Because start codons can only be occupied by a single ribosome at a time, mRNA-specific protein synthesis rates are controlled by translation elongation whenever the rate of translation initiation attempts exceeds the rate with which ribosomes elongate away from the start codon (Chu et al., 2014). The rate of ribosome movement is sequence dependent and controlled by codon usage patterns (Gardin et al., 2014), mRNA secondary structure (Tuller et al., 2011), the charge of the nascent chain within the ribosomal exit tunnel (Charneski and Hurst, 2013), and the readiness of particular amino acids to undergo peptidyl transfer (Pavlov et al., 2009). Although ribosome speed can thus be affected by many different mechanisms, models which exclusively consider tRNA:codon interactions as determinants of ribosome speed can predict protein yields with high degrees of accuracy (Chu et al., 2014). This, together with the recent finding that ribosome collision sites are strongly enriched near the start codon (Diament et al., 2018), predicts that codon-dependent ribosome speed near the start codon is an important determinant of translational control. We therefore initiated our analyses of elongation-controlled mRNAs in the yeast transcriptome by analysing codon-dependent ribosome speed immediately following the start codon, as a non-exhaustive but useful potential indicator of mRNAs which might be subject to control by translation elongation.

We used our published, model-based decoding time estimates for each codon (Chu et al., 2014) to calculate the speed of decoding of the first ten codons following the start codon (approximately the span of one ribosome) throughout the entire yeast transcriptome. In addition to the decoding speed of the observed sequences, we also computed decoding speeds of random sequences encoding the same peptides as the actual genes, and determined the proportion of sequences that were more extreme in their speed than the observed sequence. If the observed sequence was faster than the mean decoding speed of the random sequences “more extreme” meant the proportion of random sequences faster than the observed sequence, whereas if the observed sequence was slower than the mean random sequence “more extreme” meant the proportion of random sequences slower than the observed sequence. The negative logarithm of this proportion (p) gives an estimate of the probability of observing the actual decoding speed of a gene if codon usage for this gene was entirely random.

Plotting p against absolute speed for all yeast genes results in a volcano plot (Figure 1A) which displays a clear skew towards fast sequences, reflecting the well-documented general codon bias in the yeast genome (Behura and Severson, 2013). To explore the observed distribution further, we selected the fastest, slowest, and central 5% of genes and analysed these groups with respect to particular features. The fastest group was associated with significantly higher expression levels than the other two groups, and the very highest expressed proteins were found exclusively within this group (Figure 1B). The fastest group also showed strong association with specific GO terms including cytoplasmic translation, consistent with the observation that ribosomal proteins show particularly high codon usage bias (Sharp et al., 1986). In contrast, the slowest group of genes was not associated with lower expression levels compared to the central group and was much less strongly associated with specific GO terms. This indicates that fast sequences are favoured in the yeast genome through evolutionary selection on genes whose functions require particularly high expression. In contrast, we detected no significant selection for slowly decoded sequences.

This observation is relevant to the question how and why elongation-controlled mRNAs arose in the yeast genome. A feature that places an mRNA under control of translation elongation is a combination of (relatively) high initiation and (relatively) slow elongation rates. In a population of mRNAs where initiation and elongation rates evolve randomly and independently, combinations of fast initiation and slow elongation could arise simply by chance. In this case, elongation-controlled mRNAs would be enriched among those that contain particularly slow codons, where the probability of a high initiation/elongation rate ratio is highest. Our observation that such mRNAs are not widely selected for implies that either elongation-controlled mRNAs themselves are also not widely selected for, or that selection for elongation-controlled mRNAs occurs by combined selection on initiation-and elongation-rate determining features.

In recent work, Archer et al. (2016) used translation complex profiling, a variant of the ribosome profiling approach (Iwasaki and Ingolia, 2017), to study footprints on mRNAs derived solely from small ribosomal subunits (SSUs). Since our previous findings indicate that mRNAs become elongation controlled when initiating ribosomes are prevented from accessing the start codon because the previous ribosome has not yet liberated this site (Chu et al., 2014), we expected that scanning 40S subunits may form queues 5’ of the start codon on such mRNAs. Queueing 40 S subunits should be detectable in the Archer et al. dataset. If we assume that initiating ribosomes physically cover around 30 nucleotides centred around the start codon itself, the centre of any small ribosomal subunit queueing immediately upstream should be somewhere in the region spanning nucleotides −60 to −15 upstream of the AUG (note that due to the trailing mass of translation initiation factors, scanning 40 S subunits produce larger footprints with less clearly defined boundaries compared to elongating ribosomes (Archer et al., 2016).

