Dissecting the expression dynamics of RNA-binding proteins in posttranscriptional regulatory networks

RBPs Show High Abundance and Tight Regulation at the Protein Level.

To compare and understand the differences in the gene expression dynamics of RBPs with other protein coding genes in S. cerevisiae, we first compiled the set of RBPs and non-RBPs as described in Materials and Methods (Fig. 1B). This allowed us to define a set of 561 proteins in yeast as those that encode for RNA-binding proteins and the remaining 5,685 proteins (from the complete set of protein coding genes) as non-RNA-binding proteins. We also collected high-throughput data documenting various dynamic properties of messenger RNA transcripts and their translated protein products in yeast from different sources as described in Materials and Methods. These properties included the mRNA stability, mRNA copy number, ribosome occupancy, protein stability, and abundance. In addition to these attributes of mRNAs and proteins, we also obtained the data describing the cell-to-cell variation in protein expression in a genetically homogenous population of cells, typically referred to as protein expression noise.

Messenger RNA half-life is a measure of transcript stability in the cell, whereas mRNA copy number reflects its abundance. We first asked whether RBPs as a functional class show a different tendency in comparison to non-RBPs in these properties. As a result of this analysis, we found that mRNAs encoding RBPs are significantly less stable (i.e., short half-life) at the transcript level compared to those genes that do not encode RBPs (P = 3.1 × 10⁻¹⁰, Wilcoxon test) (Fig. 2A). In yeast it has been shown that, in general, mRNAs of central physiological pathways have a longer half-life and mRNAs encoding regulatory and signaling proteins have a shorter half-life (22). In line with these observations, the observed lower half-life of RBPs in our analysis is consistent with their regulatory function and quick turn over at the transcript level. However, a comparison of the mRNA copy number of the two groups of genes, which is a proxy for mRNA abundance in the cell, indicated that RBPs are encoded by genes that exhibit much higher mRNA copy number (P < 2.2 × 10⁻¹⁶, Wilcoxon test) (Fig. 2B). Exclusion of translation and ribosome-associated genes, which form a significant fraction of the total repertoire of RBPs and are known to be highly expressed, did not change our results (SI Text S2). These observations suggest that RBPs tend to be less stable but more abundant at the transcript level, suggesting that abundance is a more prominent factor than their stability. Both mRNA half-life and mRNA abundance data indicate that RBP's expression at mRNA level is likely to be transient but whenever they are transcribed they are produced at high concentrations.

Ribosome occupancy has been shown to be a measure of translational efficiency of mRNA. Higher ribosome occupancy relates to higher protein synthesis, and lower ribosome occupancy indicates low translation rate of mRNA. We next asked whether the ribosome occupancy i.e., rate of translation, of RBPs is higher than those for non-RBPs and whether their protein levels are higher within the cell. This analysis clearly revealed that RBPs have high ribosome occupancy (P = 2.5 × 10⁻¹³, Wilcoxon test) (Fig. 2C) and are also present in much higher concentrations (P < 2.2 × 10⁻¹⁶, Wilcoxon test) (Fig. 2D) with median abundances of RBPs being roughly double that observed for non-RBPs (3,895 versus 2,132 protein molecules/cell). These results indicate that RBPs are abundant and are translated rapidly, supporting the versatile nature of their involvement in multiple posttranscriptional control mechanisms at different cellular locations (SI Text S1). Exclusion of ribosome and translation-associated factors from RBPs to compare nonribosomal RBPs against non-RBPs indicated that ribosomal RBPs contribute significantly to the observed differences in the rate of translation and protein abundance of RBPs (SI Text S2). Comparing the protein concentrations of nonribosomal RBPs with non-RBPs indicated that the former are still significantly more abundant (P = 2.2 × 10⁻²).

Stability of a protein measured as its half-life can be considered as a proxy for the life time of a protein in a cell. Therefore, to understand the degradation rates of RBPs and to compare them against non-RBPs we analyzed their protein half-lives (see Materials and Methods). This analysis revealed that RBPs are significantly more stable than non-RBPs, with RBPs exhibiting a median half-life of 71 min as against non-RBPs with 46 min (P = 5 × 10⁻¹², Wilcoxon test) (Fig. 2E). Repeating the analyses with nonribosomal RBPs showed a consistent trend despite their exclusion (P = 4.8 × 10⁻²) (SI Text S2). Our observations on the increased protein stability and concentration of the RBPs compared to other proteins in the cell suggests that RBPs, whose main functional role is in the processing and localization of their mRNA targets, might be required at multiple subcellular locations and be used throughout the cell cycle. This may likely warrant their higher abundance and stability at the protein level. It is important to note that although RBPs exhibit high protein stability, they also show low transcript stability, which indicates that most RBPs that are stable at the protein level, might be avoiding cellular crowding of their transcripts by quick turnover at the transcript level. Indeed, it has been shown in yeast that most RBPs autoregulate their own activity at the transcript level (23).

