Correlations with existing data

Next, we checked if our calculations were in agreement with published data on protein and mRNA abundances. We compared our results with two previously published studies that provide information on transcript and protein copy number for numerous S. cerevisiae genes [13], [26], by performing linear regression through the origin on the log-transformed values. The adjusted values, as well as the corresponding regression coefficients, were calculated for six pairs of datasets and the results are presented in Table 2. Scatter plots are presented in Figure 2. The results show that our model explains 84% of the variability in mRNA abundance and 97% of the variability in protein abundance reported by experimental studies. Such values are reasonably good, taking into account the differences in the particular yeast strains and laboratory protocols used, as well as the fact that our calculations are based on a few simplifications that can disrupt the final outcome. Moreover, values reported for our model do not stand out from those calculated for comparisons of two experimental datasets with each other, suggesting that the observed differences constitute the internal variability of the system, not a methodology error. To measure if our results suffer from systematic shift, we calculated the fold difference values for transcript and protein abundance comparisons with two experimental datasets (see Supplementary Figure S1). In general, our calculations slightly overestimate the transcript copy number and underestimate the protein copy number, in relation to published data. This is mainly caused by the assumption we made: that one yeast cell contains, on average, 36,000 transcripts. The transcript copy number used in both reference studies is originally taken from older research [27], which quantified the relative mRNA concentrations and transformed them into absolute copy number, assuming 15,000 as the total number of transcripts per cell. This estimation seems inadequate to us in the light of current discoveries, which are explained in the Materials and Methods.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Model results vs experimental studies.

The plots show the comparison of model parameters (left) and (right) with experimentally determined mRNA and protein abundances by two independent studies [13], [26]. The axes were log transformed. Calculated values are presented in Table 2. The distribution of the log-fold differences of the mRNA and protein concentrations reported by the model and reference studies are presented in Supplementary Figure S1.

https://doi.org/10.1371/journal.pcbi.1000865.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Model determined mRNA and protein abundances versus experimental studies.

https://doi.org/10.1371/journal.pcbi.1000865.t002

Transcript copy number is also problematic due to the wide discrepancies in mRNA levels reported by different studies [28]. Above mentioned mRNA concentration dataset [27] was obtained in a serial analysis of gene expression (SAGE) experiment and it is likely that such concentration estimates have low precision for low abundance mRNAs [13], [26]. On the other hand, it is hypothesized that SAGE is more accurate for abundant mRNAs when compared with other widely used technique: high-density oligonucleotide arrays (HDA) [26], [28]. Thus, we decided to compare mRNA concentrations calculated in our model with results obtained in genome-wide HDA experiment [29]. We performed linear regression through the origin on log-transformed data on mRNA abundance for 3769 genes. Scatter plot and the distribution of fold difference values are presented in Figure 3. The obtained adjusted value was 0.30 (see Table 2), meaning that parameter is able to explain only one third of the variability in mRNA abundance reported by this experiment [29]. This discrepancy is probably caused again by the experimental error. Parameter reflects mRNA concentration obtained by means of deep-sequencing, technique considered to be far more precise in measuring mRNA levels than other hybridisation or sequence-based approaches [23]. However, it is likely, that it is less precise for low abundance mRNAs, which may be seen in Figure S1 provided by Ingolia et al. [22]. This would explain why parameter better describes variability in mRNA concentrations obtained from SAGE than HDA experiments.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Calculated transcript abundance vs experimental studies.

Left plot: the comparison of model parameter with mRNA abundances determined by high-density oligonucleotide array (HDA) experiment [29]. The axes were log transformed. Calculated value for the comparison is presented in Table 2. Right plot: distribution of the log-fold differences of the mRNA concentrations reported by the model and reference study.

https://doi.org/10.1371/journal.pcbi.1000865.g003

In addition, we estimated that the cell-wide rate of translation for S. cerevisiae at 30C is 5.5 amino acids (aa) per second, which corresponds to an average time of translation for one codon of 183 ms. This is in agreement with experimental studies, reporting rates of 8.8 aa/sec and 5.2 aa/sec for fast-growing and slow-growing yeast cells, respectively [30]. It is worth noting that the obtained value is also within the range reported for proteins from other organisms, namely 6 aa/sec for human apolipoprotein [31], 0.74 aa/sec for rabbit hemoglobin [32], 5 aa/sec for chick ovalbumin [33], and an average translation rate of 7.3 aa/sec in cockerel liver [34].