We retrieved the footprinting data generated by Archer et al. from the GWIPS-viz database (Michel et al., 2014), processed the data for each gene using a peak calling algorithm, and classified the genes into those containing an identifiable second peak of SSU footprints within the −60 to −15 region adjacent to the main SSU footprint peak over the ORF start codon, and those without such a peak (Figure 2A). Because upstream footprints can only occur on sufficiently long 5′-UTRs, we restricted this analysis to those mRNAs having 5′-UTRs longer than 45 nucleotides. All of the 3,477 genes with a 5′-UTR length above this cut-off showed a detectable small subunit peak centering round the start codon. In addition, 638 (18.3%) of these genes also showed a detectable second upstream peak, with a mean peak location around 50 nucleotides from the start codon (Figure 2A).

We propose that the mRNAs displaying a second 40 S peak correspond to the elongation-controlled transcriptome (further experimental evidence for this is given below). However, our analyses could be confounded by other mechanisms attracting small ribosomal subunits to sites upstream of the main start codon, notably by translation initiation events on upstream open reading frames (uORFs). We therefore analysed the relationship between uORFs and apparent queuing SSU peaks in more detail. The proportion of genes with uORFs is similar in the gene subsets with and without a second peak (p = 0.67 by Fisher’s Exact Test), and second SSU peaks are therefore not generally associated with uORFs. Of those genes displaying second SSU peaks and also containing uORFs, the majority of uORF locations is outside of the −15 to −60 nt analysis window, and the observed second footprint peaks can thus not be results of uORF initiation (Supplementary Figure S1). Genes which do have uORFs within the analysis window comprise less than 5% of the “second SSU” set. Spot checks with representative genes (Supplementary Figure S1) indicate that for some but not all of these the second SSU footprints may indeed arise from uORF initiation events rather than from ribosome queuing, in particular those where the main ORF AUG footprints are substantially lower than the queuing footprints. In addition to uORFs with AUG start codons, yeast contain comparable numbers of uORFs with non-AUG start codons which are however less efficiently recognised (Spealman et al., 2018), and which could therefore contribute 40S peaks in locations where they could be mis-interpreted as queueing ribosomes at lower levels than AUG uORFs.Overall this suggests the proportion of genes mis-annotated as subunit queuing genes in our analyses due to interference from uORFs is < 5%.

When mapped against the 5′-decoding speed of the yeast transcriptome, the corresponding mRNAs appear randomly distributed without any strong association with specific speed properties (Figure 2B). Based on the considerations outlined in the previous section, this suggests that such mRNAs may have arisen as the result of selection on particular initiation/elongation rate ratios, rather than selection solely for slow movement. To explicitly test the assumption that mRNAs displaying a second peak in the Archer et al. dataset correspond to elongation-controlled mRNAs, we manipulated the ribosomal decoding speed of a number of yeast genes at their normal chromosomal loci in vivo. This involved designing and synthesizing speed-optimised gene sequences, replacing the original genes in the yeast genome using a CRISPR-based approach, and finally assessing the effect of speed optimisation on mRNA and protein expression levels for this gene. For any gene where ribosome speed restricts achievable translation initiation rates, we expect increases in decoding speed to increase the protein/mRNA ratio, but this should not be the case for initiation-controlled genes.

This assay, which we had originally used to demonstrate elongation control of the yeast HIS3 gene (Chu et al., 2014), is laborious and cannot be applied to large numbers of genes, but we reasoned that it would be a good way of verifying predictions from the Archer et al. dataset using a smaller number of genes. In addition to the HIS3 gene, we applied the assay here to six additional genes, for which we could source high quality antibodies and which span a range of decoding speeds (Figure 3A). Importantly, these genes were selected before the results from the SSU foot-printing analyses had been completed, and the experimenters were blind to the results from these analyses throughout the experimental procedure.