To understand how these properties vary with different processes in which RBPs are involved, we divided RBPs into four major categories: translation, transport, RNA localization, and processing using GO annotations and compared them with non-RBPs (SI Text S3). This analysis revealed that the general trends observed for different categories are similar to those seen for RBPs as a whole although certain categories comprised relatively few RBPs.

Several RBPs have been shown to be posttranslationally modified, which adds a layer of flexibility to their function. Many of these posttranslational modifications have been shown to modify their RNA-binding properties or their subcellular localization. Indeed, at least four types of posttranslational modifications namely phosphorylation, ubiquitination, methylation, and SUMOylation have been reported for RBPs (2). High stability of RBPs indicates the potential that posttranslational modifications can offer in the diversification of their function. In fact, analysis of the number of kinase substrates in RBP and non-RBP populations using the currently available protein phosphorylation map for yeast (24), suggests that some kinases not only target higher number of RBPs compared to non-RBPs (P = 2.7 × 10⁻²) but also more kinases are associated with RBPs (P < 2.2 × 10⁻¹⁶) (see SI Text S4).

Gene expression is a highly dynamic process and because of its dynamic nature there is a large variation in a protein's abundance among different cells in a population. This variation is termed as biological noise. Genes whose expression varies to a large extent show more noise and these are typically involved in stress response, amino acid biosynthesis, and heat shock. On the other hand, genes that show consistent expression during the cell cycle such as those involved in protein degradation and ribosomal proteins tend to show low noise (25). Here, we have explored this noise data, to address whether RBPs show significant difference from non-RBPs in terms of biological noise. As shown in Fig. 2F, RBPs were found to show significantly lower noise levels in comparison to non-RBPs (P = 1.7 × 10⁻¹², Wilcoxon test). Reanalyzing the data by excluding ribosomal proteins still clearly indicated that RBPs exhibit much lower noise compared to other protein coding genes (P = 6.3 × 10⁻⁶, Wilcoxon test) (SI Text S2). This analysis unambiguously reveals that low noise is an inherent property of all RBPs and suggests that RBPs are tightly regulated at the protein level with little variation in their expression from cell to cell. An independent analysis to compare the dynamic properties of RBPs with all of the protein coding genes (including RBPs) and varying the test statistic used to calculate the significance, did not change our results. This suggests that the trends observed are generally robust and are independent of the statistical test used (SI Text S5).

The Number of Distinct Targets Bound by a RBP Is Correlated with Its Cellular Abundance.

RBPs are the key elements responsible for the posttranscriptional control of gene expression and when combined with their RNA targets, this information can be represented as a RBP-RNA network. Although, on a genomic scale, RBPs are believed to control diverse range of functions with some eukaryotic systems predominantly using posttranscriptional mechanisms for gene expression control (11, 13), large-scale elucidation of posttranscriptional networks is limited to few model organisms for a select set of RBPs. In yeast, few recent genomewide studies identified the targets for several RBPs using RIP-chip technology (23, 26). These studies revealed the important roles played by different families of RBPs and the structure of the posttranscriptional network formed by them. These high-throughput studies showed that the number of targets of a RBP can vary widely, from <10 to more than thousands. In this study we obtained this network, where nodes represent RBPs or their targets and links represent a distinct physical association between the RBP and the target RNA. We then systematically investigated the relationship between different dynamic properties of RBPs and the number of distinct RNA targets they control.

