|Home | About | Journals | Submit | Contact Us | Français|
A general means of viral attenuation involves the extensive recoding of synonymous codons in the viral genome. The mechanistic underpinnings of this approach remain unclear, however. Using quantitative proteomics and RNA sequencing, we explore the molecular basis of attenuation in a strain of bacteriophage T7 whose major capsid gene was engineered to carry 182 suboptimal codons. We do not detect transcriptional effects from recoding. Proteomic observations reveal that translation is halved for the recoded major capsid gene, and a more modest reduction applies to several coexpressed downstream genes. We observe no changes in protein abundances of other coexpressed genes that are encoded upstream. Viral burst size, like capsid protein abundance, is also decreased by half. Together, these observations suggest that, in this virus, reduced translation of an essential polycistronic transcript and diminished virion assembly form the molecular basis of attenuation.
A recent methodological advance in the development of attenuated viral vaccines has been the design of genomes with hundreds of synonymous codon changes that are collectively suboptimal (Burns et al. 2006, 2009; Mueller et al. 2006, 2010; Coleman et al. 2008; Wimmer et al. 2009; Bull et al. 2012; Nougairede et al. 2013; Shen et al. 2015). The nature of the design varies from the replacement of common codons with rare codons to merely shuffling existing codons to create uncommon codon pairs. By maintaining wild-type protein sequences, the recoding method retains the antigenic profile of the wild type while reducing viral growth rate and virulence. Based on the premise that silent codon changes have individually small effects, the method should not only allow the tunable crippling of viral growth to an arbitrary degree but also profoundly retard the reevolution of high viral fitness. Both predictions have been supported empirically. Thus, poliovirus, influenza, several arboviruses, and a bacteriophage all exhibit quantitative fitness declines after synonymous codon changes (Burns et al. 2006; Mueller et al. 2006, 2010; Coleman et al. 2008; Nougairede et al. 2013; Shen et al. 2015; Bull et al. 2012), and fitness recovery during viral growth over hundreds of generations is at best slow (Burns et al. 2006; Coleman et al. 2008; Bull et al. 2012; Nougairede et al. 2013).
The broad success of attenuation from synonymous codon changes in different viruses and with different designs could arise from a common underlying mechanism. Yet, there is ongoing debate about how synonymous codon changes affect fitness and thus what that mechanism could be. One popular hypothesis is that codon usage controls translational efficiency, in turn affecting the rate of protein synthesis (Fredrick and Ibba 2010; Plotkin and Kudla 2010; Tuller et al. 2010; Shah and Gilchrist 2011; Zur and Tuller 2016). Under this hypothesis, highly expressed and functionally important genes are encoded by optimal codons to increase translational efficiency. However, a simple codon “optimality” model cannot explain the attenuation attained by merely shuffling codons; since the abundance of the different codons is not being changed by shuffling, the suboptimality must be due to something besides codon abundance. Likewise, some highly expressed genes in cyanobacteria and Neurospora have nonoptimal encodings (Xu et al. 2013; Zhou et al. 2013). In other cases, codon usage determines expression through transcription, not translation (Zhou et al. 2016). In the face of so many seemingly contradictory observations, further advances in this research program would benefit from the identification of a common mechanism for viral attenuation, or at least benefit from the demonstration that different mechanisms are involved.
Resolving the basis of attenuation in viral systems is partly hampered by the sequence of life-history steps between the initial effects of codon changes and the final emergence of assembled virions. The initial impact may lie in transcription or translation of one or more genes, but the effect on the number of virions will depend on which proteins are limiting during assembly. The same modification of an early life-history stage may have different fitness effects in different viruses. Ultimately, it may be necessary to interpret the genome engineering in the context of a system-wide, comprehensive model of the viral life cycle. Such is our motivation.
In the bacterial virus T7, recoding the major capsid protein (gene 10A) with synonymous codons reduced the fitness of the phage (Bull et al. 2012). To engineer T7, codons in gene 10A that were highly utilized in Escherichia coli (the T7 host) were replaced with codons that were underutilized in the host. The major capsid protein, which forms the head of the T7 phage particle, is the most abundant and highly expressed phage protein (Dunn and Studier 1983). In the phage genome with the most extensive set of gene 10A synonymous codon replacements, the fitness was 35.7 doublings/hr compared to 43.2 doublings/hr in the wild type (Bull et al. 2012). This difference translates to a 180-fold decline in descendants produced per hour. In total, 182 codons were changed, just over half the number of codons in the major capsid gene; in addition, the first 20 codons were not altered to avoid disrupting translation initiation processes.