The results of these assays are displayed in Figures 3B–H. The HIS3 gene (Figure 3B) has a very short 5′-UTR of 10 nt (Nagalakshmi et al., 2008), which is too short to accommodate a queueing 40S subunit. Although the protein/mRNA ratio for the His3 protein is increased by speed optimising the gene, confirming that the mRNA is elongation-controlled, this gene therefore does not show a second SSU peak in the Archer et al. dataset. The remaining six genes all have 5′-UTR lengths that should permit queueing ribosomes to form footprints, but only one of these (RAS2) actually showed clear evidence of a second footprint. RAS2 is also the only one of the six additionally tested genes for which the replacement assay led to a clear increase in the protein/mRNA ratio, confirming the link between elongation control of protein abundance and the presence of an upstream SSU footprint. CDC10 showed a weak increase in the protein/mRNA ratio as well as displaying a weak queueing ribosome peak, however these signals were obscured by the very low expression levels for this gene. The other four genes neither showed increased mRNA/protein ratios, nor evidence of upstream SSU footprints. Thus, for the sample of genes tested with this assay, the codon replacement assay and the SSU footprinting dataset arrive at a unanimous classification of genes as either elongation or initiation controlled. The observation that 18% of yeast genes with sufficiently long 5′-UTRs show footprints for queueing small ribosomal subunits thus suggests that this number is a good first estimate for the proportion of elongation-controlled genes in the yeast genome generally.

It should be noted that this proportion is specific to the physiological conditions of our assays, i.e., logarithmic growth in rich medium. Any change in growth conditions that entails, for example, a specific reduction in translation elongation rates would lead to the transfer of additional mRNAs into the elongation-controlled pool and vice versa. Under conditions of strong regulation (e.g., when translation initiation is essentially halted by phosphorylation of eIF2 during amino acid starvation, or when translation elongation is stalled globally by activation of eEF2 kinases) all mRNAs should become controlled by either translation initiation or elongation.

While substantial regulation of translation factor activity is predicted to eventually affect all transcripts, we reasoned that limited regulation of initiation factor activity should preferentially affect initiation-controlled mRNAs, whereas limited regulation of elongation factors should preferentially affect elongation-controlled mRNAs. The co-existence of initiation- and elongation-controlled sets of mRNAs in a cell could thus split the transcriptome into two separately addressable regulons.

To test the concept that such separate regulation can indeed occur, we designed a series of reporter constructs expressing two luciferases (Figure 4A). mRNAs encoding the two luciferases were placed either under initiation control (by combining an inefficient, uORF-containing 5′-UTR with an efficient, codon-optimised ORF), or under elongation control (by combining an efficient 5′-UTR with an inefficient, slow-codon containing ORF). The use uORFs to limit translation initiation activity on the main ORF means that the initiation controlled luciferase is synthesised following re-initiation, whereas the elongation-controlled luciferase is synthesised following normal translation initiation. However, we have previously demonstrated that the hand-over from elongation to initiation control on re-initiating constructs quantitatively matches model predictions built on the assumption of normal initiation, so that we do not expect that the choice of a re-initiation system fundamentally affects the behaviour of our constructs.

In construct I (Figure 4A), we paired an initiation-controlled firefly luciferase with an elongation-controlled Renilla-luciferase gene, whereas the control regimes were inverted in construct II. In a control construct, C, both luciferases were placed under elongation control. By normalising the expression ratios observed with constructs I and II to construct C, and by ensuring that constructs I and II changed expression in opposite ways, we could reliably separate changes in expression resulting from the regulation of translation initiation or elongation from expression changes resulting from other regulatory events including transcription and protein turnover.

We measured changes in the expression ratio of the two luciferases in yeast strains containing a deletion of one of two identical isogenes for translation initiation factor 4A (tif1Δ, mimicking limited regulation of translation initiation), or a deletion of one of two identical isogenes for translation elongation factor 1A (tef1Δ, mimicking limited regulation of translation elongation). We observed that in the tif1Δ strain, where initiation activity was reduced, the expression ratio changed in favour of the elongation-controlled luciferase in both constructs I and II (Figure 4B). In contrast, in the tef1Δ strain, the expression ratio changed in favour of the initiation-controlled luciferase in each construct. We further tested the behaviour of our constructs upon addition of the translational inhibitor, cycloheximide. This inhibitor is generally used at high concentrations as a translation elongation inhibitor that interferes with the translocation step (Schneider-Poetsch et al., 2010), although at low concentrations in vivo it is thought to act as a translation initiation inhibitor before inhibiting elongation at higher concentrations (Oleinick, 1977). In our reporter constructs, the ratios of the two luciferases changed with cycloheximide addition in a concentration dependent manner consistent with this notion of dual inhibition. At low concentrations, the elongation-controlled luciferases were initially favoured but this trend became reversed at higher concentrations (Figure 4C). Thus, both genetic and chemical manipulation of translation initiation and elongation rates indicated that transcripts under distinct control regimes can indeed be separately addressed by translational control mechanisms.