We first asked whether the number of targets of a RBP is correlated with its transcript stability by grouping the RBPs into different connectivity bins i.e., groups of RBPs comprising a number of distinct RNA targets (see Materials and Methods). As a result of this analysis, we found that there was a weak but positive correlation between them, suggesting that transcript turnover of RBPs may not be dependent on their number of targets (R² = 0.18, P < 0.24) (Fig. 3A). On the other hand, a comparison of the mRNA copy number of a RBP and its number of targets revealed a strong positive correlation between them, suggesting that RBPs with a high number of targets are likely to be more highly expressed at the mRNA level (R² = 0.96, P < 1 × 10⁻³) (Fig. 3B). For instance, PAB1 is a highly connected essential RBP that can bind to the poly(A) tail of an mRNA to regulate its translational initiation through its binding with eIF4G protein (18, 27). Indeed, it was reported to bind to 1,994 distinct RNA targets and was among the genes with very high mRNA copy number (7.1 mRNA copies/cell). These observations point to a direct link between the number of distinct targets of a RBP and its available number of copies of mRNA in the cell. To test the existence of a correlation between the connectivity and the rate of translation or the absolute protein abundance profile of RBPs, we further explored the relationship between them (Fig. 3 C and D). This comparison uncovered a more general link between translational efficiency of a RBP and its degree. For instance, Pub1p is another poly(A) binding protein (28) that binds to diverse sets of transcripts involved in ribosome biogenesis, cellular metabolism, and transport (29). This protein was reported to be localized to both nucleus and cytoplasm (30). Hence to be present at different locations and to bind to a large number of transcripts it has to be translated more often and should be present in a higher number of copies. Consistent with this, we find that its transcript exhibits high ribosome occupancy. Indeed, Hogan et al. (23) demonstrated that RNA targets of highly connected RBPs were enriched for multiple processes and subcellular localizations. These results clearly unveil the strong relationship between the concentration of a RBP and the number of distinct RNA targets bound by them, indicating that RBPs responsible for controlling a wide range of targets must occur in a higher number of copies at the protein level. It is important to note that although RBPs as a group of genes are significantly higher expressed at the transcript and protein levels compared to non-RBP population, relative abundance of the RBPs is correlated to the hierarchy of a RBP, defined as the number of distinct RNA targets. It is also noteworthy to mention that the RBPs analyzed for connectivity in this section did not comprise core ribosomal proteins, strengthening the generality of these observations.

RBPs Bound to Many RNA Targets Are Less Frequently Degraded and Tightly Controlled at Protein Level.

Although RBPs with a higher number of distinct targets are expressed at a higher level compared to those that control fewer targets, it is not evident whether their protein turnover rates would hold a similar trend. Therefore, to understand whether there is any dependence between the stability of a RBP and the number of transcripts it controls, we used a similar approach as above. This analysis clearly showed that RBPs that regulate many targets are highly stable at the protein level (R² = 0.95, P < 3 × 10⁻³) (Fig. 4A). The link between protein stability and RBP's degree indicates that RBPs controlling several targets are less frequently degraded at the protein level and might be present throughout the cell cycle. Taken together, these observations raise the question: If highly connected RBPs are consistently expressed in large concentrations and are less frequently degraded, would their regulation be tightly controlled at the protein level? The fact that RBPs as a group show significantly lower noise in comparison to non-RBPs and that previous studies reported that regulatory proteins generally exhibit low noise (25) suggests that highly connected RBPs can be expected to show less noise in comparison to those that are poorly connected. Hence, we compared the connectivity of RBPs with their noise value. As shown in Fig. 4B, we found a strong correlation between the number of targets of a RBP and its protein noise. In particular, highly connected RBPs showed minimal variation in their protein expression across a population of cells (R² = 0.93, P < 4 × 10⁻³). This suggests that RBPs controlling many targets are very tightly regulated with little cell-to-cell variation in their protein expression. These observations indicate that any significant change in their availability or regulation may result in an imbalance in cellular homeostasis as it may affect a vast number of transcripts. Indeed, a comparison of the number of essential genes in RBPs showed a twofold enrichment compared to the whole genome, suggesting their central role in maintaining cellular homeostasis (SI Text S6). These lines of evidence reveal that RBPs act as an important class of regulatory molecules in the cell whose expression is tightly controlled despite their occurrence in large cellular concentrations and in multiple subcellular locations.

Data on RNA-Binding Proteins in S. cerevisiae and Their Interactions.

The complete list of annotated RBPs and the data for well-studied RBPs in S. cerevisiae were obtained from Hogan et al. (23). The total number of annotated RBPs in yeast reported in this study was 561 and mRNA targets for 41 RBPs have been systematically identified on a whole genome scale by employing the RIP-chip technology. This approach essentially consists of two steps. The first involves generation of two RNA samples, isolation of RBP-bound mRNA by immunoprecipitation of messenger-ribonucleoproteins using affinity purification, and isolation of cellular RNA representing the whole set of transcripts in the cell. The second step involves hybridization of the two isolated RNA samples using dual-color microarrays and are analyzed for enriched transcripts to detect the bound targets of a RBP (39). A total of 14,312 interactions comprising 41 RBPs and 5,025 genes in the entire genome of S. cerevisiae, which forms a network of posttranscriptional interactions between RBPs and the target RNAs obtained using this approach, were used in this study (23).