After adapting the recoded phage for 1000 generations, fitness increased to 38.7, a recovery of almost half the initial deficit (on a log scale). Thus, recoding gene 10A induced a moderately stable fitness reduction. A mere nine nucleotide changes were responsible for the fitness recovery, and seven fell outside the recoded region, shedding little light on the underlying mechanism of fitness reduction.
Here, we apply new methods to continue exploration of the T7 attenuation. Using the same strains and designs as the prior study (Bull et al. 2012), our purpose is to develop a comprehensive model of the way silent codon changes cause reduced fitness. As part of this effort, we propose and test three mechanistic models that could explain the fitness reduction in recoded T7. In testing those models, we apply proteomic methods, RNA sequencing, and various phenotypic measures in a systems approach to understanding the basis of attenuation. We find that recoding gene 10A reduces protein abundances of gene 10 and also of several downstream genes. From there, we address the impact of protein abundances on viral fitness components (burst size and lysis time), ultimately connecting these measurements to a model that describes actual fitness.
Gene 10 is translated in two forms, A and B. Form A is 344 amino acids and is formally denoted the major capsid protein. Form B (the minor capsid protein) is not essential and results from a ribosomal frameshift at the end of A and is 397 amino acids. In the engineering, all codon changes were within 10A and thus also within 10B. Moreover, since most peptide fragments coming from the minor and major capsid proteins ambiguously mapped to both proteins, abundances of 10A and 10B were not differentiated using our proteomic methods and were combined. We followed a similar procedure for our RNA-sequencing analyses. To simplify notation, we merely refer to the recoded gene as 10 and the affected A and B proteins as capsid protein.
The host for all experiments was IJ1133 [E. coli K-12, F-ΔlacX74 thiΔ (mcrC-mrr)102::Tn10]. T7 strains used in this study come from previous work (Bull et al. 2012). An isolate of T761 (a population adapted to grow optimally on IJ1133 specifically through serial passage) was first deleted of its gene 10, then recombined over a plasmid carrying a different gene 10 engineered to contain a low fraction (0.1) of preferred codons, with 182 codon changes. The recombinant, denoted here as the recoded strain, could be identified by its ability to grow without complementation. The evolved strain was initiated from the recoded phage and adapted over 800–1000 generations [strain L1 from Bull et al. (2012)]. The wild-type strain in this study was derived from the recoded strain after recombination over a plasmid containing wild-type gene 10, then grown out for 6 hr of serial transfer on IJ1133. Fitness of this “wild-type” strain was approximately the same as that of the ancestral population (T761).
Lysis time and burst size assays were performed as previously described (Heineman and Bull 2007; Bull et al. 2011). The initial infection steps were identical for both assays. Briefly, 108 phage (MOI = 0.1) were added to a 10 ml culture of exponentially growing cells (37° with agitation), incubated for 3 min, and subsequently diluted 104-fold to prevent further adsorption. For lysis times, phage were plated at various time points between 4 and 18 min (after initial infection) to monitor changes in titer; lysis time was taken as the time of the first significant increase in titer.
To determine burst size, initial density of phage-infected cells was determined by plating phage before and after treatment with chloroform 5 and 6 min after initial infection. Cells infected with phage at the time of chloroform treatment do not produce viable phage, so only free phage will form plaques, allowing for the determination of phage-infected cells at these times. Final phage titers were obtained at 15, 16, and 17 min by plating chloroform-treated samples. Burst size was then calculated as the phage titer at the end time points divided by the number of initial phage-infected cells.
E. coli was grown in LB broth to a concentration of 108 cells/ml at 37° with agitation, then infected with phage at an MOI of 2.5. At 1, 5, and 9 min postinfection, 2 ml of bacterial suspension were removed from the phage-infected cultures and pelleted in a microcentrifuge. Pellets were either flash frozen in liquid nitrogen or immediately used for downstream processes (RNA extraction or protein preparation for proteomics). RNA was isolated using Trizol reagent, following the manufacturer’s protocol. Library preparation and sequencing was performed by the University of Texas Genome Sequencing and Analysis Facility using Illumina NextSequation 500 (SR75).
Since gene 10B is a readthrough product of gene 10A, we excluded the gene 10B transcript from the reference transcriptome. RNA-sequencing reads were quantified using Kallisto (Bray et al. 2016) and E. coli K-12 (NCBI: U00096.3) and T7 (NCBI: NC_001604.1) reference genomes. For analyses of the recoded and evolved T7 strains, the gene 10 sequence was replaced with the recoded sequence (supplementary file S1 in Bull et al. 2012,) in the reference genome. A population of T7-infected E. coli has no core set of stably expressed genes with which to normalize during differential expression analysis. Therefore, all transcript abundance estimates (transcripts per million) were normalized to the total cellular transcript abundance (including both T7 and E. coli transcripts). Differential expression analysis was only possible within genes across treatments, but not between genes in the same treatment.
