soluble protein (manuscript in preparation).

With full optimization of both the head and the tail of a gene, the predicted probability of high expression tends to approach 100% (Table 2), and 15 trial optimizations of different proteins with commercial optimizations uniformly produce high-level overexpression in vivo in E. coli.

Access

The OPT.williams.edu server is free for academic non-commercial use. A patent has been issued on the methods described above. Commercial users can request a license from Columbia Ventures.

Highlights:

• server evaluates likelihood gene sequences will express at a high level in E. coli

• server proposes synonymous gene sequences with increased probability of high expression

• method is based on a statistical model trained on 6000+ over-expression experiments (Boël et al, Nature 2016)

• we present experimental evidence that OPT synonymous gene designs produce uniformly high levels of over expression for all the commercially valuable proteins tested