Data for Comparative Analysis of Expression Dynamics.

To study the expression dynamics of RBPs in comparison to other protein coding genes in the genome and to analyze their relationship with the number of RNAs controlled by RBPs, we have used a variety of datasets. These include the transcript stability, mRNA copy number, ribosome occupancy, protein half-life, protein abundance, and protein noise. Transcript stability, which is measured as the RNA half-life of a transcript, was obtained from Wang et al. (22) and contained mRNA half-lives for 4,687 genes in the entire genome. A key parameter describing the translational status of a gene is the fraction of its transcripts engaged in translation, which is defined by the ribosome occupancy (40). Likewise, the number of mRNA copies of a gene can be best described by the parameter mRNA copy number per cell. Both these parameters for genes in S. cerevisiae were obtained from Arava et al. (40) where the authors used velocity sedimentation to separate mRNAs bound to ribosomes and quantified them using microarray analysis. mRNA copy number could be obtained for 5,643 genes whereas ribosome occupancy could be mapped for 5,700 genes, allowing us to study the extent of transcript abundance and translation rates of the genes and transcripts. Stability of a protein, which is an estimate of the duration it occurs within the cell is measured as the half-life of the protein. In yeast, protein half-lives have been estimated by Belle and coworkers for ≈3,750 proteins by inhibiting translation (41). In this study, we used these data by excluding proteins whose half-lives have been obtained by extrapolation. Protein abundance, which reveals the absolute number of protein molecules per cell, was obtained from Ghaemmaghami et al. (42). We could obtain abundance values for 3,868 proteins in the entire genome. Biological noise, which is typically defined as the variation in the expression of a protein between different cells in a homogenous population of cells, was obtained from Newman et al. (25). We could obtain noise data for 2,213 genes for cells grown on rich media. The authors in this study used two distinct measures for calculating protein noise, coefficient of variation (CV), which is the ratio of the standard deviation in the expression of a protein and its mean expression and distance from median (DM), which was calculated as the difference between the CV value of a protein and a running median of all CV values. In this study, we have used DM as a measure of protein noise as it was indicated to be a more robust measure compared to CV to understand protein-to-protein variations in noise levels (25). Since DM is the distance between the CV and median value of all CVs, negative values correspond to relatively less noise whereas positive values reflect higher levels of noise in the protein expression.

Comparison of the Regulatory Properties of RBPs with Other Protein Coding Genes.

To study whether RBPs show differences in dynamic properties when compared to other protein coding genes, we defined a non-RBP set of proteins. This set essentially comprised proteins in the whole genome after excluding the list of 561 RBPs defined above. To assess whether RBPs exhibit a different trend compared to non-RBPs for each of the properties studied, we used the Wilcoxon rank-sum test or the Mann–Whitney U test available in the R statistical package to calculate the significance. The Wilcoxon test enables the comparison of two samples to assess whether they come from the same distribution or not. Since this test is nonparametric and does not assume any inherent distribution of the samples it is ideal to compare different samples. Box plots were used to represent the distribution of values for each property. Independently, analysis of the mean values of a property for RBP and non-RBP sets of proteins was also carried out and P-values were estimated using the Welch t test, which gave similar results (SI Text S5). Because the RBP set comprised a number of ribosome-associated proteins we also excluded them from this list and repeated the analysis to test the robustness of the tendencies observed, in the absence of ribosomal proteins (SI Text S2).

Analysis of the Relationship Between the Number of Targets of a RBP and Their Dynamic Properties.

To understand the link between the number of targets of a RBP and their dynamic properties, RBPs were first grouped on the basis of their number of distinct RNA targets to which they were bound. This grouping was done in such a way that each bin of RBPs contained roughly an equal number of RBPs. This resulted in five different bins corresponding to varying degrees of RBPs, with some RBPs controlling as many as 2,000 mRNAs in the RBP-RNA network. To nullify the effect of outliers in each bin, median values were calculated for different dynamic properties and correlation was estimated between median values and connectivity of RBPs. P-values were calculated using the coefficient of correlation and the number of data points, on the basis of a linear fit.

Supporting Information.

For additional details relating to SI Text S1–S6, see Dataset S1, Figs. S1–S7, and Tables S1–S3. An analysis of enrichment of cell-cycle regulated genes, performed to address the concerns of a reviewer, is detailed in SI Text S7.