Proteomics was performed as previously described in Houser et al. (2015). In brief, T7-infected E. coli cell pellets (prepared as described in the RNA sequencing section above) were resuspended in 50 mM Tris-HCl, pH 8.0, 10 mM DTT. 2,2,2-trifluoroethanol (Sigma-Aldrich) was added to 50% (v/v) final concentration and samples were incubated at 56° for 45 min. Following incubation, iodoacetamide was added to a concentration of 25 mM and samples were incubated at room temperature in the dark for 30 min. Samples were diluted 10-fold with 2 mM CaCl2, 50 mM Tris-HCl, pH 8.0. Samples were digested with trypsin (Pierce) at 37° for 5 hr. Digestion was quenched by adding formic acid to 1% (v/v). Tryptic peptides were bound, washed, and eluted from HyperSep C18 SpinTips (Thermo Fisher Scientific). Eluted peptides were dried by speed-vac and resuspended in Buffer C (5% acetonitrile, 0.1% formic acid) for analysis by LC-MS/MS.
For LC-MS/MS analysis, peptides were subjected to separation by C18 reverse phase chromatography on a Dionex Ultimate 3000 RSLCnano UHPLC system (Thermo Fisher Scientific). Peptides were loaded onto an Acclaim C18 PepMap RSLC column (Dionex; Thermo Fisher Scientific) and eluted using a 5–40% acetonitrile gradient over 250 min at 300 nl/min flow rate. Eluted peptides were directly injected into an Orbitrap Elite mass spectrometer (Thermo Fisher Scientific) by nano-electrospray and subjected to data-dependent tandem mass spectrometry, with full precursor ion scans (MS1) collected at 60,000 resolution. Monoisotopic precursor selection and charge-state screening were enabled, with ions of charge >+1 selected for collision-induced dissociation. Up to 20 fragmentation scans (MS2) were collected per MS1. Dynamic exclusion was active with 45-sec exclusion for ions selected twice within a 30-sec window.
We assigned each peptide to a protein or protein group (in the case of ambiguous peptides which map to multiple proteins) using Proteome Discoverer (Thermo Fisher Scientific) and REL606 and T7 reference proteomes (NCBI: NC_012967, NC_001604.1) concatenated with a database of contaminant proteins (http://www.biochem.mpg.de/5111795/maxquant). We selected the top three most-abundant peptides by peak area for each protein. We averaged these peptide peak areas across technical replicates to obtain a protein-abundance estimate (Silva et al. 2006). All protein-abundance estimates were normalized to the total E. coli protein content of the sample.
To determine if gene 10 C-terminal peptides were more common than N-terminal peptides in the recoded T7 strain compared to the wild-type strain, we compared the fit of the following two models [given here in notation from the lme4 R package (Bates et al. 2015)]:
where count is the number of peptides of type peptide, strain is the strain of T7, and location is the location of peptide within gene 10. We assume peptide to be a random effect, given by the (1|peptide) term. The term strain:location indicates interaction between strain and location. Thus, we compare a model in which the location of a peptide and the strain interact, and a model in which there is no such interaction. If N-terminal peptides were less prevalent than C-terminal peptides in the recoded T7 strain, we would expect Equation 2 to provide a better fit than Equation 1.
Secondary structure near the ribosome binding site (RBS) can inhibit translation initiation. On polycistronic transcripts, this secondary structure can be disrupted by ribosomes completing translation of an upstream gene, thus increasing translation initiation rates. Estimates of this relative increase in translation initiation of gene 11 due to translation of gene 10 was predicted using the Operon Calculator (Tian and Salis 2015).
To model the effects of translational coupling on protein production, we first assume a polycistronic transcript with three genes a, b, and c. We write the effective initiation rate of gene a as
where is the aggregate initiation rate of a and is the translation elongation rate of gene a. We assume that if the initiation rate ever exceeds the elongation rate, ribosomes will quickly back up on the transcript and make elongation the rate-limiting step of translation. Thus, in our model, the elongation rate can never be exceeded by the aggregate translation initiation rate. For gene a, the aggregate initiation rate is simply the de novo initiation rate because there are no genes upstream of a.
The rate at which ribosomes complete translation of gene a, equivalent to the production rate of protein A, is defined as
at steady state. In this context, the steady state assumption means that all three protein products are being produced continuously at the equilibrium rate.
For the effective translation initiation rate of gene b, we similarly write
where is the aggregate translation initiation rate due to upstream-dependent reinitiation and de novo initiation, and is the translation elongation rate of b. We define the aggregate translation initiation rate
where is the rate of upstream translating ribosomes reinitiating on gene b, and is the rate of ribosomes initiating de novo on gene b.