Furthermore, it is reported in independent studies that the total amount of protein in a yeast cell varies from g [16] to g [35]. Based on known protein sequences and the molecular mass of particular amino acids, we can calculate the mass of each yeast protein. By multiplying this by the protein copy number and summing over all expressed yeast proteins, we estimated that the total mass of proteins in a yeast cell is around g. Although this number is smaller than values reported previously, it is still consistent taking into account the fact that we excluded from the calculations all transcripts with ribosome density , as our model cannot operate on such elevated values of this parameter. Most likely, results in very high level of translation, meaning that excluded transcripts would have large values, if they could be counted by our model. Thus, excluding these transcripts strongly affects the final mass of proteins in a yeast cell, diminishing it noticeably. Moreover, we must not forget that calculated values of the parameter reflect only the total amount of proteins produced from a given transcript, whereas the cell contains many other proteins produced in the past that are still present in the cell.

General features of the yeast gene expression system

Based on our results, we can draw the following conclusions concerning gene expression in S. cerevisiae:

First, half of the genes produce less than 2.73 transcripts per cell. The distribution of the transcript copy number is skewed with a long right tail: only 55 genes have more than 100 mRNA copies. Unsurprisingly, the top 20 genes with the highest values turned out to be either ribosomal proteins (18 genes) or enzymes engaged in glycolysis (genes YKL060C and YKL152C). One mRNA molecule is translated from 0.14 to 40,110 times, and the median is 257.9. Typically, one gene produces 677 protein copies; however, the most active genes may generate more than 2 million protein copies. Only six genes are common for the sets of the top 20 genes with the highest transcript levels and protein abundance. Among the 20 most highly produced proteins, there are 14 ribosomal proteins, two genes engaged in glycolysis (YCR012W, YKL060C), a highly expressed mitochondrial aminotransferase (YHR208W), alcohol dehydrogenase (YOL086C), and two cell wall proteins (YLR110C, YKL096W-A). There is only partial correlation between transcript and protein copy number and protein production does not necessarily follow the concentration of mRNA molecules (see Figure 4). We compared mRNA () and protein () abundance calculated in our model, by performing linear regression through the origin on log transformed data. Adjusted value calculated over the entire dataset (4192 genes with known ) was 0.59. This means that over 40% (in log space) of the variation in protein abundance cannot be explained by variation in mRNA abundance, suggesting some additional, posttranscriptional mechanisms of gene expression regulation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Correlation of mRNA and protein expression levels.

The plot shows the correlation between mRNA abundance (parameter ) and the number of protein molecules produced from a given gene (parameter ). We performed linear regression through the origin on log transformed data. Adjusted value calculated over the entire dataset (4192 genes of known ) was 0.59. This means that over 40% (in log space) of the variation in protein abundance cannot be explained by variation in mRNA abundance, suggesting some additional, posttranscriptional mechanisms of gene expression regulation.

https://doi.org/10.1371/journal.pcbi.1000865.g004

Next, we analysed yeast genes for expression strategies applied to produce the highest number of protein molecules. We prepared two datasets: 200 genes with the highest values () and 200 genes with the lowest values (). We compared the rest of the translation parameters between these two sets, performing a two-sided Mann-Whitney test. The mean value of most parameters differs between the two datasets in an intuitive manner: genes coding for highly abundant proteins usually produce more transcripts, which have a shorter time of translation (both and ), as well as stronger resilience to degradation and are occupied by more ribosomes per 100 codons. All differences are statistically significant with p-value (data not shown). Only one parameter appeared not to affect the number of proteins produced: the relative rate of initiating translation once the ribosome attaches to the free 5′ end of an mRNA molecule (p-value). Moreover, the Spearman's correlation coefficient between parameters and for the entire dataset is very weak (, p-value ).