Having shown that initiation- and elongation-controlled yeast transcripts could in principle be addressed via distinct regulatory mechanisms, we wished to explore in how far this was used during natural gene expression regulation. For this purpose, we measured the luciferase ratios of our reporter assays under a number of different growth and stress conditions. We observed conspicuous divergent shifts of the luciferase ratios with changes in temperature (Figure 5A), which were consistent with relatively slower elongation rates at lower temperatures, and increased elongation rates at higher temperatures. Although this is to our knowledge the first report of regulation of translation elongation in sub-physiological temperatures in yeast, these findings mirror the known regulation of translation elongation during cooling in mammalian cells (Knight et al., 2015) and tissues (Bastide et al., 2017).

We also tested a number of stress conditions including oxidative, osmotic and cell wall stresses. Most of these did not affect the expression balance between the initiation- and elongation-controlled reporter genes, with the conspicuous exception of the two osmotic stress conditions we tested (0.5 M NaCl and 1 M sorbitol, Figure 5B). For both of these conditions, the observed expression pattern was consistent with a clear initiation block. The response to osmotic stress has been studied in detail by a number of authors and this response is known to involve regulation at both the levels of translation elongation (Teige et al., 2001) and initiation (Garre et al., 2012). The elongation response, which involves activation of the Hog1 kinase, is known to be transient persisting for less than 10 min under a low intensity stress (0.2 M NaCl, Mettetal et al., 2008) and for around 30 min for a stronger stress (0.8 M NaCl, Marques et al., 2006). The dynamics of the regulation of initiation has not been studied in detail but it is known that this component is important for re-establishing translation following the initial sharp downregulation upon onset of the stress. Since the measurements in Figure 5B are performed under steady-state stress conditions (i.e., after the initial, transient regulatory events have passed), these findings indicate that an initial translational arrest due predominantly to the inhibition of translation elongation is followed by a steady-state response in which translation elongation recovers, but translation initiation is partially inhibited compared to pre-stress conditions.

We sought to more closely examine the responses of initiation- and elongation-controlled mRNAs during the initial, elongation-regulating adaptation to osmotic stress conditions. We attempted to measure the luciferase ratios from our reporter constructs, but found that due to different protein half-lives of the firefly and Renilla luciferases this assay is not suitable for application outside of steady-state conditions (data not shown). However, Lee et al. (2011) generated detailed, genome-wide mRNA and protein abundance data at timepoints immediately following the onset of a 0.7 M NaCl stress, and we used these data to study the evolution of protein/mRNA ratios over time following the onset of the stress. For this analysis, we initially split the yeast transcriptome into elongation- and initiation-controlled sets as shown in Figure 2A (again only using the subset of mRNAs with 5′-UTR lengths >45 nt, for which footprints of queueing SSUs are observable). We then clustered each of the two sets according to the time-evolution of the protein/mRNA ratios, in order to determine subsets of mRNAs that showed similar behaviour (Figure 6). For a cluster number of four, this analysis revealed qualitatively very similar clusters for both initiation- and elongation-controlled mRNAs. However, although the behaviour was qualitatively similar, a quantitative comparison showed that, on average, within each cluster the elongation controlled mRNAs were more repressed at the initial time point (t = 10 min) than the initiation controlled mRNAs. At this time point, the Hog1 response is known to be maximal (Marques et al., 2006). For all four clusters, this difference at the initial time point was statistically significant (Figure 6). Overall, these analyses showed that under a stress condition where the regulation of translation elongation makes a strong contribution to the cellular response, the initiation- and elongation-controlled subsets of the transcriptome respond with distinct dynamics.