Lastly we define reinitiation and de novo initiation rates on gene b as follows:
where is the rate of ribosomes flowing from upstream translation of gene a, and represents the proportion of that ribosome flow reinitiating on gene b. We assume that the rate of de novo initiation depends, in part, on upstream ribosomes relaxing secondary structure around the RBS of gene b. Thus, the rate of de novo initiation is given by the upstream ribosome flow scaled by some constant (facilitated binding), and by a constant rate that does not depend on upstream ribosome flow.
We can simplify the effective initiation rate to
We simplify the effective initiation rate further by defining a coupling constant,
which accounts for both the effects of facilitated binding and ribosome reinitiation. The final effective initiation rate of b is defined as
Similar to the protein production rate of gene a, we define the protein production of gene b
at steady state.
The effective initiation rate of gene c is similar to that of gene b
where is the rate of ribosomes flowing from the end of upstream gene b, and represents the proportion of that ribosome flow reinitiating on gene c. The rate of de novo initiation dependent on upstream ribosome flow (facilitated binding) is given by and is the de novo initiation rate independent of upstream ribosome flow. Again, is a coupling constant that incorporates the effects of both facilitated binding and ribosome reinitiation.
Raw RNA reads are available at NCBI GEO (ID: GSE96573) (Barrett et al. 2013). The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE (Vizcaíno et al. 2016) partner repository with the data set identifier PXD006502. All processed data and scripts are available at https://github.com/benjaminjack/phage_attenuation.
As codon deoptimization is thought to affect translational efficiency, we propose three models in which recoding gene 10 (capsid protein) affects protein abundances in T7. In model 1, codon deoptimization slows translation of gene 10 and reduces the abundance of the capsid protein only. In model 2, deoptimization depletes the ribosome pool by creating high ribosomal densities on capsid protein transcripts, thus reducing translation of all viral proteins late in the infection cycle (Vind et al. 1993; Birch et al. 2012; Raveh et al. 2016). In model 3, codon deoptimization has intermediate effects between models 1 and 2: translation is impaired for gene 10 and downstream genes but not for upstream genes. The expectation in model 3 arises because T7 produces many polycistronic transcripts, leading to translational coupling of gene 10 and genes immediately downstream. Translational coupling has been observed in bacterial operons but has not been considered in the context of codon deoptimization (Oppenheim and Yanofsky 1980; Schümperli et al. 1982; Aksoy et al. 1984; Tian and Salis 2015). All three models assume that deoptimizing gene 10 will reduce the abundance of at least the capsid protein.
To differentiate between our three proposed models, we compared the T7 proteome during infection among wild-type, attenuated, and evolved phages using mass spectrometry-based protein quantitation. T7 is thought to encode 58–60 proteins, but only 19 are essential, and many have no known function (Dunn and Studier 1983; Molineux 2006) (e.g., some are thought to be homing endonucleases, selfish elements). All genes are encoded on the same strand, and expression order is linear; genes are numbered in order with essential genes having integral numbers (1–19) and nonessential genes having fractional numbers. The genome is divided into three expression groups. Class-I genes are the first to enter and are expressed from promoters at the entering end of the genome, transcribed by the host RNA polymerase (RNAP). The phage RNAP gene (numbered gene 1) is the last of the class-I genes and the first essential gene. All other genes are expressed from phage promoters, but nearly all transcripts are polycistronic as there is only one terminator for phage RNAP (immediately after gene 10), and there are only 17 phage promoters (Figure 1). Gene 1.1 is the first class-II gene (Dunn and Studier 1983). [Note that genes 1.1–1.3 are sometimes also considered to be part of class I, because they are transcribed by both E. coli and T7 RNAP (Molineux 2006).] Gene 6.5 is the first class-III gene.
Because the wild-type phage lyses the cell at ~11 min after infection (Heineman and Bull 2007; Bull et al. 2011), the proteome of infected hosts was sampled at 1, 5, and 9 min after phage addition to the culture; infection of cells is neither immediate nor synchronous upon phage addition to the culture, so these times are approximate postinfection values. By 9 min after infection, ~50 of the known or predicted T7 proteins were detected. All samples recovered ~4000 E. coli proteins. Since proteins have a much longer half-life than transcripts, and T7 has no known mechanism of degrading E. coli proteins (Molineux 2006), we assumed that E. coli protein abundances remained constant over the 9 min infection, and normalized the phage protein abundances to that of E. coli (Houser et al. 2015). Thus, we report all phage protein abundances as a proportion of E. coli protein content.