Analogously, we analysed two datasets of 200 genes with the highest and lowest values ( and , respectively). According to the Mann-Whitney test, transcripts of higher ribosome density typically produce more protein molecules and have shorter times of translation (both and ). All differences are statistically significant with p-value (data not shown). In contrast to the result mentioned above, the shorter time for genes of the highest ribosome density is here caused mainly by elevated , while has little influence, but the difference in is still statistically significant (p-value ). Nevertheless, no significant correlation was observed between the parameters and measured over the entire dataset (p-value). The roles of and in modifying values of and are detailed in Supplementary Figure S2.

Furthermore, we studied, in detail, 20 genes from the set of 200 genes producing the highest number of proteins but with low transcriptional activity ( for all of them). Interestingly, these genes are involved in many distinct biological processes, with the notable exception of ribosome formation. The mechanism of their regulation, deduced from the values of the translational parameters, is almost the same for all genes. For instance, two parameters seem to play the main role in sustaining the high protein synthesis rate: relatively long mean life-time of the mRNA molecule, reaching up to several dozens hours (the maximal observed mean lifetime of a yeast transcript is 61 hours), and about four times shorter time of translation initiation, caused mainly by relatively high values. On average, the observed is one order of magnitude higher than the median for all yeast genes. The shorter pause between subsequent initiations results in elevated ribosome density and increased protein production rate. On the other hand, the total time of translation, as well as the mean elongation time, are unexpectedly long (i.e., slightly above the median value of all yeast genes (see Table 3)). This indicated that in cases of long-lived mRNAs, high transcriptional rates and usage of frequent codons are not required to achieve a high rate of protein synthesis. This strategy of expression constitutes an interesting but still inscrutable example of translation regulation, and further research should be carried out.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Translational parameters of 20 genes of low transcriptional activity and high protein production rate.

https://doi.org/10.1371/journal.pcbi.1000865.t003

Translation times and codon bias

In Supplementary Table S2 we present times of translation of individual yeast codons at 30C. We compared these values with codon optimality calculated by [36]. The value of measures whether the codon is preferred in highly expressed genes compared with all other codons encoding the same amino acid. is calculated as the odds ratio of codon usage between highly and lowly expressed genes. Figure 5 shows that there is negative correlation between value and translation time of a codon. However, while optimal codons have only short times of translation, non-optimal codons may be translated at both high and low rates. Linear regression model through the origin on log transformed values confirmed this conclusion: the obtained adjusted is only 0.15. This indicates, that translation speed may be the one, but not the only criterium for selection on codon bias. This is in agreement with other reports, discussed widely in the recent review [37]. Also, it has been shown that codon usage bias in yeast is associated with translation accuracy [38] and protein structure [36].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 5. Codon optimality vs translation time.

The plot shows the coparison of translation times in 30C of individual yeast codons with codon optimality values calculated by [36]. There is negative correlation between value and translation time of a codon. However, while optimal codons (high values) have only short times of translation, non-optimal codons may be translated at both high and low rates. Adjusted value obtained in linear regression model through the origin on log transformed values indicates, that translation speed may explain only 15% of variability in values.

https://doi.org/10.1371/journal.pcbi.1000865.g005

Translational parameters and protein interactions

Interacting proteins are often precisely co-expressed, presumably to maintain proper stoichiometry among interacting components [39]. For instance, it was shown that functionally associated proteins exhibit correlated mRNA expression profiles over a set of environmental conditions [40], [41]. Other studies report the co-evolution of codon usage of functionally linked genes [39], [42] and show that codon usage is a strong predictor of protein-protein interactions [43]. Our model provides far more information on translation regulation than mRNA expression profiles or codon usage alone, thus we decided to examine calculated parameters in a set of well-known interacting proteins.

As a model, we chose the 20S proteasome complex, built of 28 proteins. There are 14 genes in the yeast genome coding for proteasome subunits 1–7 and 1–7, and each subunit is present in the complex in two copies [44]. Only subunit is nonessential for the functionality of the complex and may be replaced by the subunit under stress conditions to create a more active proteasomal isoform [45].