Abundances of the major capsid protein (a product of gene 10) were of primary interest, as gene 10 is the most highly expressed phage gene and is also the one deoptimized. Gene 10 comprises two protein products: the major capsid protein (gp10A) and the much less abundant minor capsid protein (gp10B). The minor capsid protein is produced after a frameshift and stop codon readthrough of 10A and, except for the C-terminal ~53 amino acids, is identical in sequence to the major capsid protein. Thus, our proteomic methods have limited ability to distinguish between the two protein products, so we combine abundance estimates into a single capsid protein measurement (see Materials and Methods). By 9 min after infection, capsid protein abundances in the attenuated strain were about half of those in the wild type ( paired t-test; Figure 2). The capsid protein abundance for the evolved strain was intermediate.
Recoding of gene 10 could have reduced capsid protein abundance by reducing rates of translation elongation, thereby increasing the likelihood of ribosome stalling and fall-off. If ribosome fall-off were the dominant mechanism by which protein abundance was reduced, attenuation should have been accompanied by an excess of short peptides from the N-terminal end. Alternatively, if translation was slowed down without ribosome fall-off, a uniform distribution of peptides should be observed across the capsid protein. When mapping individual peptides recovered from the mass spectrometry proteomics, no systematic change was observed in the distribution of peptides across the protein (Figure 3). Thus, the recoded phage strains produced complete capsid protein, but in smaller quantities than that of the wild type.
If recoding a highly expressed gene saturates the pool of ribosomes by slowing translation of the recoded gene, the rate of translation of all T7 genes could ultimately decline (model 2, above). Models of T7 replication in E. coli, and limited experimental data, are consistent with protein synthesis being the rate-limiting step of T7 replication (although the evidence is at best weak and indirect) (Endy et al. 1997; You et al. 2002). Moreover, depletion of free ribosomes is common in E. coli transformed with highly expressed heterologous genes (Vind et al. 1993; Scott et al. 2010; Reuveni et al. 2011; Raveh et al. 2016).
The data allow us to measure other T7 protein abundances over time. 9 min postinfection, abundances of class-III proteins gp11, gp12, gp13, gp14, and gp15 in the recoded strain were lower than that of the wild type [false discovery rate (FDR) < 0.1, FDR-corrected t-test] (Figure 4 and Table 1). These proteins are mostly or all structural: tail tubular proteins (gp11 and gp12), two internal virion proteins (gp14 and gp15), and a protein of unknown function but required for incorporation of other essential proteins in the virion (gp13). The T7 endonuclease (gp3), encoded by a class-II gene, also showed reduced abundance in the recoded strain. Abundances of these class-III proteins in the evolved strain again fall somewhere between that of the wild-type and recoded strains. Together, these results demonstrate that recoding gene 10 reduced the protein abundances of the capsid protein and several proteins encoded downstream of the capsid protein. The false discovery threshold we employed allows for one or two of those downstream proteins identified to be false positives, but the majority are likely real.
Whereas proteins encoded downstream of the recoded gene 10 showed decreased abundances, those encoded immediately upstream of gene 10 were not obviously affected. Gene 9, immediately upstream of 10, encodes the highly expressed scaffold protein but this protein showed no difference in abundance between wild-type and recoded strains (Figure 5). Under the ribosome-depletion hypothesis (model 2), we expected all genes expressed at the same time as gene 10 to be suppressed, thus including gene 9. Indeed, because of the high levels of expression of 9 and the consequent ease of measuring it with our proteomics methods, we reject the ribosome-depletion model as the basis of attenuation.
A reduction in capsid protein abundance in the attenuated virus could be explained by fewer transcripts from the gene. Of course, a reduction in transcription is not expected from a change in codon usage, and indeed, the design left the 5′ end of the gene unaltered specifically to keep transcription and translation initiation unchanged. Nonetheless, RNA sequences were obtained from phage-infected E. coli at 1, 5, and 9 min after infection to see if the decline in protein was accompanied by a decline in transcripts. Transcript abundances were normalized to the total RNA abundance within a sample, excluding rRNA and tRNA. Both phage and E. coli transcripts were included.
Gene 10 transcript abundance increased over time, being highest at the final 9-min time point. Yet no significant heterogeneity was observed among the wild-type, recoded, and evolved strain transcripts of gene 10 (Figure 6A). This observation seems to rule out a transcriptional cause of capsid protein reduction.
As validation of the RNA-sequencing methods, transcription analyses were extended to other properties of the T7 and E. coli transcriptomes. Gene 1 (T7 RNAP) transcript abundance decreased over time as a proportion of total transcripts (Figure 6B), consistent with previously observed class-I gene expression timing (Molineux 2006). Likewise, the relative abundance of T7 transcripts to E. coli increased sharply, consistent with established mechanisms by which T7 shuts off host transcription and degrades the host genome (Molineux 2006). At 1 min after infection, T7 transcripts comprised <1% of the transcript pool of infected E. coli (Figure 7). By 5 min after infection, T7 transcripts made up >75% of the transcript pool. By 9 min postinfection, this proportion reached ~95%. Although it is not possible to assess changes in absolute transcript abundances (see Materials and Methods), the data require some combination of host-cell transcripts being degraded rapidly, or T7 synthesizing transcripts so rapidly that they quickly dwarf the pool of E. coli transcripts. No strain-specific trends were detected.
The proteomics suggest that translation of some downstream genes are specifically depressed by the recoding of 10. This effect on downstream genes might involve translational coupling, in which a stalling of translation over 10 delays translation of genes further down on the same transcript. Translational coupling often occurs when multiple genes are encoded on a single transcript with little intergenic space, such as in bacterial operons and viruses (Lesage et al. 1992; Hellmuth et al. 1991; Schümperli et al. 1982; Oppenheim and Yanofsky 1980; Aksoy et al. 1984; Torgov et al. 1998). Translational coupling is a plausible process for sets of T7 genes because most transcripts include multiple genes (Dunn and Studier 1983). The T7 class-III promoters precede genes 6.5 (the first class-III gene), 9, 10, 13, and 17; some transcripts with 9 and 10 will thus include earlier genes, but many will not (Figure 1). Although a phage-specific terminator between 10 and 11 aborts most (but clearly not all) 10 transcripts before 11, all transcripts with 11 and 12 necessarily include 10. Thus, translational coupling would operate for 11 and 12 if many of the ribosomes on those genes first translated 10. Translational coupling beyond 12 is less plausible, however. The combination of an RNAse-III site between 12 and 13 and a promoter before 13 will mean that many or most transcripts with 13 do not include 12. So we expect substantially higher levels of translational coupling of 11 and 12 with 10, but far less between 10 and 13.
Polycistronic transcripts are necessary for translational coupling, but not sufficient. It must also be the case that downstream genes contain structured, inaccessible RBSs that only initiate translation in the presence of upstream translating ribosomes (Tian and Salis 2015; Rex et al. 1994; Qu et al. 2011). Ribosomes that reach the stop codon of one gene often expose the RBS of the next gene on the transcript and then reinitiate translation on that downstream gene (Spanjaard and van Duin 1989). Moreover, this exposure of the RBS also facilitates binding from the free ribosome pool (Rex et al. 1994). Support for the coupling model was evaluated from secondary structure predictions and in silico predictions of ribosome binding (Tian and Salis 2015). Due to limitations of software-based methods, we only tested the coupling of genes 10 and 11. With translational coupling, translation initiation rates are predicted to be ~6 times greater for gene 11 than they would if gene 11 occurred on a single-gene transcript. This supports a model in which at least genes 10 and 11 are translationally coupled.
As the inference of translational coupling here is indirect and tentative, additional insight was sought from a mathematical model. The model assumed three genes on a single transcript in the order a, b, and c (Figure 8A, Equations 3–14 in Materials and Methods). All genes were assumed to be the same length. In the model, protein production rates (i.e., the rate at which ribosomes complete translation) depended on translation initiation and translation elongation rates. The model allowed varying degrees of coupling, with a coupling constant y, which accounts for both ribosome reinitiation and facilitated binding. The model also allowed for translation initiation independent of any coupling effects, given by w. (In the results presented here, and See Table 2 and Materials and Methods for all parameters). In a system with no coupling, translation initiation of gene b did not depend on translation of gene a, and likewise between genes b and c. In a strongly coupled system, translation initiation of gene b depended almost entirely on the rate at which ribosomes complete translation of gene a. Translation of gene c similarly depended on gene b in a strongly coupled system. Thus, we explored how relative rates of protein production depend on both coupling between genes and elongation rates within genes.
To simulate codon deoptimization of one gene in the transcript, we varied the translation elongation rate of gene b only (Figure 8). We considered reductions in translation rate within a two- to threefold range. This reduction in translation elongation is compatible with prior studies of codon deoptimization in E. coli (Kudla et al. 2009). We also varied the coupling constant y. Under strong translational coupling (), the translation initiation rates of a gene depended mostly on the rate of ribosomes moving through the stop codon of an upstream gene. Thus, the relative rate of protein A, B, and C production (assuming that initiation is slower than elongation in each gene) depended on the translation initiation rate of gene a. In turn, when the translation elongation rate was less than the initiation rate for gene b, translation elongation became rate limiting. Under these elongation rate-limited conditions, the production rates of proteins B and C only increased as the elongation rate of b increased, while A production rates remained unaffected (Figure 8C, right panels). Conversely, in a weakly coupled model (), even if elongation rates were slower than initiation rates in gene b, production of protein C was only weakly affected by recoding of gene b (Figure 8C, left panels). In partially coupled models ( Figure 8B, middle panels), the production rate of protein C also increased as elongation rates increased until surpassing initiation rates for gene b. However, this increase in the rate of C protein production was smaller than the increase observed on a strongly coupled model. Increasing the independent translation initiation rate w decreases the dependence of protein-C production on gene b elongation rates (Figure 8C, top row to bottom row). This model demonstrated that recoding a gene in a translationally coupled set of genes affects the protein production rates of downstream, but not upstream, genes.