The analysis of the translational parameters (see Supplementary, Table S3) shows that the mean translation time () of all proteins is similar and ranges from 194 to 259 ms. As all interacting proteins are of similar length, the total time of elongation does not vary much; the biggest observed difference between two proteins was 24 s. However, the level of transcription is more variable and ranges from 3.99 to 22.61 transcripts per cell. There is a considerable divergence of ribosome density (from 0.61 to 4.32), but regulation at the level of translation initiation (similar values of and variability of reaching two orders of magnitude) keeps the initiation time at the same level for all 14 proteins. The biggest observed difference of between two proteins equals 31 s. This results in congruent total times of translation , the difference between maximal and minimal values is only two-fold with a mean value of 73 s. Nevertheless, the observed differences in the mean lifetimes of mRNA molecules are huge, reaching up to 278 min. In consequence, the number of protein molecules produced is strongly variable, ranging from 318 to 11,185 molecules per cell, and this is surprising as the stoichiometry of the 20S proteasome would rather suggest equal amounts of all subunits. Indeed, for four proteasome proteins, the value of the parameter is almost the same, about 2,600 subunits of , , , and per cell. Similar values, which do not exceed the range 2,6001,000, were reported for subunits , , , and . Subunits , , , and are produced to less than 1,100 copies, while the rate of protein synthesis of subunits and is 5,481 and 11,185 molecules per cell, respectively. To maintain the number of different subunits at the same level, the high translation rates of and may be balanced by post-translational regulation, presumably by elevated protein degradation. Conversely, the reduced translation rate of , , , and may be compensated at the level of transcription, for instance by more frequent transcription initiations. In addition, the limited number of these subunits, as well as the relatively short life-time of their mRNAs, makes them ideal candidates for regulators of the abundance of proteasome complexes.

Theoretical model of translation

The molecular mechanism of translation was well characterised previously [49]. However, for the purpose of this research, we must consider the process both at the single transcript and genome-wide levels. Quantifying the process of protein biosynthesis engages vast array of data, some of which is incomplete or missing. Thus, the following assumptions and simplifications must be made: (i) the pools of all molecules participating in translation (mRNA, tRNA, ribosomes, translation factors, and so on) are constant, and molecules diffuse without restraint; (ii) all transcripts derived from the same gene have identical sequence, i.e., there is no alternative splicing and/or posttranscriptional modification; and (iii) the elongation process is never interrupted, and it always ends by producing a full-length protein molecule (note, that experimentally estimated procesivity of translation in yeast was 99.8–99.9% [50]). When these assumptions are satisfied, the model is as follows:

Let be the set of all transcripts present in the yeast cell at the moment of observation. We can make a partition of the set into subsets, each containing transcripts of identical sequence. Thus, denotes the number of transcriptionally active genes in the cell. To each gene (subset), we attribute the index and define as the number of transcripts in the subset. The variable is reflected though by the transcriptional activity of a gene.

Let be the total observed time of synthesis of one protein molecule from a transcript belonging to the subset. We define it as:(1)where denotes the time required for translation initiation, and is the total time of the elongation process.

We define as the time interval from the point when the free 5′ end of a transcript becomes available for ribosomes to the moment when a ribosome finds the initiator AUG codon and the entire complex enters into the elongation phase. The inverse of the initiation time is initiation frequency :(2)If these frequencies are multiplied by a brief time interval , one obtains the probabilities that the initiation process will occur during time interval . We assume that the initiation of translation follows the scanning model [51], which postulates that the small ribosomal subunit enters at the 5′ end of the mRNA and moves linearly, searching for the initiator AUG codon; once it finds it, the elongation process begins. We define as the relative binding rate of free ribosomes to the 5′ end of the transcript, and assume it is proportional to the concentration of the transcript (see Eq.10). This means, that in our model the binding constants of ribosomes are the same for all mRNAs. Contrary, the process of 5′UTR scanning by the ribosome is not straightforward, as there are many intrinsic features of mRNA molecules that can considerably delay or hasten the start of elongation (for review, see [1]). Sometimes, the ribosome detaches from the mRNA molecule before reaching the initial AUG, and the process must return to the point when a ribosome binds at the 5′ end. To describe the efficiency of the scanning process by one numerical parameter, we normalised by the rate of binding of free ribosomes :(3)The calculated parameter describes the rate of successful accomplishment of initiation on the transcript once the ribosome-mRNA complex is formed. Its value reflects the relative capability of an mRNA molecule to be translated, regardless its expression level. The rates and are calculated in relation to all studied transcripts, thus they can only be compared within one particular analysis.