In previous work, the T7 with a recoded capsid gene had been found to have a fitness of 35.7 doublings/hr, compared to a value of 43.2 for the wild type. Here we consider whether and how the altered proteomics might lead to this fitness reduction. The connection from proteomics to fitness spans two steps: (i) identify the phage life-history components affected by the recoding and evaluate whether that change is compatible with the proteomics, and then (ii) assess whether the magnitude of altered fitness components is compatible with overall fitness.
In the growth conditions used for our assays, fitness is determined by cell density and three phage properties: burst size, lysis time, and adsorption rate (Wang et al. 1996; Guyader and Burch 2008; Shao and Wang 2008; Bull 2006; Patwa and Wahl 2009; Bull et al. 2011). Burst size refers to the average number of viral particles released from each infected cell. Burst sizes and lysis times were estimated here for the wild-type and recoded phages (there was no expectation that adsorption rate would be affected, which depends on the presence of tail fibers, the product of gene 17). Although the same cell line was used here as in previous studies (IJ1133), new cell preparations were used, so quantitative agreement with past estimates of burst size and lysis time is not expected, but proportional differences should scale across different cell preparations. No difference in lysis time was observed between strains, but burst size was reduced almost 50% with the recoding (Figure 9).
The proportional reduction in burst size is nearly the same as that for the reduction in capsid protein abundance at 9 min. This reduction in burst size is no doubt caused by the reduction in capsid protein. There is perhaps little basis for arguing that the reductions should match quantitatively, but the agreement between the two numbers poses no dilemma.
The second step in connecting proteomics with fitness is to consider whether a 50% reduction in burst size (with no change in lysis time) is compatible with the observed fitness reduction of 7.5 doublings/hr. The mapping of phage fitness components onto total fitness has been addressed in detail (Bull et al. 2011). From numerical trials in that study, a 50% reduction in burst size is compatible with a fitness reduction of the magnitude observed here (table 1, lines L1 and L5, in Bull et al. 2011).
In the decade since the first proposals to attenuate viruses by synonymous codon substitutions, it has been established that the method works in many viruses and offers many advantages over earlier methods of attenuation. Yet the mechanism by which silent codon changes attenuate not only remains elusive but seems less clear now than it did at the start. Nor is it clear that a single mechanism underlies the attenuation in different systems. Here, we extended previous work on a bacteriophage system in which the encoding of rare codons in the major capsid gene reduced fitness. Our goal was to refine an understanding of the molecular basis of the attenuation.
The capsid protein (encoded by gene 10) is the most abundant protein produced during the infection cycle of bacteriophage T7. Deoptimizing 50% of gene 10 codons reduced fitness (Bull et al. 2012). In exploring the underlying molecular mechanism by which the recoding has this effect, our primary result is that the protein product of gene 10 is reduced almost 50% by the end of the infection cycle, but protein abundance of genes immediately downstream of gene 10 are also depressed. The differences in protein abundance are not reflected in transcript levels, so it appears that the suppression of protein levels lies in translation. The evidence thus supports a simple interpretation of the fitness impact of recoding the major capsid gene:
Taken together, these two observations indicate that the reduction of capsid protein abundance is a consequence of recoding and that this reduction is a plausible intracellular cause of viral attenuation.
One mechanism we entertained to explain the altered proteomics of the recoded phage is saturation of the ribosomes with gene 10 transcripts. Such a model requires that the production of all T7 proteins would decline late in the infection cycle for the recoded phage. Within the limits of resolution, the proteomics rule out an overall reduction in T7 protein production, indicating that reductions are limited to the recoded 10 and a few downstream genes. This study may provide the first indication that translational effects of the recoding extend beyond the recoded genes. There was also no evidence for aborted gene 10 polypeptides in the recoded strains, as might occur from ribosomes stalled on gene 10 transcripts. An obvious next step is to extend these analyses to ribosome profiling, which would directly indicate whether the recoding does tie up ribosomes on gene 10 (Li et al. 2014).