The time (see Eq.1) is defined as a time interval from the recognition of the initiator AUG codon by the ribosome to the moment when the last peptide bond of a protein molecule is formed. Each elongation event consists of two main steps: (i) finding the correct tRNA molecule, and (ii) formation of the peptide bond and translocation. The time required for the first event is much larger than for the second. In fact, the second step is almost instantaneous [52]; thus, the times needed for transpeptidase and translocation reactions can be neglected, and time may be simplified to:(4)where is the time of translation of the codon, and is the number of codons in the coding sequence of the transcript.

Translation times for all yeast codons, as well as the values of and , can be calculated on the basis of existing data (see below). These values can also be used to calculate times and the rest of the model parameters , , and , if the numbers of ribosomes attached to the mRNA molecules are known. Here, the reasoning is as follows:

Let be the number of ribosomes attached to the transcript. We introduce the measure of ribosome density , defined as the number of ribosomes attached to the transcript per 100 codons:(5)One ribosome occupies ten codons of a mRNA molecule [53], and the E site of one ribosome can be immediately adjacent to the A site of another ribosome [54]. This means that the maximum possible value is . Next, the attachment of a ribosome to the 5′ end is possible only if it is not occupied by other ribosomes. Thus, the most efficient mRNA sequences should have . Nevertheless, the majority of observed values are much smaller, meaning that there are usually gaps of varying length between attached ribosomes. As the exact positions of ribosomes on a particular transcript cannot be deduced from the data, we must operate on the averaged gap lengths, defined as the quotient of the transcript length and number of attached ribosomes . The length of a gap measured in codons is meaningless, as each type of codon has a different translation time. However, the gap can be calculated as the sum of translation times of these codons, becoming an adequate measure of the time interval between individual translation initiation events on a given mRNA molecule. This time is actually a delay from the best possible initiation frequency and reflects the efficiency of the initiation process. In principle, this is the time (see Eq. 1) expressed in the same time units as the translation times of particular codons:(6)Note that due to unknown ribosome positions on a transcript, both the gap length and time of its translation are averaged. The rest of the parameters (, , and ) can be calculated based on , as shown in Eq. 1, 2, and 3.

Calculating model parameters

The S. cerevisiae coding sequences used in our calculations were downloaded from the Saccharomyces Genome Database [55] (accessed 25 June 2009). For each gene, we determined the values of , , , , , and on the basis of the recent research of Ingolia et al. [22], quantifying simultaneously mRNA abundance and ribosome footprints by means of deep sequencing. The study was done for the yeast strain BY4741 grown in YEPD at 30C. In the first step Ingolia et al. performed deep sequencing on a DNA library that was generated from fragmented total mRNA in order to measure abundance of different yeast transcripts. Next, they applied a new ribosome-profiling strategy based on the deep sequencing of ribosome-protected fragments. This resulted in a dataset of 4,648 reliable transcripts (for the definition of “reliability”, see Supplementary Materials of [22]) that was used as an input in our research. For each transcript in the dataset, the following values were attributed: , raw count of mRNA-seq reads aligned to transcript coding sequence (CDS); , density of mRNA-seq reads in reads per kilobase per million CDS-aligned reads (RPKM); , raw count of ribosome CDS-aligned footprints; and , density of ribosome footprints in reads per kilobase per million CDS-aligned reads. Next, the relative numbers of reads counted in RPKM were transformed into the transcript copy numbers. Normally, for each transcript , is defined as:(7)where is the length of the transcript CDS in codons, and is the sum of all mappable reads [56]. Assuming uniform distribution of the mappable reads across the transcriptome coding sequences, the probability of observing reads on the transcript CDS of length in attempts corresponds to the fraction of the transcriptome composed of the transcript:(8)where is the sum of all CDS of the transcriptome in base pairs. The meaning of was explained in the previous section. We can substitute final RPKMs to get:(9)

Although the length of the entire transcriptome was estimated as nucleotides [57], deriving is more problematic, as little is known about the accurate boundaries of non-coding elements in transcript sequences [9]. There were some attempts to determine the length of UTRs on a global scale in yeast [58], [59], but the results show that even the length of transcripts derived from the same gene of the same yeast strain cultured in the same growing conditions may vary considerably. This causes the discrepancies between reported transcript lengths by these two studies, making the analysis at the level of individual genes difficult and inaccurate.

To overcome this problem we use , the relative rate of binding of free ribosomes to the 5′end of a given transcript (see Eq.3). This rate corresponds to the fraction of transcript in the set of all transcripts. By substituting as shown in Eq. 9, we obtained the following relation:(10)where is the sum of all densities of mRNA-seq reads. Thus:(11)

The next step was to calculate (the absolute number of translationally active ribosomes attached to the transcript), and (the measure of ribosome density, as defined in Eq.5). The dataset used provides information only on ribosome footprints aligned to the coding sequences. However, in practice, there were some exceptions to this rule, caused mostly by the presence of ORFs in the 5′UTR sequences [22]. Due to the lack of data and aforementioned difficulties in determining exact transcript length, this fact is not taken into account in our analysis. Furthermore, we defined as the number of all ribosomes in a yeast cell and as the fraction of ribosomes participating in the process of translation at the moment of observation. In contrast to raw mRNA-seq reads, the distribution of ribosome footprints is not uniform across the transcriptome, due to differences in genes translational activity. Thus, the probability of observing a ribosome attached to the transcript corresponds to the fraction of all ribosome footprints composed of the raw footprints in the transcript, . This probability is equal to the ratio of all ribosomes engaged in translation of transcripts of type and the number of all occupied ribosomes in the cell:(12)Thus, ribosome density for the transcript can be calculated as:(13)

Global parameters estimation

Three parameters must be estimated to transform relative numbers of transcripts and ribosomes attached to them into absolute measures. These parameters are , the total number of mRNA transcripts in a yeast cell; , the total number of ribosomes in a yeast cell; and , the fraction of ribosomes participating in the translation process at the moment of observation. There are many studies concerning the quantitative measurement of yeast cells, and we used the Bionumbers database [60] to extract these data.

Two reports provide an independent, yet coherent, estimation of the total number of ribosomes: 187,00056,000 [9] and 200,000 [57] molecules per cell. In this study, we decided to set to 200,000. The value of 85% was established for the parameter , as stated in experimental studies [61], [62]. The number of all transcripts in a cell is more problematic. Many contemporary studies assume that a yeast cell contains 15,000 mRNAs per cell on average [27], [63], which is based on estimations done over 30 years ago [64]. Current research, based on more up-to-date techniques (e.g., in situ hybridisation or GATC-PCR) argues that the number should be at least doubled [65] or even quadrupled [62]. We decided to use the value of situated between these estimates and equal to 36,000. This number was also confirmed by other studies [65].

Assuming , , and , we obtained the mean ribosome density equal to 1.66 ribosomes per 100 codons. This is in agreement with experimental analysis, which reports that, on average, there is one ribosome per 156 nucleotides, corresponding to a density of 1.92 ribosomes per 100 codons [61]. Moreover, it was estimated that mRNA constitutes 5% of the total amount of RNA present in a cell, and the RNA∶DNA ratio is 50∶1 [57]. Assuming the yeast genome size of nucleotides, the expected length of the entire transcriptome would be nucleotides. Thus, the length of all transcribed coding sequences can be defined as:(14)The meanings of , , and are explained above. Thus, the calculated length of all coding sequences equals nucleotides. This would suggest that non-coding elements constitute, on average, more than 50% of a transcript. In conclusion, it seems that the chosen parameter values generate reasonable measures of the global characteristics of the yeast cell.