The means by which synonymous codon replacement attenuates may be more straightforward for phage T7 than for eukaryotic viruses. Several lines of evidence suggest that the eukaryotic virus attenuation by synonymous codon changes is from the creation of CpG dinucleotides (Burns et al. 2009; Atkinson et al. 2014; Tulloch et al. 2014); indeed, especially powerful evidence to support this interpretation is that evolutionary reversions of attenuated viruses disproportionately reverse CpGs (Burns et al. 2006). In contrast, T7 evolutionary reversions did not exhibit any signature suggestive of a dinucleotide basis for attenuation (Bull et al. 2012). Nonetheless, we expect that the mechanism of attenuation in T7 will apply across other viruses, even if it is not the only mechanism operating in those viruses.
If our interpretation is correct for the mechanism underlying the effect of the recoding, an evolutionary response to overcome the effect might be duplication of the promoter immediately upstream of gene 10. Such a duplication would increase the number of 10 transcripts and, in the absence of ribosome saturation, would increase the amount of capsid protein. A duplication of a class-II promoter was observed during adaptation of a different phage (Springman et al. 2005), adding credence to the possibility of such an outcome. Yet the mutational origin of a promoter duplication may be highly dependent on surrounding sequences, so perhaps not feasible for all promoters.
We propose that translational coupling may explain why expression of genes downstream of 10 is suppressed by the recoding. Part of that inference is based on a mathematical model of translation. That model is necessarily simplified, however, and there are some obvious improvements needed to increase its realism. First, it assumes a fixed quantity of transcripts, when we know from our RNA-sequencing results that T7 transcripts increase rapidly during infection. Second, the model assumes a per-gene translation elongation rate, but does not model individual codons. A more sophisticated model that includes codon-level detail and the T7 life cycle would be needed to predict the fitness effects of codon deoptimization. Several life-cycle and molecular models of T7 have achieved limited success in predicting the phenotypic effects of genome manipulations (Endy et al. 1997; You et al. 2002; Kosuri et al. 2007; Birch et al. 2012), but none enable codon modifications or complex translation mechanisms such as coupling. The proteomics and RNA-sequencing data generated in this study should be useful in future high-resolution modeling studies that scale from the molecular level to that of viral fitness.
Although it is tempting to interpret the slowed translation as a consequence of using rare codons, which then use rare tRNAs, some recoding strategies used in other genomes suggest alternative possibilities. Codon deoptimization of GFP in E. coli initially yielded a range of protein expression effects, but these effects were eventually attributed to changes in mRNA secondary structure in the first 28 codons of the GFP sequence (Kudla et al. 2009). Codon changes beyond these first 28 codons had a weak effect on protein expression (Kudla et al. 2009). In our designs, we explicitly excluded these 5′-end codons from modification. Moreover, some attenuation designs with influenza virus, poliovirus, and arboviruses have achieved attenuation by merely shuffling codons within a gene or genome to create rare codon pairs (Coleman et al. 2008; Burns et al. 2009; Mueller et al. 2010; Nougairede et al. 2013; Shen et al. 2015). Since the numbers of each codon are not changed in those designs, the mechanism cannot be one of simple tRNA abundance. One possibility is suggested by recent work in yeast and Salmonella, whereby the degree of codon clustering is important to rapid translation (Cannarozzi et al. 2010; Gamble et al. 2016; Chevance and Hughes 2017). We did not attempt to control for codon-pair bias in our recoded T7 constructs. Further experiments will be needed to determine if the E. coli translation machinery is sensitive to changes in codon-pair bias.
The approach developed here should help elucidate other mechanisms of viral attenuation. For example, the timing of gene expression appears important to fitness: a reciprocal exchange of some middle and late genes had some major fitness effects, and those effects were not recovered on long-term adaptation (Cecchini et al. 2013). Ultimately, we envision a future in which an understanding of viral life history at the molecular level enables facile engineering of arbitrary fitness and alternative vaccine designs.
We thank I. J. Molineux for advice on and insight into T7. This work was supported in part by National Institutes of Health grants R01 GM088344, National Science Foundation Cooperative agreement no. DBI-0939454 (BEACON Center), and Army Research Office grant W911NF-12-1-0390. The Texas Advanced Computing Center provided high-performance computing resources. J.J.B. is supported as the University of Texas Miescher Regents Professor.
Author contributions: J.J.B., C.O.W., and D.R.B. conceived and designed the experiments and provided reagents and materials. D.R.B., M.L.P., and B.L.S. performed the experiments. B.R.J., D.R.B., M.L.P., and B.L.S. analyzed the data. B.R.J., C.O.W., and J.J.B. developed the model. J.J.B., C.O.W., D.R.B., M.L.P., B.L.S., and B.R.J. wrote the paper. The authors report no conflicts of interest.
Communicating editor: B. J. Andrews