Determining absolute times of translation

In the previous section, we calculated the values of , , , , and for each gene. To determine the absolute times of translation , , and , we need to know the times of translation for individual codons. To achieve this goal, we adapted a model proposed for Escherichia coli [66] to the yeast system. Here, we briefly present the model and all of the necessary changes we made. For a description of the derivation, see the original paper.

The transport mechanism in the cytoplasm is diffusion, thus the aa-tRNAs act as a random walker, and the ribosomes on mRNAs with vacant A sites are the targets. We assume a yeast cytoplasm volume [67]. We divide it into walker occupation sites, where:(15)and is a measure of the walker size. The values of used previously [66] were determined separately for individual E. coli aa-tRNA molecules [68]. As we are not aware of any similar reports for S. cerevisiae, we decided to use for all yeast codons, which is the mean of the E. coli values. The average time that elapses before the arrival of a walker is defined as:(16)where is the characteristic time of the walker, associated with its transition from one cellular occupation site to the other. It depends on the size of the walker and its diffusion coefficient :(17)The measures of were taken directly from [66]. As this value depends only on the accepted amino acid, we assumed that the difference in size between yeast and E.coli tRNA molecules is negligible. In Eq.16, stands for the probability that a tRNA-aa molecule of type arrives at an open A site in the time interval and is proportional to the number of walker occupation sites containing the walker:(18)We assume that the number of the molecules of the walker is proportional to the number of corresponding tRNA genes of type , which is reasonable, as it was shown that in yeast the concentration of the various tRNA species is largely determined by tRNA gene copy number [69]. In particular, the calculated correlation coefficient between tRNA gene copy number and experimentally determined tRNA abundance for a subset of 21 tRNA species equaled 0.91. According to [57], the RNA-DNA ratio is 50∶1 and tRNA constitutes 15% of the total amount of RNA in a yeast cell. Assuming a genome size of nucleotides, the total cellular tRNA size is nucleotides. When divided by the average tRNA molecule size (74.5 nt) we obtain the number of tRNA molecules in a cell equal to 2,818,792. Next, this number was multiplied by the fraction of all tRNA genes composed of the tRNA genes of type , yielding the absolute amount of particular tRNA molecules in a cell. Gene copy number and predicted decoding specificities of yeast tRNAs were taken from Table 1 of [69]. The values of all presented parameters for individual tRNAs are gathered in Supplementary Table S4.

All 61 codons that code for the 20 amino acids have one or more aa-tRNAs and varying numbers of near-cognates. Near-cognates are defined as having a single mismatch in the codon-anticodon loop in either the 2nd or 3rd position. Since some cognate tRNAs have a mismatch in the 3rd position, these tRNAs are excluded from the set of near-cognates [70]. The theoretical background of the model is based on the observation that the translation rate of a codon reflects the competition between its non-cognate, near-cognate and cognate aa-tRNAs [71], and that such nonspecific binding of the tRNAs to the ribosomal A site is rate-limiting to the elongation cycle for every codon [72]. The model of Fluitt et al [66] introduces two competition measures, and , being the quotients of the sum of arrival frequencies of near-cognates vs. cognates and non-cognates vs. cognates, respectively. For each codon, we determined its cognates, near-, and non-cognates (based on [69]) and calculated the competition measures and (see Supplementary Table S2).

According to [66], the average time to add an amino acid coded by the codon to the nascent peptide chain can be calculated as:(19)where is the average time to insert an amino acid from a cognate aa-tRNA, and and are the average time delays caused by the binding attempts of near- and non-cognate tRNAs, respectively. Based on existing data and assumption that the activation energies for the various reactions do not vary much, Fluitt et al [66] showed how to calculate the values of , and at any given temperature. Table 4 contains these values for S. cerevisiae at 20, 24, 30, and 37C. Next, we calculated translation rates for all yeast codons at the four different temperatures (see Supplementary Table S2). However, as the main part of our analysis is based on the ribosome footprints measured at 30C, in further calculations we use only the values of estimated at this temperature. The last step was to calculate times for individual genes, as described in Eq.4.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Time of tRNAs insertions at four different temperatures.

https://doi.org/10.1371/journal.pcbi.1000865.t004

Ribosome queuing

It has been found that subsequent ribosomes are loaded onto the transcript sufficiently fast to make them interfere with each other, leading to ribosome queuing [73]. This phenomenon is usually caused by the presence of rare codons clusters in CDS, although other sequence features may also be very important [74]. Such elongation pauses may have distinct consequences, for instance ORF shifting or ribosome dissociation, often followed by decay of the mRNA and partly completed protein products [75]. Moreover, stalled ribosomes generate a false picture of a transcript translational activity, elevating the observed ribosome density in relation to the actual frequency of translation initiation events. For these reasons, we decided to reduce the dataset to the transcripts on which ribosome queuing does not occur. We wrote a simple program that simulates the ribosomes translocation along a transcript sequence. A ribosome moves from one codon to another only if it has spent a required amount of time for translation of the current codon (taken from Supplementary Table S2) and the subsequent codon is vacant. The successive ribosome attempts to attach to the initial AUG codon after the elapse of time interval , calculated as shown in Eq.6. The cumulative time of the movement is calculated for each ribosome separately. If this time is identical for each ribosome, translation is believed to pass without ribosome queuing. If the time is different, namely, the first ribosome moves faster than the rest, it means that some sequence features allow ribosome stacking under the assumed conditions (i.e., temperature and translational parameters). If the attachment of subsequent ribosomes is prevented by very slow translation of the first few codons, we consider it a particular case of ribosome queuing and reject all such transcripts.

Calculation of protein abundances

To enrich our dataset, we estimated the total number of proteins produced from a given transcript. Considering the mRNA molecules as a decaying quantity, we defined as the mean lifetime of the transcript expressed in time units:(20)where is the half-life of the transcript. Assuming that each translation event happens independently, we calculated the abundance of the protein as the number of translation initiation events that happen during the life-time of the transcript multiplied by the its copy number:(21)The dataset of mRNA relative half-lives is provided in the Supplementary Materials of [25]. In our calculations, we used the times t0 measured at exponential growth in YPD medium for 5,718 ORFs. It was determined experimentally by independent studies that the absolute mRNA half-life of the yeast gene YOR202W (HIS3) ranges from 7 (at 24C) [76] to 11 min (at 30C) [77]. Assuming the mean value of 9 min for this gene, we can quantify the half-lives for the rest of the genes in the dataset, as well as the values of and .

Calculations summary

Based on the presented reasoning, we calculated translational parameters for the majority of yeast genes. In particular, parameter was calculated on the basis of yeast coding sequences downloaded from [55]. Parameter was obtained from Eq.11, where values of and were taken from the experimental study [22], and was set to 36,000, as estimated by [65]. Parameter was obtained from Eq.13, where values of , were taken from [22], was set to 200,000, based on [57], and was set to 0.85 as stated in [61], [62]. Parameter was calculated from Eq.5. Parameter was calculated from Eq.4, based on yeast coding sequences downloaded from [55] and the values of translation times of codons , calculated as shown in Eq.19. The values of , and (at 30C) used in Eq.19 were calculated as shown in [66], and the values of and were calculated separately for each codon as shown in [66], by substituing the number of its cognates, near-, and non-cognates tRNAs determined on the basis of [69]. Parameter was obtained from Eq.6, and from Eq.2. Parameter was obtained from Eq.10, where values of and were taken from the experimental study [22]. Parameter was obtained from Eq.3 and then normalised by its maximal value reported for the gene YLL040C. Total time of translation was calculated as stated in Eq.1. Mean time required for elongation of one codon of the transcript () was calculated by dividing elongation time by the length of this transcript in codons . Parameter was obtained on the basis of relative half-lives for yeast transcripts reported by [25] and mRNA half-life of the yeast gene YOR202W, assumed to be on average 9 min [76], [77]. Parameter was obtained from Eq.20, and from Eq.21. The meaning of all variables was presented at the beginning of this section.